ZDM presents a traditional block device for ZBC/ZAC zoned devices. User space utilities in zdm-tools for creating, repairing and restore DM instances at: https://github.com/Seagate Signed-off-by: Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx> --- Documentation/device-mapper/zoned.txt | 75 + MAINTAINERS | 7 + drivers/md/Kconfig | 11 + drivers/md/Makefile | 1 + drivers/md/dm-zoned.c | 2535 ++++++++++++ drivers/md/dm-zoned.h | 714 ++++ drivers/md/libzoned.c | 7149 +++++++++++++++++++++++++++++++++ 7 files changed, 10492 insertions(+) create mode 100644 Documentation/device-mapper/zoned.txt create mode 100644 drivers/md/dm-zoned.c create mode 100644 drivers/md/dm-zoned.h create mode 100644 drivers/md/libzoned.c diff --git a/Documentation/device-mapper/zoned.txt b/Documentation/device-mapper/zoned.txt new file mode 100644 index 0000000..2d1db28 --- /dev/null +++ b/Documentation/device-mapper/zoned.txt @@ -0,0 +1,75 @@ +Overview of Host Aware ZBC/ZAC Device Mapper + - Zone size (256MiB) + - Reset WP, Open Zone, Close Zone, Get Zone Info ... + +The Zoned DM treats a zoned device as a collection of 1024 zones 256GiB, +referred to internally as 'megazones' as with zoned devices the last +megazone may be less than 1024 zones in size. If the last zone is of +different size it is ignored. + +What that means is that drives which report SAME=0 are not supported +and likely won't be supported within this architecture. However drives +which report a SAME code of: all same, last differs or same length +different types, would all be supported by this architecture. + +The initial implementation focuses on drives with same sized zones of +256MB which is 65536 4k blocks. In future the zone size of 256MB will +be relaxed to allow any size of zone as long as they are all the same. + +Internally all addressing is on 4k boundaries. Currently a 4k PAGE_SIZE is +assumed. Architectures with 8k (or other) PAGE_SIZE have not been tested +and are likly broken at the moment. + +Host Managed drives should work if the zone type at the start of the partition +is Conventional, or Preferred. + + +Megazones: + Each megazone is managed independently and partitioned into + meta data and data. The Device Mapper Meta Data is logically + located in the 2nd and 3rd zones of the megazone. The 1st and + 2nd zones are reserved for the megazone's minimally relocatable + super block which must be the first block of the first or + second zone. The most recent is determined by the generation + number embedded in the super block. The meta data is sized + for two zones and logically located in sectors non-addressable + to the upper device. The actual storage of the meta data is pooled + with the data using the same mapping scheme. + The device mapper internally is a COW device with a 4k per block + addressing scheme. There are some fix-ups to handle non-4k aligned + requests to support applications which read and write in 512 byte + blocks, however it is still desirable to submit patches for these + subsystems assuming the respective maintainers are willing to + accept 4k alignment patches. + +Address space: + The zoned device mapper presents a smaller block device than + the amount of data available on the physical media. The extra + space is used to hold the meta data needed for managing the + data being stored on the drive performing COW block [re]mapping. + The 'shrink' is done by appropriately sizing the device via + dmsetup. + See the zdmadm utility will detect and size the device appropriaty. + +Map Strategy: + Map incoming sector onto device mapper sector space. + +Read Strategy: + Check each block requested in the bio to determine if the data + blocks are consecutively stored on disk. Pass as much per-bio + as possible through to the backing block device. + +Write Strategy: + Allocate space for entire bio on the backing block device + redirecting all incoming write requests to the most recently + written zone until the zone is filled or the bio is too large + to fit and a new zone is opened. Note that if the zone is not + filled this zone will likely become used by meta data writes + which are typically single blocks. + +Sync Strategy: + On SYNC bios all the meta data need to restore the zoned device + mapper for disk is written to one of the well known zones at + the beginning of the mega zone. Data consistency is only + 'guaranteed' to be on-disk and consistent following sync + events [same as ext4]. diff --git a/MAINTAINERS b/MAINTAINERS index c6ad972..5a79f49 100644 --- a/MAINTAINERS +++ b/MAINTAINERS @@ -12182,6 +12182,13 @@ L: zd1211-devs@xxxxxxxxxxxxxxxxxxxxx (subscribers-only) S: Maintained F: drivers/net/wireless/zydas/zd1211rw/ +ZDM ZONED DEVICE MAPPER TARGET +M: Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx> +L: dm-devel@xxxxxxxxxx +S: Maintained +F: drivers/md/dm-zoned.* +F: drivers/md/libzoned.c + ZPOOL COMPRESSED PAGE STORAGE API M: Dan Streetman <ddstreet@xxxxxxxx> L: linux-mm@xxxxxxxxx diff --git a/drivers/md/Kconfig b/drivers/md/Kconfig index 0a2e727..d0894e4 100644 --- a/drivers/md/Kconfig +++ b/drivers/md/Kconfig @@ -345,6 +345,17 @@ config DM_ERA over time. Useful for maintaining cache coherency when using vendor snapshots. +config DM_ZONED + tristate "ZDM: Zoned based device target (EXPERIMENTAL)" + depends on BLK_DEV_DM + default n + select LIBCRC32C + ---help--- + dm-zoned provides a rand access block device on top of a + ZBC/ZAC block device. + Forward writing within zones, garbage collection within zones. + Use zdmadm to create, repair and/or restore ZDM instances. + config DM_MIRROR tristate "Mirror target" depends on BLK_DEV_DM diff --git a/drivers/md/Makefile b/drivers/md/Makefile index 62a6576..953bdda 100644 --- a/drivers/md/Makefile +++ b/drivers/md/Makefile @@ -60,6 +60,7 @@ obj-$(CONFIG_DM_CACHE_SMQ) += dm-cache-smq.o obj-$(CONFIG_DM_CACHE_CLEANER) += dm-cache-cleaner.o obj-$(CONFIG_DM_ERA) += dm-era.o obj-$(CONFIG_DM_LOG_WRITES) += dm-log-writes.o +obj-$(CONFIG_DM_ZONED) += dm-zoned.o ifeq ($(CONFIG_DM_UEVENT),y) dm-mod-objs += dm-uevent.o diff --git a/drivers/md/dm-zoned.c b/drivers/md/dm-zoned.c new file mode 100644 index 0000000..4e04464 --- /dev/null +++ b/drivers/md/dm-zoned.c @@ -0,0 +1,2535 @@ +/* + * Kernel Device Mapper for abstracting ZAC/ZBC devices as normal + * block devices for linux file systems. + * + * Copyright (C) 2015 Seagate Technology PLC + * + * Written by: + * Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx> + * + * This file is licensed under the terms of the GNU General Public + * License version 2. This program is licensed "as is" without any + * warranty of any kind, whether express or implied. + */ + +#include "dm.h" +#include <linux/dm-io.h> +#include <linux/init.h> +#include <linux/mempool.h> +#include <linux/module.h> +#include <linux/slab.h> +#include <linux/vmalloc.h> +#include <linux/random.h> +#include <linux/crc32c.h> +#include <linux/crc16.h> +#include <linux/sort.h> +#include <linux/ctype.h> +#include <linux/types.h> +#include <linux/blkzoned_api.h> +#include <linux/timer.h> +#include <linux/delay.h> +#include <linux/kthread.h> +#include <linux/freezer.h> +#include <linux/proc_fs.h> +#include <linux/seq_file.h> +#include <linux/kfifo.h> +#include <linux/bsearch.h> +#include "dm-zoned.h" + +#define PRIu64 "llu" +#define PRIx64 "llx" +#define PRId32 "d" +#define PRIx32 "x" +#define PRIu32 "u" + +#define BIOSET_RESV 4 + +/** + * _zdisk() - Return a pretty ZDM name. + * @znd: ZDM Instance + * + * Return: ZDM/backing device pretty name. + */ +static inline char *_zdisk(struct zoned *znd) +{ + return znd->bdev_name; +} + +#define Z_ERR(znd, fmt, arg...) \ + pr_err("dm-zoned(%s): " fmt "\n", _zdisk(znd), ## arg) + +#define Z_INFO(znd, fmt, arg...) \ + pr_info("dm-zoned(%s): " fmt "\n", _zdisk(znd), ## arg) + +#define Z_DBG(znd, fmt, arg...) \ + pr_debug("dm-zoned(%s): " fmt "\n", _zdisk(znd), ## arg) + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ +static void do_io_work(struct work_struct *work); +static int block_io(struct zoned *, enum dm_io_mem_type, void *, sector_t, + unsigned int, int, int); +static int znd_async_io(struct zoned *znd, + enum dm_io_mem_type dtype, + void *data, + sector_t block, unsigned int nDMsect, int rw, int queue, + io_notify_fn callback, void *context); +static int zoned_bio(struct zoned *znd, struct bio *bio); +static int zoned_map_write(struct zoned *znd, struct bio*, u64 s_zdm); +static sector_t get_dev_size(struct dm_target *ti); +static int dmz_reset_wp(struct zoned *znd, u64 z_id); +static int dmz_open_zone(struct zoned *znd, u64 z_id); +static int dmz_close_zone(struct zoned *znd, u64 z_id); +static u32 dmz_report_count(struct zoned *znd, void *report, size_t bufsz); +static int dmz_report_zones(struct zoned *znd, u64 z_id, + struct page *pgs, size_t bufsz); +static void activity_timeout(unsigned long data); +static void zoned_destroy(struct zoned *); +static int gc_can_cherrypick(struct zoned *znd, u32 sid, int delay, gfp_t gfp); +static void bg_work_task(struct work_struct *work); +static void on_timeout_activity(struct zoned *znd, int delay); +static int zdm_create_proc_entries(struct zoned *znd); +static void zdm_remove_proc_entries(struct zoned *znd); + +/** + * get_bdev_bd_inode() - Get primary backing device inode + * @znd: ZDM Instance + * + * Return: backing device inode + */ +static inline struct inode *get_bdev_bd_inode(struct zoned *znd) +{ + return znd->dev->bdev->bd_inode; +} + +#include "libzoned.c" + +#define BIO_CACHE_SECTORS (IO_VCACHE_PAGES * Z_BLOCKS_PER_DM_SECTOR) + +/** + * bio_stream() - Decode stream id from BIO. + * @znd: ZDM Instance + * + * Return: stream_id + */ +static inline u32 bio_stream(struct bio *bio) +{ + u32 stream_id = 0x40; + + /* + * Since adding stream id to a BIO is not yet in mainline we just + * use this heuristic to try to skip unnecessary co-mingling of data. + */ + + if (bio->bi_rw & REQ_META) + stream_id = 0xff; + + return stream_id; +} + +/** + * zoned_map_discard() - Return a pretty ZDM name. + * @znd: ZDM Instance + * @bio: struct bio hold discard information + * @s_zdm: tlba being discarded. + * + * Return: 0 on success, otherwise error code. + */ +static int zoned_map_discard(struct zoned *znd, struct bio *bio, u64 s_zdm) +{ + int rcode = DM_MAPIO_SUBMITTED; + u64 blks = bio->bi_iter.bi_size / Z_C4K; + int err; + + if (znd->is_empty) + goto cleanup_out; + + err = z_mapped_discard(znd, s_zdm, blks, CRIT); + if (err < 0) { + rcode = err; + goto out; + } + +cleanup_out: + bio->bi_iter.bi_sector = 8; + bio_endio(bio); +out: + return rcode; +} + +/** + * is_non_wp_zone() - Test zone # to see if it flagged as conventional. + * @znd: ZDM Instance + * @z_id: Zone # + * + * Return: 1 if conventional zone. 0 if sequentional write zone. + */ +static int is_non_wp_zone(struct zoned *znd, u64 z_id) +{ + u32 gzoff = z_id % 1024; + struct meta_pg *wpg = &znd->wp[z_id >> 10]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + + return (wp & Z_WP_NON_SEQ) ? 1 : 0; +} + +/** + */ +struct zone_action { + struct work_struct work; + struct zoned *znd; + u64 s_addr; + unsigned long bi_rw; + int wp_err; +}; + +/** + * do_zone_action_work() - Issue a 'zone action' to the backing device. + * @work: Work to do. + */ +static void do_zone_action_work(struct work_struct *work) +{ + struct zone_action *za = container_of(work, struct zone_action, work); + struct zoned *znd = za->znd; + struct block_device *bdev = znd->dev->bdev; + const gfp_t gfp = GFP_KERNEL; + + /* Explicitly work on device lbas not partition offsets. */ + if (bdev != bdev->bd_contains) + bdev = bdev->bd_contains; + + za->wp_err = blkdev_issue_zone_action(bdev, za->bi_rw, za->s_addr, gfp); +} + +/** + * dmz_zone_action() - Issue a 'zone action' to the backing device (via worker). + * @znd: ZDM Instance + * @z_id: Zone # to open. + * @rw: One of REQ_OPEN_ZONE, REQ_CLOSE_ZONE, or REQ_RESET_ZONE. + * + * Return: 0 on success, otherwise error. + */ +static int dmz_zone_action(struct zoned *znd, u64 z_id, unsigned long rw) +{ + int wp_err = 0; + + if (is_non_wp_zone(znd, z_id)) + return wp_err; + + if (znd->bdev_is_zoned) { + u64 z_offset = zone_to_sector(z_id + znd->zdstart); + struct zone_action za = { + .znd = znd, + .bi_rw = rw, + .s_addr = z_offset, + .wp_err = 0, + }; + if (znd->ata_passthrough) + za.bi_rw |= REQ_META; + /* + * Issue the synchronous I/O from a different thread + * to avoid generic_make_request recursion. + */ + INIT_WORK_ONSTACK(&za.work, do_zone_action_work); + queue_work(znd->zone_action_wq, &za.work); + flush_workqueue(znd->zone_action_wq); + destroy_work_on_stack(&za.work); + wp_err = za.wp_err; + + if (wp_err) { + Z_ERR(znd, "Zone Cmd: LBA: %" PRIx64 + " [Z:%" PRIu64 "] -> %d failed.", + za.s_addr, z_id, wp_err); + Z_ERR(znd, "ZAC/ZBC support disabled."); + znd->bdev_is_zoned = 0; + wp_err = -ENOTSUPP; + } + } + return wp_err; +} + +/** + * dmz_reset_wp() - Reset write pointer for zone z_id. + * @znd: ZDM Instance + * @z_id: Zone # to reset. + * + * Return: 0 on success, otherwise error. + */ +static int dmz_reset_wp(struct zoned *znd, u64 z_id) +{ + return dmz_zone_action(znd, z_id, REQ_RESET_ZONE); +} + +/** + * dmz_open_zone() - Open zone for writing. + * @znd: ZDM Instance + * @z_id: Zone # to open. + * + * Return: 0 on success, otherwise error. + */ +static int dmz_open_zone(struct zoned *znd, u64 z_id) +{ + if (!znd->issue_open_zone) + return 0; + return dmz_zone_action(znd, z_id, REQ_OPEN_ZONE); + +} + +/** + * dmz_close_zone() - Close zone to writing. + * @znd: ZDM Instance + * @z_id: Zone # to close. + * + * Return: 0 on success, otherwise error. + */ +static int dmz_close_zone(struct zoned *znd, u64 z_id) +{ + if (!znd->issue_close_zone) + return 0; + return dmz_zone_action(znd, z_id, REQ_CLOSE_ZONE); +} + +/** + * dmz_report_count() - Read number of zones returned. + * @znd: ZDM Instance + * @rpt_in: Report data. + * @bufsz: size of space allocated for report + * + * Return: Number of zones in report. + */ +static u32 dmz_report_count(struct zoned *znd, void *rpt_in, size_t bufsz) +{ + u32 count; + u32 max_count = (bufsz - sizeof(struct bdev_zone_report)) + / sizeof(struct bdev_zone_descriptor); + + if (znd->ata_passthrough) { + struct bdev_zone_report_le *report = rpt_in; + + /* ZAC: ata results are little endian */ + if (max_count > le32_to_cpu(report->descriptor_count)) + report->descriptor_count = cpu_to_le32(max_count); + count = le32_to_cpu(report->descriptor_count); + } else { + struct bdev_zone_report *report = rpt_in; + + /* ZBC: scsi results are big endian */ + if (max_count > be32_to_cpu(report->descriptor_count)) + report->descriptor_count = cpu_to_be32(max_count); + count = be32_to_cpu(report->descriptor_count); + } + return count; +} + +/** + * dmz_report_zones() - issue report zones from z_id zones after zdstart + * @znd: ZDM Instance + * @z_id: Zone past zdstart + * @report: structure filled + * @bufsz: kmalloc()'d space reserved for report + * + * Return: -ENOTSUPP or 0 on success + */ +static int dmz_report_zones(struct zoned *znd, u64 z_id, + struct page *pgs, size_t bufsz) +{ + int wp_err = -ENOTSUPP; + + if (znd->bdev_is_zoned) { + u8 opt = ZOPT_NON_SEQ_AND_RESET; + unsigned long bi_rw = 0; + struct block_device *bdev = znd->dev->bdev; + u64 s_addr = zone_to_sector(z_id + znd->zdstart); + + if (bdev != bdev->bd_contains) + s_addr -= znd->start_sect << Z_SHFT4K; + + if (znd->ata_passthrough) + bi_rw = REQ_META; + + wp_err = blkdev_issue_zone_report(bdev, bi_rw, s_addr, opt, + pgs, bufsz, GFP_KERNEL); + if (wp_err) { + Z_ERR(znd, "Report Zones: LBA: %" PRIx64 + " [Z:%" PRIu64 " -> %d failed.", + s_addr, z_id + znd->zdstart, wp_err); + Z_ERR(znd, "ZAC/ZBC support disabled."); + znd->bdev_is_zoned = 0; + wp_err = -ENOTSUPP; + } + } + return wp_err; +} + +/** + * is_conventional() - Determine if zone is conventional. + * @dentry: Zone descriptor entry. + * + * Return: 1 if zone type is conventional. + */ +static inline int is_conventional(struct bdev_zone_descriptor *dentry) +{ + return (ZTYP_CONVENTIONAL == (dentry->type & 0x0F)) ? 1 : 0; +} + +/** + * is_zone_reset() - Determine if zone is reset / ready for writing. + * @dentry: Zone descriptor entry. + * + * Return: 1 if zone condition is empty or zone type is conventional. + */ +static inline int is_zone_reset(struct bdev_zone_descriptor *dentry) +{ + u8 type = dentry->type & 0x0F; + u8 cond = (dentry->flags & 0xF0) >> 4; + + return (ZCOND_ZC1_EMPTY == cond || ZTYP_CONVENTIONAL == type) ? 1 : 0; +} + +/** + * get_wp_from_descriptor() - Decode write pointer as # of blocks from start + * @znd: ZDM Instance + * @dentry_in: Zone descriptor entry. + * + * Return: Write Pointer as number of blocks from start of zone. + */ +static inline u32 get_wp_from_descriptor(struct zoned *znd, void *dentry_in) +{ + u32 wp = 0; + + /* + * If ATA passthrough was used then ZAC results are little endian. + * otherwise ZBC results are big endian. + */ + + if (znd->ata_passthrough) { + struct bdev_zone_descriptor_le *lil = dentry_in; + + wp = le64_to_cpu(lil->lba_wptr) - le64_to_cpu(lil->lba_start); + } else { + struct bdev_zone_descriptor *big = dentry_in; + + wp = be64_to_cpu(big->lba_wptr) - be64_to_cpu(big->lba_start); + } + return wp; +} + +/** + * _dec_wp_avail_by_lost() - Update free count due to lost/unusable blocks. + * @wpg: Write pointer metadata page. + * @gzoff: Zone entry in page. + * @lost: Number of blocks 'lost'. + */ +static inline +void _dec_wp_avail_by_lost(struct meta_pg *wpg, u32 gzoff, u32 lost) +{ + wpg->zf_est[gzoff] = cpu_to_le32(le32_to_cpu(wpg->zf_est[gzoff])+lost); +} + +/** + * zoned_wp_sync() - Re-Sync expected WP location with drive + * @znd: ZDM Instance + * @reset_non_empty: Reset the non-empty zones. + * + * Return: 0 on success, otherwise error. + */ +static int zoned_wp_sync(struct zoned *znd, int reset_non_empty) +{ + int rcode = 0; + u32 rcount = 0; + u32 iter; + struct bdev_zone_report *report = NULL; + int order = REPORT_ORDER; + size_t bufsz = REPORT_FILL_PGS * Z_C4K; + struct page *pgs = alloc_pages(GFP_KERNEL, order); + + if (pgs) + report = page_address(pgs); + + if (!report) { + rcode = -ENOMEM; + goto out; + } + + for (iter = 0; iter < znd->data_zones; iter++) { + u32 entry = iter % 4096; + u32 gzno = iter >> 10; + u32 gzoff = iter & ((1 << 10) - 1); + struct meta_pg *wpg = &znd->wp[gzno]; + struct bdev_zone_descriptor *dentry; + u32 wp_flgs; + u32 wp_at; + u32 wp; + + if (entry == 0) { + int err = dmz_report_zones(znd, iter, pgs, bufsz); + + if (err) { + Z_ERR(znd, "report zones-> %d", err); + if (err != -ENOTSUPP) + rcode = err; + goto out; + } + rcount = dmz_report_count(znd, report, bufsz); + } + + dentry = &report->descriptors[entry]; + if (reset_non_empty && !is_conventional(dentry)) { + int err = 0; + + if (!is_zone_reset(dentry)) + err = dmz_reset_wp(znd, iter); + + if (err) { + Z_ERR(znd, "reset wp-> %d", err); + if (err != -ENOTSUPP) + rcode = err; + goto out; + } + wp = wp_at = 0; + wpg->wp_alloc[gzoff] = cpu_to_le32(0); + wpg->zf_est[gzoff] = cpu_to_le32(Z_BLKSZ); + continue; + } + + wp = get_wp_from_descriptor(znd, dentry); + wp >>= Z_SHFT4K; /* 512 sectors to 4k sectors */ + wp_at = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_VALUE_MASK; + wp_flgs = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_FLAGS_MASK; + + if (is_conventional(dentry)) { + wp = wp_at; + wp_flgs |= Z_WP_NON_SEQ; + } else { + wp_flgs &= ~Z_WP_NON_SEQ; + } + + if (wp > wp_at) { + u32 lost = wp - wp_at; + + wp_at = wp; + _dec_wp_avail_by_lost(wpg, gzoff, lost); + + Z_ERR(znd, "Z#%u z:%x [wp:%x rz:%x] lost %u blocks.", + iter, gzoff, wp_at, wp, lost); + } + wpg->wp_alloc[gzoff] = cpu_to_le32(wp_at|wp_flgs); + } + +out: + if (pgs) + __free_pages(pgs, order); + + return rcode; +} + +#if USE_KTHREAD + +static inline int stop_or_data(struct zoned *znd) +{ + if (kthread_should_stop()) + return 1; + return kfifo_is_empty(&znd->bio_fifo) ? 0 : 1; +} + +static int znd_bio_kthread(void *arg) +{ + int err = 0; + struct zoned *znd = (struct zoned *)arg; + + Z_ERR(znd, "znd_bio_kthread [started]"); + + while (!kthread_should_stop()) { + struct bio *bio = NULL; + int rcode; + + if (kfifo_is_empty(&znd->bio_fifo)) { + wake_up(&znd->wait_fifo); + wait_event_freezable(znd->wait_bio, stop_or_data(znd)); + continue; + } + if (!kfifo_get(&znd->bio_fifo, &bio)) { + wake_up(&znd->wait_fifo); + continue; + } + if (kfifo_avail(&znd->bio_fifo) > 5) + wake_up(&znd->wait_fifo); + + rcode = zoned_bio(znd, bio); + if (rcode == DM_MAPIO_REMAPPED) + rcode = DM_MAPIO_SUBMITTED; + + if (rcode < 0) { + znd->meta_result = err = rcode; + goto out; + } + } + err = 0; + Z_ERR(znd, "znd_bio_kthread [stopped]"); + +out: + return err; +} + +/** + * zoned_map() - Handle an incoming BIO + * @ti: Device Mapper Target Instance + * @bio: The BIO to disposition. + * + * Return: 0 on success, otherwise error. + */ +static int zoned_map(struct dm_target *ti, struct bio *bio) +{ + struct zoned *znd = ti->private; + int rcode = DM_MAPIO_REQUEUE; + + while (kfifo_put(&znd->bio_fifo, bio) == 0) { + wake_up_process(znd->bio_kthread); + wake_up(&znd->wait_bio); + wait_event_freezable(znd->wait_fifo, + kfifo_avail(&znd->bio_fifo) > 0); + } + rcode = DM_MAPIO_SUBMITTED; + wake_up_process(znd->bio_kthread); + wake_up(&znd->wait_bio); + + return rcode; +} + +#else + +/** + * zoned_map() - Handle an incoming BIO + * @ti: Device Mapper Target Instance + * @bio: The BIO to disposition. + * + * Return: 0 on success, otherwise error. + */ +static int zoned_map(struct dm_target *ti, struct bio *bio) +{ + struct zoned *znd = ti->private; + int err = zoned_bio(znd, bio); + + if (err < 0) { + znd->meta_result = err; + err = 0; + } + + return err; +} + +#endif + +/** + * zoned_actual_size() - Set number of 4k blocks available on block device. + * @ti: Device Mapper Target Instance + * @znd: ZDM Instance + * + * Return: 0 on success, otherwise error. + */ +static void zoned_actual_size(struct dm_target *ti, struct zoned *znd) +{ + znd->nr_blocks = i_size_read(get_bdev_bd_inode(znd)) / Z_C4K; +} + +/** + * zoned_ctr() - Create a ZDM Instance from DM Target Instance and args. + * @ti: Device Mapper Target Instance + * @argc: Number of args to handle. + * @argv: args to handle. + * + * Return: 0 on success, otherwise error. + */ +static int zoned_ctr(struct dm_target *ti, unsigned argc, char **argv) +{ + const int reset_non_empty = 0; + int create = 0; + int force = 0; + int zbc_probe = 1; + int zac_probe = 1; + int r; + struct zoned *znd; + long long first_data_zone = 0; + long long mz_md_provision = MZ_METADATA_ZONES; + + BUILD_BUG_ON(Z_C4K != (sizeof(struct map_sect_to_lba) * Z_UNSORTED)); + BUILD_BUG_ON(Z_C4K != (sizeof(struct io_4k_block))); + BUILD_BUG_ON(Z_C4K != (sizeof(struct mz_superkey))); + + znd = ZDM_ALLOC(NULL, sizeof(*znd), KM_00, NORMAL); + if (!znd) { + ti->error = "Error allocating zoned structure"; + return -ENOMEM; + } + + znd->trim = 1; + + if (argc < 1) { + ti->error = "Invalid argument count"; + return -EINVAL; + } + + for (r = 1; r < argc; r++) { + if (isdigit(*argv[r])) { + int krc = kstrtoll(argv[r], 0, &first_data_zone); + + if (krc != 0) { + DMERR("Failed to parse %s: %d", argv[r], krc); + first_data_zone = 0; + } + } + if (!strcasecmp("create", argv[r])) + create = 1; + if (!strcasecmp("load", argv[r])) + create = 0; + if (!strcasecmp("force", argv[r])) + force = 1; + if (!strcasecmp("nozbc", argv[r])) + zbc_probe = 0; + if (!strcasecmp("nozac", argv[r])) + zac_probe = 0; + if (!strcasecmp("discard", argv[r])) + znd->trim = 1; + if (!strcasecmp("nodiscard", argv[r])) + znd->trim = 0; + + if (!strncasecmp("reserve=", argv[r], 8)) { + long long mz_resv; + int krc = kstrtoll(argv[r] + 8, 0, &mz_resv); + + if (krc == 0) { + if (mz_resv > mz_md_provision) + mz_md_provision = mz_resv; + } else { + DMERR("Reserved 'FAILED TO PARSE.' %s: %d", + argv[r]+8, krc); + mz_resv = 0; + } + } + } + + znd->ti = ti; + ti->private = znd; + znd->zdstart = first_data_zone; + znd->mz_provision = mz_md_provision; + + r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &znd->dev); + if (r) { + ti->error = "Error opening backing device"; + zoned_destroy(znd); + return -EINVAL; + } + + if (znd->dev->bdev) { + bdevname(znd->dev->bdev, znd->bdev_name); + znd->start_sect = get_start_sect(znd->dev->bdev) >> 3; + } + + /* + * Set if this target needs to receive flushes regardless of + * whether or not its underlying devices have support. + */ + ti->flush_supported = true; + + /* + * Set if this target needs to receive discards regardless of + * whether or not its underlying devices have support. + */ + ti->discards_supported = true; + + /* + * Set if the target required discard bios to be split + * on max_io_len boundary. + */ + ti->split_discard_bios = false; + + /* + * Set if this target does not return zeroes on discarded blocks. + */ + ti->discard_zeroes_data_unsupported = false; + + /* + * Set if this target wants discard bios to be sent. + */ + ti->num_discard_bios = 1; + + if (!znd->trim) { + ti->discards_supported = false; + ti->num_discard_bios = 0; + } + + zoned_actual_size(ti, znd); + + r = do_init_zoned(ti, znd); + if (r) { + ti->error = "Error in zoned init"; + zoned_destroy(znd); + return -EINVAL; + } + + znd->issue_open_zone = 1; + znd->issue_close_zone = 1; + znd->filled_zone = NOZONE; + + if (zac_probe) + znd->bdev_is_zoned = znd->ata_passthrough = 1; + if (zbc_probe) + znd->bdev_is_zoned = 1; + + r = zoned_init_disk(ti, znd, create, force); + if (r) { + ti->error = "Error in zoned init from disk"; + zoned_destroy(znd); + return -EINVAL; + } + r = zoned_wp_sync(znd, reset_non_empty); + if (r) { + ti->error = "Error in zoned re-sync WP"; + zoned_destroy(znd); + return -EINVAL; + } + + znd->bio_set = bioset_create(BIOSET_RESV, 0); + if (!znd->bio_set) + return -ENOMEM; + +#if USE_KTHREAD + r = kfifo_alloc(&znd->bio_fifo, KFIFO_SIZE, GFP_KERNEL); + if (r) + return r; + + znd->bio_kthread = kthread_run(znd_bio_kthread, znd, "zdm-io-%s", + znd->bdev_name); + if (IS_ERR(znd->bio_kthread)) { + r = PTR_ERR(znd->bio_kthread); + ti->error = "Couldn't alloc kthread"; + zoned_destroy(znd); + return r; + } +#endif + + r = zdm_create_proc_entries(znd); + if (r) { + ti->error = "Failed to create /proc entries"; + zoned_destroy(znd); + return -EINVAL; + } + + mod_timer(&znd->timer, jiffies + msecs_to_jiffies(5000)); + + return 0; +} + +/** + * zoned_dtr() - Deconstruct a ZDM Instance from DM Target Instance. + * @ti: Device Mapper Target Instance + * + * Return: 0 on success, otherwise error. + */ +static void zoned_dtr(struct dm_target *ti) +{ + struct zoned *znd = ti->private; + + if (znd->z_sballoc) { + struct mz_superkey *key_blk = znd->z_sballoc; + struct zdm_superblock *sblock = &key_blk->sblock; + + sblock->flags = cpu_to_le32(0); + sblock->csum = sb_crc32(sblock); + } + +#if USE_KTHREAD + wake_up_process(znd->bio_kthread); + wait_event(znd->wait_bio, kfifo_is_empty(&znd->bio_fifo)); + kthread_stop(znd->bio_kthread); + kfifo_free(&znd->bio_fifo); +#endif + if (znd->bio_set) + bioset_free(znd->bio_set); + zdm_remove_proc_entries(znd); + zoned_destroy(znd); +} + + +/** + * do_io_work() - Read or write a data from a block device. + * @work: Work to be done. + */ +static void do_io_work(struct work_struct *work) +{ + struct z_io_req_t *req = container_of(work, struct z_io_req_t, work); + struct dm_io_request *io_req = req->io_req; + unsigned long error_bits = 0; + + req->result = dm_io(io_req, 1, req->where, &error_bits); + if (error_bits) + DMERR("ERROR: dm_io error: %lx", error_bits); +} + +/** + * znd_async_io() - Issue I/O via dm_io async or sync (using worker thread). + * @znd: ZDM Instance + * @dtype: Type of memory in data + * @data: Data for I/O + * @block: bLBA for I/O + * @nDMsect: Number of 512 byte blocks to read/write. + * @rw: READ or WRITE + * @queue: if true then use worker thread for I/O and wait. + * @callback: callback to use on I/O complete. + * context: context to be passed to callback. + * + * Return 0 on success, otherwise error. + */ +static int znd_async_io(struct zoned *znd, + enum dm_io_mem_type dtype, + void *data, + sector_t block, unsigned int nDMsect, int rw, int queue, + io_notify_fn callback, void *context) +{ + unsigned long error_bits = 0; + int rcode; + struct dm_io_region where = { + .bdev = znd->dev->bdev, + .sector = block, + .count = nDMsect, + }; + struct dm_io_request io_req = { + .bi_rw = rw, + .mem.type = dtype, + .mem.offset = 0, + .mem.ptr.vma = data, + .client = znd->io_client, + .notify.fn = callback, + .notify.context = context, + }; + + switch (dtype) { + case DM_IO_KMEM: + io_req.mem.ptr.addr = data; + break; + case DM_IO_BIO: + io_req.mem.ptr.bio = data; + where.count = nDMsect; + break; + case DM_IO_VMA: + io_req.mem.ptr.vma = data; + break; + default: + Z_ERR(znd, "page list not handled here .. see dm-io."); + break; + } + + if (queue) { + struct z_io_req_t req; + + /* + * Issue the synchronous I/O from a different thread + * to avoid generic_make_request recursion. + */ + INIT_WORK_ONSTACK(&req.work, do_io_work); + req.where = &where; + req.io_req = &io_req; + queue_work(znd->io_wq, &req.work); + flush_workqueue(znd->io_wq); + destroy_work_on_stack(&req.work); + + rcode = req.result; + if (rcode < 0) + Z_ERR(znd, "ERROR: dm_io error: %d", rcode); + goto done; + } + rcode = dm_io(&io_req, 1, &where, &error_bits); + if (error_bits || rcode < 0) + Z_ERR(znd, "ERROR: dm_io error: %d -- %lx", rcode, error_bits); + +done: + return rcode; +} + +/** + * block_io() - Issue sync I/O maybe using using a worker thread. + * @znd: ZDM Instance + * @dtype: Type of memory in data + * @data: Data for I/O + * @s: bLBA for I/O [512 byte resolution] + * @n: Number of 512 byte blocks to read/write. + * @rw: READ or WRITE + * @queue: if true then use worker thread for I/O and wait. + * + * Return 0 on success, otherwise error. + */ +static int block_io(struct zoned *znd, + enum dm_io_mem_type dtype, + void *data, sector_t s, unsigned int n, int rw, int queue) +{ + return znd_async_io(znd, dtype, data, s, n, rw, queue, NULL, NULL); +} + +/** + * read_block() - Issue sync read maybe using using a worker thread. + * @ti: Device Mapper Target Instance + * @dtype: Type of memory in data + * @data: Data for I/O + * @lba: bLBA for I/O [4k resolution] + * @count: Number of 4k blocks to read/write. + * @queue: if true then use worker thread for I/O and wait. + * + * Return 0 on success, otherwise error. + */ +static int read_block(struct dm_target *ti, enum dm_io_mem_type dtype, + void *data, u64 lba, unsigned int count, int queue) +{ + struct zoned *znd = ti->private; + sector_t block = lba << Z_SHFT4K; + unsigned int nDMsect = count << Z_SHFT4K; + int rc; + + if (lba >= znd->nr_blocks) { + Z_ERR(znd, "Error reading past end of media: %llx.", lba); + rc = -EIO; + return rc; + } + + rc = block_io(znd, dtype, data, block, nDMsect, READ, queue); + if (rc) { + Z_ERR(znd, "read error: %d -- R: %llx [%u dm sect] (Q:%d)", + rc, lba, nDMsect, queue); + dump_stack(); + } + + return rc; +} + +/** + * writef_block() - Issue sync write maybe using using a worker thread. + * @ti: Device Mapper Target Instance + * @dtype: Type of memory in data + * @data: Data for I/O + * @lba: bLBA for I/O [4k resolution] + * @count: Number of 4k blocks to read/write. + * @queue: if true then use worker thread for I/O and wait. + * + * Return 0 on success, otherwise error. + */ +static int writef_block(struct dm_target *ti, int rw, enum dm_io_mem_type dtype, + void *data, u64 lba, unsigned int count, int queue) +{ + struct zoned *znd = ti->private; + sector_t block = lba << Z_SHFT4K; + unsigned int nDMsect = count << Z_SHFT4K; + int rc; + + rc = block_io(znd, dtype, data, block, nDMsect, WRITE, queue); + if (rc) { + Z_ERR(znd, "write error: %d W: %llx [%u dm sect] (Q:%d)", + rc, lba, nDMsect, queue); + dump_stack(); + } + + return rc; +} + +/** + * write_block() - Issue sync write maybe using using a worker thread. + * @ti: Device Mapper Target Instance + * @dtype: Type of memory in data + * @data: Data for I/O + * @lba: bLBA for I/O [4k resolution] + * @count: Number of 4k blocks to read/write. + * @queue: if true then use worker thread for I/O and wait. + * + * Return 0 on success, otherwise error. + */ +static int write_block(struct dm_target *ti, enum dm_io_mem_type dtype, + void *data, u64 lba, unsigned int count, int queue) +{ + return writef_block(ti, WRITE, dtype, data, lba, count, queue); +} + +/** + * struct zsplit_hook - Extra data attached to a hooked bio + * @znd: ZDM Instance to update on BIO completion. + * @endio: BIO's original bi_end_io handler + * @private: BIO's original bi_private data. + */ +struct zsplit_hook { + struct zoned *znd; + bio_end_io_t *endio; + void *private; +}; + +/** + * hook_bio() - Wrapper for hooking bio's endio function. + * @znd: ZDM Instance + * @bio: Bio to clone and hook + * @endiofn: End IO Function to hook with. + */ +static int hook_bio(struct zoned *znd, struct bio *split, bio_end_io_t *endiofn) +{ + struct zsplit_hook *hook = kmalloc(sizeof(*hook), GFP_NOIO); + + if (!hook) + return -ENOMEM; + + /* + * On endio report back to ZDM Instance and restore + * original the bi_private and bi_end_io. + * Since all of our splits are also chain'd we also + * 'know' that bi_private will be the bio we sharded + * and that bi_end_io is the bio_chain_endio helper. + */ + hook->znd = znd; + hook->private = split->bi_private; /* = bio */ + hook->endio = split->bi_end_io; /* = bio_chain_endio */ + + /* + * Now on complete the bio will call endiofn which is 'zsplit_endio' + * and we can record the update WP location and restore the + * original bi_private and bi_end_io + */ + split->bi_private = hook; + split->bi_end_io = endiofn; + + return 0; +} + + +/** + * TODO: On write error such as this we can incr wp-used but we need + * to re-queue/re-map the write to a new location on disk? + * + * sd 0:0:1:0: [sdb] tag#1 FAILED Result: hostbyte=DID_SOFT_ERROR + * driverbyte=DRIVER_OK + * sd 0:0:1:0: [sdb] tag#1 + * CDB: Write(16) 8a 00 00 00 00 00 06 d4 92 a0 00 00 00 08 00 00 + * blk_update_request: I/O error, dev sdb, sector 114594464 + * exec scsi cmd failed,opcode:133 + * sdb: command 1 failed + * sd 0:0:1:0: [sdb] tag#1 + * CDB: Write(16) 8a 00 00 00 00 00 00 01 0a 70 00 00 00 18 00 00 + * mpt3sas_cm0: sas_address(0x4433221105000000), phy(5) + * mpt3sas_cm0: enclosure_logical_id(0x500605b0074854d0),slot(6) + * mpt3sas_cm0: enclosure level(0x0000), connector name( ) + * mpt3sas_cm0: handle(0x000a), ioc_status(success)(0x0000), smid(17) + * mpt3sas_cm0: request_len(12288), underflow(12288), resid(-1036288) + * mpt3sas_cm0: tag(65535), transfer_count(1048576), sc->result(0x00000000) + * mpt3sas_cm0: scsi_status(check condition)(0x02), + * scsi_state(autosense valid )(0x01) + * mpt3sas_cm0: [sense_key,asc,ascq]: [0x06,0x29,0x00], count(18) + * Aborting journal on device dm-0-8. + * EXT4-fs error (device dm-0): + * ext4_journal_check_start:56: Detected aborted journal + * EXT4-fs (dm-0): Remounting filesystem read-only + */ + +/** + * _common_endio() - Bio endio tracking for update internal WP. + * @bio: Bio being completed. + * + * Bios that are split for writing are usually split to land on a zone + * boundary. Forward the bio along the endio path and update the WP. + */ +static void _common_endio(struct zoned *znd, struct bio *bio) +{ + u64 lba = bio->bi_iter.bi_sector >> Z_SHFT4K; + u32 blks = bio->bi_iter.bi_size / Z_C4K; + + if (bio_data_dir(bio) == WRITE && lba > znd->start_sect) { + lba -= znd->start_sect; + if (lba > 0) + increment_used_blks(znd, lba - 1, blks + 1); + } +} + +/** + * zoned_endio() - DM bio completion notification. + * @ti: DM Target instance. + * @bio: Bio being completed. + * @err: Error associated with bio. + * + * Non-split and non-dm_io bios end notification is here. + * Update the WP location for WRITE bios. + */ +static int zoned_endio(struct dm_target *ti, struct bio *bio, int err) +{ + struct zoned *znd = ti->private; + + _common_endio(znd, bio); + return 0; +} + +/** + * zsplit_endio() - Bio endio tracking for update internal WP. + * @bio: Bio being completed. + * + * Bios that are split for writing are usually split to land on a zone + * boundary. Forward the bio along the endio path and update the WP. + */ +static void zsplit_endio(struct bio *bio) +{ + struct zsplit_hook *hook = bio->bi_private; + struct bio *parent = hook->private; + struct zoned *znd = hook->znd; + + _common_endio(znd, bio); + + bio->bi_private = hook->private; + bio->bi_end_io = hook->endio; + + /* On split bio's we are responsible for de-ref'ing and freeing */ + bio_put(bio); + if (parent) + bio_endio(parent); + + /* release our temporary private data */ + kfree(hook); +} + +/** + * zsplit_bio() - Split and chain a bio. + * @znd: ZDM Instance + * @bio: Bio to split + * @sectors: Number of sectors. + * + * Return: split bio. + */ +static struct bio *zsplit_bio(struct zoned *znd, struct bio *bio, int sectors) +{ + struct bio *split = bio; + + if (bio_sectors(bio) > sectors) { + split = bio_split(bio, sectors, GFP_NOIO, znd->bio_set); + if (!split) + goto out; + bio_chain(split, bio); + if (bio_data_dir(bio) == WRITE) + hook_bio(znd, split, zsplit_endio); + } +out: + return split; +} + +/** + * zm_cow() - Read Modify Write to write less than 4k size blocks. + * @znd: ZDM Instance + * @bio: Bio to write + * @s_zdm: tLBA + * @blks: number of blocks to RMW (should be 1). + * @origin: Current bLBA + * + * Return: 0 on success, otherwise error. + */ +static int zm_cow(struct zoned *znd, struct bio *bio, u64 s_zdm, u32 blks, + u64 origin) +{ + struct dm_target *ti = znd->ti; + int count = 1; + int use_wq = 1; + unsigned int bytes = bio_cur_bytes(bio); + u8 *data = bio_data(bio); + u8 *io = NULL; + u16 ua_off = bio->bi_iter.bi_sector & 0x0007; + u16 ua_size = bio->bi_iter.bi_size & 0x0FFF; /* in bytes */ + u32 mapped = 0; + u64 disk_lba = 0; + + if (!znd->cow_block) + znd->cow_block = ZDM_ALLOC(znd, Z_C4K, PG_02, CRIT); + + io = znd->cow_block; + if (!io) + return -EIO; + + disk_lba = z_acquire(znd, Z_AQ_STREAM_ID, blks, &mapped); + if (!disk_lba || !mapped) + return -ENOSPC; + + while (bytes) { + int ioer; + unsigned int iobytes = Z_C4K; + + /* ---------------------------------------------------------- */ + if (origin) { + if (s_zdm != znd->cow_addr) { + Z_ERR(znd, "Copy block from %llx <= %llx", + origin, s_zdm); + ioer = read_block(ti, DM_IO_KMEM, io, origin, + count, use_wq); + if (ioer) + return -EIO; + + znd->cow_addr = s_zdm; + } else { + Z_ERR(znd, "Cached block from %llx <= %llx", + origin, s_zdm); + } + } else { + memset(io, 0, Z_C4K); + } + + if (ua_off) + iobytes -= ua_off * 512; + + if (bytes < iobytes) + iobytes = bytes; + + Z_ERR(znd, "Moving %u bytes from origin [offset:%u]", + iobytes, ua_off * 512); + + memcpy(io + (ua_off * 512), data, iobytes); + + /* ---------------------------------------------------------- */ + + ioer = write_block(ti, DM_IO_KMEM, io, disk_lba, count, use_wq); + if (ioer) + return -EIO; + + MutexLock(&znd->mz_io_mutex); + ioer = z_mapped_addmany(znd, s_zdm, disk_lba, mapped, CRIT); + mutex_unlock(&znd->mz_io_mutex); + if (ioer) { + Z_ERR(znd, "%s: Journal MANY failed.", __func__); + return -EIO; + } + increment_used_blks(znd, disk_lba, mapped); + + data += iobytes; + bytes -= iobytes; + ua_size -= (ua_size > iobytes) ? iobytes : ua_size; + ua_off = 0; + disk_lba++; + + if (bytes && (ua_size || ua_off)) { + s_zdm++; + origin = current_mapping(znd, s_zdm, CRIT); + } + } + bio_endio(bio); + + return DM_MAPIO_SUBMITTED; +} + +/** + * Write 4k blocks from cache to lba. + * Move any remaining 512 byte blocks to the start of cache and update + * the @_blen count is updated + */ +static int zm_write_cache(struct zoned *znd, struct io_dm_block *dm_vbuf, + u64 lba, u32 *_blen) +{ + int use_wq = 1; + int cached = *_blen; + int blks = cached >> 3; + int sectors = blks << 3; + int remainder = cached - sectors; + int err; + + err = write_block(znd->ti, DM_IO_VMA, dm_vbuf, lba, blks, use_wq); + if (!err) { + if (remainder) + memcpy(dm_vbuf[0].data, + dm_vbuf[sectors].data, remainder * 512); + *_blen = remainder; + } + return err; +} + +/** + * zm_write_bios() - Map and write bios. + * @znd: ZDM Instance + * @bio: Bio to be written. + * @s_zdm: tLBA for mapping. + * + * Return: DM_MAPIO_SUBMITTED or negative on error. + */ +static int zm_write_bios(struct zoned *znd, struct bio *bio, u64 s_zdm) +{ + struct bio *split = NULL; + u32 acqflgs = Z_AQ_STREAM_ID | bio_stream(bio); + u64 lba = 0; + u32 mapped = 0; + int err = -EIO; + int done = 0; + int sectors; + u32 blks; + + do { + blks = dm_div_up(bio->bi_iter.bi_size, Z_C4K); + lba = z_acquire(znd, acqflgs, blks, &mapped); + if (!lba && mapped) + lba = z_acquire(znd, acqflgs, mapped, &mapped); + + if (!lba) { + if (atomic_read(&znd->gc_throttle) == 0) { + err = -ENOSPC; + goto out; + } + + Z_ERR(znd, "Throttle input ... Mandatory GC."); + if (delayed_work_pending(&znd->gc_work)) { + mod_delayed_work(znd->gc_wq, &znd->gc_work, 0); + flush_delayed_work(&znd->gc_work); + } + continue; + } + + sectors = mapped << Z_SHFT4K; + split = zsplit_bio(znd, bio, sectors); + if (split == bio) + done = 1; + + if (!split) { + err = -ENOMEM; + goto out; + } + split->bi_iter.bi_sector = lba << Z_SHFT4K; + generic_make_request(split); + MutexLock(&znd->mz_io_mutex); + err = z_mapped_addmany(znd, s_zdm, lba, mapped, CRIT); + mutex_unlock(&znd->mz_io_mutex); + if (err) { + Z_ERR(znd, "%s: Journal MANY failed.", __func__); + err = DM_MAPIO_REQUEUE; + goto out; + } + s_zdm += mapped; + } while (!done); + err = DM_MAPIO_SUBMITTED; + +out: + return err; +} + +/** + * zm_write_pages() - Copy bio pages to 4k aligned buffer. Write and map buffer. + * @znd: ZDM Instance + * @bio: Bio to be written. + * @s_zdm: tLBA for mapping. + * + * Return: DM_MAPIO_SUBMITTED or negative on error. + */ +static int zm_write_pages(struct zoned *znd, struct bio *bio, u64 s_zdm) +{ + u32 blks = dm_div_up(bio->bi_iter.bi_size, Z_C4K); + u64 lba = 0; + u32 blen = 0; /* total: IO_VCACHE_PAGES * 8 */ + u32 written = 0; + int avail = 0; + u32 acqflgs = Z_AQ_STREAM_ID | bio_stream(bio); + int err; + struct bvec_iter start; + struct bvec_iter iter; + struct bio_vec bv; + struct io_4k_block *io_vcache; + struct io_dm_block *dm_vbuf = NULL; + + MutexLock(&znd->vcio_lock); + io_vcache = get_io_vcache(znd, CRIT); + if (!io_vcache) { + Z_ERR(znd, "%s: FAILED to get SYNC CACHE.", __func__); + err = -ENOMEM; + goto out; + } + + dm_vbuf = (struct io_dm_block *)io_vcache; + + /* USE: dm_vbuf for dumping bio pages to disk ... */ + start = bio->bi_iter; /* struct implicit copy */ + do { + u64 alloc_ori = 0; + u32 mcount = 0; + u32 mapped = 0; + +reacquire: + /* + * When lba is zero no blocks were not allocated. + * Retry with the smaller request + */ + lba = z_acquire(znd, acqflgs, blks - written, &mapped); + if (!lba && mapped) + lba = z_acquire(znd, acqflgs, mapped, &mapped); + + if (!lba) { + if (atomic_read(&znd->gc_throttle) == 0) { + err = -ENOSPC; + goto out; + } + + Z_ERR(znd, "Throttle input ... Mandatory GC."); + if (delayed_work_pending(&znd->gc_work)) { + mod_delayed_work(znd->gc_wq, &znd->gc_work, 0); + flush_delayed_work(&znd->gc_work); + } + goto reacquire; + } + + /* this may be redundant .. if we have lba we have mapped > 0 */ + if (lba && mapped) + avail += mapped * 8; /* claimed pages in dm blocks */ + + alloc_ori = lba; + + /* copy [upto mapped] pages to buffer */ + __bio_for_each_segment(bv, bio, iter, start) { + int issue_write = 0; + unsigned int boff; + void *src; + + if (avail <= 0) { + Z_ERR(znd, "%s: TBD: Close Z# %llu", + __func__, alloc_ori >> 16); + start = iter; + break; + } + + src = kmap_atomic(bv.bv_page); + boff = bv.bv_offset; + memcpy(dm_vbuf[blen].data, src + boff, bv.bv_len); + kunmap_atomic(src); + blen += bv.bv_len / 512; + avail -= bv.bv_len / 512; + + if ((blen >= (mapped * 8)) || + (blen >= (BIO_CACHE_SECTORS - 8))) + issue_write = 1; + + /* + * If there is less than 1 4k block in out cache, + * send the available blocks to disk + */ + if (issue_write) { + int blks = blen / 8; + + err = zm_write_cache(znd, dm_vbuf, lba, &blen); + if (err) { + Z_ERR(znd, "%s: bio-> %" PRIx64 + " [%d of %d blks] -> %d", + __func__, lba, blen, blks, err); + bio->bi_error = err; + bio_endio(bio); + goto out; + } + lba += blks; + written += blks; + mcount += blks; + mapped -= blks; + + if (mapped == 0) { + bio_advance_iter(bio, &iter, bv.bv_len); + start = iter; + break; + } + if (mapped < 0) { + Z_ERR(znd, "ERROR: Bad write %" + PRId32 " beyond alloc'd space", + mapped); + } + } + } /* end: __bio_for_each_segment */ + if ((mapped > 0) && ((blen / 8) > 0)) { + int blks = blen / 8; + + err = zm_write_cache(znd, dm_vbuf, lba, &blen); + if (err) { + Z_ERR(znd, "%s: bio-> %" PRIx64 + " [%d of %d blks] -> %d", + __func__, lba, blen, blks, err); + bio->bi_error = err; + bio_endio(bio); + goto out; + } + lba += blks; + written += blks; + mcount += blks; + mapped -= blks; + + if (mapped < 0) { + Z_ERR(znd, "ERROR: [2] Bad write %" + PRId32 " beyond alloc'd space", + mapped); + } + } + MutexLock(&znd->mz_io_mutex); + err = z_mapped_addmany(znd, s_zdm, alloc_ori, mcount, CRIT); + mutex_unlock(&znd->mz_io_mutex); + if (err) { + Z_ERR(znd, "%s: Journal MANY failed.", __func__); + err = DM_MAPIO_REQUEUE; + /* + * FIXME: + * Ending the BIO here is causing a GFP: + - DEBUG_PAGEALLOC + - in Workqueue: + - writeback bdi_writeback_workfn (flush-252:0) + - backtrace: + - __map_bio+0x7a/0x280 + - __split_and_process_bio+0x2e3/0x4e0 + - ? __split_and_process_bio+0x22/0x4e0 + - ? generic_start_io_acct+0x5/0x210 + - dm_make_request+0x6b/0x100 + - generic_make_request+0xc0/0x110 + - .... + - + - bio->bi_error = err; + - bio_endio(bio); + */ + goto out; + } + increment_used_blks(znd, alloc_ori, mcount); + + if (written < blks) + s_zdm += written; + + if (written == blks && blen > 0) + Z_ERR(znd, "%s: blen: %d un-written blocks!!", + __func__, blen); + } while (written < blks); + bio_endio(bio); + err = DM_MAPIO_SUBMITTED; + +out: + put_io_vcache(znd, io_vcache); + mutex_unlock(&znd->vcio_lock); + + return err; +} + +/** + * is_empty_page() - Scan memory range for any set bits. + * @pg: The start of memory to be scanned. + * @len: Number of bytes to check (should be long aligned) + * Return: 0 if any bits are set, 1 if all bits are 0. + */ +static int is_empty_page(void *pg, size_t len) +{ + unsigned long *chk = pg; + size_t count = len / sizeof(*chk); + size_t entry; + + for (entry = 0; entry < count; entry++) { + if (chk[entry]) + return 0; + } + return 1; +} + +/** + * is_zero_bio() - Scan bio to see if all bytes are 0. + * @bio: The bio to be scanned. + * Return: 1 if all bits are 0. 0 if any bits in bio are set. + */ +static int is_zero_bio(struct bio *bio) +{ + int is_empty = 0; + struct bvec_iter iter; + struct bio_vec bv; + + /* Scan bio to determine if it is zero'd */ + bio_for_each_segment(bv, bio, iter) { + unsigned int boff; + void *src; + + src = kmap_atomic(bv.bv_page); + boff = bv.bv_offset; + is_empty = is_empty_page(src + boff, bv.bv_len); + kunmap_atomic(src); + + if (!is_empty) + break; + } /* end: __bio_for_each_segment */ + + return is_empty; +} + +/** + * is_bio_aligned() - Test bio and bio_vec for 4k aligned pages. + * @bio: Bio to be tested. + * Return: 1 if bio is 4k aligned, 0 if not. + */ +static int is_bio_aligned(struct bio *bio) +{ + int aligned = 1; + struct bvec_iter iter; + struct bio_vec bv; + + bio_for_each_segment(bv, bio, iter) { + if ((bv.bv_offset & 0x0FFF) || (bv.bv_len & 0x0FFF)) { + aligned = 0; + break; + } + } + return aligned; +} + +/** + * zoned_map_write() - Write a bio by the fastest safe method. + * @znd: ZDM Instance + * @bio: Bio to be written + * @s_zdm: tLBA for mapping. + * + * Bios that are less than 4k need RMW. + * Bios that are single pages are deduped and written or discarded. + * Bios that are multiple pages with 4k aligned bvecs are written as bio(s). + * Biso that are multiple pages and mis-algined are copied to an algined buffer + * and submitted and new I/O. + */ +static int zoned_map_write(struct zoned *znd, struct bio *bio, u64 s_zdm) +{ + u32 blks = dm_div_up(bio->bi_iter.bi_size, Z_C4K); + u16 ua_off = bio->bi_iter.bi_sector & 0x0007; + u16 ua_size = bio->bi_iter.bi_size & 0x0FFF; /* in bytes */ + int rcode = -EIO; + + if (ua_size || ua_off) { + u64 origin; + + origin = current_mapping(znd, s_zdm, CRIT); + if (origin) { + znd->is_empty = 0; + rcode = zm_cow(znd, bio, s_zdm, blks, origin); + } + return rcode; + } + + /* on RAID 4/5/6 all writes are 4k */ + if (blks == 1) { + if (is_zero_bio(bio)) { + rcode = zoned_map_discard(znd, bio, s_zdm); + } else { + znd->is_empty = 0; + rcode = zm_write_bios(znd, bio, s_zdm); + } + return rcode; + } + + /* + * For larger bios test for 4k alignment. + * When bios are mis-algined we must copy out the + * the mis-algined pages into a new bio and submit. + * [The 4k alignment requests on our queue may be ignored + * by mis-behaving layers that are not 4k safe]. + */ + znd->is_empty = 0; + if (is_bio_aligned(bio)) + rcode = zm_write_bios(znd, bio, s_zdm); + else + rcode = zm_write_pages(znd, bio, s_zdm); + + return rcode; +} + +/** + * zm_read_bios() - Read bios from device + * @znd: ZDM Instance + * @bio: Bio to read + * @s_zdm: tLBA to read from. + * + * Return DM_MAPIO_SUBMITTED or negative on error. + */ +static int zm_read_bios(struct zoned *znd, struct bio *bio, u64 s_zdm) +{ + struct bio *split = NULL; + int rcode = DM_MAPIO_SUBMITTED; + u64 lba; + u64 blba; + u32 blks; + int sectors; + int count; + u16 ua_off; + u16 ua_size; + + do { + blks = dm_div_up(bio->bi_iter.bi_size, Z_C4K); + ua_off = bio->bi_iter.bi_sector & 0x0007; + ua_size = bio->bi_iter.bi_size & 0x0FFF; /* in bytes */ + + lba = blba = current_mapping(znd, s_zdm, CRIT); + if (blba) { + for (count = 1; count < blks; count++) { + lba = current_mapping(znd, s_zdm + count, CRIT); + if (lba != (blba + count)) + break; + } + } else { + for (count = 1; count < blks; count++) { + lba = current_mapping(znd, s_zdm + count, CRIT); + if (lba != 0ul) + break; + } + } + s_zdm += count; + sectors = (count << Z_SHFT4K); + if (ua_size) + sectors += (ua_size >> SECTOR_SHIFT) - 8; + + split = zsplit_bio(znd, bio, sectors); + if (!split) { + rcode = -ENOMEM; + goto out; + } + if (blba) { + split->bi_iter.bi_sector = (blba << Z_SHFT4K) + ua_off; + generic_make_request(split); + } else { + zero_fill_bio(split); + bio_endio(split); + } + } while (split != bio); + +out: + return rcode; +} + +/** + * zoned_bio() - Handle and incoming BIO. + * @znd: ZDM Instance + */ +static int zoned_bio(struct zoned *znd, struct bio *bio) +{ + bool is_write = (bio_data_dir(bio) == WRITE); + u64 s_zdm = (bio->bi_iter.bi_sector >> Z_SHFT4K) + znd->md_end; + int rcode = DM_MAPIO_SUBMITTED; + struct request_queue *q; + int force_sync_now = 0; + + /* map to backing device ... NOT dm-zoned device */ + bio->bi_bdev = znd->dev->bdev; + + q = bdev_get_queue(bio->bi_bdev); + q->queue_flags |= QUEUE_FLAG_NOMERGES; + + if (is_write && znd->meta_result) { + if (!(bio->bi_rw & REQ_DISCARD)) { + rcode = znd->meta_result; + Z_ERR(znd, "MAP ERR (meta): %d", rcode); + goto out; + } + } + + /* check for REQ_FLUSH flag */ + if (bio->bi_rw & REQ_FLUSH) { + bio->bi_rw &= ~(REQ_FLUSH); + set_bit(DO_FLUSH, &znd->flags); + force_sync_now = 1; + } + if (bio->bi_rw & REQ_SYNC) { + set_bit(DO_SYNC, &znd->flags); + force_sync_now = 1; + } + + Z_DBG(znd, "%s: s:%"PRIx64" sz:%u -> %s", __func__, s_zdm, + bio->bi_iter.bi_size, is_write ? "W" : "R"); + + if (bio->bi_iter.bi_size) { + if (bio->bi_rw & REQ_DISCARD) { + rcode = zoned_map_discard(znd, bio, s_zdm); + } else if (is_write) { + const int wait = 1; + u32 blks = bio->bi_iter.bi_size >> 12; + + rcode = zoned_map_write(znd, bio, s_zdm); + if (rcode == DM_MAPIO_SUBMITTED) { + z_discard_partial(znd, blks, CRIT); + if (znd->z_gc_free < 5) + gc_immediate(znd, !wait, CRIT); + } + } else { + rcode = zm_read_bios(znd, bio, s_zdm); + } + znd->age = jiffies; + } + + if (test_bit(DO_FLUSH, &znd->flags) || + test_bit(DO_SYNC, &znd->flags) || + test_bit(DO_JOURNAL_MOVE, &znd->flags) || + test_bit(DO_MEMPOOL, &znd->flags)) { + if (!test_bit(DO_METAWORK_QD, &znd->flags) && + !work_pending(&znd->meta_work)) { + set_bit(DO_METAWORK_QD, &znd->flags); + queue_work(znd->meta_wq, &znd->meta_work); + } + } + + if (force_sync_now && work_pending(&znd->meta_work)) + flush_workqueue(znd->meta_wq); + +out: + if (rcode == DM_MAPIO_REMAPPED || rcode == DM_MAPIO_SUBMITTED) + goto done; + + if (rcode < 0 || rcode == DM_MAPIO_REQUEUE) { + Z_ERR(znd, "MAP ERR: %d", rcode); + Z_ERR(znd, "%s: s:%"PRIx64" sz:%u -> %s", __func__, s_zdm, + bio->bi_iter.bi_size, is_write ? "W" : "R"); + dump_stack(); + } else { + Z_ERR(znd, "MAP ERR: %d", rcode); + Z_ERR(znd, "%s: s:%"PRIx64" sz:%u -> %s", __func__, s_zdm, + bio->bi_iter.bi_size, is_write ? "W" : "R"); + dump_stack(); + rcode = -EIO; + } +done: + return rcode; +} + +/** + * _do_mem_purge() - conditionally trigger a reduction of cache memory + * @znd: ZDM Instance + */ +static inline int _do_mem_purge(struct zoned *znd) +{ + int do_work = 0; + + if (atomic_read(&znd->incore) > 3) { + set_bit(DO_MEMPOOL, &znd->flags); + if (!work_pending(&znd->meta_work)) + do_work = 1; + } + return do_work; +} + +/** + * on_timeout_activity() - Periodic background task execution. + * @znd: ZDM Instance + * @mempurge: If memory purge should be scheduled. + * @delay: Delay metric for periodic GC + * + * NOTE: Executed as a worker task queued froma timer. + */ +static void on_timeout_activity(struct zoned *znd, int delay) +{ + int max_tries = 1; + + if (test_bit(ZF_FREEZE, &znd->flags)) + return; + + if (is_expired_msecs(znd->age, DISCARD_IDLE_MSECS)) + max_tries = 20; + + do { + int count; + + count = z_discard_partial(znd, Z_BLKSZ, NORMAL); + if (count != 1 || --max_tries < 0) + break; + + if (test_bit(ZF_FREEZE, &znd->flags)) + return; + + } while (is_expired_msecs(znd->age, DISCARD_IDLE_MSECS)); + + gc_queue_with_delay(znd, delay, NORMAL); + + if (!test_bit(DO_GC_NO_PURGE, &znd->flags) && _do_mem_purge(znd)) + queue_work(znd->meta_wq, &znd->meta_work); +} + +/** + * bg_work_task() - periodic background worker + * @work: context for worker thread + */ +static void bg_work_task(struct work_struct *work) +{ + struct zoned *znd; + const int delay = 1; + + if (!work) + return; + + znd = container_of(work, struct zoned, bg_work); + on_timeout_activity(znd, delay); +} + +/** + * activity_timeout() - Handler for timer used to trigger background worker. + * @data: context for timer. + */ +static void activity_timeout(unsigned long data) +{ + struct zoned *znd = (struct zoned *) data; + + if (!work_pending(&znd->bg_work)) + queue_work(znd->bg_wq, &znd->bg_work); + + if (!test_bit(ZF_FREEZE, &znd->flags)) + mod_timer(&znd->timer, jiffies + msecs_to_jiffies(2500)); +} + +/** + * get_dev_size() - Report accessible size of device to upper layer. + * @ti: DM Target + * + * Return: Size in 512 byte sectors + */ +static sector_t get_dev_size(struct dm_target *ti) +{ + struct zoned *znd = ti->private; + u64 sz = i_size_read(get_bdev_bd_inode(znd)); /* size in bytes. */ + u64 lut_resv = znd->gz_count * znd->mz_provision; + + /* + * NOTE: `sz` should match `ti->len` when the dm_table + * is setup correctly + */ + sz -= (lut_resv * Z_SMR_SZ_BYTES); + + return to_sector(sz); +} + +/** + * zoned_iterate_devices() - Iterate over devices call fn() at each. + * @ti: DM Target + * @fn: Function for each callout + * @data: Context for fn(). + */ +static int zoned_iterate_devices(struct dm_target *ti, + iterate_devices_callout_fn fn, void *data) +{ + struct zoned *znd = ti->private; + int rc; + + rc = fn(ti, znd->dev, 0, get_dev_size(ti), data); + return rc; +} + +/** + * zoned_io_hints() - The place to tweek queue limits for DM targets + * @ti: DM Target + * @limits: queue_limits for this DM target + */ +static void zoned_io_hints(struct dm_target *ti, struct queue_limits *limits) +{ + struct zoned *znd = ti->private; + u64 io_opt_sectors = limits->io_opt >> SECTOR_SHIFT; + + /* + * If the system-determined stacked limits are compatible with the + * zoned device's blocksize (io_opt is a factor) do not override them. + */ + if (io_opt_sectors < 8 || do_div(io_opt_sectors, 8)) { + blk_limits_io_min(limits, 0); + blk_limits_io_opt(limits, 8 << SECTOR_SHIFT); + } + + limits->logical_block_size = + limits->physical_block_size = + limits->io_min = Z_C4K; + if (znd->trim) { + limits->discard_alignment = Z_C4K; + limits->discard_granularity = Z_C4K; + limits->max_discard_sectors = 1 << 30; + limits->max_hw_discard_sectors = 1 << 30; + limits->discard_zeroes_data = 1; + } +} + +/** + * zoned_status() - Report status of DM Target + * @ti: DM Target + * @type: Type of status to report. + * @status_flags: Flags + * @result: Fill in with status. + * @maxlen: Maximum number of bytes for result. + */ +static void zoned_status(struct dm_target *ti, status_type_t type, + unsigned status_flags, char *result, unsigned maxlen) +{ + struct zoned *znd = (struct zoned *) ti->private; + + switch (type) { + case STATUSTYPE_INFO: + result[0] = '\0'; + break; + + case STATUSTYPE_TABLE: + scnprintf(result, maxlen, "%s Z#%u", znd->dev->name, + znd->zdstart); + break; + } +} + +/* -------------------------------------------------------------------------- */ +/* --ProcFS Support Routines------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +#if defined(CONFIG_PROC_FS) + +/** + * struct zone_info_entry - Proc zone entry. + * @zone: Zone Index + * @info: Info (WP/Used). + */ +struct zone_info_entry { + u32 zone; + u32 info; +}; + +/** + * Startup writing to our proc entry + */ +static void *proc_wp_start(struct seq_file *seqf, loff_t *pos) +{ + struct zoned *znd = seqf->private; + + if (*pos == 0) + znd->wp_proc_at = *pos; + return &znd->wp_proc_at; +} + +/** + * Increment to our next 'grand zone' 4k page. + */ +static void *proc_wp_next(struct seq_file *seqf, void *v, loff_t *pos) +{ + struct zoned *znd = seqf->private; + u32 zone = ++znd->wp_proc_at; + + return zone < znd->data_zones ? &znd->wp_proc_at : NULL; +} + +/** + * Stop ... a place to free resources that we don't hold .. [noop]. + */ +static void proc_wp_stop(struct seq_file *seqf, void *v) +{ +} + +/** + * Write as many entries as possbile .... + */ +static int proc_wp_show(struct seq_file *seqf, void *v) +{ + int err = 0; + struct zoned *znd = seqf->private; + u32 zone = znd->wp_proc_at; + u32 out = 0; + + while (zone < znd->data_zones) { + u32 gzno = zone >> GZ_BITS; + u32 gzoff = zone & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + struct zone_info_entry entry; + + entry.zone = zone; + entry.info = le32_to_cpu(wpg->wp_alloc[gzoff]); + + err = seq_write(seqf, &entry, sizeof(entry)); + if (err) { + /* + * write failure is temporary .. + * just return and try again + */ + err = 0; + goto out; + } + out++; + zone = ++znd->wp_proc_at; + } + +out: + if (err) + Z_ERR(znd, "%s: %llu -> %d", __func__, znd->wp_proc_at, err); + + return err; +} + +/** + * zdm_wp_ops() - Seq_file operations for retrieving WP via proc fs + */ +static const struct seq_operations zdm_wp_ops = { + .start = proc_wp_start, + .next = proc_wp_next, + .stop = proc_wp_stop, + .show = proc_wp_show +}; + +/** + * zdm_wp_open() - Need to migrate our private data to the seq_file + */ +static int zdm_wp_open(struct inode *inode, struct file *file) +{ + /* seq_open will populate file->private_data with a seq_file */ + int err = seq_open(file, &zdm_wp_ops); + + if (!err) { + struct zoned *znd = PDE_DATA(inode); + struct seq_file *seqf = file->private_data; + + seqf->private = znd; + } + return err; +} + +/** + * zdm_wp_fops() - File operations for retrieving WP via proc fs + */ +static const struct file_operations zdm_wp_fops = { + .open = zdm_wp_open, + .read = seq_read, + .llseek = seq_lseek, + .release = seq_release, +}; + + +/** + * Startup writing to our proc entry + */ +static void *proc_used_start(struct seq_file *seqf, loff_t *pos) +{ + struct zoned *znd = seqf->private; + + if (*pos == 0) + znd->wp_proc_at = *pos; + return &znd->wp_proc_at; +} + +/** + * Increment to our next zone + */ +static void *proc_used_next(struct seq_file *seqf, void *v, loff_t *pos) +{ + struct zoned *znd = seqf->private; + u32 zone = ++znd->wp_proc_at; + + return zone < znd->data_zones ? &znd->wp_proc_at : NULL; +} + +/** + * Stop ... a place to free resources that we don't hold .. [noop]. + */ +static void proc_used_stop(struct seq_file *seqf, void *v) +{ +} + +/** + * proc_used_show() - Write as many 'used' entries as possbile. + * @seqf: seq_file I/O handler + * @v: An unused parameter. + */ +static int proc_used_show(struct seq_file *seqf, void *v) +{ + int err = 0; + struct zoned *znd = seqf->private; + u32 zone = znd->wp_proc_at; + u32 out = 0; + + while (zone < znd->data_zones) { + u32 gzno = zone >> GZ_BITS; + u32 gzoff = zone & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + struct zone_info_entry entry; + + entry.zone = zone; + entry.info = le32_to_cpu(wpg->zf_est[gzoff]); + + err = seq_write(seqf, &entry, sizeof(entry)); + if (err) { + /* + * write failure is temporary .. + * just return and try again + */ + err = 0; + goto out; + } + out++; + zone = ++znd->wp_proc_at; + } + +out: + if (err) + Z_ERR(znd, "%s: %llu -> %d", __func__, znd->wp_proc_at, err); + + return err; +} + +/** + * zdm_used_ops() - Seq_file Ops for retrieving 'used' state via proc fs + */ +static const struct seq_operations zdm_used_ops = { + .start = proc_used_start, + .next = proc_used_next, + .stop = proc_used_stop, + .show = proc_used_show +}; + +/** + * zdm_used_open() - Need to migrate our private data to the seq_file + */ +static int zdm_used_open(struct inode *inode, struct file *file) +{ + /* seq_open will populate file->private_data with a seq_file */ + int err = seq_open(file, &zdm_used_ops); + + if (!err) { + struct zoned *znd = PDE_DATA(inode); + struct seq_file *seqf = file->private_data; + + seqf->private = znd; + } + return err; +} + +/** + * zdm_used_fops() - File operations for retrieving 'used' state via proc fs + */ +static const struct file_operations zdm_used_fops = { + .open = zdm_used_open, + .read = seq_read, + .llseek = seq_lseek, + .release = seq_release, +}; + +/** + * zdm_status_show() - Dump the status structure via proc fs + */ +static int zdm_status_show(struct seq_file *seqf, void *unused) +{ + struct zoned *znd = seqf->private; + struct zdm_ioc_status status; + u32 zone; + + memset(&status, 0, sizeof(status)); + + for (zone = 0; zone < znd->data_zones; zone++) { + u32 gzno = zone >> GZ_BITS; + u32 gzoff = zone & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp_at = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_VALUE_MASK; + + status.b_used += wp_at; + status.b_available += Z_BLKSZ - wp_at; + } + status.mc_entries = znd->mc_entries; + status.dc_entries = znd->dc_entries; + status.b_discard = znd->discard_count; + + /* fixed array of ->fwd_tm and ->rev_tm */ + status.m_zones = znd->data_zones; + + status.memstat = znd->memstat; + memcpy(status.bins, znd->bins, sizeof(status.bins)); + status.mlut_blocks = atomic_read(&znd->incore); + + return seq_write(seqf, &status, sizeof(status)); +} + +/** + * zdm_status_open() - Open seq_file from file. + * @inode: Our data is stuffed here, Retrieve it. + * @file: file objected used by seq_file. + */ +static int zdm_status_open(struct inode *inode, struct file *file) +{ + return single_open(file, zdm_status_show, PDE_DATA(inode)); +} + +/** + * zdm_used_fops() - File operations to chain to zdm_status_open. + */ +static const struct file_operations zdm_status_fops = { + .open = zdm_status_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + +/** + * zdm_info_show() - Report some information as text. + * @seqf: Sequence file for writing + * @unused: Not used. + */ +static int zdm_info_show(struct seq_file *seqf, void *unused) +{ + struct zoned *znd = seqf->private; + + seq_printf(seqf, "On device: %s\n", _zdisk(znd)); + seq_printf(seqf, "Data Zones: %u\n", znd->data_zones); + seq_printf(seqf, "Empty Zones: %u\n", znd->z_gc_free); + seq_printf(seqf, "Cached Pages: %u\n", znd->mc_entries); + seq_printf(seqf, "Discard Pages: %u\n", znd->dc_entries); + seq_printf(seqf, "ZTL Pages: %d\n", atomic_read(&znd->incore)); + seq_printf(seqf, " in ZTL: %d\n", znd->in_zlt); + seq_printf(seqf, " in LZY: %d\n", znd->in_lzy); + seq_printf(seqf, "RAM in Use: %lu\n", znd->memstat); + seq_printf(seqf, "Zones GC'd: %u\n", znd->gc_events); + seq_printf(seqf, "GC Throttle: %d\n", atomic_read(&znd->gc_throttle)); +#if ALLOC_DEBUG + seq_printf(seqf, "Max Allocs: %u\n", znd->hw_allocs); +#endif + + return 0; +} + +/** + * zdm_info_open() - Open seq_file from file. + * @inode: Our data is stuffed here, Retrieve it. + * @file: file objected used by seq_file. + */ +static int zdm_info_open(struct inode *inode, struct file *file) +{ + return single_open(file, zdm_info_show, PDE_DATA(inode)); +} + +/** + * zdm_used_fops() - File operations to chain to zdm_info_open. + */ +static const struct file_operations zdm_info_fops = { + .open = zdm_info_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + +/** + * zdm_create_proc_entries() - Create proc entries for ZDM utilities + * @znd: ZDM Instance + */ +static int zdm_create_proc_entries(struct zoned *znd) +{ + snprintf(znd->proc_name, sizeof(znd->proc_name), "zdm_%s", _zdisk(znd)); + + znd->proc_fs = proc_mkdir(znd->proc_name, NULL); + if (!znd->proc_fs) + return -ENOMEM; + + proc_create_data(PROC_WP, 0, znd->proc_fs, &zdm_wp_fops, znd); + proc_create_data(PROC_FREE, 0, znd->proc_fs, &zdm_used_fops, znd); + proc_create_data(PROC_DATA, 0, znd->proc_fs, &zdm_status_fops, znd); + proc_create_data(PROC_STATUS, 0, znd->proc_fs, &zdm_info_fops, znd); + + return 0; +} + +/** + * zdm_remove_proc_entries() - Remove proc entries + * @znd: ZDM Instance + */ +static void zdm_remove_proc_entries(struct zoned *znd) +{ + remove_proc_subtree(znd->proc_name, NULL); +} + +#else /* !CONFIG_PROC_FS */ + +static int zdm_create_proc_entries(struct zoned *znd) +{ + (void)znd; + return 0; +} +static void zdm_remove_proc_entries(struct zoned *znd) +{ + (void)znd; +} +#endif /* CONFIG_PROC_FS */ + + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static void start_worker(struct zoned *znd) +{ + clear_bit(ZF_FREEZE, &znd->flags); + atomic_set(&znd->suspended, 0); + mod_timer(&znd->timer, jiffies + msecs_to_jiffies(5000)); +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static void stop_worker(struct zoned *znd) +{ + set_bit(ZF_FREEZE, &znd->flags); + atomic_set(&znd->suspended, 1); + zoned_io_flush(znd); +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static void zoned_postsuspend(struct dm_target *ti) +{ + struct zoned *znd = ti->private; + + stop_worker(znd); +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static void zoned_resume(struct dm_target *ti) +{ + /* TODO */ +} + +/* + * + */ +static int zoned_message(struct dm_target *ti, unsigned argc, char **argv) +{ + struct zoned *znd = ti->private; + int iter; + + for (iter = 0; iter < argc; iter++) + Z_ERR(znd, "Message: %s not handled.", argv[argc]); + + return 0; +} + + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static int zoned_preresume(struct dm_target *ti) +{ + struct zoned *znd = ti->private; + + start_worker(znd); + return 0; +} +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static struct target_type zoned_target = { + .name = "zoned", + .module = THIS_MODULE, + .version = {1, 0, 0}, + .ctr = zoned_ctr, + .dtr = zoned_dtr, + .map = zoned_map, + .end_io = zoned_endio, + .postsuspend = zoned_postsuspend, + .preresume = zoned_preresume, + .resume = zoned_resume, + .status = zoned_status, + .message = zoned_message, + .iterate_devices = zoned_iterate_devices, + .io_hints = zoned_io_hints +}; + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static int __init dm_zoned_init(void) +{ + int rcode = dm_register_target(&zoned_target); + + if (rcode) + DMERR("zoned target registration failed: %d", rcode); + + return rcode; +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ +static void __exit dm_zoned_exit(void) +{ + dm_unregister_target(&zoned_target); +} + +module_init(dm_zoned_init); +module_exit(dm_zoned_exit); + +MODULE_DESCRIPTION(DM_NAME " zoned target for Host Aware/Managed drives."); +MODULE_AUTHOR("Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx>"); +MODULE_LICENSE("GPL"); diff --git a/drivers/md/dm-zoned.h b/drivers/md/dm-zoned.h new file mode 100644 index 0000000..c73ddea --- /dev/null +++ b/drivers/md/dm-zoned.h @@ -0,0 +1,714 @@ +/* + * Kernel Device Mapper for abstracting ZAC/ZBC devices as normal + * block devices for linux file systems. + * + * Copyright (C) 2015 Seagate Technology PLC + * + * Written by: + * Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx> + * + * This file is licensed under the terms of the GNU General Public + * License version 2. This program is licensed "as is" without any + * warranty of any kind, whether express or implied. + */ + +#ifndef _DM_ZONED_H +#define _DM_ZONED_H + +#define ALLOC_DEBUG 0 +#define USE_KTHREAD 0 +#define KFIFO_SIZE (1 << 14) + +#define NORMAL 0 +#define CRIT GFP_NOIO + +#define DM_MSG_PREFIX "zoned" + +#define PROC_WP "wp.bin" +#define PROC_FREE "free.bin" +#define PROC_DATA "data.bin" +#define PROC_STATUS "status" + +#define ZDM_RESERVED_ZNR 0 +#define ZDM_CRC_STASH_ZNR 1 /* first 64 blocks */ +#define ZDM_RMAP_ZONE 2 +#define ZDM_SECTOR_MAP_ZNR 3 +#define ZDM_DATA_START_ZNR 4 + +#define Z_WP_GC_FULL (1u << 31) +#define Z_WP_GC_ACTIVE (1u << 30) +#define Z_WP_GC_TARGET (1u << 29) +#define Z_WP_GC_READY (1u << 28) +#define Z_WP_GC_BITS (0xFu << 28) + +#define Z_WP_GC_PENDING (Z_WP_GC_FULL|Z_WP_GC_ACTIVE) +#define Z_WP_NON_SEQ (1u << 27) +#define Z_WP_RRECALC (1u << 26) +#define Z_WP_RESV_02 (1u << 25) +#define Z_WP_RESV_03 (1u << 24) + +#define Z_WP_VALUE_MASK (~0u >> 8) +#define Z_WP_FLAGS_MASK (~0u << 24) +#define Z_WP_STREAM_MASK Z_WP_FLAGS_MASK + +#define Z_AQ_GC (1u << 31) +#define Z_AQ_META (1u << 30) +#define Z_AQ_NORMAL (1u << 29) +#define Z_AQ_STREAM_ID (1u << 28) +#define Z_AQ_STREAM_MASK (0xFF) +#define Z_AQ_META_STREAM (Z_AQ_META | Z_AQ_STREAM_ID | 0xFE) + +#define Z_C4K (4096ul) +#define Z_UNSORTED (Z_C4K / sizeof(struct map_sect_to_lba)) +#define Z_BLOCKS_PER_DM_SECTOR (Z_C4K/512) +#define MZ_METADATA_ZONES (8ul) +#define Z_SHFT4K (3) + + +#define LBA_SB_START 1 + +#define SUPERBLOCK_LOCATION 0 +#define SUPERBLOCK_MAGIC 0x5a6f4e65ul /* ZoNe */ +#define SUPERBLOCK_CSUM_XOR 146538381 +#define MIN_ZONED_VERSION 1 +#define Z_VERSION 1 +#define MAX_ZONED_VERSION 1 +#define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION + +#define UUID_LEN 16 + +#define Z_TYPE_SMR 2 +#define Z_TYPE_SMR_HA 1 +#define Z_VPD_INFO_BYTE 8 + +#define MAX_CACHE_INCR 320ul +#define CACHE_COPIES 3 +#define MAX_MZ_SUPP 64 +#define FWD_TM_KEY_BASE 4096ul + +#define IO_VCACHE_ORDER 8 +#define IO_VCACHE_PAGES (1 << IO_VCACHE_ORDER) /* 256 pages => 1MiB */ + +#ifdef __cplusplus +extern "C" { +#endif + + +enum superblock_flags_t { + SB_DIRTY = 1, +}; + +struct z_io_req_t { + struct dm_io_region *where; + struct dm_io_request *io_req; + struct work_struct work; + int result; +}; + +#define Z_LOWER48 (~0ul >> 16) +#define Z_UPPER16 (~Z_LOWER48) + +#define STREAM_SIZE 256 + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +/** + * enum pg_flag_enum - Map Pg flags + * @IS_DIRTY: Block is modified from on-disk copy. + * @IS_STALE: ?? + * + * @IS_FWD: Is part of Forward ZLT or CRC table. + * @IS_REV: Is part of Reverse ZLT or CRC table. + * @IS_CRC: Is part of CRC table. + * @IS_LUT: Is part of ZTL table. + * + * @R_IN_FLIGHT: Async read in progress. + * @W_IN_FLIGHT: ASync write in progress. + * @DELAY_ADD: Spinlock was busy for zltlst, added to lazy for transit. + * @STICKY: ??? + * @IS_READA: Block will pulled for Read Ahead. Cleared when used. + * @IS_DROPPED: Has clean/expired from zlt_lst and is waiting free(). + * + * @IS_LAZY: Had been added to 'lazy' lzy_lst. + * @IN_ZLT: Had been added to 'inpool' zlt_lst. + * + */ +enum pg_flag_enum { + IS_DIRTY, + IS_STALE, + IS_FLUSH, + + IS_FWD, + IS_REV, + IS_CRC, + IS_LUT, + + R_IN_FLIGHT, + W_IN_FLIGHT, + DELAY_ADD, + STICKY, + IS_READA, + IS_DROPPED, + IS_LAZY, + IN_ZLT, +}; + +/** + * enum gc_flags_enum - Garbage Collection [GC] states. + */ +enum gc_flags_enum { + DO_GC_NEW, + DO_GC_PREPARE, /* -> READ or COMPLETE state */ + DO_GC_WRITE, + DO_GC_META, /* -> PREPARE state */ + DO_GC_COMPLETE, +}; + +/** + * enum znd_flags_enum - zoned state/feature/action flags + */ +enum znd_flags_enum { + ZF_FREEZE, + ZF_POOL_FWD, + ZF_POOL_REV, + ZF_POOL_CRCS, + + ZF_RESV_1, + ZF_RESV_2, + ZF_RESV_3, + ZF_RESV_4, + + DO_JOURNAL_MOVE, + DO_MEMPOOL, + DO_SYNC, + DO_FLUSH, + DO_JOURNAL_LOAD, + DO_GC_NO_PURGE, + DO_METAWORK_QD, +}; + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +struct zoned; + +/** + * struct gc_state - A page of map table + * @znd: ZDM Instance + * @gc_flags: See gc_flags_enum + * @r_ptr: Next read in zone. + * @w_ptr: Next write in target zone. + * @nblks: Number of blocks in I/O + * @result: GC operation result. + * @z_gc: Zone undergoing compacation + * @tag: System wide serial number (debugging). + * + * Longer description of this structure. + */ +struct gc_state { + struct zoned *znd; + unsigned long gc_flags; + + u32 r_ptr; + u32 w_ptr; + + u32 nblks; /* 1-65536 */ + u32 z_gc; + + int result; +}; + +/** + * struct map_addr - A page of map table + * @dm_s: full map on dm layer + * @zone_id: z_id match zone_list_t.z_id + * @pg_idx: entry in lut (0-1023) + * @lut_s: sector table lba + * @lut_r: reverse table lba + * + * Longer description of this structure. + */ +struct map_addr { + u64 dm_s; + u64 lut_s; + u64 lut_r; + + u32 zone_id; + u32 pg_idx; +}; + +/** + * struct map_sect_to_lba - Sector to LBA mapping. + * @tlba: tlba + * @physical: blba or number of blocks + * + * Longer description of this structure. + */ +struct map_sect_to_lba { + __le64 tlba; /* record type [16 bits] + logical sector # */ + __le64 bval; /* csum 16 [16 bits] + 'physical' block lba */ +} __packed; + +/** + * enum gc_flags_enum - Garbage Collection [GC] states. + */ +enum map_type_enum { + IS_MAP, + IS_DISCARD, + MAP_COUNT, +}; + +/** + * struct map_cache - A page of map table + * @mclist: + * @jdata: 4k page of data + * @refcount: + * @cached_lock: + * @busy_locked: + * @jcount: + * @jsorted: + * @jsize: + * @map_content: See map_type_enum + * + * Longer description of this structure. + */ +struct map_cache { + struct list_head mclist; + struct map_sect_to_lba *jdata; + atomic_t refcount; + struct mutex cached_lock; + atomic_t busy_locked; + u32 jcount; + u32 jsorted; + u32 jsize; + u32 map_content; /* IS_MAP or IS_DISC */ +}; + + +union map_pg_data { + __le32 *addr; + __le16 *crc; +}; + + +/** + * struct map_pg - A page of map table + * + * @mdata: 4k page of table entries + * @refcount: + * @age: most recent access in jiffies + * @lba: logical position (use lookups to find actual) + * @md_lock: lock mdata i/o + * @last_write: last known position on disk + * @flags: is dirty flag, is fwd, is rev, is crc, is lut + * @inpool: + * + * Longer description of this structure. + */ +struct map_pg { + union map_pg_data data; + atomic_t refcount; + u64 age; + u64 lba; + u64 last_write; + struct mutex md_lock; + unsigned long flags; + struct list_head zltlst; + + /* in flight tracking */ + struct list_head lazy; + struct zoned *znd; + struct map_pg *crc_pg; + int hotness; + int index; + __le16 md_crc; +}; + +/** + * struct map_crc - Map to backing crc16. + * + * @table: Maybe via table + * @pg_no: Page [index] of table entry if applicable. + * @pg_idx: Offset within page (From: zoned::md_crcs when table is null) + * + * Longer description of this structure. + */ +struct map_crc { + struct map_pg **table; + int pg_no; + int pg_idx; +}; + +/** + * struct map_info - Map to backing lookup table. + * + * @table: backing table + * @crc: backing crc16 detail. + * @index: index [page] of table entry. Use map_addr::pg_idx for offset. + * @bit_type: IS_LUT or IS_CRC + * @bit_dir: IS_FWD or IS_REV + * + * Longer description of this structure. + */ +struct mpinfo { + struct map_pg **table; + struct map_crc crc; + int index; + int bit_type; + int bit_dir; +}; + + +/** + * struct meta_pg - A page of zone WP mapping. + * + * @wp_alloc: Bits 23-0: wp alloc location. Bits 31-24: GC Flags, Type Flags + * @zf_est: Bits 23-0: free block count. Bits 31-24: Stream Id + * @wp_used: Bits 23-0: wp written location. Bits 31-24: Ratio ReCalc flag. + * @lba: pinned LBA in conventional/preferred zone. + * @wplck: spinlock held during data updates. + * @flags: IS_DIRTY flag + * + * One page is used per 1024 zones on media. + * For an 8TB drive this uses 30 entries or about 360k RAM. + */ +struct meta_pg { + __le32 *wp_alloc; + __le32 *zf_est; + __le32 *wp_used; + u64 lba; + spinlock_t wplck; + unsigned long flags; +}; + +/** + * struct zdm_superblock - A page of map table + * @uuid: + * @nr_zones: + * @magic: + * @zdstart: + * @version: + * @packed_meta: + * @flags: + * @csum: + * + * Longer description of this structure. + */ +struct zdm_superblock { + u8 uuid[UUID_LEN]; /* 16 */ + __le64 nr_zones; /* 8 */ + __le64 magic; /* 8 */ + __le32 resvd; /* 4 */ + __le32 zdstart; /* 4 */ + __le32 version; /* 4 */ + __le32 packed_meta; /* 4 */ + __le32 flags; /* 4 */ + __le32 csum; /* 4 */ +} __packed; /* 56 */ + +#define MAX_CACHE_SYNC 400 + +/** + * struct mz_superkey - A page of map table + * @sig0: 8 - Native endian + * @sig1: 8 - Little endian + * @sblock: 56 - + * @stream: 1024 - + * @reserved: 2982 - + * @gc_resv: 4 - + * @meta_resv: 4 - + * @generation: 8 - + * @key_crc: 2 - + * @magic: 8 - + * + * Longer description of this structure. + */ +struct mz_superkey { + u64 sig0; + __le64 sig1; + struct zdm_superblock sblock; + __le32 stream[STREAM_SIZE]; + __le32 gc_resv; + __le32 meta_resv; + __le16 n_crcs; + __le16 crcs[MAX_CACHE_SYNC]; + __le16 md_crc; + __le16 wp_crc[64]; + __le16 zf_crc[64]; + __le16 discards; + __le16 maps; + u8 reserved[1908]; + __le32 crc32; + __le64 generation; + __le64 magic; +} __packed; + +/** + * struct io_4k_block - Sector to LBA mapping. + * @data: A 4096 byte block + * + * Longer description of this structure. + */ +struct io_4k_block { + u8 data[Z_C4K]; +}; + +/** + * struct io_dm_block - Sector to LBA mapping. + * @data: A 512 byte block + * + * Longer description of this structure. + */ +struct io_dm_block { + u8 data[512]; +}; + +struct stale_tracking { + u32 binsz; + u32 count; + int bins[STREAM_SIZE]; +}; + + +/* + * + * Partition -----------------------------------------------------------------+ + * Table ---+ | + * | | + * SMR Drive |^-------------------------------------------------------------^| + * CMR Zones ^^^^^^^^^ + * meta data ||||||||| + * + * Remaining partition is filesystem data + * + */ + +/** + * struct zoned - A page of map table + * @ti: dm_target entry + * @dev: dm_dev entry + * @mclist: list of pages of in-memory LBA mappings. + * @mclck: in memory map-cache lock (spinlock) + + * @zltpool: pages of lookup table entries + * @zlt_lck: zltpool: memory pool lock + * @lzy_pool: + * @lzy_lock: + * @fwd_tm: + * @rev_tm: + + * @bg_work: background worker (periodic GC/Mem purge) + * @bg_wq: background work queue + * @stats_lock: + * @gc_active: Current GC state + * @gc_lock: GC Lock + * @gc_work: GC Worker + * @gc_wq: GC Work Queue + * @data_zones: # of data zones on device + * @gz_count: # of 256G mega-zones + * @nr_blocks: 4k blocks on backing device + * @md_start: LBA at start of metadata pool + * @data_lba: LBA at start of data pool + * @zdstart: ZONE # at start of data pool (first Possible WP ZONE) + * @start_sect: where ZDM partition starts (RAW LBA) + * @flags: See: enum znd_flags_enum + * @gc_backlog: + * @gc_io_buf: + * @io_vcache[32]: + * @io_vcache_flags: + * @z_sballoc: + * @super_block: + * @z_mega: + * @meta_wq: + * @gc_postmap: + * @io_client: + * @io_wq: + * @zone_action_wq: + * @timer: + * @bins: Memory usage accounting/reporting. + * @bdev_name: + * @memstat: + * @suspended: + * @gc_mz_pref: + * @mz_provision: Number of zones per 1024 of over-provisioning. + * @ata_passthrough: + * @is_empty: For fast discards on initial format + * + * Longer description of this structure. + */ +struct zoned { + struct dm_target *ti; + struct dm_dev *dev; + + struct list_head zltpool; + struct list_head lzy_pool; + struct list_head mclist[MAP_COUNT]; + spinlock_t mclck[MAP_COUNT]; + + struct work_struct bg_work; + struct workqueue_struct *bg_wq; + + spinlock_t zlt_lck; + spinlock_t lzy_lck; + spinlock_t stats_lock; + spinlock_t mapkey_lock; /* access LUT and CRC array of pointers */ + spinlock_t ct_lock; /* access LUT and CRC array of pointers */ + + struct mutex gc_wait; + struct mutex pool_mtx; + struct mutex mz_io_mutex; + struct mutex vcio_lock; + +#if USE_KTHREAD /* experimental++ */ + struct task_struct *bio_kthread; + DECLARE_KFIFO_PTR(bio_fifo, struct bio *); + wait_queue_head_t wait_bio; + wait_queue_head_t wait_fifo; +#endif + struct bio_set *bio_set; + + struct gc_state *gc_active; + spinlock_t gc_lock; + struct delayed_work gc_work; + struct workqueue_struct *gc_wq; + + u64 nr_blocks; + u64 start_sect; + + u64 md_start; + u64 md_end; + u64 data_lba; + + unsigned long flags; + + u64 r_base; + u64 s_base; + u64 c_base; + u64 c_mid; + u64 c_end; + + u64 sk_low; /* unused ? */ + u64 sk_high; /* unused ? */ + + struct meta_pg *wp; + + struct map_pg **fwd_tm; /* nr_blocks / 1024 */ + struct map_pg **rev_tm; /* nr_blocks / 1024 */ + struct map_pg **fwd_crc; /* (nr_blocks / 1024) / 2048 */ + struct map_pg **rev_crc; /* (nr_blocks / 1024) / 2048 */ + __le16 *md_crcs; /* one of crc16's for fwd, 1 for rev */ + u32 crc_count; + u32 map_count; + + void *z_sballoc; + struct mz_superkey *bmkeys; + struct zdm_superblock *super_block; + + struct work_struct meta_work; + sector_t last_w; + u8 *cow_block; + u64 cow_addr; + u32 data_zones; + u32 gz_count; + u32 zdstart; + u32 z_gc_free; + atomic_t incore; + u32 discard_count; + u32 z_current; + u32 z_meta_resv; + u32 z_gc_resv; + u32 gc_events; + int mc_entries; + int dc_entries; + int in_zlt; + int in_lzy; + int meta_result; + struct stale_tracking stale; + + int gc_backlog; + void *gc_io_buf; + struct mutex gc_vcio_lock; + struct io_4k_block *io_vcache[32]; + unsigned long io_vcache_flags; + u64 age; + struct workqueue_struct *meta_wq; + struct map_cache gc_postmap; + struct dm_io_client *io_client; + struct workqueue_struct *io_wq; + struct workqueue_struct *zone_action_wq; + struct timer_list timer; + + u32 bins[40]; + char bdev_name[BDEVNAME_SIZE]; + char proc_name[BDEVNAME_SIZE+4]; + struct proc_dir_entry *proc_fs; + loff_t wp_proc_at; + loff_t used_proc_at; + + size_t memstat; + atomic_t suspended; + atomic_t gc_throttle; + +#if ALLOC_DEBUG + atomic_t allocs; + int hw_allocs; + void **alloc_trace; +#endif + u32 filled_zone; + u16 mz_provision; + unsigned bdev_is_zoned:1; + unsigned ata_passthrough:1; + unsigned issue_open_zone:1; + unsigned issue_close_zone:1; + unsigned is_empty:1; + unsigned trim:1; + +}; + +/** + * struct zdm_ioc_request - Sector to LBA mapping. + * @result_size: + * @megazone_nr: + * + * Longer description of this structure. + */ +struct zdm_ioc_request { + u32 result_size; + u32 megazone_nr; +}; + +/** + * struct zdm_ioc_status - Sector to LBA mapping. + * @b_used: Number of blocks used + * @b_available: Number of blocks free + * @b_discard: Number of blocks stale + * @m_zones: Number of zones. + * @mc_entries: Mem cache blocks in use + * @dc_entries: Discard cache blocks in use. + * @mlut_blocks: + * @crc_blocks: + * @memstat: Total memory in use by ZDM via *alloc() + * @bins: Allocation by subsystem. + * + * This status structure is used to pass run-time information to + * user spaces tools (zdm-tools) for diagnostics and tuning. + */ +struct zdm_ioc_status { + u64 b_used; + u64 b_available; + u64 b_discard; + u64 m_zones; + u32 mc_entries; + u32 dc_entries; + u64 mlut_blocks; + u64 crc_blocks; + u64 memstat; + u32 bins[40]; +}; + +#ifdef __cplusplus +} +#endif + +#endif /* _DM_ZONED_H */ diff --git a/drivers/md/libzoned.c b/drivers/md/libzoned.c new file mode 100644 index 0000000..f887654 --- /dev/null +++ b/drivers/md/libzoned.c @@ -0,0 +1,7149 @@ +/* + * Kernel Device Mapper for abstracting ZAC/ZBC devices as normal + * block devices for linux file systems. + * + * Copyright (C) 2015 Seagate Technology PLC + * + * Written by: + * Shaun Tancheff <shaun.tancheff@xxxxxxxxxxx> + * + * This file is licensed under the terms of the GNU General Public + * License version 2. This program is licensed "as is" without any + * warranty of any kind, whether express or implied. + */ + +#define BUILD_NO 108 + +#define EXTRA_DEBUG 0 +#define ENABLE_PG_FREE_VIA_LAZY 1 + +#define MZ_MEMPOOL_SZ 512 +#define READ_AHEAD 16 /* number of LUT entries to read-ahead */ +#define MEMCACHE_HWMARK 5 +#define MEM_PURGE_MSECS 15000 +#define MEM_HOT_BOOST_INC 5000 +#define DISCARD_IDLE_MSECS 2000 +#define DISCARD_MAX_INGRESS 150 +#define MAX_WSET 2048 + + +/* acceptable 'free' count for MZ free levels */ +#define GC_PRIO_DEFAULT 0xFF00 +#define GC_PRIO_LOW 0x7FFF +#define GC_PRIO_HIGH 0x0400 +#define GC_PRIO_CRIT 0x0040 + +/* When less than 20 zones are free use aggressive gc in the megazone */ +#define GC_COMPACT_AGGRESSIVE 32 + +/* + * For performance tuning: + * Q? smaller strips give smoother performance + * a single drive I/O is 8 (or 32?) blocks? + * A? Does not seem to ... + */ +#define GC_MAX_STRIPE 256 +#define REPORT_ORDER 7 +#define REPORT_FILL_PGS 65 /* 65 -> min # pages for 4096 descriptors */ +#define SYNC_IO_ORDER 2 +#define SYNC_IO_SZ ((1 << SYNC_IO_ORDER) * PAGE_SIZE) + +#define MZTEV_UNUSED (cpu_to_le32(0xFFFFFFFFu)) +#define MZTEV_NF (cpu_to_le32(0xFFFFFFFEu)) + +#define Z_TABLE_MAGIC 0x123456787654321Eul +#define Z_KEY_SIG 0xFEDCBA987654321Ful + +#define Z_CRC_4K 4096 +#define ZONE_SECT_BITS 19 +#define Z_BLKBITS 16 +#define Z_BLKSZ (1ul << Z_BLKBITS) +#define MAX_ZONES_PER_MZ 1024 +#define Z_SMR_SZ_BYTES (Z_C4K << Z_BLKBITS) + +#define GC_READ (1ul << 15) +#define BAD_ADDR (~0ul) +#define MC_INVALID (cpu_to_le64(BAD_ADDR)) +#define NOZONE (~0u) + +#define GZ_BITS 10 +#define GZ_MMSK ((1u << GZ_BITS) - 1) + +#define CRC_BITS 11 +#define CRC_MMSK ((1u << CRC_BITS) - 1) + +#define MD_CRC_INIT (cpu_to_le16(0x5249u)) + +static int map_addr_calc(struct zoned *, u64 dm_s, struct map_addr *out); +static int zoned_io_flush(struct zoned *znd); +static int zoned_wp_sync(struct zoned *znd, int reset_non_empty); +static void cache_if_dirty(struct zoned *znd, struct map_pg *pg, int wq); +static int write_if_dirty(struct zoned *, struct map_pg *, int use_wq, int snc); +static void gc_work_task(struct work_struct *work); +static void meta_work_task(struct work_struct *work); +static u64 mcache_greatest_gen(struct zoned *, int, u64 *, u64 *); +static u64 mcache_find_gen(struct zoned *, u64 base, int, u64 *out); +static int find_superblock(struct zoned *znd, int use_wq, int do_init); +static int sync_mapped_pages(struct zoned *znd, int sync, int drop); +static int unused_phy(struct zoned *znd, u64 lba, u64 orig_s, gfp_t gfp); +static struct io_4k_block *get_io_vcache(struct zoned *znd, gfp_t gfp); +static int put_io_vcache(struct zoned *znd, struct io_4k_block *cache); +static struct map_pg *get_map_entry(struct zoned *, u64 lba, gfp_t gfp); +static void put_map_entry(struct map_pg *); +static int cache_pg(struct zoned *znd, struct map_pg *pg, gfp_t gfp, + struct mpinfo *mpi); +static int move_to_map_tables(struct zoned *znd, struct map_cache *mcache); +static void update_stale_ratio(struct zoned *znd, u32 zone); +static int zoned_create_disk(struct dm_target *ti, struct zoned *znd); +static int do_init_zoned(struct dm_target *ti, struct zoned *znd); +static int z_discard_partial(struct zoned *znd, u32 blks, gfp_t gfp); +static u64 z_discard_range(struct zoned *znd, u64 addr, gfp_t gfp); +static u64 z_lookup_cache(struct zoned *znd, u64 addr, int type); +static u64 z_lookup_table(struct zoned *znd, u64 addr, gfp_t gfp); +static u64 current_mapping(struct zoned *znd, u64 addr, gfp_t gfp); +static int z_mapped_add_one(struct zoned *znd, u64 dm_s, u64 lba, gfp_t gfp); +static int z_mapped_discard(struct zoned *znd, u64 tlba, u64 count, gfp_t gfp); +static int z_mapped_addmany(struct zoned *znd, u64 dm_s, u64 lba, u64, + gfp_t gfp); +static int z_to_map_list(struct zoned *znd, u64 dm_s, u64 lba, gfp_t gfp); +static int z_to_discard_list(struct zoned *znd, u64 dm_s, u64 blks, gfp_t gfp); +static int discard_merge(struct zoned *znd, u64 tlba, u64 blks); +static int z_mapped_sync(struct zoned *znd); +static int z_mapped_init(struct zoned *znd); +static u64 z_acquire(struct zoned *znd, u32 flags, u32 nblks, u32 *nfound); +static __le32 sb_crc32(struct zdm_superblock *sblock); +static int update_map_entry(struct zoned *, struct map_pg *, + struct map_addr *, u64, int); +static int read_block(struct dm_target *, enum dm_io_mem_type, + void *, u64, unsigned int, int); +static int write_block(struct dm_target *, enum dm_io_mem_type, + void *, u64, unsigned int, int); +static int writef_block(struct dm_target *ti, int rw, enum dm_io_mem_type dtype, + void *data, u64 lba, unsigned int count, int queue); +static int zoned_init_disk(struct dm_target *ti, struct zoned *znd, + int create, int force); + +#define MutexLock(m) test_and_lock((m), __LINE__) +#define SpinLock(s) test_and_spin((s), __LINE__) + +static __always_inline void test_and_lock(struct mutex *m, int lineno) +{ + if (!mutex_trylock(m)) { + pr_debug("mutex stall at %d\n", lineno); + mutex_lock(m); + } +} + +static __always_inline void test_and_spin(spinlock_t *lock, int lineno) +{ + if (!spin_trylock(lock)) { + pr_debug("spin stall at %d\n", lineno); + spin_lock(lock); + } +} + +/** + * ref_pg() - Decrement refcount on page of ZLT + * @pg: Page of ZLT map + */ +static __always_inline void deref_pg(struct map_pg *pg) +{ + atomic_dec(&pg->refcount); +} + +/** + * ref_pg() - Increment refcount on page of ZLT + * @pg: Page of ZLT map + */ +static __always_inline void ref_pg(struct map_pg *pg) +{ + atomic_inc(&pg->refcount); +#if 0 /* ALLOC_DEBUG */ + if (atomic_read(&pg->refcount) > 20) { + pr_err("Excessive %d who dunnit?\n", + atomic_read(&pg->refcount)); + dump_stack(); + } +#endif +} + +/** + * getref_pg() - Read the refcount + * @pg: Page of ZLT map + */ +static __always_inline int getref_pg(struct map_pg *pg) +{ + return atomic_read(&pg->refcount); +} + + +/** + * mcache_ref() - Increment the reference count of mcache() + * @mcache: Page of map cache + */ +static __always_inline void mcache_ref(struct map_cache *mcache) +{ + atomic_inc(&mcache->refcount); +} + +/** + * mcache_deref() - Decrement the reference count of mcache() + * @mcache: Page of map cache + */ +static __always_inline void mcache_deref(struct map_cache *mcache) +{ + atomic_dec(&mcache->refcount); +} + +/** + * mcache_getref() - Read the refcount + * @mcache: Page of map cache + */ +static __always_inline int mcache_getref(struct map_cache *mcache) +{ + return atomic_read(&mcache->refcount); +} + +/** + * mcache_busy() - Increment the busy test of mcache() + * @mcache: Page of map cache + */ +static __always_inline void mcache_busy(struct map_cache *mcache) +{ + atomic_inc(&mcache->busy_locked); +} + +/** + * mcache_unbusy() - Decrement the busy test count of mcache() + * @mcache: Page of map cache + */ +static __always_inline void mcache_unbusy(struct map_cache *mcache) +{ + atomic_dec(&mcache->busy_locked); +} + +/** + * mcache_is_busy() - Test if mcache is busy + * @mcache: Page of map cache + * Return non-zero if busy otherwise 0. + */ +static __always_inline int mcache_is_busy(struct map_cache *mcache) +{ + return atomic_read(&mcache->busy_locked); +} + +/** + * crc16_md() - 16 bit CRC on metadata blocks + * @data: Block of metadata. + * @len: Number of bytes in block. + * + * Return: 16 bit CRC. + */ +static inline u16 crc16_md(void const *data, size_t len) +{ + const u16 init = 0xFFFF; + const u8 *p = data; + + return crc16(init, p, len); +} + +/** + * crc_md_le16() - 16 bit CRC on metadata blocks in little endian + * @data: Block of metadata. + * @len: Number of bytes in block. + * + * Return: 16 bit CRC. + */ +static inline __le16 crc_md_le16(void const *data, size_t len) +{ + u16 crc = crc16_md(data, len); + + return cpu_to_le16(crc); +} + +/** + * crcpg() - 32 bit CRC [NOTE: 32c is HW assisted on Intel] + * @data: Block of metadata [4K bytes]. + * + * Return: 32 bit CRC. + */ +static inline u32 crcpg(void *data) +{ + return crc32c(~0u, data, Z_CRC_4K) ^ SUPERBLOCK_CSUM_XOR; +} + +/** + * le64_to_lba48() - Return the lower 48 bits of LBA + * @enc: 64 bit LBA + flags + * @flg: optional 16 bits of classification. + * + * Return: 48 bits of LBA [and flg]. + */ +static inline u64 le64_to_lba48(__le64 enc, u16 *flg) +{ + const u64 lba64 = le64_to_cpu(enc); + + if (flg) + *flg = (lba64 >> 48) & 0xFFFF; + + return lba64 & Z_LOWER48; +} + +/** + * lba48_to_le64() - Encode 48 bits of lba + 16 bits of flags. + * @flags: flags to encode. + * @lba48: LBA to encode + * + * Return: Little endian u64. + */ +static inline __le64 lba48_to_le64(u16 flags, u64 lba48) +{ + u64 high_bits = flags; + + return cpu_to_le64((high_bits << 48) | (lba48 & Z_LOWER48)); +} + +/** + * sb_test_flag() - Test if flag is set in Superblock. + * @sb: zdm_superblock. + * @bit_no: superblock flag + * + * Return: non-zero if flag is set. + */ +static inline int sb_test_flag(struct zdm_superblock *sb, int bit_no) +{ + u32 flags = le32_to_cpu(sb->flags); + + return (flags & (1 << bit_no)) ? 1 : 0; +} + +/** + * sb_set_flag() - Set a flag in superblock. + * @sb: zdm_superblock. + * @bit_no: superblock flag + */ +static inline void sb_set_flag(struct zdm_superblock *sb, int bit_no) +{ + u32 flags = le32_to_cpu(sb->flags); + + flags |= (1 << bit_no); + sb->flags = cpu_to_le32(flags); +} + +/** + * zone_to_sector() - Calculate starting LBA of zone + * @zone: zone number (0 based) + * + * Return: LBA at start of zone. + */ +static inline u64 zone_to_sector(u64 zone) +{ + return zone << ZONE_SECT_BITS; +} + +/** + * is_expired_msecs() - Determine if age + msecs is older than now. + * @age: jiffies at last access + * @msecs: msecs of extra time. + * + * Return: non-zero if block is expired. + */ +static inline int is_expired_msecs(u64 age, u32 msecs) +{ + u64 expire_at = age + msecs_to_jiffies(msecs); + int expired = time_after64(jiffies_64, expire_at); + + return expired; +} + +/** + * is_expired() - Determine if age is older than MEM_PURGE_MSECS. + * @age: jiffies at last access + * + * Return: non-zero if block is expired. + */ +static inline int is_expired(u64 age) +{ + return is_expired_msecs(age, MEM_PURGE_MSECS); +} + +/** + * _calc_zone() - Determine zone number from addr + * @addr: 4k sector number + * + * Return: znum or 0xFFFFFFFF if addr is in metadata space. + */ +static inline u32 _calc_zone(struct zoned *znd, u64 addr) +{ + u32 znum = NOZONE; + + if (addr < znd->data_lba) + return znum; + + addr -= znd->data_lba; + znum = addr >> Z_BLKBITS; + + return znum; +} + +/** + * lazy_pool_add - Set a flag and add map page to the lazy pool + * @znd: ZDM Instance + * @expg: Map table page. + * @bit: Flag to set on page. + * + * Lazy pool is used for deferred adding and delayed removal. + */ +static __always_inline +void lazy_pool_add(struct zoned *znd, struct map_pg *expg, int bit) +{ + unsigned long flags; + + spin_lock_irqsave(&znd->lzy_lck, flags); + if (!test_bit(IS_LAZY, &expg->flags)) { + set_bit(IS_LAZY, &expg->flags); + list_add(&expg->lazy, &znd->lzy_pool); + znd->in_lzy++; + } + set_bit(bit, &expg->flags); + spin_unlock_irqrestore(&znd->lzy_lck, flags); +} + +/** + * lazy_pool_splice - Add a list of pages to the lazy pool + * @znd: ZDM Instance + * @list: List of map table page to add. + * + * Lazy pool is used for deferred adding and delayed removal. + */ +static __always_inline +void lazy_pool_splice(struct zoned *znd, struct list_head *list) +{ + unsigned long flags; + + spin_lock_irqsave(&znd->lzy_lck, flags); + list_splice_tail(list, &znd->lzy_pool); + spin_unlock_irqrestore(&znd->lzy_lck, flags); +} + +/** + * pool_add() - Add metadata block to zltlst + * @znd: ZDM instance + * @expg: current metadata block to add to zltlst list. + */ +static inline int pool_add(struct zoned *znd, struct map_pg *expg) +{ + int rcode = 0; + + if (expg) { + unsigned long flags; + + if (spin_trylock_irqsave(&znd->zlt_lck, flags)) { + if (test_bit(IN_ZLT, &expg->flags)) { + Z_ERR(znd, "Double list_add from:"); + dump_stack(); + } else { + set_bit(IN_ZLT, &expg->flags); + list_add(&expg->zltlst, &znd->zltpool); + znd->in_zlt++; + } + spin_unlock_irqrestore(&znd->zlt_lck, flags); + rcode = 0; + } else { + lazy_pool_add(znd, expg, DELAY_ADD); + } + } + + return rcode; +} + +/** + * to_table_entry() - Deconstrct metadata page into mpinfo + * @znd: ZDM instance + * @lba: Address (4k resolution) + * @expg: current metadata block in zltlst list. + * + * Return: Index into mpinfo.table. + */ +static int to_table_entry(struct zoned *znd, u64 lba, struct mpinfo *mpi) +{ + int index = -1; + + if (lba >= znd->s_base && lba < znd->r_base) { + mpi->table = znd->fwd_tm; + index = lba - znd->s_base; + mpi->bit_type = IS_LUT; + mpi->bit_dir = IS_FWD; + mpi->crc.table = znd->fwd_crc; + mpi->crc.pg_no = index >> CRC_BITS; + mpi->crc.pg_idx = index & CRC_MMSK; + if (index < 0 || index >= znd->map_count) { + Z_ERR(znd, "%s: FWD BAD IDX %"PRIx64" %d of %d", + __func__, lba, index, znd->map_count); + dump_stack(); + } + } else if (lba >= znd->r_base && lba < znd->c_base) { + mpi->table = znd->rev_tm; + index = lba - znd->r_base; + mpi->bit_type = IS_LUT; + mpi->bit_dir = IS_REV; + mpi->crc.table = znd->rev_crc; + mpi->crc.pg_no = index >> CRC_BITS; + mpi->crc.pg_idx = index & CRC_MMSK; + if (index < 0 || index >= znd->map_count) { + Z_ERR(znd, "%s: REV BAD IDX %"PRIx64" %d of %d", + __func__, lba, index, znd->map_count); + dump_stack(); + } + } else if (lba >= znd->c_base && lba < znd->c_mid) { + mpi->table = znd->fwd_crc; + index = lba - znd->c_base; + mpi->bit_type = IS_CRC; + mpi->bit_dir = IS_FWD; + mpi->crc.table = NULL; + mpi->crc.pg_no = 0; + mpi->crc.pg_idx = index & CRC_MMSK; + if (index < 0 || index >= znd->crc_count) { + Z_ERR(znd, "%s: CRC BAD IDX %"PRIx64" %d of %d", + __func__, lba, index, znd->crc_count); + dump_stack(); + } + } else if (lba >= znd->c_mid && lba < znd->c_end) { + mpi->table = znd->rev_crc; + index = lba - znd->c_mid; + mpi->bit_type = IS_CRC; + mpi->bit_dir = IS_REV; + mpi->crc.table = NULL; + mpi->crc.pg_no = 1; + mpi->crc.pg_idx = (1 << CRC_BITS) + (index & CRC_MMSK); + if (index < 0 || index >= znd->crc_count) { + Z_ERR(znd, "%s: CRC BAD IDX %"PRIx64" %d of %d", + __func__, lba, index, znd->crc_count); + dump_stack(); + } + } else { + Z_ERR(znd, "** Corrupt lba %" PRIx64 " not in range.", lba); + znd->meta_result = -EIO; + dump_stack(); + } + mpi->index = index; + return index; +} + +/** + * is_ready_for_gc() - Test zone flags for GC sanity and ready flag. + * @znd: ZDM instance + * @z_id: Address (4k resolution) + * + * Return: non-zero if zone is suitable for GC. + */ +static inline int is_ready_for_gc(struct zoned *znd, u32 z_id) +{ + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + u32 used = le32_to_cpu(wpg->wp_used[gzoff]) & Z_WP_VALUE_MASK; + + if (((wp & Z_WP_GC_BITS) == Z_WP_GC_READY) && (used == Z_BLKSZ)) + return 1; + return 0; +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +/* + * generic-ish n-way alloc/free + * Use kmalloc for small (< 4k) allocations. + * Use vmalloc for multi-page alloctions + * Except: + * Use multipage allocations for dm_io'd pages that a frequently hit. + * + * NOTE: ALL allocations are zero'd before returning. + * alloc/free count is tracked for dynamic analysis. + */ + +#define GET_ZPG 0x040000 +#define GET_KM 0x080000 +#define GET_VM 0x100000 + +#define PG_01 (GET_ZPG | 1) +#define PG_02 (GET_ZPG | 2) +#define PG_05 (GET_ZPG | 5) +#define PG_06 (GET_ZPG | 6) +#define PG_08 (GET_ZPG | 8) +#define PG_09 (GET_ZPG | 9) +#define PG_10 (GET_ZPG | 10) +#define PG_11 (GET_ZPG | 11) +#define PG_13 (GET_ZPG | 13) +#define PG_17 (GET_ZPG | 17) +#define PG_27 (GET_ZPG | 27) + +#define KM_00 (GET_KM | 0) +#define KM_07 (GET_KM | 7) +#define KM_14 (GET_KM | 14) +#define KM_15 (GET_KM | 15) +#define KM_16 (GET_KM | 16) +#define KM_18 (GET_KM | 18) +#define KM_19 (GET_KM | 19) +#define KM_20 (GET_KM | 20) +#define KM_25 (GET_KM | 25) +#define KM_26 (GET_KM | 26) +#define KM_28 (GET_KM | 28) +#define KM_29 (GET_KM | 29) +#define KM_30 (GET_KM | 30) + +#define VM_03 (GET_VM | 3) +#define VM_04 (GET_VM | 4) +#define VM_12 (GET_VM | 12) +#define VM_21 (GET_VM | 21) +#define VM_22 (GET_VM | 22) + +#define ZDM_FREE(z, _p, sz, id) \ + do { zdm_free((z), (_p), (sz), (id)); (_p) = NULL; } while (0) + +#define ZDM_ALLOC(z, sz, id, gfp) zdm_alloc((z), (sz), (id), (gfp)) +#define ZDM_CALLOC(z, n, sz, id, gfp) zdm_calloc((z), (n), (sz), (id), (gfp)) + +/** + * zdm_free_debug() - Extra cleanup for memory debugging. + * @znd: ZDM instance + * @p: memory to be released. + * @sz: allocated size. + * @id: allocation bin. + * + * Additional alloc/free debugging and statistics handling. + */ +static inline void zdm_free_debug(struct zoned *znd, void *p, size_t sz, int id) +{ +#if ALLOC_DEBUG + int iter; + + if (atomic_read(&znd->allocs) > znd->hw_allocs) + znd->hw_allocs = atomic_read(&znd->allocs); + atomic_dec(&znd->allocs); + for (iter = 0; iter < znd->hw_allocs; iter++) { + if (p == znd->alloc_trace[iter]) { + znd->alloc_trace[iter] = NULL; + break; + } + } + if (iter == znd->hw_allocs) { + Z_ERR(znd, "Free'd something *NOT* allocated? %d", id); + dump_stack(); + } +#endif + + SpinLock(&znd->stats_lock); + if (sz > znd->memstat) + Z_ERR(znd, "Free'd more mem than allocated? %d", id); + + if (sz > znd->bins[id]) { + Z_ERR(znd, "Free'd more mem than allocated? %d", id); + dump_stack(); + } + znd->memstat -= sz; + znd->bins[id] -= sz; + spin_unlock(&znd->stats_lock); +} + +/** + * zdm_free() - Unified free by allocation 'code' + * @znd: ZDM instance + * @p: memory to be released. + * @sz: allocated size. + * @code: allocation size + * + * This (ugly) unified scheme helps to find leaks and monitor usage + * via ioctl tools. + */ +static void zdm_free(struct zoned *znd, void *p, size_t sz, u32 code) +{ + int id = code & 0x00FFFF; + int flag = code & 0xFF0000; + + if (p) { + + if (znd) + zdm_free_debug(znd, p, sz, id); + + memset(p, 0, sz); /* DEBUG */ + + switch (flag) { + case GET_ZPG: + free_page((unsigned long)p); + break; + case GET_KM: + kfree(p); + break; + case GET_VM: + vfree(p); + break; + default: + Z_ERR(znd, + "zdm_free %p scheme %x not mapped.", p, code); + break; + } + } else { + Z_ERR(znd, "double zdm_free %p [%d]", p, id); + dump_stack(); + } +} + +/** + * zdm_free_debug() - Extra tracking for memory debugging. + * @znd: ZDM instance + * @p: memory to be released. + * @sz: allocated size. + * @id: allocation bin. + * + * Additional alloc/free debugging and statistics handling. + */ +static inline +void zdm_alloc_debug(struct zoned *znd, void *p, size_t sz, int id) +{ +#if ALLOC_DEBUG + int iter; + int count; + + atomic_inc(&znd->allocs); + + count = atomic_read(&znd->allocs); + for (iter = 0; iter < count; iter++) { + if (!znd->alloc_trace[iter]) { + znd->alloc_trace[iter] = p; + break; + } + } +#endif + + SpinLock(&znd->stats_lock); + znd->memstat += sz; + znd->bins[id] += sz; + spin_unlock(&znd->stats_lock); +} + + +/** + * zdm_alloc() - Unified alloc by 'code': + * @znd: ZDM instance + * @sz: allocated size. + * @code: allocation size + * @gfp: kernel allocation flags. + * + * There a few things (like dm_io) that seem to need pages and not just + * kmalloc'd memory. + * + * This (ugly) unified scheme helps to find leaks and monitor usage + * via ioctl tools. + */ +static void *zdm_alloc(struct zoned *znd, size_t sz, int code, gfp_t gfp) +{ + void *pmem = NULL; + int id = code & 0x00FFFF; + int flag = code & 0xFF0000; + gfp_t gfp_mask; + +#if USE_KTHREAD + gfp_mask = GFP_KERNEL; +#else + gfp_mask = gfp ? GFP_ATOMIC : GFP_KERNEL; +#endif + + if (flag == GET_VM) + might_sleep(); + + if (gfp_mask != GFP_ATOMIC) + might_sleep(); + + switch (flag) { + case GET_ZPG: + pmem = (void *)get_zeroed_page(gfp_mask); + if (!pmem && gfp_mask == GFP_ATOMIC) { + Z_ERR(znd, "No atomic for %d, try noio.", id); + pmem = (void *)get_zeroed_page(GFP_NOIO); + } + break; + case GET_KM: + pmem = kzalloc(sz, gfp_mask); + if (!pmem && gfp_mask == GFP_ATOMIC) { + Z_ERR(znd, "No atomic for %d, try noio.", id); + pmem = kzalloc(sz, GFP_NOIO); + } + break; + case GET_VM: + WARN_ON(gfp); + pmem = vzalloc(sz); + break; + default: + Z_ERR(znd, "zdm alloc scheme for %u unknown.", code); + break; + } + + if (pmem) { + if (znd) + zdm_alloc_debug(znd, pmem, sz, id); + } else { + Z_ERR(znd, "Out of memory. %d", id); + dump_stack(); + } + + return pmem; +} + +/** + * zdm_calloc() - Unified alloc by 'code': + * @znd: ZDM instance + * @n: number of elements in array. + * @sz: allocation size of each element. + * @c: allocation strategy (VM, KM, PAGE, N-PAGES). + * @q: kernel allocation flags. + * + * calloc is just an zeroed memory array alloc. + * all zdm_alloc schemes are for zeroed memory so no extra memset needed. + */ +static void *zdm_calloc(struct zoned *znd, size_t n, size_t sz, int c, gfp_t q) +{ + return zdm_alloc(znd, sz * n, c, q); +} + +/** + * get_io_vcache() - Get a pre-allocated pool of memory for IO. + * @znd: ZDM instance + * @gfp: Allocation flags if no pre-allocated pool can be found. + * + * Return: Pointer to pool memory or NULL. + */ +static struct io_4k_block *get_io_vcache(struct zoned *znd, gfp_t gfp) +{ + struct io_4k_block *cache = NULL; + int avail; + + might_sleep(); + + for (avail = 0; avail < ARRAY_SIZE(znd->io_vcache); avail++) { + if (!test_and_set_bit(avail, &znd->io_vcache_flags)) { + cache = znd->io_vcache[avail]; + if (!cache) + znd->io_vcache[avail] = cache = + ZDM_CALLOC(znd, IO_VCACHE_PAGES, + sizeof(*cache), VM_12, gfp); + if (cache) + break; + } + } + return cache; +} + +/** + * put_io_vcache() - Get a pre-allocated pool of memory for IO. + * @znd: ZDM instance + * @cache: Allocated cache entry. + * + * Return: Pointer to pool memory or NULL. + */ +static int put_io_vcache(struct zoned *znd, struct io_4k_block *cache) +{ + int err = -ENOENT; + int avail; + + if (cache) { + for (avail = 0; avail < ARRAY_SIZE(znd->io_vcache); avail++) { + if (cache == znd->io_vcache[avail]) { + WARN_ON(!test_and_clear_bit(avail, + &znd->io_vcache_flags)); + err = 0; + break; + } + } + } + return err; +} + +/** + * map_value() - translate a lookup table entry to a Sector #, or LBA. + * @znd: ZDM instance + * @delta: little endian map entry. + * + * Return: LBA or 0 if invalid. + */ +static inline u64 map_value(struct zoned *znd, __le32 delta) +{ + u64 mval = 0ul; + + if ((delta != MZTEV_UNUSED) && (delta != MZTEV_NF)) + mval = le32_to_cpu(delta); + + return mval; +} + +/** + * map_encode() - Encode a Sector # or LBA to a lookup table entry value. + * @znd: ZDM instance + * @to_addr: address to encode. + * @value: encoded value + * + * Return: 0. + */ +static int map_encode(struct zoned *znd, u64 to_addr, __le32 *value) +{ + int err = 0; + + *value = MZTEV_UNUSED; + if (to_addr != BAD_ADDR) + *value = cpu_to_le32((u32)to_addr); + + return err; +} + +/** + * release_memcache() - Free all the memcache blocks. + * @znd: ZDM instance + * + * Return: 0. + */ +static int release_memcache(struct zoned *znd) +{ + int no; + + for (no = 0; no < MAP_COUNT; no++) { + struct list_head *head = &(znd->mclist[no]); + struct map_cache *mcache; + struct map_cache *_mc; + + if (list_empty(head)) + return 0; + + SpinLock(&znd->mclck[no]); + list_for_each_entry_safe(mcache, _mc, head, mclist) { + list_del(&mcache->mclist); + ZDM_FREE(znd, mcache->jdata, Z_C4K, PG_08); + ZDM_FREE(znd, mcache, sizeof(*mcache), KM_07); + } + spin_unlock(&znd->mclck[no]); + + } + return 0; +} + +/** + * warn_bad_lba() - Warn if a give LBA is not valid (Esp if beyond a WP) + * @znd: ZDM instance + * @lba48: 48 bit lba. + * + * Return: non-zero if lba is not valid. + */ +static inline int warn_bad_lba(struct zoned *znd, u64 lba48) +{ +#define FMT_ERR "LBA %" PRIx64 " is not valid: Z# %u, off:%x wp:%x" + int rcode = 0; + u32 zone; + + if (lba48 < znd->data_lba) + return rcode; + + zone = _calc_zone(znd, lba48); + if (zone < znd->data_zones) { /* FIXME: MAYBE? md_end */ + u32 gzoff = zone & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[zone >> GZ_BITS]; + u32 wp_at = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_VALUE_MASK; + u16 off = (lba48 - znd->data_lba) % Z_BLKSZ; + + if (off >= wp_at) { + rcode = 1; + Z_ERR(znd, FMT_ERR, lba48, zone, off, wp_at); + dump_stack(); + } + } else { + rcode = 1; + Z_ERR(znd, "LBA is not valid - Z# %u, count %u", + zone, znd->data_zones); + } + + return rcode; +} + +/** + * mapped_free() - Release a page of lookup table entries. + * @znd: ZDM instance + * @mapped: mapped page struct to free. + */ +static void mapped_free(struct zoned *znd, struct map_pg *mapped) +{ + if (mapped) { + MutexLock(&mapped->md_lock); + WARN_ON(test_bit(IS_DIRTY, &mapped->flags)); + if (mapped->data.addr) { + ZDM_FREE(znd, mapped->data.addr, Z_C4K, PG_27); + atomic_dec(&znd->incore); + } + mutex_unlock(&mapped->md_lock); + ZDM_FREE(znd, mapped, sizeof(*mapped), KM_20); + } +} + +/** + * flush_map() - write dirty map entres to disk. + * @znd: ZDM instance + * @map: Array of mapped pages. + * @count: number of elements in range. + * Return: non-zero on error. + */ +static int flush_map(struct zoned *znd, struct map_pg **map, u32 count) +{ + const int use_wq = 1; + const int sync = 1; + u32 ii; + int err = 0; + + if (!map) + return err; + + for (ii = 0; ii < count; ii++) { + if (map[ii] && map[ii]->data.addr) { + cache_if_dirty(znd, map[ii], use_wq); + err |= write_if_dirty(znd, map[ii], use_wq, sync); + } + } + + return err; +} + +/** + * zoned_io_flush() - flush all pending IO. + * @znd: ZDM instance + */ +static int zoned_io_flush(struct zoned *znd) +{ + int err = 0; + + set_bit(ZF_FREEZE, &znd->flags); + atomic_inc(&znd->gc_throttle); + + mod_delayed_work(znd->gc_wq, &znd->gc_work, 0); + flush_delayed_work(&znd->gc_work); + + clear_bit(DO_GC_NO_PURGE, &znd->flags); + set_bit(DO_JOURNAL_MOVE, &znd->flags); + set_bit(DO_MEMPOOL, &znd->flags); + set_bit(DO_SYNC, &znd->flags); + set_bit(DO_FLUSH, &znd->flags); + queue_work(znd->meta_wq, &znd->meta_work); + flush_workqueue(znd->meta_wq); + flush_workqueue(znd->bg_wq); + + mod_delayed_work(znd->gc_wq, &znd->gc_work, 0); + flush_delayed_work(&znd->gc_work); + atomic_dec(&znd->gc_throttle); + + spin_lock(&znd->lzy_lck); + INIT_LIST_HEAD(&znd->lzy_pool); + spin_unlock(&znd->lzy_lck); + + spin_lock(&znd->zlt_lck); + INIT_LIST_HEAD(&znd->zltpool); + spin_unlock(&znd->zlt_lck); + + err = flush_map(znd, znd->fwd_tm, znd->map_count); + if (err) + goto out; + + err = flush_map(znd, znd->rev_tm, znd->map_count); + if (err) + goto out; + + err = flush_map(znd, znd->fwd_crc, znd->crc_count); + if (err) + goto out; + + err = flush_map(znd, znd->rev_crc, znd->crc_count); + if (err) + goto out; + + set_bit(DO_SYNC, &znd->flags); + set_bit(DO_FLUSH, &znd->flags); + queue_work(znd->meta_wq, &znd->meta_work); + flush_workqueue(znd->meta_wq); + +out: + return err; +} + +/** + * release_table_pages() - flush and free all table map entries. + * @znd: ZDM instance + */ +static void release_table_pages(struct zoned *znd) +{ + int entry; + + if (znd->fwd_tm) { + for (entry = 0; entry < znd->map_count; entry++) { + mapped_free(znd, znd->fwd_tm[entry]); + znd->fwd_tm[entry] = NULL; + } + } + if (znd->rev_tm) { + for (entry = 0; entry < znd->map_count; entry++) { + mapped_free(znd, znd->rev_tm[entry]); + znd->rev_tm[entry] = NULL; + } + } + if (znd->fwd_crc) { + for (entry = 0; entry < znd->crc_count; entry++) { + mapped_free(znd, znd->fwd_crc[entry]); + znd->fwd_crc[entry] = NULL; + } + } + if (znd->rev_crc) { + for (entry = 0; entry < znd->crc_count; entry++) { + mapped_free(znd, znd->rev_crc[entry]); + znd->rev_crc[entry] = NULL; + } + } +} + +/** + * _release_wp() - free all WP alloc/usage/used data. + * @znd: ZDM instance + * @wp: Object to free. + */ +static void _release_wp(struct zoned *znd, struct meta_pg *wp) +{ + u32 gzno; + + for (gzno = 0; gzno < znd->gz_count; gzno++) { + struct meta_pg *wpg = &wp[gzno]; + + if (wpg->wp_alloc) + ZDM_FREE(znd, wpg->wp_alloc, Z_C4K, PG_06); + if (wpg->zf_est) + ZDM_FREE(znd, wpg->zf_est, Z_C4K, PG_06); + if (wpg->wp_used) + ZDM_FREE(znd, wpg->wp_used, Z_C4K, PG_06); + } + ZDM_FREE(znd, wp, znd->gz_count * sizeof(*wp), VM_21); + znd->wp = NULL; +} + +/** + * zoned_destroy() - Teardown a zoned device mapper instance. + * @znd: ZDM instance + */ +static void zoned_destroy(struct zoned *znd) +{ + int purge; + size_t ppgsz = sizeof(struct map_pg *); + size_t mapsz = ppgsz * znd->map_count; + size_t crcsz = ppgsz * znd->crc_count; + + del_timer_sync(&znd->timer); + + if (zoned_io_flush(znd)) + Z_ERR(znd, "sync/flush failure"); + + release_table_pages(znd); + release_memcache(znd); + + if (znd->dev) { + dm_put_device(znd->ti, znd->dev); + znd->dev = NULL; + } + if (znd->io_wq) { + destroy_workqueue(znd->io_wq); + znd->io_wq = NULL; + } + if (znd->zone_action_wq) { + destroy_workqueue(znd->zone_action_wq); + znd->zone_action_wq = NULL; + } + if (znd->bg_wq) { + destroy_workqueue(znd->bg_wq); + znd->bg_wq = NULL; + } + if (znd->gc_wq) { + destroy_workqueue(znd->gc_wq); + znd->gc_wq = NULL; + } + if (znd->meta_wq) { + destroy_workqueue(znd->meta_wq); + znd->meta_wq = NULL; + } + if (znd->io_client) + dm_io_client_destroy(znd->io_client); + if (znd->wp) + _release_wp(znd, znd->wp); + if (znd->md_crcs) + ZDM_FREE(znd, znd->md_crcs, Z_C4K * 2, VM_22); + if (znd->gc_io_buf) + ZDM_FREE(znd, znd->gc_io_buf, Z_C4K * GC_MAX_STRIPE, VM_04); + if (znd->gc_postmap.jdata) { + size_t sz = Z_BLKSZ * sizeof(*znd->gc_postmap.jdata); + + ZDM_FREE(znd, znd->gc_postmap.jdata, sz, VM_03); + } + if (znd->fwd_tm) + ZDM_FREE(znd, znd->fwd_tm, mapsz, VM_21); + if (znd->rev_tm) + ZDM_FREE(znd, znd->rev_tm, mapsz, VM_22); + if (znd->fwd_crc) + ZDM_FREE(znd, znd->fwd_crc, crcsz, VM_21); + if (znd->rev_crc) + ZDM_FREE(znd, znd->rev_crc, crcsz, VM_22); + + for (purge = 0; purge < ARRAY_SIZE(znd->io_vcache); purge++) { + size_t vcsz = IO_VCACHE_PAGES * sizeof(struct io_4k_block *); + + if (znd->io_vcache[purge]) { + if (test_and_clear_bit(purge, &znd->io_vcache_flags)) + Z_ERR(znd, "sync cache entry %d still in use!", + purge); + ZDM_FREE(znd, znd->io_vcache[purge], vcsz, VM_12); + } + } + if (znd->z_sballoc) + ZDM_FREE(znd, znd->z_sballoc, Z_C4K, PG_05); + + ZDM_FREE(NULL, znd, sizeof(*znd), KM_00); +} + + +/** + * _init_streams() - Setup initial conditions for streams and reserved zones. + * @znd: ZDM instance + */ +static void _init_streams(struct zoned *znd) +{ + u64 stream; + + for (stream = 0; stream < ARRAY_SIZE(znd->bmkeys->stream); stream++) + znd->bmkeys->stream[stream] = cpu_to_le32(NOZONE); + + znd->z_meta_resv = cpu_to_le32(znd->data_zones - 2); + znd->z_gc_resv = cpu_to_le32(znd->data_zones - 1); + znd->z_gc_free = znd->data_zones - 2; +} + +/** + * _init_mdcrcs() - Setup initial values for empty CRC blocks. + * @znd: ZDM instance + */ +static void _init_mdcrcs(struct zoned *znd) +{ + int idx; + + for (idx = 0; idx < Z_C4K; idx++) + znd->md_crcs[idx] = MD_CRC_INIT; +} + +/** + * _init_wp() - Setup initial usage for empty data zones. + * @znd: ZDM instance + */ +static void _init_wp(struct zoned *znd, u32 gzno, struct meta_pg *wpg) +{ + u32 gzcount = 1 << GZ_BITS; + u32 iter; + + if (znd->data_zones < ((gzno+1) << GZ_BITS)) + gzcount = znd->data_zones & GZ_MMSK; + + /* mark as empty */ + for (iter = 0; iter < gzcount; iter++) + wpg->zf_est[iter] = cpu_to_le32(Z_BLKSZ); + + /* mark as n/a -- full */ + gzcount = 1 << GZ_BITS; + for (; iter < gzcount; iter++) { + wpg->wp_alloc[iter] = cpu_to_le32(~0u); + wpg->wp_used[iter] = cpu_to_le32(~0u); + } +} + +/** + * _alloc_wp() - Allocate needed WP (meta_pg) objects + * @znd: ZDM instance + */ +struct meta_pg *_alloc_wp(struct zoned *znd) +{ + struct meta_pg *wp; + u32 gzno; + + wp = ZDM_CALLOC(znd, znd->gz_count, sizeof(*znd->wp), VM_21, NORMAL); + if (!wp) + goto out; + for (gzno = 0; gzno < znd->gz_count; gzno++) { + struct meta_pg *wpg = &wp[gzno]; + + spin_lock_init(&wpg->wplck); + wpg->lba = 2048ul + (gzno * 2); + wpg->wp_alloc = ZDM_ALLOC(znd, Z_C4K, PG_06, NORMAL); + wpg->zf_est = ZDM_ALLOC(znd, Z_C4K, PG_06, NORMAL); + wpg->wp_used = ZDM_ALLOC(znd, Z_C4K, PG_06, NORMAL); + if (!wpg->wp_alloc || !wpg->zf_est || !wpg->wp_used) { + _release_wp(znd, wp); + wp = NULL; + goto out; + } + _init_wp(znd, gzno, wpg); + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + } + +out: + return wp; +} + +/** + * alloc_map_tables() - Allocate map table entries. + * @znd: ZDM Target + * @mapct: Number of map entries needed. + * @crcct: Number of CRC entries needed. + */ +static int alloc_map_tables(struct zoned *znd, u64 mapct, u64 crcct) +{ + const size_t ptrsz = sizeof(struct map_pg *); + int rcode = 0; + + znd->fwd_tm = ZDM_CALLOC(znd, mapct, ptrsz, VM_21, NORMAL); + znd->rev_tm = ZDM_CALLOC(znd, mapct, ptrsz, VM_22, NORMAL); + znd->fwd_crc = ZDM_CALLOC(znd, crcct, ptrsz, VM_21, NORMAL); + znd->rev_crc = ZDM_CALLOC(znd, crcct, ptrsz, VM_22, NORMAL); + if (!znd->fwd_tm || !znd->rev_tm || !znd->fwd_crc || !znd->rev_crc) + rcode = -ENOMEM; + + return rcode; +} + +/** + * do_init_zoned() - Initialize a zoned device mapper instance + * @ti: DM Target Info + * @znd: ZDM Target + * + * Return: 0 on success. + * + * Setup the zone pointer table and do a one time calculation of some + * basic limits. + * + * While start of partition may not be zone aligned + * md_start, data_lba and md_end are all zone aligned. + * From start of partition [or device] to md_start some conv/pref + * space is required for superblocks, memcache, zone pointers, crcs + * and optionally pinned forward lookup blocks. + * + * 0 < znd->md_start <= znd->data_lba <= znd->md_end + * + * incoming [FS sectors] are linearly mapped after md_end. + * And blocks following data_lba are serialzied into zones either with + * explicit stream id support from BIOs [FUTURE], or implictly by LBA + * or type of data. + */ +static int do_init_zoned(struct dm_target *ti, struct zoned *znd) +{ + u64 size = i_size_read(get_bdev_bd_inode(znd)); + u64 bdev_nr_sect4k = size / Z_C4K; + u64 data_zones = (bdev_nr_sect4k >> Z_BLKBITS) - znd->zdstart; + u64 mzcount = dm_div_up(data_zones, MAX_ZONES_PER_MZ); + u64 cache = MAX_CACHE_INCR * CACHE_COPIES * MAX_MZ_SUPP; + u64 z0_lba = dm_round_up(znd->start_sect, Z_BLKSZ); + u64 mapct = dm_div_up(data_zones << Z_BLKBITS, 1024); + u64 crcct = dm_div_up(mapct, 2048); + int type; + u32 mz_min = 0; /* cache */ + int rcode = 0; + + INIT_LIST_HEAD(&znd->zltpool); + INIT_LIST_HEAD(&znd->lzy_pool); + + for (type = 0; type < MAP_COUNT; type++) { + INIT_LIST_HEAD(&znd->mclist[type]); + spin_lock_init(&znd->mclck[type]); + } + spin_lock_init(&znd->zlt_lck); + spin_lock_init(&znd->lzy_lck); + spin_lock_init(&znd->gc_lock); + spin_lock_init(&znd->stats_lock); + spin_lock_init(&znd->mapkey_lock); + spin_lock_init(&znd->ct_lock); + + mutex_init(&znd->pool_mtx); + mutex_init(&znd->gc_wait); + mutex_init(&znd->gc_postmap.cached_lock); + mutex_init(&znd->gc_vcio_lock); + mutex_init(&znd->vcio_lock); + mutex_init(&znd->mz_io_mutex); + + znd->data_zones = data_zones; + znd->gz_count = mzcount; + znd->crc_count = crcct; + znd->map_count = mapct; + + /* z0_lba - lba of first full zone in disk addr space */ + + znd->md_start = z0_lba - znd->start_sect; + if (znd->md_start < cache) { + znd->md_start += Z_BLKSZ; + mz_min++; + } + + /* md_start - lba of first full zone in partition addr space */ + znd->s_base = znd->md_start; + mz_min += mzcount; + if (mz_min < znd->zdstart) + set_bit(ZF_POOL_FWD, &znd->flags); + + znd->r_base = znd->s_base + (mzcount << Z_BLKBITS); + mz_min += mzcount; + if (mz_min < znd->zdstart) + set_bit(ZF_POOL_REV, &znd->flags); + + znd->c_base = znd->r_base + (mzcount << Z_BLKBITS); + znd->c_mid = znd->c_base + (mzcount * 0x20); + znd->c_end = znd->c_mid + (mzcount * 0x20); + mz_min++; + + if (mz_min < znd->zdstart) { + Z_ERR(znd, "Conv space for CRCs: Seting ZF_POOL_CRCS"); + set_bit(ZF_POOL_CRCS, &znd->flags); + } + + if (test_bit(ZF_POOL_FWD, &znd->flags)) { + znd->sk_low = znd->sk_high = 0; + } else { + znd->sk_low = znd->s_base; + znd->sk_high = znd->sk_low + (mzcount * 0x40); + } + + /* logical *ending* lba for meta [bumped up to next zone alignment] */ + znd->md_end = znd->c_base + (1 << Z_BLKBITS); + + /* actual starting lba for data pool */ + znd->data_lba = znd->md_end; + if (!test_bit(ZF_POOL_CRCS, &znd->flags)) + znd->data_lba = znd->c_base; + if (!test_bit(ZF_POOL_REV, &znd->flags)) + znd->data_lba = znd->r_base; + if (!test_bit(ZF_POOL_FWD, &znd->flags)) + znd->data_lba = znd->s_base; + + /* NOTE: md_end == data_lba => all meta is in conventional zones. */ + Z_INFO(znd, "ZDM #%u", BUILD_NO); + Z_INFO(znd, "Start Sect: %" PRIx64 " ZDStart# %u md- start %" PRIx64 + " end %" PRIx64 " Data LBA %" PRIx64, + znd->start_sect, znd->zdstart, znd->md_start, + znd->md_end, znd->data_lba); + + Z_INFO(znd, "%s: size:%" PRIu64 " zones: %u, gzs %u, resvd %d", + __func__, size, znd->data_zones, znd->gz_count, + znd->gz_count * znd->mz_provision); + +#if ALLOC_DEBUG + znd->alloc_trace = vzalloc(sizeof(*znd->alloc_trace) * 65536); + if (!znd->alloc_trace) { + ti->error = "couldn't allocate in-memory mem debug trace"; + rcode = -ENOMEM; + goto out; + } +#endif + + znd->z_sballoc = ZDM_ALLOC(znd, Z_C4K, PG_05, NORMAL); + if (!znd->z_sballoc) { + ti->error = "couldn't allocate in-memory superblock"; + rcode = -ENOMEM; + goto out; + } + + rcode = alloc_map_tables(znd, mapct, crcct); + if (rcode) + goto out; + + znd->gc_postmap.jdata = ZDM_CALLOC(znd, Z_BLKSZ, + sizeof(*znd->gc_postmap.jdata), VM_03, NORMAL); + znd->md_crcs = ZDM_ALLOC(znd, Z_C4K * 2, VM_22, NORMAL); + znd->gc_io_buf = ZDM_CALLOC(znd, GC_MAX_STRIPE, Z_C4K, VM_04, NORMAL); + znd->wp = _alloc_wp(znd); + znd->io_vcache[0] = ZDM_CALLOC(znd, IO_VCACHE_PAGES, + sizeof(struct io_4k_block), VM_12, NORMAL); + znd->io_vcache[1] = ZDM_CALLOC(znd, IO_VCACHE_PAGES, + sizeof(struct io_4k_block), VM_12, NORMAL); + + if (!znd->gc_postmap.jdata || !znd->md_crcs || !znd->gc_io_buf || + !znd->wp || !znd->io_vcache[0] || !znd->io_vcache[1]) { + rcode = -ENOMEM; + goto out; + } + znd->gc_postmap.jsize = Z_BLKSZ; + _init_mdcrcs(znd); + + znd->io_client = dm_io_client_create(); + if (!znd->io_client) { + rcode = -ENOMEM; + goto out; + } + + znd->meta_wq = create_singlethread_workqueue("znd_meta_wq"); + if (!znd->meta_wq) { + ti->error = "couldn't start metadata worker thread"; + rcode = -ENOMEM; + goto out; + } + + znd->gc_wq = create_singlethread_workqueue("znd_gc_wq"); + if (!znd->gc_wq) { + ti->error = "couldn't start GC workqueue."; + rcode = -ENOMEM; + goto out; + } + + znd->bg_wq = create_singlethread_workqueue("znd_bg_wq"); + if (!znd->bg_wq) { + ti->error = "couldn't start background workqueue."; + rcode = -ENOMEM; + goto out; + } + + znd->io_wq = create_singlethread_workqueue("kzoned_dm_io_wq"); + if (!znd->io_wq) { + ti->error = "couldn't start DM I/O thread"; + rcode = -ENOMEM; + goto out; + } + + znd->zone_action_wq = create_singlethread_workqueue("zone_act_wq"); + if (!znd->zone_action_wq) { + ti->error = "couldn't start zone action worker"; + rcode = -ENOMEM; + goto out; + } + +#if USE_KTHREAD + init_waitqueue_head(&znd->wait_bio); + init_waitqueue_head(&znd->wait_fifo); +#endif + + INIT_WORK(&znd->meta_work, meta_work_task); + INIT_WORK(&znd->bg_work, bg_work_task); + INIT_DELAYED_WORK(&znd->gc_work, gc_work_task); + setup_timer(&znd->timer, activity_timeout, (unsigned long)znd); + znd->last_w = BAD_ADDR; + set_bit(DO_SYNC, &znd->flags); + +out: + return rcode; +} + +/** + * check_metadata_version() - Test ZDM version for compatibility. + * @sblock: Super block + * + * Return 0 if valud, or -EINVAL if version is not recognized. + */ +static int check_metadata_version(struct zdm_superblock *sblock) +{ + u32 metadata_version = le32_to_cpu(sblock->version); + + if (metadata_version < MIN_ZONED_VERSION + || metadata_version > MAX_ZONED_VERSION) { + DMERR("Unsupported metadata version %u found.", + metadata_version); + DMERR("Only versions between %u and %u supported.", + MIN_ZONED_VERSION, MAX_ZONED_VERSION); + return -EINVAL; + } + + return 0; +} + +/** + * sb_crc32() - CRC check for superblock. + * @sblock: Superblock to check. + */ +static __le32 sb_crc32(struct zdm_superblock *sblock) +{ + const __le32 was = sblock->csum; + u32 crc; + + sblock->csum = 0; + crc = crc32c(~(u32) 0u, sblock, sizeof(*sblock)) ^ SUPERBLOCK_CSUM_XOR; + + sblock->csum = was; + return cpu_to_le32(crc); +} + +/** + * sb_check() - Check the superblock to see if it is valid and not corrupt. + * @sblock: Superblock to check. + */ +static int sb_check(struct zdm_superblock *sblock) +{ + __le32 csum_le; + + if (le64_to_cpu(sblock->magic) != SUPERBLOCK_MAGIC) { + DMERR("sb_check failed: magic %" PRIx64 ": wanted %lx", + le64_to_cpu(sblock->magic), SUPERBLOCK_MAGIC); + return -EILSEQ; + } + + csum_le = sb_crc32(sblock); + if (csum_le != sblock->csum) { + DMERR("sb_check failed: csum %u: wanted %u", + csum_le, sblock->csum); + return -EILSEQ; + } + + return check_metadata_version(sblock); +} + +/** + * zoned_create_disk() - Initialize the on-disk format of a zoned device mapper. + * @ti: DM Target Instance + * @znd: ZDM Instance + */ +static int zoned_create_disk(struct dm_target *ti, struct zoned *znd) +{ + const int reset_non_empty = 1; + struct zdm_superblock *sblock = znd->super_block; + int err; + + memset(sblock, 0, sizeof(*sblock)); + generate_random_uuid(sblock->uuid); + sblock->magic = cpu_to_le64(SUPERBLOCK_MAGIC); + sblock->version = cpu_to_le32(Z_VERSION); + sblock->zdstart = cpu_to_le32(znd->zdstart); + + err = zoned_wp_sync(znd, reset_non_empty); + + return err; +} + + +/** + * zoned_repair() - Attempt easy on-line fixes. + * @znd: ZDM Instance + * + * Repair an otherwise good device mapper instance that was not cleanly removed. + */ +static int zoned_repair(struct zoned *znd) +{ + Z_INFO(znd, "Is Dirty .. zoned_repair consistency fixer TODO!!!."); + return -ENOMEM; +} + + +/** + * zoned_init_disk() - Init from exising or re-create DM Target (ZDM) + * @ti: DM Target Instance + * @znd: ZDM Instance + * @create: Create if not found. + * @force: Force create even if it looks like a ZDM was here. + * + * Locate the existing SB on disk and re-load or create the device-mapper + * instance based on the existing disk state. + */ +static int zoned_init_disk(struct dm_target *ti, struct zoned *znd, + int create, int force) +{ + struct mz_superkey *key_blk = znd->z_sballoc; + + int jinit = 1; + int n4kblks = 1; + int use_wq = 1; + int rc = 0; + u32 zdstart = znd->zdstart; + + memset(key_blk, 0, sizeof(*key_blk)); + + znd->super_block = &key_blk->sblock; + + znd->bmkeys = key_blk; + znd->bmkeys->sig0 = Z_KEY_SIG; + znd->bmkeys->sig1 = cpu_to_le64(Z_KEY_SIG); + znd->bmkeys->magic = cpu_to_le64(Z_TABLE_MAGIC); + + _init_streams(znd); + + znd->stale.binsz = STREAM_SIZE; + if (znd->data_zones < STREAM_SIZE) + znd->stale.binsz = znd->data_zones; + else if ((znd->data_zones / STREAM_SIZE) > STREAM_SIZE) + znd->stale.binsz = dm_div_up(znd->data_zones, STREAM_SIZE); + znd->stale.count = dm_div_up(znd->data_zones, znd->stale.binsz); + + Z_ERR(znd, "Bin: Sz %u, count %u", znd->stale.binsz, znd->stale.count); + + if (create && force) { + Z_ERR(znd, "Force Creating a clean instance."); + } else if (find_superblock(znd, use_wq, 1)) { + u64 sb_lba = 0; + u64 generation; + +/* FIXME: Logic [zdstart] not well tested or ... logicial */ + + Z_INFO(znd, "Found existing superblock"); + if (zdstart != znd->zdstart) { + if (force) { + Z_ERR(znd, " (force) zdstart: %u <- %u", + zdstart, znd->zdstart); + } else { + znd->zdstart = zdstart; + jinit = 0; + } + } + + generation = mcache_greatest_gen(znd, use_wq, &sb_lba, NULL); + Z_DBG(znd, "Generation: %" PRIu64 " @ %" PRIx64, + generation, sb_lba); + + rc = read_block(ti, DM_IO_KMEM, key_blk, sb_lba, + n4kblks, use_wq); + if (rc) { + ti->error = "Superblock read error."; + return rc; + } + } + + rc = sb_check(znd->super_block); + if (rc) { + jinit = 0; + if (create) { + DMWARN("Check failed .. creating superblock."); + zoned_create_disk(ti, znd); + znd->super_block->nr_zones = + cpu_to_le64(znd->data_zones); + DMWARN("in-memory superblock created."); + znd->is_empty = 1; + } else { + ti->error = "Superblock check failed."; + return rc; + } + } + + if (sb_test_flag(znd->super_block, SB_DIRTY)) { + int repair_check = zoned_repair(znd); + + if (!force) { + /* if repair failed -- don't load from disk */ + if (repair_check) + jinit = 0; + } else if (repair_check && jinit) { + Z_ERR(znd, "repair failed, force enabled loading ..."); + } + } + + if (jinit) { + Z_ERR(znd, "INIT: Reloading DM Zoned metadata from DISK"); + znd->zdstart = le32_to_cpu(znd->super_block->zdstart); + set_bit(DO_JOURNAL_LOAD, &znd->flags); + queue_work(znd->meta_wq, &znd->meta_work); + Z_ERR(znd, "Waiting for load to complete."); + flush_workqueue(znd->meta_wq); + } + + Z_ERR(znd, "ZONED: Build No %d marking superblock dirty.", BUILD_NO); + + /* write the 'dirty' flag back to disk. */ + sb_set_flag(znd->super_block, SB_DIRTY); + znd->super_block->csum = sb_crc32(znd->super_block); + + return 0; +} + +/** + * compare_tlba() - Compare on tlba48 ignoring high 16 bits. + * @x1: Page of map cache + * @x2: Page of map cache + * + * Return: -1 less than, 1 greater than, 0 if equal. + */ +static int compare_tlba(const void *x1, const void *x2) +{ + const struct map_sect_to_lba *r1 = x1; + const struct map_sect_to_lba *r2 = x2; + const u64 v1 = le64_to_lba48(r1->tlba, NULL); + const u64 v2 = le64_to_lba48(r2->tlba, NULL); + + return (v1 < v2) ? -1 : ((v1 > v2) ? 1 : 0); +} + +/** + * _lsearch_tlba() - Do a linear search over a page of map_cache entries. + * @mcache: Page of map cache entries to search. + * @dm_s: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + */ +static int _lsearch_tlba(struct map_cache *mcache, u64 dm_s) +{ + int at = -1; + int jentry; + + for (jentry = mcache->jcount; jentry > 0; jentry--) { + u64 logi = le64_to_lba48(mcache->jdata[jentry].tlba, NULL); + + if (logi == dm_s) { + at = jentry - 1; + goto out; + } + } + +out: + return at; +} + +/** + * _lsearch_extent() - Do a linear search over a page of discard entries. + * @mcache: Page of map cache entries to search. + * @dm_s: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + * + * NOTE: In this case the match is if the tlba is include in the extent + */ +static int _lsearch_extent(struct map_cache *mcache, u64 dm_s) +{ + int at = -1; + int jentry; + + for (jentry = mcache->jcount; jentry > 0; jentry--) { + u64 addr = le64_to_lba48(mcache->jdata[jentry].tlba, NULL); + u64 blocks = le64_to_lba48(mcache->jdata[jentry].bval, NULL); + + if ((dm_s >= addr) && dm_s < (addr+blocks)) { + at = jentry - 1; + goto out; + } + } + +out: + return at; +} + +/** + * _bsrch_tlba() - Do a binary search over a page of map_cache entries. + * @mcache: Page of map cache entries to search. + * @tlba: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + */ +static int _bsrch_tlba(struct map_cache *mcache, u64 tlba) +{ + int at = -1; + void *base = &mcache->jdata[1]; + void *map; + struct map_sect_to_lba find; + + find.tlba = lba48_to_le64(0, tlba); + + if (mcache->jcount < 0 || mcache->jcount > mcache->jsize) + return at; + + map = bsearch(&find, base, mcache->jcount, sizeof(find), compare_tlba); + if (map) + at = (map - base) / sizeof(find); + + return at; +} + +/** + * compare_extent() - Compare on tlba48 ignoring high 16 bits. + * @x1: Page of map cache + * @x2: Page of map cache + * + * Return: -1 less than, 1 greater than, 0 if equal. + */ +static int compare_ext(const void *x1, const void *x2) +{ + const struct map_sect_to_lba *r1 = x1; + const struct map_sect_to_lba *r2 = x2; + const u64 key = le64_to_lba48(r1->tlba, NULL); + const u64 val = le64_to_lba48(r2->tlba, NULL); + const u64 ext = le64_to_lba48(r2->bval, NULL); + + if (key < val) + return -1; + if (key < (val+ext)) + return 0; + return 1; +} + + +/** + * _bsrch_extent() - Do a binary search over a page of map_cache extents. + * @mcache: Page of map cache extent entries to search. + * @tlba: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + */ +static int _bsrch_extent(struct map_cache *mcache, u64 tlba) +{ + int at = -1; + void *base = &mcache->jdata[1]; + void *map; + struct map_sect_to_lba find; + + find.tlba = lba48_to_le64(0, tlba); + map = bsearch(&find, base, mcache->jcount, sizeof(find), compare_ext); + if (map) + at = (map - base) / sizeof(find); + +#if 1 + { + int cmp = _lsearch_extent(mcache, tlba); + + if (cmp != at) { + pr_err("_bsrch_extent is failing got %d need %d.\n", + at, cmp); + at = cmp; + } + } +#endif + + return at; +} + +/** + * _bsrch_tlba() - Mutex Lock and binary search. + * @mcache: Page of map cache entries to search. + * @tlba: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + */ +static int _bsrch_tlba_lck(struct map_cache *mcache, u64 tlba) +{ + int at = -1; + + MutexLock(&mcache->cached_lock); + at = _bsrch_tlba(mcache, tlba); + mutex_unlock(&mcache->cached_lock); + + return at; +} + +/** + * _bsrch_extent_lck() - Mutex Lock and binary search. + * @mcache: Page of map cache extent entries to search. + * @tlba: tlba being sought. + * + * Return: 0 to jcount - 1 if found. -1 if not found + */ +static int _bsrch_extent_lck(struct map_cache *mcache, u64 tlba) +{ + int at = -1; + + MutexLock(&mcache->cached_lock); + at = _bsrch_extent(mcache, tlba); + mutex_unlock(&mcache->cached_lock); + + return at; +} + +/** + * mcache_bsearch() - Do a binary search over a page of map_cache entries. + * @mcache: Page of map cache entries to search. + * @tlba: tlba being sought. + */ +static inline int mcache_bsearch(struct map_cache *mcache, u64 tlba) +{ + int at = (mcache->map_content == IS_MAP) + ? _bsrch_tlba(mcache, tlba) + : _bsrch_extent(mcache, tlba); + + return at; +} + + +/** + * mcache_lsearch() - Do a linear search over a page of map/discard entries. + * @mcache: Page of map cache entries to search. + * @dm_s: tlba being sought. + */ +static inline int mcache_lsearch(struct map_cache *mcache, u64 tlba) +{ + int at = (mcache->map_content == IS_MAP) + ? _lsearch_tlba(mcache, tlba) + : _lsearch_extent(mcache, tlba); + + return at; +} + + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +static u64 z_lookup_key_range(struct zoned *znd, u64 addr) +{ + u64 found = 0ul; + + if (test_bit(ZF_POOL_FWD, &znd->flags)) + return found; + + if ((znd->sk_low <= addr) && (addr < znd->sk_high)) + found = FWD_TM_KEY_BASE + (addr - znd->sk_low); + + return found; +} + +/** + * increment_used_blks() - Update the 'used' WP when data hits disk. + * @znd: ZDM Instance + * @lba: blba if bio completed. + * @blks: number of blocks of bio completed. + * + * Called from a BIO end_io function so should not sleep or deadlock + * The 'critical' piece here is ensuring that the wp is advanced to 0x10000 + * Secondarily is triggering filled_zone which ultimatly sets the + * Z_WP_GC_READY in wp_alloc. While important this flag could be set + * during other non-critical passes over wp_alloc and wp_used such + * as during update_stale_ratio(). + */ +static void increment_used_blks(struct zoned *znd, u64 lba, u32 blks) +{ + u32 zone = _calc_zone(znd, lba); + u32 wwp = ((lba - znd->data_lba) - (zone << Z_BLKBITS)) + blks; + + if (zone < znd->data_zones) { + u32 gzno = zone >> GZ_BITS; + u32 gzoff = zone & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 used, uflags; + + used = le32_to_cpu(wpg->wp_used[gzoff]); + uflags = used & Z_WP_FLAGS_MASK; + used &= Z_WP_VALUE_MASK; + + if (wwp > used) { + wpg->wp_used[gzoff] = cpu_to_le32(wwp | uflags); + /* signal zone closure */ + if (wwp == Z_BLKSZ) + znd->filled_zone = zone; + } + } +} + +/** + * current_mapping() - Lookup a logical sector address to find the disk LBA + * @znd: ZDM instance + * @nodisk: Optional ignore the discard cache + * @addr: Logical LBA of page. + * @gfp: Memory allocation rule + * + * Return: Disk LBA or 0 if not found. + */ +static u64 _current_mapping(struct zoned *znd, int nodisc, u64 addr, gfp_t gfp) +{ + u64 found = 0ul; + + if (addr < znd->data_lba) + found = addr; + if (!found) + found = z_lookup_key_range(znd, addr); + if (!found && !nodisc) { + found = z_lookup_cache(znd, addr, IS_DISCARD); + if (found) { + found = 0; + goto out; + } + } + if (!found) + found = z_lookup_cache(znd, addr, IS_MAP); + if (!found) + found = z_lookup_table(znd, addr, gfp); + +out: + return found; +} + +/** + * current_mapping() - Lookup a logical sector address to find the disk LBA + * @znd: ZDM instance + * @addr: Logical LBA of page. + * @gfp: Memory allocation rule + * + * NOTE: Discard cache is checked. + * + * Return: Disk LBA or 0 if not found. + */ +static u64 current_mapping(struct zoned *znd, u64 addr, gfp_t gfp) +{ + const int nodisc = 0; + + return _current_mapping(znd, nodisc, addr, gfp); +} + +/** + * z_mapped_add_one() - Add an entry to the map cache block mapping. + * @znd: ZDM Instance + * @tlba: Address being discarded. + * @blks: number of blocks being discard. + * @gfp: Current memory allocation scheme. + * + * Add a new extent entry. + */ +static int z_mapped_add_one(struct zoned *znd, u64 dm_s, u64 lba, gfp_t gfp) +{ + int err = 0; + + if (dm_s < znd->data_lba) + return err; + +#if 0 /* FIXME: Do the RIGHT thing ... whatever that is. */ + + /* + * location of the SLT key sectors need to be + * stashed into the sector lookup table block map + * Does dm_s point in the sector lookup table block map + */ + if (!test_bit(ZF_POOL_FWD, &znd->flags)) { + if ((znd->sk_low <= dm_s) && (dm_s < znd->sk_high)) + lba = FWD_TM_KEY_BASE + (dm_s - znd->sk_low); + } +#endif + + /* + * If this mapping over-writes a discard entry then we need to punch a + * hole in the discard entry. + * - Find the discard entry. + * - Translate a chunk of the discard into map entries. + * - Break [or trim] the discard entry. + */ + if (lba) { + int is_locked = mutex_is_locked(&znd->mz_io_mutex); + + if (is_locked) + mutex_unlock(&znd->mz_io_mutex); + z_discard_range(znd, dm_s, gfp); + if (is_locked) + MutexLock(&znd->mz_io_mutex); + } + + do { + err = z_to_map_list(znd, dm_s, lba, gfp); + } while (-EBUSY == err); + + return err; +} + +/** + * z_discard_small() - Translate discard extents to map cache block mapping. + * @znd: ZDM Instance + * @tlba: Address being discarded. + * @blks: number of blocks being discard. + * @gfp: Current memory allocation scheme. + * + * Add a new extent entry. + */ +static int z_discard_small(struct zoned *znd, u64 tlba, u64 count, gfp_t gfp) +{ + const int nodisc = 1; + int err = 0; + u64 ii; + + for (ii = 0; ii < count; ii++) { + u64 lba = _current_mapping(znd, nodisc, tlba+ii, gfp); + + if (lba) + do { + err = z_to_map_list(znd, tlba+ii, 0ul, gfp); + } while (-EBUSY == err); + if (err) + break; + } + return err; +} + +/** + * z_discard_large() - Add a discard extent to the mapping cache + * @znd: ZDM Instance + * @tlba: Address being discarded. + * @blks: number of blocks being discard. + * @gfp: Current memory allocation scheme. + * + * Add a new extent entry. + */ +static int z_discard_large(struct zoned *znd, u64 tlba, u64 blks, gfp_t gfp) +{ + int rcode = 0; + + /* for a large discard ... add it to the discard queue */ + do { + rcode = z_to_discard_list(znd, tlba, blks, gfp); + } while (-EBUSY == rcode); + + return rcode; +} + +/** + * mapped_discard() - Add a discard extent to the mapping cache + * @znd: ZDM Instance + * @tlba: Address being discarded. + * @blks: number of blocks being discard. + * @merge: Attempt to merge with an existing extent. + * @gfp: Current memory allocation scheme. + * + * First attempt to merge with an existing entry. + * Otherwise attempt to push small entires into the map_cache 1:1 mapping + * Otherwise add a new extent entry. + */ +static +int mapped_discard(struct zoned *znd, u64 tlba, u64 blks, int merge, gfp_t gfp) +{ + int err = 0; + + if (merge && discard_merge(znd, tlba, blks)) + goto done; + + if (blks < DISCARD_MAX_INGRESS || znd->dc_entries > 39) { + /* io_mutex is avoid adding map cache entries during SYNC */ + MutexLock(&znd->mz_io_mutex); + err = z_discard_small(znd, tlba, blks, gfp); + mutex_unlock(&znd->mz_io_mutex); + goto done; + } + + err = z_discard_large(znd, tlba, blks, gfp); + +done: + return err; +} + +/** + * z_mapped_discard() - Add a discard extent to the mapping cache + * @znd: ZDM Instance + * @tlba: Address being discarded. + * @blks: number of blocks being discard. + * @gfp: Current memory allocation scheme. + */ +static int z_mapped_discard(struct zoned *znd, u64 tlba, u64 blks, gfp_t gfp) +{ + const int merge = 1; + + return mapped_discard(znd, tlba, blks, merge, gfp); +} + +/** + * mcache_alloc() - Allocate a new chunk of map cache + * @znd: ZDM Instance + * @gfp: Current memory allocation scheme. + */ +static struct map_cache *mcache_alloc(struct zoned *znd, gfp_t gfp) +{ + struct map_cache *mcache; + + mcache = ZDM_ALLOC(znd, sizeof(*mcache), KM_07, gfp); + if (mcache) { + mutex_init(&mcache->cached_lock); + mcache->jcount = 0; + mcache->jsorted = 0; + mcache->jdata = ZDM_CALLOC(znd, Z_UNSORTED, + sizeof(*mcache->jdata), PG_08, gfp); + + if (mcache->jdata) { + u64 logical = Z_LOWER48; + u64 physical = Z_LOWER48; + + mcache->jdata[0].tlba = cpu_to_le64(logical); + mcache->jdata[0].bval = cpu_to_le64(physical); + mcache->jsize = Z_UNSORTED - 1; + + } else { + Z_ERR(znd, "%s: in memory journal is out of space.", + __func__); + ZDM_FREE(znd, mcache, sizeof(*mcache), KM_07); + mcache = NULL; + } + } + return mcache; +} + +/** + * mcache_first_get() - Get the first chunk from the list. + * @znd: ZDM Instance + * @mcache: Map cache chunk to add. + * @type: Type flag of mcache entry (MAP or DISCARD extent). + * + * The mcache chunk will have it's refcount elevated in the xchg. + * Caller should use xchg_next to navigate through the list or + * deref_cache when exiting from the list walk early. + */ +static struct map_cache *mcache_first_get(struct zoned *znd, int type) +{ + unsigned long flags; + struct map_cache *mc; + + spin_lock_irqsave(&znd->mclck[type], flags); + mc = list_first_entry_or_null(&znd->mclist[type], typeof(*mc), mclist); + if (mc && (&mc->mclist != &znd->mclist[type])) + mcache_ref(mc); + else + mc = NULL; + spin_unlock_irqrestore(&znd->mclck[type], flags); + + return mc; +} + +/** + * mcache_put_get_next() - Get the next map cache page and put current. + * @znd: ZDM Instance + * @mcache: Map cache chunk to add. + * @type: Type flag of mcache entry (MAP or DISCARD extent). + * + * The current mcache chunk will have an elevated refcount, + * that will dropped, while the new the returned chunk will + * have it's refcount elevated in the xchg. + */ +static inline struct map_cache *mcache_put_get_next(struct zoned *znd, + struct map_cache *mcache, + int type) +{ + unsigned long flags; + struct map_cache *next; + + spin_lock_irqsave(&znd->mclck[type], flags); + next = list_next_entry(mcache, mclist); + if (next && (&next->mclist != &znd->mclist[type])) + mcache_ref(next); + else + next = NULL; + mcache_deref(mcache); + spin_unlock_irqrestore(&znd->mclck[type], flags); + + return next; +} + +/** + * mclist_add() - Add a new chunk to the map_cache pool. + * @znd: ZDM Instance + * @mcache: Map cache chunk to add. + * @type: Type flag of mcache entry (MAP or DISCARD extent). + */ +static inline void mclist_add(struct zoned *znd, + struct map_cache *mcache, int type) +{ + unsigned long flags; + + spin_lock_irqsave(&znd->mclck[type], flags); + list_add(&mcache->mclist, &znd->mclist[type]); + spin_unlock_irqrestore(&znd->mclck[type], flags); +} + +/** + * mclist_del() - Release an (empty) chunk of map cache. + * @znd: ZDM Instance + * @mcache: Map cache chunk to add. + * @type: Type flag of mcache entry (MAP or DISCARD extent). + */ +static inline int mclist_del(struct zoned *znd, + struct map_cache *mcache, int type) +{ + unsigned long flags; + int deleted = 0; + + if (spin_trylock_irqsave(&znd->mclck[type], flags)) { + MutexLock(&mcache->cached_lock); + if (mcache->jcount == 0 && mcache_getref(mcache) == 1) { + list_del(&mcache->mclist); + deleted = 1; + } + mutex_unlock(&mcache->cached_lock); + spin_unlock_irqrestore(&znd->mclck[type], flags); + } + return deleted; +} + +/** + * mcache_put() - Dereference an mcache page + * @mcache: Map cache page. + */ +static inline void mcache_put(struct map_cache *mcache) +{ + mcache_deref(mcache); +} + +/** + * memcache_sort() - Sort an unsorted map cache page + * @znd: ZDM instance + * @mcache: Map cache page. + * + * Sort a map cache entry if is sorted. + * Lock using mutex if not already locked. + */ +static void memcache_sort(struct zoned *znd, struct map_cache *mcache) +{ + mcache_ref(mcache); + if (mcache->jsorted < mcache->jcount) { + sort(&mcache->jdata[1], mcache->jcount, + sizeof(*mcache->jdata), compare_tlba, NULL); + mcache->jsorted = mcache->jcount; + } + mcache_deref(mcache); +} + +/** + * memcache_lock_and_sort() - Attempt to sort an unsorted map cache page + * @znd: ZDM instance + * @mcache: Map cache page. + * + * Sort a map cache entry if is not busy being purged to the ZLT. + * + * Return: -EBUSY if busy, or 0 if sorted. + */ +static int memcache_lock_and_sort(struct zoned *znd, struct map_cache *mcache) +{ + if (mcache_is_busy(mcache)) + return -EBUSY; + + if (mcache->jsorted == mcache->jcount) + return 0; + + mutex_lock_nested(&mcache->cached_lock, SINGLE_DEPTH_NESTING); + memcache_sort(znd, mcache); + mutex_unlock(&mcache->cached_lock); + return 0; +} + +/** + * z_lookup_cache() - Scan mcache entries for addr + * @znd: ZDM Instance + * @addr: Address [tLBA] to find. + * @type: mcache type (MAP or DISCARD cache). + */ +static u64 z_lookup_cache(struct zoned *znd, u64 addr, int type) +{ + struct map_cache *mcache; + u64 found = 0ul; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int at; + int err; + + /* sort, if needed. only err is -EBUSY so do a linear find. */ + err = memcache_lock_and_sort(znd, mcache); + if (!err) + at = mcache_bsearch(mcache, addr); + else + at = mcache_lsearch(mcache, addr); + + if (at != -1) { + struct map_sect_to_lba *data = &mcache->jdata[at + 1]; + + found = le64_to_lba48(data->bval, NULL); + } + if (found) { + mcache_put(mcache); + goto out; + } + + mcache = mcache_put_get_next(znd, mcache, type); + } +out: + return found; +} + +/** + * mcache_insert - Insertion sort (for mostly sorted data) + * @znd: ZDM Instance + * @mcache: Map Cache block + * @tlba: upper stack sector + * @blba: lower mapped lba + */ +static void mcache_insert(struct zoned *znd, struct map_cache *mcache, + u64 tlba, u64 blba) +{ + u16 fido = ++mcache->jcount; + u16 idx; + + WARN_ON(mcache->jcount > mcache->jsize); + + for (idx = fido; idx > 1; --idx) { + u16 prev = idx - 1; + u64 a0 = le64_to_lba48(mcache->jdata[prev].tlba, NULL); + + if (a0 < tlba) { + mcache->jdata[idx].tlba = lba48_to_le64(0, tlba); + mcache->jdata[idx].bval = lba48_to_le64(0, blba); + if ((mcache->jsorted + 1) == mcache->jcount) + mcache->jsorted = mcache->jcount; + break; + } + /* move UP .. tlba < a0 */ + mcache->jdata[idx].tlba = mcache->jdata[prev].tlba; + mcache->jdata[idx].bval = mcache->jdata[prev].bval; + } + if (idx == 1) { + mcache->jdata[idx].tlba = lba48_to_le64(0, tlba); + mcache->jdata[idx].bval = lba48_to_le64(0, blba); + mcache->jsorted = mcache->jcount; + } +} + + +/** + * mc_delete_entry - Overwrite a map_cache entry (for mostly sorted data) + * @mcache: Map Cache block + * @entry: Entry to remove. [1 .. count] + */ +void mc_delete_entry(struct map_cache *mcache, int entry) +{ + int iter; + + MutexLock(&mcache->cached_lock); + + mcache->jdata[entry].tlba = lba48_to_le64(0, 0ul); + mcache->jdata[entry].bval = lba48_to_le64(0, 0ul); + + for (iter = entry; iter < mcache->jcount; iter++) { + int next = iter + 1; + + mcache->jdata[iter].tlba = mcache->jdata[next].tlba; + mcache->jdata[iter].bval = mcache->jdata[next].bval; + } + if (mcache->jcount > 0) { + if (mcache->jsorted == mcache->jcount) + mcache->jsorted--; + mcache->jcount--; + } else { + pr_err("mcache delete ... discard beyound empty\n"); + } + mutex_unlock(&mcache->cached_lock); +} + +/** + * _discard_split_entry() - Split a discard into two parts. + * @znd: ZDM Instance + * @mcahce: mcache holding the current entry + * @at: location of entry in mcache (0 .. count - 1) + * @addr: address to punch a hole at + * @gfp: current memory allocation rules. + * + * Discard 'extents' need to broken up when a sub range is being allocated + * for writing. Ideally the punch only requires the extent to be shrunk. + */ +static int _discard_split_entry(struct zoned *znd, struct map_cache *mcache, + int at, u64 addr, gfp_t gfp) +{ + const int merge = 0; + int err = 0; + struct map_sect_to_lba *entry = &mcache->jdata[at + 1]; + u64 tlba = le64_to_lba48(entry->tlba, NULL); + u64 blocks = le64_to_lba48(entry->bval, NULL); + u64 dblks = blocks; + const u64 chunk = addr - tlba; + + /* + * Now chunk is a number of blocks until addr starts. + * If the chunk is small push the discard immediately. + * Otherwise punch a hole and see if we need to create + * a new entry with what remains. + */ + if (chunk > 0 && chunk < DISCARD_MAX_INGRESS) { + + err = mapped_discard(znd, tlba, chunk, merge, gfp); + if (err) + goto out; + + blocks -= chunk; + tlba += chunk; + + entry->tlba = lba48_to_le64(0, tlba); + entry->bval = lba48_to_le64(0, blocks); + + if (blocks == 0ul) { + mc_delete_entry(mcache, at + 1); + goto out; + } + dblks = blocks; + } else { + dblks = blocks - chunk; /* nblks from addr to end */ + } + + /* + * When chunk >= DISCARD_MAX_INGRESS + * then the current entry needs to be ended at 'chunk'. + * Note that tlba != addr in this case and 'chunk' will + * be until we get around to committing to truncating + * the entry extent. Specifcily we need to ensure that + * the hole doens't expose blocks covered by the current + * extent. + */ + if (dblks > DISCARD_MAX_INGRESS) + dblks = DISCARD_MAX_INGRESS - 1; + + err = mapped_discard(znd, addr, dblks, merge, gfp); + if (err) + goto out; + + /* + * Now we have moved the hole into the map cache we are free + * to update the entry. + * If tlba == addr the we have continued to discard from the + * from of this extent. + * + * If tlba != addr then we need discarded from the middle or + * the end. Either there is a hole and we need a second extent + * entry or we ran to the end of the extent and we can just + * truncate the extent. + */ + + if (tlba == addr) { + blocks -= dblks; + tlba += dblks; + entry->tlba = lba48_to_le64(0, tlba); + entry->bval = lba48_to_le64(0, blocks); + if (blocks == 0ul) + mc_delete_entry(mcache, at + 1); + } else { + u64 start = addr + dblks; + + /* + * Because we are doing insertion sorting + * and tlba < addr we can safely add a new + * discard extent entry. + * As the current entry is 'stable' we don't need to + * re-acquire it. + */ + blocks -= (chunk + dblks); + if (blocks > 0ul) { + err = mapped_discard(znd, start, blocks, merge, gfp); + if (err) + goto out; + } + entry->bval = lba48_to_le64(0, chunk); + } + +out: + return err; +} + + +/** + * z_discard_range() - Migrate a discard range to memcache/ZLT + * @znd: ZDM Instance + * @addr: tLBA target for breaking up a discard range. + * @gfp: Memory allocation scheme + * + * When an address is being written to a check is made to see if the tLBA + * overlaps with an exisitng discard pool. If so the discard pool is + * split and multiple smaller discard entries are pushed into the memcache. + * + * The overall purpose of the discard pool is to reduce the amount of intake + * on the memcache to avoid starving the I/O requests. + */ +static u64 z_discard_range(struct zoned *znd, u64 addr, gfp_t gfp) +{ + struct map_cache *mcache; + u64 found = 0ul; + int type = IS_DISCARD; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int at; + int err; + + /* sort, if needed. only err is -EBUSY so do a linear find. */ + err = memcache_lock_and_sort(znd, mcache); + if (!err) + at = mcache_bsearch(mcache, addr); + else + at = mcache_lsearch(mcache, addr); + + if (at != -1) { + /* break apart the entry */ + err = _discard_split_entry(znd, mcache, at, addr, gfp); + if (err) { + mcache_put(mcache); + goto out; + } + found = 1; + } + + if (found) { + mcache_put(mcache); + goto out; + } + + mcache = mcache_put_get_next(znd, mcache, type); + } +out: + return found; +} + +/** + * z_discard_partial() - Migrate a discard range to memcache/ZLT + * @znd: ZDM Instance + * @minblks: Target number of blocks to cull. + * @gfp: Memory allocation scheme + * + * When an address is being written to a check is made to see if the tLBA + * overlaps with an exisitng discard pool. If so the discard pool is + * split and multiple smaller discard entries are pushed into the memcache. + * + * The overall purpose of the discard pool is to reduce the amount of intake + * on the memcache to avoid starving the I/O requests. + */ +static int z_discard_partial(struct zoned *znd, u32 minblks, gfp_t gfp) +{ + struct map_cache *mcache; + int completions = 0; + int type = IS_DISCARD; + int lumax = minblks; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int at = -1; + int err; + + if (mcache->jcount > 1) + at = 0; + + if (at != -1) { + const int nodisc = 1; + s32 e = at + 1; + u64 tlba = le64_to_lba48(mcache->jdata[e].tlba, NULL); + u64 blks = le64_to_lba48(mcache->jdata[e].bval, NULL); + u64 chunk = DISCARD_MAX_INGRESS - 1; + u32 dcount = 0; + u32 dstop = minblks; + + while (blks > 0 && dcount < dstop && --lumax > 0) { + u64 lba = _current_mapping(znd, nodisc, + tlba, gfp); + if (lba) { + do { + err = z_to_map_list(znd, tlba, + 0ul, gfp); + } while (-EBUSY == err); + if (err) + break; + dcount++; + } + tlba++; + blks--; + } + + if (chunk > blks) + chunk = blks; + + if (chunk == 0) { + mc_delete_entry(mcache, e); + completions++; + } else if (chunk < DISCARD_MAX_INGRESS) { + err = mapped_discard(znd, tlba, chunk, 0, gfp); + if (err) { + mcache_put(mcache); + goto out; + } + blks -= chunk; + tlba += chunk; + + mcache->jdata[e].tlba = lba48_to_le64(0, tlba); + mcache->jdata[e].bval = lba48_to_le64(0, blks); + if (blks == 0ul) + mc_delete_entry(mcache, e); + } + completions++; + } + if (completions) { + int deleted = 0; + + if (completions > 1 && + znd->dc_entries > MEMCACHE_HWMARK && + mcache->jcount == 0 && + mcache_getref(mcache) == 1) + deleted = mclist_del(znd, mcache, type); + + mcache_put(mcache); + + if (deleted) { + ZDM_FREE(znd, mcache->jdata, Z_C4K, PG_08); + ZDM_FREE(znd, mcache, sizeof(*mcache), KM_07); + mcache = NULL; + znd->dc_entries--; + } + goto out; + } + mcache = mcache_put_get_next(znd, mcache, type); + } +out: + return completions; +} + +/** + * lba_in_zone() - Scan a map cache entry for any targets pointing to zone. + * @znd: ZDM instance + * @mcache: Map cache page. + * @zone: Zone (that was gc'd). + * + * Scan the map cache page to see if any entries are pointing to a zone. + * + * Return: 0 if none are found. 1 if any are found. + */ +static int lba_in_zone(struct zoned *znd, struct map_cache *mcache, u32 zone) +{ + int jentry; + + if (zone >= znd->data_zones) + goto out; + + for (jentry = mcache->jcount; jentry > 0; jentry--) { + u64 lba = le64_to_lba48(mcache->jdata[jentry].bval, NULL); + + if (lba && _calc_zone(znd, lba) == zone) + return 1; + } +out: + return 0; +} + +/** + * gc_verify_cache() - Test map cache to see if any entries point to zone. + * @znd: ZDM instance + * @zone: Zone (that was gc'd). + * + * Scan the map cache to see if any entries are pointing to a zone that + * was just GC'd. + * If any are found it indicates a bug. + * + * Return: 0 on success or 1 on error and sets meta_result to ENOSPC + * to effectivly freeze ZDM. + */ +static int gc_verify_cache(struct zoned *znd, u32 zone) +{ + struct map_cache *mcache = NULL; + int err = 0; + int type = IS_MAP; + + mcache = mcache_first_get(znd, type); + while (mcache) { + MutexLock(&mcache->cached_lock); + if (lba_in_zone(znd, mcache, zone)) { + Z_ERR(znd, "GC: **ERR** %" PRIx32 + " LBA in cache <= Corrupt", zone); + err = 1; + znd->meta_result = -ENOSPC; + } + mutex_unlock(&mcache->cached_lock); + mcache = mcache_put_get_next(znd, mcache, type); + } + + return err; +} + + +/** + * __cached_to_tables() - Migrate map cache entries to ZLT + * @znd: ZDM instance + * @type: Which type (MAP) of cache entries to migrate. + * @zone: zone to force migration for partial memcache block + * + * Scan the memcache and move any full blocks to lookup tables + * If a (the) partial memcache block contains lbas that map to zone force + * early migration of the memcache block to ensure it is properly accounted + * for and migrated during and upcoming GC pass. + * + * Return: 0 on success or -errno value + */ +static int __cached_to_tables(struct zoned *znd, int type, u32 zone) +{ + struct map_cache *mcache = NULL; + int err = 0; + int try_free = 0; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int deleted = 0; + struct map_cache *jskip; + + mcache_busy(mcache); + MutexLock(&mcache->cached_lock); + if (mcache->jcount == mcache->jsize) { + memcache_sort(znd, mcache); + err = move_to_map_tables(znd, mcache); + if (!err && (mcache->jcount == 0)) + try_free++; + } else { + if (lba_in_zone(znd, mcache, zone)) { + Z_ERR(znd, + "Moving %d Runts because z: %u", + mcache->jcount, zone); + + memcache_sort(znd, mcache); + err = move_to_map_tables(znd, mcache); + } + } + mutex_unlock(&mcache->cached_lock); + mcache_unbusy(mcache); + if (err) { + mcache_put(mcache); + Z_ERR(znd, "%s: Sector map failed.", __func__); + goto out; + } + + mcache_ref(mcache); + jskip = mcache_put_get_next(znd, mcache, type); + + if (try_free > 0 && znd->mc_entries > MEMCACHE_HWMARK && + mcache->jcount == 0 && mcache_getref(mcache) == 1) { + deleted = mclist_del(znd, mcache, type); + try_free--; + } + + mcache_deref(mcache); + + if (deleted) { + ZDM_FREE(znd, mcache->jdata, Z_C4K, PG_08); + ZDM_FREE(znd, mcache, sizeof(*mcache), KM_07); + mcache = NULL; + znd->mc_entries--; + } + mcache = jskip; + } +out: + + return err; +} + + +/** + * _cached_to_tables() - Migrate memcache entries to lookup tables + * @znd: ZDM instance + * @zone: zone to force migration for partial memcache block + * + * Scan the memcache and move any full blocks to lookup tables + * If a (the) partial memcache block contains lbas that map to zone force + * early migration of the memcache block to ensure it is properly accounted + * for and migrated during and upcoming GC pass. + * + * Return: 0 on success or -errno value + */ +static int _cached_to_tables(struct zoned *znd, u32 zone) +{ + int err; + + err = __cached_to_tables(znd, IS_MAP, zone); + return err; +} + + +/** + * z_flush_bdev() - Request backing device flushed to disk. + * @znd: ZDM instance + * + * Return: 0 on success or -errno value + */ +static int z_flush_bdev(struct zoned *znd, gfp_t gfp) +{ + int err; + sector_t bi_done; + + err = blkdev_issue_flush(znd->dev->bdev, gfp, &bi_done); + if (err) + Z_ERR(znd, "%s: flush failing sector %lu!", __func__, bi_done); + + return err; +} + +/** + * pg_delete - Free a map_pg with spin locks held. + * @znd: ZDM Instance + * @expg: Page being released. + * + * Forced inline as it is 'optional' and because it is called with + * spin locks enabled and only from a single caller. + */ +static __always_inline int pg_delete(struct zoned *znd, struct map_pg *expg) +{ + struct map_pg **zlt; + spinlock_t *lock; + int req_flush = 0; + int dropped = 0; + int index = expg->index; + int is_lut = test_bit(IS_LUT, &expg->flags); + int is_fwd = test_bit(IS_FWD, &expg->flags); + u32 count = is_lut ? znd->map_count : znd->crc_count; + + if (is_lut) { + zlt = is_fwd ? znd->fwd_tm : znd->rev_tm; + lock = &znd->mapkey_lock; + } else { + zlt = is_fwd ? znd->fwd_crc : znd->rev_crc; + lock = &znd->ct_lock; + } + + if (!spin_trylock(lock)) + return req_flush; + + if (index > -1 && index < count && expg == zlt[index] && + test_bit(IS_DROPPED, &expg->flags)) { + zlt[index] = NULL; + MutexLock(&expg->md_lock); + if (test_and_clear_bit(IS_LAZY, &expg->flags)) { + clear_bit(IS_DROPPED, &expg->flags); + clear_bit(DELAY_ADD, &expg->flags); + if (expg->data.addr) { + void *pg = expg->data.addr; + + expg->data.addr = NULL; + ZDM_FREE(znd, pg, Z_C4K, PG_27); + atomic_dec(&znd->incore); + } else { + Z_ERR(znd, "** No data pg? %llx", expg->lba); + } + list_del(&expg->lazy); + znd->in_lzy--; + req_flush = !test_bit(IS_FLUSH, &expg->flags); + dropped = 1; + } else { + Z_ERR(znd, "Detected double list del."); + } + mutex_unlock(&expg->md_lock); + if (dropped) + ZDM_FREE(znd, expg, sizeof(*expg), KM_20); + } + spin_unlock(lock); + return req_flush; +} + +/** + * manage_lazy_activity() - Migrate delayed 'add' entries to the 'ZTL' + * @znd: ZDM Instance + * + * The lzy list is used to perform less critical activities that could + * be done via the ZTL primary list but gives a second chance when + * - Adding if the spin lock would lock. + * - Deleting ... if the cache entry turns out to be 'hotter' than + * the default we can catch it and make it 'hotter' before the + * hotness indicator is lost. + */ +static int manage_lazy_activity(struct zoned *znd) +{ + struct map_pg *expg; + struct map_pg *_tpg; + int want_flush = 0; + const u32 msecs = MEM_PURGE_MSECS; + + spin_lock(&znd->lzy_lck); + expg = list_first_entry_or_null(&znd->lzy_pool, typeof(*expg), lazy); + if (!expg || (&expg->lazy == &znd->lzy_pool)) + goto out; + + _tpg = list_next_entry(expg, lazy); + while (&expg->lazy != &znd->lzy_pool) { + /* + * this should never happen: + */ + if (test_bit(IS_DIRTY, &expg->flags)) + set_bit(DELAY_ADD, &expg->flags); + /* + * Migrage pg to zltlst list + */ + if (test_bit(DELAY_ADD, &expg->flags)) { + if (spin_trylock(&znd->zlt_lck)) { + if (!test_bit(IN_ZLT, &expg->flags)) { + list_del(&expg->lazy); + znd->in_lzy--; + clear_bit(IS_LAZY, &expg->flags); + clear_bit(IS_DROPPED, &expg->flags); + clear_bit(DELAY_ADD, &expg->flags); + set_bit(IN_ZLT, &expg->flags); + list_add(&expg->zltlst, &znd->zltpool); + znd->in_zlt++; + } else { + Z_ERR(znd, "** ZLT double add? %llx", + expg->lba); + } + spin_unlock(&znd->zlt_lck); + } +#if ENABLE_PG_FREE_VIA_LAZY + } else { + /* + * Delete page + */ + if (!test_bit(IN_ZLT, &expg->flags) && + test_bit(IS_DROPPED, &expg->flags)) { + if (is_expired_msecs(expg->age, msecs)) + want_flush |= pg_delete(znd, expg); + } +#endif + } + expg = _tpg; + _tpg = list_next_entry(expg, lazy); + } + +out: + spin_unlock(&znd->lzy_lck); + return want_flush; +} + + + +/** + * mark_clean_flush_zlt() - Mark all non-dirty ZLT blocks as 'FLUSH' + * @znd: ZDM instance + * + * After a FLUSH/FUA these blocks are on disk and redundant FLUSH + * can be skipped if the block is later ejected. + */ +static void mark_clean_flush_zlt(struct zoned *znd) +{ + struct map_pg *expg = NULL; + struct map_pg *_tpg; + + spin_lock(&znd->zlt_lck); + if (list_empty(&znd->zltpool)) + goto out; + + expg = list_last_entry(&znd->zltpool, typeof(*expg), zltlst); + if (!expg || &expg->zltlst == (&znd->zltpool)) + goto out; + + _tpg = list_prev_entry(expg, zltlst); + while (&expg->zltlst != &znd->zltpool) { + ref_pg(expg); + if (!test_bit(IS_DIRTY, &expg->flags)) + set_bit(IS_FLUSH, &expg->flags); + deref_pg(expg); + expg = _tpg; + _tpg = list_prev_entry(expg, zltlst); + } + +out: + spin_unlock(&znd->zlt_lck); +} + +/** + * mark_clean_flush_lzy() - Mark all non-dirty ZLT blocks as 'FLUSH' + * @znd: ZDM instance + * + * After a FLUSH/FUA these blocks are on disk and redundant FLUSH + * can be skipped if the block is later ejected. + */ +static void mark_clean_flush_lzy(struct zoned *znd) +{ + struct map_pg *expg = NULL; + struct map_pg *_tpg; + + spin_lock(&znd->lzy_lck); + expg = list_first_entry_or_null(&znd->lzy_pool, typeof(*expg), lazy); + if (!expg || (&expg->lazy == &znd->lzy_pool)) + goto out; + + _tpg = list_next_entry(expg, lazy); + while (&expg->lazy != &znd->lzy_pool) { + ref_pg(expg); + if (!test_bit(IS_DIRTY, &expg->flags)) + set_bit(IS_FLUSH, &expg->flags); + deref_pg(expg); + expg = _tpg; + _tpg = list_next_entry(expg, lazy); + } + +out: + spin_unlock(&znd->lzy_lck); +} + +/** + * mark_clean_flush() - Mark all non-dirty ZLT/LZY blocks as 'FLUSH' + * @znd: ZDM instance + * + * After a FLUSH/FUA these blocks are on disk and redundant FLUSH + * can be skipped if the block is later ejected. + */ +static void mark_clean_flush(struct zoned *znd) +{ + mark_clean_flush_zlt(znd); + mark_clean_flush_lzy(znd); +} + +/** + * do_sync_metadata() - Write ZDM state to disk. + * @znd: ZDM instance + * + * Return: 0 on success or -errno value + */ +static int do_sync_metadata(struct zoned *znd, int sync, int drop) +{ + int err = 0; + int want_flush; + + want_flush = manage_lazy_activity(znd); + if (want_flush) + set_bit(DO_FLUSH, &znd->flags); + + /* if drop is non-zero, DO_FLUSH may be set on return */ + err = sync_mapped_pages(znd, sync, drop); + if (err) { + Z_ERR(znd, "Uh oh: sync_mapped_pages -> %d", err); + goto out; + } + + /* + * If we are lucky then this sync will get us to a 'clean' + * state and the follow on bdev flush is redunant and skipped + * + * If not we will suffer a performance stall because we were + * ejected blocks. + * + * TODO: On Sync/Flush/FUA we can mark all of our clean ZLT + * as flushed and we can bypass elevating the drop count + * to trigger a flush for such already flushed blocks. + */ + want_flush = test_bit(DO_FLUSH, &znd->flags); + err = z_mapped_sync(znd); + if (err) { + Z_ERR(znd, "Uh oh. z_mapped_sync -> %d", err); + goto out; + } + + if (test_and_clear_bit(DO_FLUSH, &znd->flags)) { + err = z_flush_bdev(znd, GFP_KERNEL); + if (err) { + Z_ERR(znd, "Uh oh. flush_bdev failed. -> %d", err); + goto out; + } + want_flush = 1; + } + if (want_flush) + mark_clean_flush(znd); + +out: + return err; +} + +/** + * do_init_from_journal() - Restore ZDM state from disk. + * @znd: ZDM instance + * + * Return: 0 on success or -errno value + */ +static int do_init_from_journal(struct zoned *znd) +{ + int err = 0; + + if (test_and_clear_bit(DO_JOURNAL_LOAD, &znd->flags)) + err = z_mapped_init(znd); + + return err; +} + +/** + * do_journal_to_table() - Migrate memcache entries to lookup tables + * @znd: ZDM instance + * + * Return: 0 on success or -errno value + */ +static int do_journal_to_table(struct zoned *znd) +{ + int err = 0; + + if (test_and_clear_bit(DO_JOURNAL_MOVE, &znd->flags) || + test_bit(DO_SYNC, &znd->flags)) + err = _cached_to_tables(znd, znd->data_zones); + + return err; +} + +/** + * do_sync_to_disk() - Write ZDM state to disk. + * @znd: ZDM instance + * + * Return: 0 on success or -errno value + */ +static int do_sync_to_disk(struct zoned *znd) +{ + int err = 0; + int drop = 0; + int sync = 0; + + if (test_and_clear_bit(DO_SYNC, &znd->flags) || + test_bit(DO_FLUSH, &znd->flags)) + sync = 1; + + if (test_and_clear_bit(DO_MEMPOOL, &znd->flags)) { + int pool_size = MZ_MEMPOOL_SZ * 4; + +#if ENABLE_PG_FREE_VIA_LAZY + /** + * Trust our cache miss algo + */ + pool_size = MZ_MEMPOOL_SZ * 3; +#endif + + if (is_expired_msecs(znd->age, MEM_PURGE_MSECS * 2)) + pool_size = 3; + else if (is_expired_msecs(znd->age, MEM_PURGE_MSECS)) + pool_size = MZ_MEMPOOL_SZ; + + if (atomic_read(&znd->incore) > pool_size) + drop = atomic_read(&znd->incore) - pool_size; + } + if (sync || drop) + err = do_sync_metadata(znd, sync, drop); + + return err; +} + +/** + * meta_work_task() - Worker thread from metadata activity. + * @work: Work struct containing ZDM instance. + */ +static void meta_work_task(struct work_struct *work) +{ + int err = 0; + int locked = 0; + struct zoned *znd; + + if (!work) + return; + + znd = container_of(work, struct zoned, meta_work); + if (!znd) + return; + + err = do_init_from_journal(znd); + + if (test_bit(DO_JOURNAL_MOVE, &znd->flags) || + test_bit(DO_FLUSH, &znd->flags) || + test_bit(DO_SYNC, &znd->flags)) { + MutexLock(&znd->mz_io_mutex); + locked = 1; + } + + /* + * Reduce memory pressure on journal list of arrays + * by pushing them into the sector map lookup tables + */ + if (!err) + err = do_journal_to_table(znd); + + /* force a consistent set of meta data out to disk */ + if (!err) + err = do_sync_to_disk(znd); + + if (locked) + mutex_unlock(&znd->mz_io_mutex); + + znd->age = jiffies_64; + if (err < 0) + znd->meta_result = err; + + clear_bit(DO_METAWORK_QD, &znd->flags); +} + +/** + * _update_disc() - Update an existing map cache entry. + * @znd: ZDM instance + * @mcache: Map cache block + * @at: Data entry in map cache block + * @dm_s: tBLA mapping from + * @lba: New lba + * @gfp: Allocation flags to use. + * + * If overwriting an entry will trigger an 'unused' to be pushed recording + * the old block being stale. + */ +static int _update_entry(struct zoned *znd, struct map_cache *mcache, int at, + u64 dm_s, u64 lba, gfp_t gfp) +{ + struct map_sect_to_lba *data; + u64 lba_was; + int err = 0; + + data = &mcache->jdata[at + 1]; + lba_was = le64_to_lba48(data->bval, NULL); + lba &= Z_LOWER48; + data->bval = lba48_to_le64(0, lba); + + if (lba != lba_was) { + Z_DBG(znd, "Remap %" PRIx64 " -> %" PRIx64 + " (was %" PRIx64 "->%" PRIx64 ")", + dm_s, lba, le64_to_lba48(data->tlba, NULL), lba_was); + err = unused_phy(znd, lba_was, 0, gfp); + if (err == 1) + err = 0; + } + + return err; +} + +/** + * _update_disc() - Update an existing discard cache entry + * @znd: ZDM instance + * @mcache: Map cache block + * @at: Data entry in map cache block + * @dm_s: tBLA mapping from + * @blks: number of blocks + */ +static int _update_disc(struct zoned *znd, struct map_cache *mcache, int at, + u64 dm_s, u64 blks) +{ + struct map_sect_to_lba *data; + u64 oldsz; + int err = 0; + + data = &mcache->jdata[at + 1]; + oldsz = le64_to_lba48(data->bval, NULL); + data->bval = lba48_to_le64(0, blks); + + return err; +} + +/** + * mc_update() - Update an existing cache entry + * @znd: ZDM instance + * @mcache: Map cache block + * @at: Data entry in map cache block + * @dm_s: tBLA mapping from + * @lba: Value (lba or number of blocks) + * @type: List (type) to be adding to (MAP or DISCARD) + * @gfp: Allocation (kmalloc) flags + */ +static int mc_update(struct zoned *znd, struct map_cache *mcache, int at, + u64 dm_s, u64 lba, int type, gfp_t gfp) +{ + int rcode; + + if (type == IS_MAP) + rcode = _update_entry(znd, mcache, at, dm_s, lba, gfp); + else + rcode = _update_disc(znd, mcache, at, dm_s, lba); + + return rcode; +} + +/** + * _mapped_list() - Add a entry to the discard or map cache + * @znd: ZDM instance + * @dm_s: tBLA mapping from + * @lba: Value (lba or number of blocks) + * @type: List (type) to be adding to (MAP or DISCARD) + * @gfp: Allocation (kmalloc) flags + */ +static +int _mapped_list(struct zoned *znd, u64 dm_s, u64 lba, int type, gfp_t gfp) +{ + struct map_cache *mcache = NULL; + struct map_cache *mc_add = NULL; + int handled = 0; + int list_count = 0; + int err = 0; + int add_to_list = 0; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int at; + + MutexLock(&mcache->cached_lock); + memcache_sort(znd, mcache); + at = _bsrch_tlba(mcache, dm_s); + if (at != -1) { + mcache_ref(mcache); + err = mc_update(znd, mcache, at, dm_s, lba, type, gfp); + mcache_deref(mcache); + handled = 1; + } else if (!mc_add) { + if (mcache->jcount < mcache->jsize) { + mc_add = mcache; + mcache_ref(mc_add); + } + } + mutex_unlock(&mcache->cached_lock); + if (handled) { + if (mc_add) + mcache_deref(mc_add); + mcache_put(mcache); + goto out; + } + mcache = mcache_put_get_next(znd, mcache, type); + list_count++; + } + + /* ------------------------------------------------------------------ */ + /* ------------------------------------------------------------------ */ + + if (!mc_add) { + mc_add = mcache_alloc(znd, gfp); + if (!mc_add) { + Z_ERR(znd, "%s: in memory journal is out of space.", + __func__); + err = -ENOMEM; + goto out; + } + mcache_ref(mc_add); + if (type == IS_MAP) { + if (list_count > MEMCACHE_HWMARK) + set_bit(DO_JOURNAL_MOVE, &znd->flags); + znd->mc_entries = list_count + 1; + } else { + znd->dc_entries = list_count + 1; + } + add_to_list = 1; + } + + /* ------------------------------------------------------------------ */ + /* ------------------------------------------------------------------ */ + + if (mc_add) { + MutexLock(&mc_add->cached_lock); + + if (!mc_add->jdata) { + mc_add->jdata = ZDM_CALLOC(znd, Z_UNSORTED, + sizeof(*mc_add->jdata), PG_08, gfp); + } + if (!mc_add->jdata) { + Z_ERR(znd, "%s: in memory journal is out of space.", + __func__); + err = -ENOMEM; + goto out; + } + + if (mc_add->jcount < mc_add->jsize) { + mcache_insert(znd, mc_add, dm_s, lba); + if (add_to_list) + mclist_add(znd, mc_add, type); + } else { + err = -EBUSY; + } + mutex_unlock(&mc_add->cached_lock); + mcache_put(mc_add); + } +out: + return err; +} + +/** + * z_to_map_list() - Add a discard extent entry to the discard cache + * @znd: ZDM instance + * @dm_s: tBLA mapping from + * @blks: Number of blocks in the extent. + * @gfp: Allocation (kmalloc) flags + */ +static int z_to_discard_list(struct zoned *znd, u64 dm_s, u64 blks, gfp_t gfp) +{ + return _mapped_list(znd, dm_s, blks, IS_DISCARD, gfp); +} + +/** + * z_to_map_list() - Add a map cache entry to the map cache + * @znd: ZDM instance + * @dm_s: tBLA mapping from + * @lba: bLBA mapping to + * @gfp: Allocation (kmalloc) flags + */ +static int z_to_map_list(struct zoned *znd, u64 dm_s, u64 lba, gfp_t gfp) +{ + return _mapped_list(znd, dm_s, lba, IS_MAP, gfp); +} + +/** + * discard_merge() - Merge a discard request with a existing entry. + * @znd: ZDM Instance + * @tlba: Starting address + * @blks: Number of blocks in discard. + * @gfp: Current memory allocation scheme. + * + */ +static int discard_merge(struct zoned *znd, u64 tlba, u64 blks) +{ + struct map_cache *mcache; + u64 ends = tlba - 1; + u64 next = tlba + blks; + int type = IS_DISCARD; + int merged = 0; + + mcache = mcache_first_get(znd, type); + while (mcache) { + int at; + int err; + + /* + * 1) Modify existing discard entry, if it exists + * 2) Otherwise: If an existing entry *starts* where this + * entry *ends* extend that entry. + * 3) Otherwise: If an existing entry ends where this + * entry starts, extend *that* entry. + */ + err = memcache_lock_and_sort(znd, mcache); + if (err == -EBUSY) + at = _lsearch_tlba(mcache, tlba); + else + at = _bsrch_tlba_lck(mcache, tlba); + + if (at != -1) { + mcache_ref(mcache); + err = _update_disc(znd, mcache, at, tlba, blks); + mcache_deref(mcache); + merged = 1; + at = -1; + } + + /* Existing entry stars where this discard ends */ + if (!merged) { + if (err == -EBUSY) + at = _lsearch_tlba(mcache, next); + else + at = _bsrch_tlba_lck(mcache, next); + } + at = _bsrch_tlba(mcache, ends); + if (at != -1) { + struct map_sect_to_lba *data; + u64 oldsz; + + mcache_ref(mcache); + data = &mcache->jdata[at + 1]; + data->tlba = lba48_to_le64(0, tlba); + oldsz = le64_to_lba48(data->bval, NULL); + data->bval = lba48_to_le64(0, oldsz + blks); + mcache_deref(mcache); + merged = 1; + at = -1; + } + + /* + * Find a discard that includes 'ends', if so + * determine if the containing discard should be extended + * or if this discard is fully contained within the current + * discard entry. + */ + if (!merged) { + if (err == -EBUSY) + at = _lsearch_extent(mcache, ends); + else + at = _bsrch_extent_lck(mcache, ends); + } + if (at != -1) { + struct map_sect_to_lba *data; + u64 oldsz; + u64 origin; + u64 extend; + + mcache_ref(mcache); + data = &mcache->jdata[at + 1]; + origin = le64_to_lba48(data->tlba, NULL); + oldsz = le64_to_lba48(data->bval, NULL); + extend = (tlba - origin) + blks; + if (extend > oldsz) + data->bval = lba48_to_le64(0, extend); + mcache_deref(mcache); + merged = 1; + at = -1; + } + if (merged) { + mcache_put(mcache); + goto done; + } + mcache = mcache_put_get_next(znd, mcache, type); + } +done: + return merged; +} + +/** + * next_generation() - Increment generation number for superblock. + * @znd: ZDM instance + */ +static inline u64 next_generation(struct zoned *znd) +{ + u64 generation = le64_to_cpu(znd->bmkeys->generation); + + if (generation == 0) + generation = 2; + + generation++; + if (generation == 0) + generation++; + + return generation; +} + +/** + * z_mapped_sync() - Write cache entries and bump superblock generation. + * @znd: ZDM instance + */ +static int z_mapped_sync(struct zoned *znd) +{ + struct dm_target *ti = znd->ti; + struct map_cache *mcache; + int nblks = 1; + int use_wq = 0; + int rc = 1; + int discards = 0; + int maps = 0; + int jwrote = 0; + int cached = 0; + int idx = 0; + int no; + int need_sync_io = 1; + u64 lba = LBA_SB_START; + u64 generation = next_generation(znd); + u64 modulo = CACHE_COPIES; + u64 incr = MAX_CACHE_INCR; + struct io_4k_block *io_vcache; + + MutexLock(&znd->vcio_lock); + io_vcache = get_io_vcache(znd, NORMAL); + if (!io_vcache) { + Z_ERR(znd, "%s: FAILED to get IO CACHE.", __func__); + rc = -ENOMEM; + goto out; + } + + lba = 0x2048ul; + for (idx = 0; idx < znd->gz_count; idx++) { + struct meta_pg *wpg = &znd->wp[idx]; + + if (test_bit(IS_DIRTY, &wpg->flags)) { + cached = 0; + memcpy(io_vcache[cached].data, wpg->wp_alloc, Z_C4K); + znd->bmkeys->wp_crc[idx] = + crc_md_le16(io_vcache[cached].data, Z_CRC_4K); + cached++; + memcpy(io_vcache[cached].data, wpg->zf_est, Z_C4K); + znd->bmkeys->zf_crc[idx] = + crc_md_le16(io_vcache[cached].data, Z_CRC_4K); + cached++; + rc = write_block(ti, DM_IO_VMA, io_vcache, lba, + cached, use_wq); + if (rc) + goto out; + clear_bit(IS_DIRTY, &wpg->flags); + + Z_DBG(znd, "%d# -- WP: %04x | ZF: %04x", + idx, znd->bmkeys->wp_crc[idx], + znd->bmkeys->zf_crc[idx]); + } + lba += 2; + } + + lba = (generation % modulo) * incr; + if (lba == 0) + lba++; + + znd->bmkeys->generation = cpu_to_le64(generation); + znd->bmkeys->gc_resv = cpu_to_le32(znd->z_gc_resv); + znd->bmkeys->meta_resv = cpu_to_le32(znd->z_meta_resv); + + idx = 0; + for (no = 0; no < MAP_COUNT; no++) { + mcache = mcache_first_get(znd, no); + while (mcache) { + u64 phy = le64_to_lba48(mcache->jdata[0].bval, NULL); + u16 jcount = mcache->jcount & 0xFFFF; + + if (jcount) { + mcache->jdata[0].bval = + lba48_to_le64(jcount, phy); + znd->bmkeys->crcs[idx] = + crc_md_le16(mcache->jdata, Z_CRC_4K); + idx++; + memcpy(io_vcache[cached].data, + mcache->jdata, Z_C4K); + cached++; + + if (no == IS_DISCARD) + discards++; + else + maps++; + } + + if (cached == IO_VCACHE_PAGES) { + rc = write_block(ti, DM_IO_VMA, io_vcache, lba, + cached, use_wq); + if (rc) { + Z_ERR(znd, "%s: cache-> %" PRIu64 + " [%d blks] %p -> %d", + __func__, lba, nblks, + mcache->jdata, rc); + mcache_put(mcache); + goto out; + } + lba += cached; + jwrote += cached; + cached = 0; + } + mcache = mcache_put_get_next(znd, mcache, no); + } + } + jwrote += cached; + if (discards > 40) + Z_ERR(znd, "**WARNING** large discard cache %d", discards); + if (maps > 40) + Z_ERR(znd, "**WARNING** large map cache %d", maps); + + znd->bmkeys->md_crc = crc_md_le16(znd->md_crcs, 2 * Z_CRC_4K); + znd->bmkeys->n_crcs = cpu_to_le16(jwrote); + znd->bmkeys->discards = cpu_to_le16(discards); + znd->bmkeys->maps = cpu_to_le16(maps); + znd->bmkeys->crc32 = 0; + znd->bmkeys->crc32 = cpu_to_le32(crc32c(~0u, znd->bmkeys, Z_CRC_4K)); + if (cached < (IO_VCACHE_PAGES - 3)) { + memcpy(io_vcache[cached].data, znd->bmkeys, Z_C4K); + cached++; + memcpy(io_vcache[cached].data, znd->md_crcs, Z_C4K * 2); + cached += 2; + need_sync_io = 0; + } + + do { + if (cached > 0) { + int rw = WRITE; + + if (!need_sync_io && + test_and_clear_bit(DO_FLUSH, &znd->flags)) + rw |= (REQ_FLUSH | REQ_FUA); + rc = writef_block(ti, rw, DM_IO_VMA, io_vcache, lba, + cached, use_wq); + if (rc) { + Z_ERR(znd, "%s: mcache-> %" PRIu64 + " [%d blks] %p -> %d", + __func__, lba, cached, io_vcache, rc); + goto out; + } + lba += cached; + } + + cached = 0; + if (need_sync_io) { + memcpy(io_vcache[cached].data, znd->bmkeys, Z_C4K); + cached++; + memcpy(io_vcache[cached].data, znd->md_crcs, Z_C4K * 2); + cached += 2; + need_sync_io = 0; + } + } while (cached > 0); + +out: + put_io_vcache(znd, io_vcache); + mutex_unlock(&znd->vcio_lock); + return rc; +} + +/** + * is_key_page() - Probe block for magic and crc to see if it is recognized. + * @_data: ZDM instance + */ +static inline int is_key_page(void *_data) +{ + int is_key = 0; + struct mz_superkey *data = _data; + + /* Starts with Z_KEY_SIG and ends with magic */ + + if (le64_to_cpu(data->sig1) == Z_KEY_SIG && + le64_to_cpu(data->magic) == Z_TABLE_MAGIC) { + __le32 orig = data->crc32; + __le32 crc_check; + + data->crc32 = 0; + crc_check = cpu_to_le32(crc32c(~0u, data, Z_CRC_4K)); + data->crc32 = orig; + if (crc_check == orig) + is_key = 1; + } + return is_key; +} + +/** + * zoned_personality() - Update zdstart value from superblock + * @znd: ZDM instance + * @sblock: Lba to start scanning for superblock. + */ +static inline +void zoned_personality(struct zoned *znd, struct zdm_superblock *sblock) +{ + znd->zdstart = le32_to_cpu(sblock->zdstart); +} + +/** + * find_superblock_at() - Find superblock following lba + * @znd: ZDM instance + * @lba: Lba to start scanning for superblock. + * @use_wq: If a work queue is needed to scanning. + * @do_init: Set zdstart from found superblock. + */ +static +int find_superblock_at(struct zoned *znd, u64 lba, int use_wq, int do_init) +{ + struct dm_target *ti = znd->ti; + int found = 0; + int nblks = 1; + int rc = -ENOMEM; + u32 count = 0; + u64 *data = ZDM_ALLOC(znd, Z_C4K, PG_10, NORMAL); + + if (!data) { + Z_ERR(znd, "No memory for finding generation .."); + return 0; + } + if (lba == 0) + lba++; + do { + rc = read_block(ti, DM_IO_KMEM, data, lba, nblks, use_wq); + if (rc) { + Z_ERR(znd, "%s: read @%" PRIu64 " [%d blks] %p -> %d", + __func__, lba, nblks, data, rc); + goto out; + } + if (is_key_page(data)) { + struct mz_superkey *kblk = (struct mz_superkey *) data; + struct zdm_superblock *sblock = &kblk->sblock; + int err = sb_check(sblock); + + if (!err) { + found = 1; + if (do_init) + zoned_personality(znd, sblock); + } + goto out; + } + if (data[0] == 0 && data[1] == 0) { + /* No SB here. */ + Z_ERR(znd, "FGen: Invalid block %" PRIx64 "?", lba); + goto out; + } + lba++; + count++; + if (count > MAX_CACHE_SYNC) { + Z_ERR(znd, "FSB: Too deep to be useful."); + goto out; + } + } while (!found); + +out: + ZDM_FREE(znd, data, Z_C4K, PG_10); + return found; +} + +/** + * find_superblock() - Find (any) superblock + * @znd: ZDM instance + * @use_wq: If a work queue is needed to scanning. + * @do_init: Set/Retrieve zdstart from found super block. + */ +static int find_superblock(struct zoned *znd, int use_wq, int do_init) +{ + int found = 0; + int iter = 0; + u64 lba = LBA_SB_START; + u64 last = MAX_CACHE_INCR * CACHE_COPIES; + + do { + found = find_superblock_at(znd, lba, use_wq, do_init); + if (found) + break; + iter++; + lba = MAX_CACHE_INCR * iter; + } while (lba < last); + + return found; +} + +/** + * mcache_find_gen() - Find the super block following lba and get gen# + * @znd: ZDM instance + * @lba: LBA to start scanning for the super block. + * @use_wq: If a work queue is needed to scanning. + * @sb_lba: LBA where the super block was found. + */ +static u64 mcache_find_gen(struct zoned *znd, u64 lba, int use_wq, u64 *sb_lba) +{ + struct dm_target *ti = znd->ti; + u64 generation = 0; + int nblks = 1; + int rc = 1; + int done = 0; + u32 count = 0; + u64 *data = ZDM_ALLOC(znd, Z_C4K, PG_11, NORMAL); + + if (!data) { + Z_ERR(znd, "No memory for finding generation .."); + return 0; + } + do { + rc = read_block(ti, DM_IO_KMEM, data, lba, nblks, use_wq); + if (rc) { + Z_ERR(znd, "%s: mcache-> %" PRIu64 + " [%d blks] %p -> %d", + __func__, lba, nblks, data, rc); + goto out; + } + if (is_key_page(data)) { + struct mz_superkey *kblk = (struct mz_superkey *) data; + + generation = le64_to_cpu(kblk->generation); + done = 1; + if (sb_lba) + *sb_lba = lba; + goto out; + } + lba++; + count++; + if (count > MAX_CACHE_SYNC) { + Z_ERR(znd, "FGen: Too deep to be useful."); + goto out; + } + } while (!done); + +out: + ZDM_FREE(znd, data, Z_C4K, PG_11); + return generation; +} + +/** + * cmp_gen() - compare two u64 numbers considering rollover + * @left: a u64 + * @right: a u64 + * Return: -1, 0, 1 if left < right, equal, or > respectively. + */ +static inline int cmp_gen(u64 left, u64 right) +{ + int result = 0; + + if (left != right) { + u64 delta = (left > right) ? left - right : right - left; + + result = (left > right) ? -1 : 1; + if (delta > 0xFFFFFFFF) { + if (left == BAD_ADDR) + result = 1; + } else { + if (right > left) + result = 1; + } + } + + return result; +} + +/** + * mcache_greatest_gen() - Pick the lba where the super block should start. + * @znd: ZDM instance + * @use_wq: If a workqueue is needed for IO. + * @sb: LBA of super block itself. + * @at_lba: LBA where sync data starts (in front of the super block). + */ +static +u64 mcache_greatest_gen(struct zoned *znd, int use_wq, u64 *sb, u64 *at_lba) +{ + u64 lba = LBA_SB_START; + u64 gen_no[CACHE_COPIES] = { 0ul, 0ul, 0ul }; + u64 gen_lba[CACHE_COPIES] = { 0ul, 0ul, 0ul }; + u64 gen_sb[CACHE_COPIES] = { 0ul, 0ul, 0ul }; + u64 incr = MAX_CACHE_INCR; + int locations = ARRAY_SIZE(gen_lba); + int pick = 0; + int idx; + + for (idx = 0; idx < locations; idx++) { + u64 *pAt = &gen_sb[idx]; + + gen_lba[idx] = lba; + gen_no[idx] = mcache_find_gen(znd, lba, use_wq, pAt); + if (gen_no[idx]) + pick = idx; + lba = idx * incr; + } + + for (idx = 0; idx < locations; idx++) { + if (cmp_gen(gen_no[pick], gen_no[idx]) > 0) + pick = idx; + } + + if (gen_no[pick]) { + if (at_lba) + *at_lba = gen_lba[pick]; + if (sb) + *sb = gen_sb[pick]; + } + + return gen_no[pick]; +} + +/** + * count_stale_blocks() - Number of stale blocks covered by meta_pg. + * @znd: ZDM instance + * @gzno: Meta page # to scan. + * @wpg: Meta page to scan. + */ +static u64 count_stale_blocks(struct zoned *znd, u32 gzno, struct meta_pg *wpg) +{ + u32 gzcount = 1 << GZ_BITS; + u32 iter; + u64 stale = 0; + + if ((gzno << GZ_BITS) > znd->data_zones) + gzcount = znd->data_zones & GZ_MMSK; + + /* mark as empty */ + for (iter = 0; iter < gzcount; iter++) { + u32 wp = le32_to_cpu(wpg->wp_alloc[iter]) & Z_WP_VALUE_MASK; + u32 nf = le32_to_cpu(wpg->zf_est[iter]) & Z_WP_VALUE_MASK; + + if (wp > (Z_BLKSZ - nf)) + stale += (wp - (Z_BLKSZ - nf)); + } + + return stale; +} + +static int do_load_cache(struct zoned *znd, int type, u64 lba, int idx, int wq) +{ + u16 count; + __le16 crc; + int rc = 0; + int blks = 1; + struct map_cache *mcache = mcache_alloc(znd, NORMAL); + + if (!mcache) { + rc = -ENOMEM; + goto out; + } + + rc = read_block(znd->ti, DM_IO_KMEM, mcache->jdata, lba, blks, wq); + if (rc) { + Z_ERR(znd, "%s: mcache-> %" PRIu64 + " [%d blks] %p -> %d", + __func__, lba, blks, mcache->jdata, rc); + + ZDM_FREE(znd, mcache->jdata, Z_C4K, PG_08); + ZDM_FREE(znd, mcache, sizeof(*mcache), KM_07); + goto out; + } + crc = crc_md_le16(mcache->jdata, Z_CRC_4K); + if (crc != znd->bmkeys->crcs[idx]) { + rc = -EIO; + Z_ERR(znd, "%s: bad crc %" PRIu64, __func__, lba); + goto out; + } + (void)le64_to_lba48(mcache->jdata[0].bval, &count); + mcache->jcount = count; + mclist_add(znd, mcache, type); + +out: + return rc; +} + +static int do_load_map_cache(struct zoned *znd, u64 lba, int idx, int wq) +{ + return do_load_cache(znd, IS_MAP, lba, idx, wq); +} + +static int do_load_discard_cache(struct zoned *znd, u64 lba, int idx, int wq) +{ + return do_load_cache(znd, IS_DISCARD, lba, idx, wq); +} + + + +/** + * z_mapped_init() - Re-Load an existing ZDM instance from the block device. + * @znd: ZDM instance + * + * FIXME: Discard extent read-back does not match z_mapped_sync writing + */ +static int z_mapped_init(struct zoned *znd) +{ + struct dm_target *ti = znd->ti; + int nblks = 1; + int wq = 0; + int rc = 1; + int idx = 0; + int jcount = 0; + u64 sblba = 0; + u64 lba = 0; + u64 generation; + __le32 crc_chk; + struct io_4k_block *io_vcache; + + MutexLock(&znd->vcio_lock); + io_vcache = get_io_vcache(znd, NORMAL); + + if (!io_vcache) { + Z_ERR(znd, "%s: FAILED to get SYNC CACHE.", __func__); + rc = -ENOMEM; + goto out; + } + + generation = mcache_greatest_gen(znd, wq, &sblba, &lba); + if (generation == 0) { + rc = -ENODATA; + goto out; + } + + if (lba == 0) + lba++; + + /* read superblock */ + rc = read_block(ti, DM_IO_VMA, io_vcache, sblba, nblks, wq); + if (rc) + goto out; + + memcpy(znd->bmkeys, io_vcache, sizeof(*znd->bmkeys)); + + /* read in map cache */ + for (idx = 0; idx < le16_to_cpu(znd->bmkeys->maps); idx++) { + rc = do_load_map_cache(znd, lba++, jcount++, wq); + if (rc) + goto out; + } + + /* read in discard cache */ + for (idx = 0; idx < le16_to_cpu(znd->bmkeys->discards); idx++) { + rc = do_load_discard_cache(znd, lba++, jcount++, wq); + if (rc) + goto out; + } + + /* skip re-read of superblock */ + if (lba == sblba) + lba++; + + /* read in CRC pgs */ + rc = read_block(ti, DM_IO_VMA, znd->md_crcs, lba, 2, wq); + if (rc) + goto out; + + crc_chk = znd->bmkeys->crc32; + znd->bmkeys->crc32 = 0; + znd->bmkeys->crc32 = cpu_to_le32(crc32c(~0u, znd->bmkeys, Z_CRC_4K)); + + if (crc_chk != znd->bmkeys->crc32) { + Z_ERR(znd, "Bad Block Map KEYS!"); + Z_ERR(znd, "Key CRC: Ex: %04x vs %04x <- calculated", + le32_to_cpu(crc_chk), + le32_to_cpu(znd->bmkeys->crc32)); + rc = -EIO; + goto out; + } + + if (jcount != le16_to_cpu(znd->bmkeys->n_crcs)) { + Z_ERR(znd, " ... mcache entries: found = %u, expected = %u", + jcount, le16_to_cpu(znd->bmkeys->n_crcs)); + rc = -EIO; + goto out; + } + + crc_chk = crc_md_le16(znd->md_crcs, Z_CRC_4K * 2); + if (crc_chk != znd->bmkeys->md_crc) { + Z_ERR(znd, "CRC of CRC PGs: Ex %04x vs %04x <- calculated", + le16_to_cpu(znd->bmkeys->md_crc), + le16_to_cpu(crc_chk)); + rc = -EIO; + goto out; + } + + /* + * Read write pointers / free counters. + */ + lba = 0x2048ul; + znd->discard_count = 0; + for (idx = 0; idx < znd->gz_count; idx++) { + struct meta_pg *wpg = &znd->wp[idx]; + __le16 crc_wp; + __le16 crc_zf; + + rc = read_block(ti, DM_IO_KMEM, wpg->wp_alloc, lba, 1, wq); + if (rc) + goto out; + crc_wp = crc_md_le16(wpg->wp_alloc, Z_CRC_4K); + if (znd->bmkeys->wp_crc[idx] != crc_wp) + Z_ERR(znd, "WP @ %d does not match written.", idx); + + rc = read_block(ti, DM_IO_KMEM, wpg->zf_est, lba + 1, 1, wq); + if (rc) + goto out; + crc_zf = crc_md_le16(wpg->zf_est, Z_CRC_4K); + if (znd->bmkeys->zf_crc[idx] != crc_zf) + Z_ERR(znd, "ZF @ %d does not match written.", idx); + + Z_DBG(znd, "%d# -- WP: %04x [%04x] | ZF: %04x [%04x]", + idx, znd->bmkeys->wp_crc[idx], crc_wp, + znd->bmkeys->zf_crc[idx], crc_zf); + + + if (znd->bmkeys->wp_crc[idx] == crc_wp && + znd->bmkeys->zf_crc[idx] == crc_zf) + znd->discard_count += count_stale_blocks(znd, idx, wpg); + + lba += 2; + } + znd->z_gc_resv = le32_to_cpu(znd->bmkeys->gc_resv); + znd->z_meta_resv = le32_to_cpu(znd->bmkeys->meta_resv); + +out: + put_io_vcache(znd, io_vcache); + mutex_unlock(&znd->vcio_lock); + return rc; +} + +/** + * z_mapped_addmany() - Add multiple entries into the map_cache + * @znd: ZDM instance + * @dm_s: tLBA + * @lba: lba on backing device. + * @c: Number of (contiguous) map entries to add. + * @fp: Allocation flags (for _ALLOC) + */ +static int z_mapped_addmany(struct zoned *znd, u64 dm_s, u64 lba, + u64 count, gfp_t gfp) +{ + int rc = 0; + sector_t blk; + + for (blk = 0; blk < count; blk++) { + rc = z_mapped_add_one(znd, dm_s + blk, lba + blk, gfp); + if (rc) + goto out; + } + +out: + return rc; +} + +/** + * alloc_pg() - Allocate a map page + * @znd: ZDM instance + * @entry: entry (in mpi table) to update on allocation. + * @lba: LBA associated with the page of ZLT + * @mpi: Map page information (lookup table entry, bit flags, etc) + * @ahead: Flag to set READA flag on page + * @gfp: Allocation flags (for _ALLOC) + */ +static struct map_pg *alloc_pg(struct zoned *znd, int entry, u64 lba, + struct mpinfo *mpi, int ahead, gfp_t gfp) +{ + spinlock_t *lock = &znd->ct_lock; + struct map_pg *found = ZDM_ALLOC(znd, sizeof(*found), KM_20, gfp); + + if (found) { + if (mpi->bit_type == IS_LUT) + lock = &znd->mapkey_lock; + + found->lba = lba; + mutex_init(&found->md_lock); + set_bit(mpi->bit_dir, &found->flags); + set_bit(mpi->bit_type, &found->flags); + found->age = jiffies_64; + found->index = entry; + INIT_LIST_HEAD(&found->zltlst); + INIT_LIST_HEAD(&found->lazy); + found->znd = znd; + found->crc_pg = NULL; /* */ + ref_pg(found); + if (ahead) + set_bit(IS_READA, &found->flags); + + /* allocation done. check and see if there as a + * concurrent race + */ + spin_lock(lock); + if (mpi->table[entry] == NULL) + mpi->table[entry] = found; + else + ZDM_FREE(znd, found, sizeof(*found), KM_20); + spin_unlock(lock); + } else { + Z_ERR(znd, "NO MEM for mapped_t !!!"); + } + return found; +} + +/** + * _maybe_undrop() - If a page is on its way out of cache pull it back. + * @znd: ZDM instance + * @pg: Page to claim + * + * When a table page is being dropped from the cache it may transition + * through the lazy pool. It a page is caught in the lazy pool it is + * deemed to be 'warm'. Not hot enough to be frequently hit but clearly + * too warm to be dropped quickly. Give it a boost to keep in in cache + * longer. + */ +static __always_inline int _maybe_undrop(struct zoned *znd, struct map_pg *pg) +{ + int undrop = 0; + + if (test_bit(IS_DROPPED, &pg->flags)) { + spin_lock(&znd->lzy_lck); + if (test_bit(IS_DROPPED, &pg->flags) && + test_bit(IS_LAZY, &pg->flags)) { + clear_bit(IS_DROPPED, &pg->flags); + set_bit(DELAY_ADD, &pg->flags); + } + spin_unlock(&znd->lzy_lck); + if (pg->hotness < MEM_PURGE_MSECS) + pg->hotness += MEM_HOT_BOOST_INC; + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + undrop = 1; + } + ref_pg(pg); + return undrop; +} + +/** + * _load_backing_pages() - Cache a backing page + * @znd: ZDM instance + * @lba: Logical LBA of page. + * @gfp: Memory allocation rule + * + * When metadata is pooled with data the FWD table lookup can + * be recursive (the page needed to resolve the FWD entry is + * itself on disk). The recursion is never deep but it can + * be avoided or mitigated by keep such 'key' pages in cache. + */ +static int _load_backing_pages(struct zoned *znd, u64 lba, gfp_t gfp) +{ + const int raflg = 1; + int rc = 0; + int entry; + struct mpinfo mpi; + struct map_addr maddr; + struct map_pg *found; + + if (lba < znd->data_lba) + goto out; + + map_addr_calc(znd, lba, &maddr); + entry = to_table_entry(znd, maddr.lut_s, &mpi); + if (entry > -1) { + spinlock_t *lock = &znd->ct_lock; + + if (mpi.bit_type == IS_LUT) + lock = &znd->mapkey_lock; + + spin_lock(lock); + found = mpi.table[entry]; + if (found) + _maybe_undrop(znd, found); + spin_unlock(lock); + if (!found) + found = alloc_pg(znd, entry, lba, &mpi, raflg, gfp); + if (found) { + if (!found->data.addr) { + rc = cache_pg(znd, found, gfp, &mpi); + if (rc < 0 && rc != -EBUSY) + znd->meta_result = rc; + } + deref_pg(found); + + if (getref_pg(found) != 0) + Z_ERR(znd, "Backing page with elevated ref: %u", + getref_pg(found)); + } + } + +out: + return rc; +} + +/** + * _load_crc_page() - Cache a page of CRC + * @znd: ZDM instance + * @lba: Logical LBA of page. + * @gfp: Memory allocation rule + * + * When a table page is cached the page containing its CRC is also pulled + * into cache. Rather than defer it to cache_pg() it's brought into the + * cache here. + */ +static int _load_crc_page(struct zoned *znd, struct mpinfo *mpi, gfp_t gfp) +{ + const int raflg = 0; + int rc = 0; + int entry; + struct map_pg *found; + spinlock_t *lock = &znd->ct_lock; + u64 base = (mpi->bit_dir == IS_REV) ? znd->c_mid : znd->c_base; + + base += mpi->crc.pg_no; + entry = to_table_entry(znd, base, mpi); + + if (mpi->bit_type != IS_CRC) + return rc; + if (entry >= znd->crc_count) + return rc; + + spin_lock(lock); + found = mpi->table[entry]; + if (found) + _maybe_undrop(znd, found); + spin_unlock(lock); + if (!found) + found = alloc_pg(znd, entry, base, mpi, raflg, gfp); + + if (found) { + if (!found->data.crc) { + rc = cache_pg(znd, found, gfp, mpi); + if (rc < 0 && rc != -EBUSY) + znd->meta_result = rc; + } + deref_pg(found); + } + return rc; +} + +/** + * put_map_entry() - Decrement refcount of mapped page. + * @pg: mapped page + */ +static inline void put_map_entry(struct map_pg *pg) +{ + if (pg) + deref_pg(pg); +} + +/** + * get_map_entry() - Find a page of LUT or CRC table map. + * @znd: ZDM instance + * @lba: Logical LBA of page. + * @gfp: Memory allocation rule + * + * Return: struct map_pg * or NULL on error. + * + * Page will be loaded from disk it if is not already in core memory. + */ +static struct map_pg *get_map_entry(struct zoned *znd, u64 lba, gfp_t gfp) +{ + struct mpinfo mpi; + struct map_pg *ahead[READ_AHEAD]; + int entry; + int iter; + int range; + u32 count; + spinlock_t *lock; + int empty_val; + + memset(&ahead, 0, sizeof(ahead)); + + entry = to_table_entry(znd, lba, &mpi); + if (entry < 0) + return NULL; + + if (mpi.bit_type == IS_LUT) { + lock = &znd->mapkey_lock; + count = znd->map_count; + empty_val = 0xff; + + if (mpi.bit_dir == IS_FWD) { + /* + * pre-load any backing pages + * to unwind recursive page lookups. + */ + _load_backing_pages(znd, lba, gfp); + } + _load_crc_page(znd, &mpi, gfp); + + range = entry + ARRAY_SIZE(ahead); + if (range > count) + range = count; + } else { + lock = &znd->ct_lock; + count = znd->crc_count; + empty_val = 0; + + /* CRC's cover 2k pages .. so only pull one extra */ + range = entry + 1; + if (range > count) + range = count; + } + + iter = 0; + while (entry < range) { + int want_cached = 1; + struct map_pg *found; + + entry = to_table_entry(znd, lba, &mpi); + if (entry < 0) + break; + + spin_lock(lock); + found = mpi.table[entry]; + if (found && + _maybe_undrop(znd, found) && + test_bit(IS_READA, &found->flags) + && iter > 0) + want_cached = 0; + + if (found) { + if (want_cached) + found->age = jiffies_64 + + msecs_to_jiffies(found->hotness); + else + found->age += + msecs_to_jiffies(MEM_HOT_BOOST_INC); + } + spin_unlock(lock); + + if (want_cached && !found) + found = alloc_pg(znd, entry, lba, &mpi, iter, gfp); + + if (found) { + if (want_cached) + ahead[iter] = found; + else + deref_pg(found); + } + iter++; + entry++; + lba++; + } + + /* + * Each entry in ahead has an elevated refcount. + * Only allow the target of get_map_entry() to remain elevated. + */ + for (iter = 0; iter < ARRAY_SIZE(ahead); iter++) { + struct map_pg *pg = ahead[iter]; + + if (pg) { + if (!pg->data.addr) { + int rc; + + to_table_entry(znd, pg->lba, &mpi); + rc = cache_pg(znd, pg, gfp, &mpi); + + if (rc < 0 && rc != -EBUSY) { + znd->meta_result = rc; + ahead[iter] = NULL; + } + } + if (iter > 0) + deref_pg(pg); + } + } + + return ahead[0]; +} + +/** + * metadata_dirty_fling() - Force a ZLT block into cache and flag it dirty. + * @znd: ZDM Instance + * @dm_s: Current lba to consider. + * + * Used when data and ZDM's metadata are co-mingled. If dm_s is a block + * of ZDM's metadata it needs to be relocated. Since we re-locate + * blocks that are in dirty and in the cache ... if this block is + * metadata, force it into the cache and flag it as dirty. + */ +static int metadata_dirty_fling(struct zoned *znd, u64 dm_s) +{ + struct map_pg *pg = NULL; + int is_flung = 0; + + /* + * When co + * nothing in the GC reverse map should point to + * a block *before* a data pool block + */ + if (dm_s < znd->data_lba) + return is_flung; + + if (dm_s >= znd->r_base && dm_s < znd->c_end) { + pg = get_map_entry(znd, dm_s, NORMAL); + if (!pg) + Z_ERR(znd, "Failed to fling: %" PRIx64, dm_s); + } + if (pg) { + MutexLock(&pg->md_lock); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + clear_bit(IS_READA, &pg->flags); + set_bit(IS_DIRTY, &pg->flags); + clear_bit(IS_FLUSH, &pg->flags); + is_flung = 1; + mutex_unlock(&pg->md_lock); + + if (pg->lba != dm_s) + Z_ERR(znd, "Excess churn? lba %"PRIx64 + " [last: %"PRIx64"]", dm_s, pg->lba); + put_map_entry(pg); + } + return is_flung; +} + +/** + * z_do_copy_more() - GC transition to read more blocks. + * @gc_state: GC State to be updated. + */ +static inline void z_do_copy_more(struct gc_state *gc_entry) +{ + unsigned long flags; + struct zoned *znd = gc_entry->znd; + + spin_lock_irqsave(&znd->gc_lock, flags); + set_bit(DO_GC_PREPARE, &gc_entry->gc_flags); + spin_unlock_irqrestore(&znd->gc_lock, flags); +} + +/** + * gc_post_add() - Add a tLBA and current bLBA origin. + * @znd: ZDM Instance + * @dm_s: tLBA + * @lba: bLBA + * + * Return: 1 if tLBA is added, 0 if block was stale. + * + * Stale block checks are performed before tLBA is added. + * Add a non-stale block to the list of blocks for moving and + * metadata updating. + */ +static int gc_post_add(struct zoned *znd, u64 addr, u64 lba) +{ + struct map_cache *post = &znd->gc_postmap; + int handled = 0; + + if (z_lookup_cache(znd, addr, IS_DISCARD)) + return handled; + + if (metadata_dirty_fling(znd, addr)) + return handled; + + if (post->jcount < post->jsize) { + u16 idx = ++post->jcount; + + WARN_ON(post->jcount > post->jsize); + + post->jdata[idx].tlba = lba48_to_le64(0, addr); + post->jdata[idx].bval = lba48_to_le64(0, lba); + handled = 1; + } else { + Z_ERR(znd, "*CRIT* post overflow L:%" PRIx64 "-> S:%" PRIx64, + lba, addr); + } + return handled; +} + + +/** + * _fwd_to_cache() - Helper to pull the forward map block into cache. + * @znd: ZDM Instance + * @addr: Address (from reverse map table entry). + */ +static __always_inline int _fwd_to_cache(struct zoned *znd, u64 addr) +{ + int err = 0; + struct map_pg *pg; + struct map_addr maddr; + + if (addr < znd->nr_blocks) { + map_addr_calc(znd, addr, &maddr); + pg = get_map_entry(znd, maddr.lut_s, NORMAL); + if (!pg) + err = -ENOMEM; + put_map_entry(pg); + } else { + Z_ERR(znd, "Invalid rmap entry: %" PRIx64, addr); + } + + return err; +} + +/** + * z_zone_gc_metadata_to_ram() - Load affected metadata blocks to ram. + * @gc_entry: Compaction event in progress + * + * Return: 0, otherwise errno. + * + * Use the reverse ZLT to find the forward ZLT entries that need to be + * remapped in this zone. + * When complete the znd->gc_postmap have a map of all the non-stale + * blocks remaining in the zone. + */ +static int z_zone_gc_metadata_to_ram(struct gc_state *gc_entry) +{ + struct zoned *znd = gc_entry->znd; + u64 from_lba = (gc_entry->z_gc << Z_BLKBITS) + znd->md_end; + struct map_pg *rev_pg; + struct map_addr origin; + int count; + int rcode = 0; + + /* pull all of the affected struct map_pg and crc pages into memory: */ + for (count = 0; count < Z_BLKSZ; count++) { + __le32 ORencoded; + u64 blba = from_lba + count; + + map_addr_calc(znd, blba, &origin); + rev_pg = get_map_entry(znd, origin.lut_r, NORMAL); + if (rev_pg && rev_pg->data.addr) { + ref_pg(rev_pg); + MutexLock(&rev_pg->md_lock); + ORencoded = rev_pg->data.addr[origin.pg_idx]; + mutex_unlock(&rev_pg->md_lock); + if (ORencoded != MZTEV_UNUSED) { + u64 tlba = map_value(znd, ORencoded); + int err = _fwd_to_cache(znd, tlba); + + if (err) + rcode = err; + gc_post_add(znd, tlba, blba); + } + deref_pg(rev_pg); + } + put_map_entry(rev_pg); + } + + return rcode; +} + +/** + * append_blks() - Read (more) blocks into buffer. + * @znd: ZDM Instance + * @lba: Starting blba + * @io_buf: Buffer to read into + * @count: Number of blocks to read. + */ +static int append_blks(struct zoned *znd, u64 lba, + struct io_4k_block *io_buf, int count) +{ + int rcode = 0; + int rc; + u32 chunk; + struct io_4k_block *io_vcache; + + MutexLock(&znd->gc_vcio_lock); + io_vcache = get_io_vcache(znd, NORMAL); + if (!io_vcache) { + Z_ERR(znd, "%s: FAILED to get SYNC CACHE.", __func__); + rc = -ENOMEM; + goto out; + } + + for (chunk = 0; chunk < count; chunk += IO_VCACHE_PAGES) { + u32 nblks = count - chunk; + + if (nblks > IO_VCACHE_PAGES) + nblks = IO_VCACHE_PAGES; + + rc = read_block(znd->ti, DM_IO_VMA, io_vcache, lba, nblks, 0); + if (rc) { + Z_ERR(znd, "Reading error ... disable zone: %u", + (u32)(lba >> 16)); + rcode = -EIO; + goto out; + } + memcpy(&io_buf[chunk], io_vcache, nblks * Z_C4K); + lba += nblks; + } +out: + put_io_vcache(znd, io_vcache); + mutex_unlock(&znd->gc_vcio_lock); + return rcode; +} + +/** + * z_zone_gc_read() - Read (up to) a buffer worth of data from zone. + * @gc_entry: Active GC state + */ +static int z_zone_gc_read(struct gc_state *gc_entry) +{ + struct zoned *znd = gc_entry->znd; + struct io_4k_block *io_buf = znd->gc_io_buf; + struct map_cache *post = &znd->gc_postmap; + unsigned long flags; + u64 start_lba; + int nblks; + int rcode = 0; + int fill = 0; + int jstart; + int jentry; + + spin_lock_irqsave(&znd->gc_lock, flags); + jstart = gc_entry->r_ptr; + spin_unlock_irqrestore(&znd->gc_lock, flags); + + if (!jstart) + jstart++; + + MutexLock(&post->cached_lock); + + /* A discard may have puched holes in the postmap. re-sync lba */ + jentry = jstart; + while (jentry <= post->jcount && (Z_LOWER48 == + le64_to_lba48(post->jdata[jentry].bval, NULL))) { + jentry++; + } + /* nothing left to move */ + if (jentry > post->jcount) + goto out_finished; + + /* skip over any discarded blocks */ + if (jstart != jentry) + jstart = jentry; + + start_lba = le64_to_lba48(post->jdata[jentry].bval, NULL); + post->jdata[jentry].bval = lba48_to_le64(GC_READ, start_lba); + nblks = 1; + jentry++; + + while (jentry <= post->jcount && (nblks+fill) < GC_MAX_STRIPE) { + u64 dm_s = le64_to_lba48(post->jdata[jentry].tlba, NULL); + u64 lba = le64_to_lba48(post->jdata[jentry].bval, NULL); + + if (Z_LOWER48 == dm_s || Z_LOWER48 == lba) { + jentry++; + continue; + } + + post->jdata[jentry].bval = lba48_to_le64(GC_READ, lba); + + /* if the block is contiguous add it to the read */ + if (lba == (start_lba + nblks)) { + nblks++; + } else { + if (nblks) { + int err; + + err = append_blks(znd, start_lba, + &io_buf[fill], nblks); + if (err) { + rcode = err; + goto out; + } + fill += nblks; + } + start_lba = lba; + nblks = 1; + } + jentry++; + } + + /* Issue a copy of 'nblks' blocks */ + if (nblks > 0) { + int err; + + err = append_blks(znd, start_lba, &io_buf[fill], nblks); + if (err) { + rcode = err; + goto out; + } + fill += nblks; + } + +out_finished: + Z_DBG(znd, "Read %d blocks from %d", fill, gc_entry->r_ptr); + + spin_lock_irqsave(&znd->gc_lock, flags); + gc_entry->nblks = fill; + gc_entry->r_ptr = jentry; + if (fill > 0) + set_bit(DO_GC_WRITE, &gc_entry->gc_flags); + else + set_bit(DO_GC_COMPLETE, &gc_entry->gc_flags); + + spin_unlock_irqrestore(&znd->gc_lock, flags); + +out: + mutex_unlock(&post->cached_lock); + + return rcode; +} + +/** + * z_zone_gc_write() - Write (up to) a buffer worth of data to WP. + * @gc_entry: Active GC state + * @stream_id: Stream Id to prefer for allocation. + */ +static int z_zone_gc_write(struct gc_state *gc_entry, u32 stream_id) +{ + struct zoned *znd = gc_entry->znd; + struct dm_target *ti = znd->ti; + struct io_4k_block *io_buf = znd->gc_io_buf; + struct map_cache *post = &znd->gc_postmap; + unsigned long flags; + u32 aq_flags = Z_AQ_GC | Z_AQ_STREAM_ID | stream_id; + u64 lba; + u32 nblks; + u32 out = 0; + int err = 0; + int jentry; + + spin_lock_irqsave(&znd->gc_lock, flags); + jentry = gc_entry->w_ptr; + nblks = gc_entry->nblks; + spin_unlock_irqrestore(&znd->gc_lock, flags); + + if (!jentry) + jentry++; + + MutexLock(&post->cached_lock); + while (nblks > 0) { + u32 nfound = 0; + u32 added = 0; + + /* + * When lba is zero blocks were not allocated. + * Retry with the smaller request + */ + lba = z_acquire(znd, aq_flags, nblks, &nfound); + if (!lba) { + if (nfound) { + u32 avail = nfound; + + nfound = 0; + lba = z_acquire(znd, aq_flags, avail, &nfound); + } + } + + if (!lba) { + err = -ENOSPC; + goto out; + } + + err = write_block(ti, DM_IO_VMA, &io_buf[out], lba, nfound, 0); + if (err) { + Z_ERR(znd, "Write %d blocks to %"PRIx64". ERROR: %d", + nfound, lba, err); + goto out; + } + out += nfound; + + while ((jentry <= post->jcount) && (added < nfound)) { + u16 rflg; + u64 orig = le64_to_lba48( + post->jdata[jentry].bval, &rflg); + u64 dm_s = le64_to_lba48( + post->jdata[jentry].tlba, NULL); + + if ((Z_LOWER48 == dm_s || Z_LOWER48 == orig)) { + jentry++; + + if (rflg & GC_READ) { + Z_ERR(znd, "ERROR: %" PRIx64 + " read and not written %" PRIx64, + orig, dm_s); + lba++; + added++; + } + continue; + } + rflg &= ~GC_READ; + post->jdata[jentry].bval = lba48_to_le64(rflg, lba); + lba++; + added++; + jentry++; + } + nblks -= nfound; + } + Z_DBG(znd, "Write %d blocks from %d", gc_entry->nblks, gc_entry->w_ptr); + set_bit(DO_GC_META, &gc_entry->gc_flags); + +out: + spin_lock_irqsave(&znd->gc_lock, flags); + gc_entry->nblks = 0; + gc_entry->w_ptr = jentry; + spin_unlock_irqrestore(&znd->gc_lock, flags); + mutex_unlock(&post->cached_lock); + + return err; +} + +/** + * gc_finalize() - Final sanity check on GC'd block map. + * @gc_entry: Active GC state + * + * gc_postmap is expected to be empty (all blocks original + * scheduled to be moved to a new zone have been accounted for... + */ +static int gc_finalize(struct gc_state *gc_entry) +{ + int err = 0; + struct zoned *znd = gc_entry->znd; + struct map_cache *post = &znd->gc_postmap; + int jentry; + + MutexLock(&post->cached_lock); + for (jentry = post->jcount; jentry > 0; jentry--) { + u64 dm_s = le64_to_lba48(post->jdata[jentry].tlba, NULL); + u64 lba = le64_to_lba48(post->jdata[jentry].bval, NULL); + + if (dm_s != Z_LOWER48 || lba != Z_LOWER48) { + Z_ERR(znd, "GC: Failed to move %" PRIx64 + " from %"PRIx64" [%d]", + dm_s, lba, jentry); + err = -EIO; + } + } + mutex_unlock(&post->cached_lock); + post->jcount = jentry; + post->jsorted = 0; + + return err; +} + +/** + * clear_gc_target_flag() - Clear any zone tagged as a GC target. + * @znd: ZDM Instance + * + * FIXME: Can we reduce the weight of this ? + * Ex. execute as zones are closed and specify the zone to clear + * at GC completion/cleanup. + */ +static void clear_gc_target_flag(struct zoned *znd) +{ + int z_id; + + for (z_id = 0; z_id < znd->data_zones; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp; + + SpinLock(&wpg->wplck); + wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + if (wp & Z_WP_GC_TARGET) { + wp &= ~Z_WP_GC_TARGET; + wpg->wp_alloc[gzoff] = cpu_to_le32(wp); + } + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + spin_unlock(&wpg->wplck); + } +} + +/** + * z_zone_gc_metadata_update() - Update ZLT as needed. + * @gc_entry: Active GC state + * + * Dispose or account for all blocks originally scheduled to be + * moved. Update ZLT (via map cache) for all moved blocks. + */ +static int z_zone_gc_metadata_update(struct gc_state *gc_entry) +{ + struct zoned *znd = gc_entry->znd; + struct map_cache *post = &znd->gc_postmap; + u32 used = post->jcount; + int err = 0; + int jentry; + + for (jentry = post->jcount; jentry > 0; jentry--) { + int discard = 0; + int mapping = 0; + struct map_pg *mapped = NULL; + u64 dm_s = le64_to_lba48(post->jdata[jentry].tlba, NULL); + u64 lba = le64_to_lba48(post->jdata[jentry].bval, NULL); + + if ((znd->r_base <= dm_s) && dm_s < (znd->r_base + Z_BLKSZ)) { + u64 off = dm_s - znd->r_base; + + mapped = znd->rev_tm[off]; + mapping = 1; + } else if ((znd->s_base <= dm_s) && + (dm_s < (znd->s_base + Z_BLKSZ))) { + u64 off = dm_s - znd->s_base; + + mapped = znd->fwd_tm[off]; + mapping = 1; + } + + if (mapping && !mapped) + Z_ERR(znd, "MD: dm_s: %" PRIx64 " -> lba: %" PRIx64 + " no mapping in ram.", dm_s, lba); + + if (mapped) { + u32 in_z; + + ref_pg(mapped); + MutexLock(&mapped->md_lock); + in_z = _calc_zone(znd, mapped->last_write); + if (in_z != gc_entry->z_gc) { + Z_ERR(znd, "MD: %" PRIx64 + " Discarded - %" PRIx64 + " already flown to: %x", + dm_s, mapped->last_write, in_z); + discard = 1; + } else if (mapped->data.addr && + test_bit(IS_DIRTY, &mapped->flags)) { + Z_ERR(znd, + "MD: %" PRIx64 " Discarded - %"PRIx64 + " is in-flight", + dm_s, mapped->last_write); + discard = 2; + } + if (!discard) + mapped->last_write = lba; + mutex_unlock(&mapped->md_lock); + deref_pg(mapped); + } + + MutexLock(&post->cached_lock); + if (discard == 1) { + Z_ERR(znd, "Dropped: %" PRIx64 " -> %"PRIx64, + le64_to_cpu(post->jdata[jentry].tlba), + le64_to_cpu(post->jdata[jentry].bval)); + + post->jdata[jentry].tlba = MC_INVALID; + post->jdata[jentry].bval = MC_INVALID; + } + if (post->jdata[jentry].tlba == MC_INVALID && + post->jdata[jentry].bval == MC_INVALID) { + used--; + } else if (lba) { + increment_used_blks(znd, lba, 1); + } + mutex_unlock(&post->cached_lock); + } + err = move_to_map_tables(znd, post); + if (err) + Z_ERR(znd, "Move to tables post GC failure"); + + clear_gc_target_flag(znd); + + return err; +} + +/** + * _blkalloc() - Attempt to reserve blocks at z_at in ZDM znd + * @znd: ZDM instance. + * @z_at: Zone to write data to + * @flags: Acquisition type. + * @nblks: Number of blocks desired. + * @nfound: Number of blocks allocated or available. + * + * Attempt allocation of @nblks within fron the current WP of z_at + * When nblks are not available 0 is returned and @nfound is the + * contains the number of blocks *available* but not *allocated*. + * When nblks are available the starting LBA in 4k space is returned and + * nblks are allocated *allocated* and *nfound is the number of blocks + * remaining in zone z_at from the LBA returned. + * + * Return: LBA if request is met, otherwise 0. nfound will contain the + * available blocks remaining. + */ +static sector_t _blkalloc(struct zoned *znd, u32 z_at, u32 flags, + u32 nblks, u32 *nfound) +{ +#define ALLOC_STICKY (Z_WP_GC_TARGET|Z_WP_NON_SEQ|Z_WP_RRECALC) + + sector_t found = 0; + u32 avail = 0; + int do_open_zone = 0; + u32 gzno = z_at >> GZ_BITS; + u32 gzoff = z_at & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp; + u32 wptr; + u32 gc_tflg; + + if (gzno >= znd->gz_count || z_at >= znd->data_zones) { + Z_ERR(znd, "Invalid zone for allocation: %u", z_at); + dump_stack(); + return 0ul; + } + + SpinLock(&wpg->wplck); + wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + gc_tflg = wp & ALLOC_STICKY; + wptr = wp & ~ALLOC_STICKY; + if (wptr < Z_BLKSZ) + avail = Z_BLKSZ - wptr; + +#if 0 /* DEBUG START: Testing zm_write_pages() */ + if (avail > 7) + avail = 7; +#endif /* DEBUG END: Testing zm_write_pages() */ + + *nfound = avail; + if (nblks <= avail) { + u64 lba = ((u64)z_at << Z_BLKBITS) + znd->data_lba; + u32 zf_est = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + + found = lba + wptr; + *nfound = nblks; + if (wptr == 0) + do_open_zone = 1; + + wptr += nblks; + zf_est -= nblks; + if (wptr == Z_BLKSZ) + znd->discard_count += zf_est; + + wptr |= gc_tflg; + if (flags & Z_AQ_GC) + wptr |= Z_WP_GC_TARGET; + + if (flags & Z_AQ_STREAM_ID) + zf_est |= (flags & Z_AQ_STREAM_MASK) << 24; + else + zf_est |= le32_to_cpu(wpg->zf_est[gzoff]) + & Z_WP_STREAM_MASK; + + wpg->wp_alloc[gzoff] = cpu_to_le32(wptr); + wpg->zf_est[gzoff] = cpu_to_le32(zf_est); + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + } + spin_unlock(&wpg->wplck); + + if (do_open_zone) + dmz_open_zone(znd, z_at); + + return found; +} + +/** + * update_stale_ratio() - Update the stale ratio for the finished bin. + * @znd: ZDM instance + * @zone: Zone that needs update. + */ +static void update_stale_ratio(struct zoned *znd, u32 zone) +{ + u64 total_stale = 0; + u64 free_zones = 1; + u32 bin = zone / znd->stale.binsz; + u32 z_id = bin * znd->stale.binsz; + u32 s_end = z_id + znd->stale.binsz; + + if (s_end > znd->data_zones) + s_end = znd->data_zones; + + for (; z_id < s_end; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 stale = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_VALUE_MASK; + u32 wflg = le32_to_cpu(wpg->wp_alloc[gzoff]); + + if (wflg & Z_WP_RRECALC) { + SpinLock(&wpg->wplck); + wflg = le32_to_cpu(wpg->wp_alloc[gzoff]) + & ~Z_WP_RRECALC; + wpg->wp_alloc[gzoff] = cpu_to_le32(wflg); + spin_unlock(&wpg->wplck); + } + + if (wp == Z_BLKSZ) + total_stale += stale; + else + free_zones++; + } + + total_stale /= free_zones; + znd->stale.bins[bin] = (total_stale > ~0u) ? ~0u : total_stale; +} + +/** + * z_zone_compact_queue() - Queue zone compaction. + * @znd: ZDM instance + * @z_gc: Zone to queue. + * @delay: Delay queue metric + * @gfp: Allocation scheme. + * + * Return: 1 on success, 0 if not queued/busy, negative on error. + */ +static +int z_zone_compact_queue(struct zoned *znd, u32 z_gc, int delay, gfp_t gfp) +{ + unsigned long flags; + int do_queue = 0; + int err = 0; + struct gc_state *gc_entry; + + gc_entry = ZDM_ALLOC(znd, sizeof(*gc_entry), KM_16, gfp); + if (!gc_entry) { + Z_ERR(znd, "No Memory for compact!!"); + return -ENOMEM; + } + gc_entry->znd = znd; + gc_entry->z_gc = z_gc; + set_bit(DO_GC_NEW, &gc_entry->gc_flags); + znd->gc_backlog++; + + spin_lock_irqsave(&znd->gc_lock, flags); + if (!znd->gc_active) { + znd->gc_active = gc_entry; + do_queue = 1; + } else { + Z_ERR(znd, "GC: Tried to queue but already active"); + } + spin_unlock_irqrestore(&znd->gc_lock, flags); + + if (do_queue) { + unsigned long tval = msecs_to_jiffies(delay); + + if (queue_delayed_work(znd->gc_wq, &znd->gc_work, tval)) + err = 1; + } else { + ZDM_FREE(znd, gc_entry, sizeof(*gc_entry), KM_16); + znd->gc_backlog--; + Z_ERR(znd, "GC: FAILED to queue work"); + } + + return err; +} + +/** + * zone_zfest() - Queue zone compaction. + * @znd: ZDM instance + * @z_id: Zone to queue. + */ +static u32 zone_zfest(struct zoned *znd, u32 z_id) +{ + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + + return le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; +} + +/** + * gc_request_queued() - Called periodically to initiate GC + * + * @znd: ZDM instance + * @bin: Bin with stale zones to scan for GC + * @delay: Metric for delay queuing. + * @gfp: Default memory allocation scheme. + * + */ +static int gc_request_queued(struct zoned *znd, int bin, int delay, gfp_t gfp) +{ + unsigned long flags; + int queued = 0; + u32 top_roi = NOZONE; + u32 stale = 0; + u32 z_gc = bin * znd->stale.binsz; + u32 s_end = z_gc + znd->stale.binsz; + + if (znd->meta_result) + goto out; + + if (test_bit(ZF_FREEZE, &znd->flags)) { + Z_ERR(znd, "Is frozen -- GC paused."); + goto out; + } + + spin_lock_irqsave(&znd->gc_lock, flags); + if (znd->gc_active) + queued = 1; + spin_unlock_irqrestore(&znd->gc_lock, flags); + if (queued) + goto out; + + if (s_end > znd->data_zones) + s_end = znd->data_zones; + + /* scan for most stale zone in STREAM [top_roi] */ + for (; z_gc < s_end; z_gc++) { + u32 gzno = z_gc >> GZ_BITS; + u32 gzoff = z_gc & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp_v = le32_to_cpu(wpg->wp_alloc[gzoff]); + u32 nfree = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + u32 wp_f = wp_v & Z_WP_FLAGS_MASK; + + wp_v &= Z_WP_VALUE_MASK; + if (wp_v == 0) + continue; + if ((wp_f & Z_WP_GC_PENDING) != 0) + continue; + + if (wp_v == Z_BLKSZ) { + stale += nfree; + if ((wp_f & Z_WP_GC_BITS) == Z_WP_GC_READY) { + if (top_roi == NOZONE) + top_roi = z_gc; + else if (nfree > zone_zfest(znd, top_roi)) + top_roi = z_gc; + } + } + } + + if (!delay && top_roi == NOZONE) + Z_ERR(znd, "No GC candidate in bin: %u -> %u", z_gc, s_end); + + /* determine the cut-off for GC based on MZ overall staleness */ + if (top_roi != NOZONE) { + u32 state_metric = GC_PRIO_DEFAULT; + u32 n_empty = znd->z_gc_free; + int pctfree = n_empty * 100 / znd->data_zones; + + /* + * -> at less than 5 zones free switch to critical + * -> at less than 5% zones free switch to HIGH + * -> at less than 25% free switch to LOW + * -> high level is 'cherry picking' near empty zones + */ + if (znd->z_gc_free < 5) + state_metric = GC_PRIO_CRIT; + else if (pctfree < 5) + state_metric = GC_PRIO_HIGH; + else if (pctfree < 25) + state_metric = GC_PRIO_LOW; + + if (zone_zfest(znd, top_roi) > state_metric) { + int rc; + + rc = z_zone_compact_queue(znd, top_roi, delay * 5, gfp); + if (rc == 1) + queued = 1; + else if (rc < 0) + Z_ERR(znd, "GC: Z#%u !Q: ERR: %d", top_roi, rc); + } + + if (!delay && !queued) + Z_ERR(znd, "GC: Z#%u !Q .. M: %u E: %u PCT: %d ZF: %u", + top_roi, state_metric, n_empty, pctfree, + zone_zfest(znd, top_roi)); + } +out: + return queued; +} + +/** + * z_zone_gc_compact() - Primary compaction worker. + * @gc_entry: GC State + */ +static int z_zone_gc_compact(struct gc_state *gc_entry) +{ + unsigned long flags; + int err = 0; + struct zoned *znd = gc_entry->znd; + u32 z_gc = gc_entry->z_gc; + u32 gzno = z_gc >> GZ_BITS; + u32 gzoff = z_gc & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + + znd->age = jiffies_64; + + if (test_bit(DO_GC_NEW, &gc_entry->gc_flags)) { + err = z_flush_bdev(znd, GFP_KERNEL); + if (err) { + gc_entry->result = err; + goto out; + } + } + + /* If a SYNC is in progress and we can delay then postpone*/ + if (mutex_is_locked(&znd->mz_io_mutex) && + atomic_read(&znd->gc_throttle) == 0) + return -EAGAIN; + + if (test_and_clear_bit(DO_GC_NEW, &gc_entry->gc_flags)) { + u32 wp; + + SpinLock(&wpg->wplck); + wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + wp |= Z_WP_GC_FULL; + wpg->wp_alloc[gzoff] = cpu_to_le32(wp); + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + spin_unlock(&wpg->wplck); + + err = _cached_to_tables(znd, gc_entry->z_gc); + if (err) { + Z_ERR(znd, "Failed to purge journal: %d", err); + gc_entry->result = err; + goto out; + } + + if (znd->gc_postmap.jcount > 0) { + Z_ERR(znd, "*** Unexpected data in postmap!!"); + znd->gc_postmap.jcount = 0; + } + + err = z_zone_gc_metadata_to_ram(gc_entry); + if (err) { + Z_ERR(znd, + "Pre-load metadata to memory failed!! %d", err); + gc_entry->result = err; + goto out; + } + + do { + err = memcache_lock_and_sort(znd, &znd->gc_postmap); + } while (err == -EBUSY); + + set_bit(DO_GC_PREPARE, &gc_entry->gc_flags); + + if (znd->gc_throttle.counter == 0) + return -EAGAIN; + } + +next_in_queue: + znd->age = jiffies_64; + + if (test_and_clear_bit(DO_GC_PREPARE, &gc_entry->gc_flags)) { + + MutexLock(&znd->mz_io_mutex); + mutex_unlock(&znd->mz_io_mutex); + + err = z_zone_gc_read(gc_entry); + if (err < 0) { + Z_ERR(znd, "z_zone_gc_chunk issue failure: %d", err); + gc_entry->result = err; + goto out; + } + if (znd->gc_throttle.counter == 0) + return -EAGAIN; + } + + if (test_and_clear_bit(DO_GC_WRITE, &gc_entry->gc_flags)) { + u32 sid = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_STREAM_MASK; + + MutexLock(&znd->mz_io_mutex); + mutex_unlock(&znd->mz_io_mutex); + + err = z_zone_gc_write(gc_entry, sid >> 24); + if (err) { + Z_ERR(znd, "z_zone_gc_write issue failure: %d", err); + gc_entry->result = err; + goto out; + } + if (znd->gc_throttle.counter == 0) + return -EAGAIN; + } + + if (test_and_clear_bit(DO_GC_META, &gc_entry->gc_flags)) { + z_do_copy_more(gc_entry); + goto next_in_queue; + } + + znd->age = jiffies_64; + if (test_and_clear_bit(DO_GC_COMPLETE, &gc_entry->gc_flags)) { + u32 non_seq; + u32 reclaimed; + + err = z_zone_gc_metadata_update(gc_entry); + gc_entry->result = err; + if (err) { + Z_ERR(znd, "Metadata error ... disable zone: %u", + gc_entry->z_gc); + } + err = gc_finalize(gc_entry); + if (err) { + Z_ERR(znd, "GC: Failed to finalize: %d", err); + gc_entry->result = err; + goto out; + } + + gc_verify_cache(znd, gc_entry->z_gc); + + err = _cached_to_tables(znd, gc_entry->z_gc); + if (err) { + Z_ERR(znd, "Failed to purge journal: %d", err); + gc_entry->result = err; + goto out; + } + + err = z_flush_bdev(znd, GFP_KERNEL); + if (err) { + gc_entry->result = err; + goto out; + } + + /* Release the zones for writing */ + dmz_reset_wp(znd, gc_entry->z_gc); + + SpinLock(&wpg->wplck); + non_seq = le32_to_cpu(wpg->wp_alloc[gzoff]) & Z_WP_NON_SEQ; + reclaimed = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + wpg->wp_alloc[gzoff] = cpu_to_le32(non_seq); + wpg->wp_used[gzoff] = cpu_to_le32(0u); + wpg->zf_est[gzoff] = cpu_to_le32(Z_BLKSZ); + znd->discard_count -= reclaimed; + znd->z_gc_free++; + if (znd->z_gc_resv & Z_WP_GC_ACTIVE) + znd->z_gc_resv = gc_entry->z_gc; + else if (znd->z_meta_resv & Z_WP_GC_ACTIVE) + znd->z_meta_resv = gc_entry->z_gc; + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + spin_unlock(&wpg->wplck); + + znd->gc_events++; + update_stale_ratio(znd, gc_entry->z_gc); + spin_lock_irqsave(&znd->gc_lock, flags); + znd->gc_backlog--; + znd->gc_active = NULL; + spin_unlock_irqrestore(&znd->gc_lock, flags); + + ZDM_FREE(znd, gc_entry, sizeof(*gc_entry), KM_16); + + set_bit(DO_JOURNAL_MOVE, &znd->flags); + set_bit(DO_MEMPOOL, &znd->flags); + set_bit(DO_SYNC, &znd->flags); + } +out: + return 0; +} + +/** + * gc_work_task() - Worker thread for GC activity. + * @work: Work struct holding the ZDM instance to do work on ... + */ +static void gc_work_task(struct work_struct *work) +{ + struct gc_state *gc_entry = NULL; + unsigned long flags; + struct zoned *znd; + int err; + + if (!work) + return; + + znd = container_of(to_delayed_work(work), struct zoned, gc_work); + if (!znd) + return; + + spin_lock_irqsave(&znd->gc_lock, flags); + if (znd->gc_active) + gc_entry = znd->gc_active; + spin_unlock_irqrestore(&znd->gc_lock, flags); + + if (!gc_entry) { + Z_ERR(znd, "ERROR: gc_active not set!"); + return; + } + + err = z_zone_gc_compact(gc_entry); + if (-EAGAIN == err) { + unsigned long tval = msecs_to_jiffies(10); + + queue_delayed_work(znd->gc_wq, &znd->gc_work, tval); + } else { + const int delay = 10; + + on_timeout_activity(znd, delay); + } +} + +/** + * is_reserved() - Check to see if a zone is 'special' + * @znd: ZDM Instance + * @z_pref: Zone to be tested. + */ +static inline int is_reserved(struct zoned *znd, const u32 z_pref) +{ + const u32 gc = znd->z_gc_resv & Z_WP_VALUE_MASK; + const u32 meta = znd->z_meta_resv & Z_WP_VALUE_MASK; + + return (gc == z_pref || meta == z_pref) ? 1 : 0; +} + +/** + * gc_can_cherrypick() - Queue a GC for zone in this bin if ... it will be easy + * @znd: ZDM Instance + * @bin: The bin (0 to 255) + * @delay: Delay metric + * @gfp: Allocation flags to use. + */ +static int gc_can_cherrypick(struct zoned *znd, u32 bin, int delay, gfp_t gfp) +{ + u32 z_id = bin * znd->stale.binsz; + u32 s_end = z_id + znd->stale.binsz; + + if (s_end > znd->data_zones) + s_end = znd->data_zones; + + for (; z_id < s_end; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + u32 nfree = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + + if (wp & Z_WP_RRECALC) + update_stale_ratio(znd, z_id); + + if (((wp & Z_WP_GC_BITS) == Z_WP_GC_READY) && + ((wp & Z_WP_VALUE_MASK) == Z_BLKSZ) && + (nfree == Z_BLKSZ)) { + if (z_zone_compact_queue(znd, z_id, delay, gfp)) + return 1; + } + } + + return 0; +} + +/** + * gc_queue_with_delay() - Scan to see if a GC can/should be queued. + * @znd: ZDM Instance + * @delay: Delay metric + * @gfp: Allocation flags to use. + * + * Return 1 if gc in progress or queued. 0 otherwise. + */ +static int gc_queue_with_delay(struct zoned *znd, int delay, gfp_t gfp) +{ + int gc_idle = 0; + unsigned long flags; + + spin_lock_irqsave(&znd->gc_lock, flags); + gc_idle = !znd->gc_active; + spin_unlock_irqrestore(&znd->gc_lock, flags); + + if (gc_idle) { + int bin = 0; + int ratio = 0; + u32 iter; + + /* Find highest ratio stream */ + for (iter = 0; iter < znd->stale.count; iter++) + if (znd->stale.bins[iter] > ratio) + ratio = znd->stale.bins[iter], bin = iter; + + /* Cherrypick a zone in the stream */ + if (gc_idle && gc_can_cherrypick(znd, bin, delay, gfp)) + gc_idle = 0; + + /* Otherwise cherrypick *something* */ + for (iter = 0; gc_idle && (iter < znd->stale.count); iter++) + if (gc_idle && (bin != iter) && + gc_can_cherrypick(znd, iter, delay, gfp)) + gc_idle = 0; + + /* Otherwise compact a zone in the stream */ + if (gc_idle && gc_request_queued(znd, bin, delay, gfp)) + gc_idle = 0; + + /* Otherwise compact *something* */ + for (iter = 0; gc_idle && (iter < znd->stale.count); iter++) + if (gc_idle && gc_request_queued(znd, iter, delay, gfp)) + gc_idle = 0; + } + return !gc_idle; +} + +/** + * gc_immediate() - Free up some space as soon as possible. + * @znd: ZDM Instance + * @gfp: Allocation flags to use. + */ +static int gc_immediate(struct zoned *znd, int wait, gfp_t gfp) +{ + const int delay = 0; + int can_retry = 0; + int lock_owner = 0; + int queued; + + /* + * Lock on the first entry. On subsequent entries + * only lock of a 'wait' was requested. + * + * Note: We really should be using a wait queue here but this + * mutex hack should do as a temporary workaround. + */ + atomic_inc(&znd->gc_throttle); + if (atomic_read(&znd->gc_throttle) == 1) { + mutex_lock(&znd->gc_wait); + lock_owner = 1; + } else if (atomic_read(&znd->gc_throttle) > 1) { + if (wait) { + mutex_lock(&znd->gc_wait); + can_retry = 1; + mutex_unlock(&znd->gc_wait); + } + goto out; + } + flush_delayed_work(&znd->gc_work); + queued = gc_queue_with_delay(znd, delay, gfp); + if (!queued) { + Z_ERR(znd, " ... GC immediate .. failed to queue GC!!."); + z_discard_partial(znd, DISCARD_MAX_INGRESS, GFP_KERNEL); + can_retry = 1; + goto out; + } + can_retry = flush_delayed_work(&znd->gc_work); + +out: + if (lock_owner) + mutex_unlock(&znd->gc_wait); + atomic_dec(&znd->gc_throttle); + + return can_retry; +} + +/** + * set_current() - Make zone the preferred zone for allocation. + * @znd: ZDM Instance + * @flags: BLock allocation scheme (including stream id) + * @zone: The zone to make preferred. + * + * Once a zone is opened for allocation, future allocations will prefer + * the same zone, until the zone is full. + * Each stream id has it's own preferred zone. + * + * NOTE: z_current is being deprecated if favor of assuming a default + * stream id when nothing is provided. + */ +static inline void set_current(struct zoned *znd, u32 flags, u32 zone) +{ + if (flags & Z_AQ_STREAM_ID) { + u32 stream_id = flags & Z_AQ_STREAM_MASK; + + znd->bmkeys->stream[stream_id] = cpu_to_le32(zone); + } + znd->z_current = zone; + if (znd->z_gc_free > 0) + znd->z_gc_free--; + else + Z_ERR(znd, "Dec z_gc_free below 0?"); +} + +/** + * next_open_zone() - Grab the next available zone + * @znd: ZDM Instance + * @z_at: Zone to start scanning from (presumable just filled). + * + * Return: NOZONE if no zone exists with space for writing. + * + * Scan through the available zones for an empty zone. + * If no empty zone is available the a zone that is not full is + * used instead. + */ +static u32 next_open_zone(struct zoned *znd, u32 z_at) +{ + u32 zone = NOZONE; + u32 z_id; + + if (znd->data_zones < z_at) + z_at = znd->data_zones; + + /* scan higher lba zones */ + for (z_id = z_at; z_id < znd->data_zones; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + + if ((wp & Z_WP_VALUE_MASK) == 0) { + u32 check = gzno << GZ_BITS | gzoff; + + if (!is_reserved(znd, check)) { + zone = check; + goto out; + } + } + } + + /* scan lower lba zones */ + for (z_id = 0; z_id < z_at; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + + if ((wp & Z_WP_VALUE_MASK) == 0) { + u32 check = gzno << GZ_BITS | gzoff; + + if (!is_reserved(znd, check)) { + zone = check; + goto out; + } + } + } + + /* No empty zones .. start co-mingling streams */ + for (z_id = 0; z_id < znd->data_zones; z_id++) { + u32 gzno = z_id >> GZ_BITS; + u32 gzoff = z_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + + if ((wp & Z_WP_VALUE_MASK) < Z_BLKSZ) { + u32 check = gzno << GZ_BITS | gzoff; + + if (!is_reserved(znd, check)) { + zone = check; + goto out; + } + } + } + +out: + return zone; +} + +/** + * zone_filled_cleanup() - Update wp_alloc GC Readu flags based on wp_used. + * @znd: ZDM Instance + */ +static void zone_filled_cleanup(struct zoned *znd) +{ + if (znd->filled_zone != NOZONE) { + u32 zone = znd->filled_zone; + u32 gzno; + u32 gzoff; + u32 wp; + u32 used; + struct meta_pg *wpg; + + znd->filled_zone = NOZONE; + + gzno = zone >> GZ_BITS; + gzoff = zone & GZ_MMSK; + wpg = &znd->wp[gzno]; + + SpinLock(&wpg->wplck); + wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + used = le32_to_cpu(wpg->wp_used[gzoff]) & Z_WP_VALUE_MASK; + if (used == Z_BLKSZ) { + if (Z_BLKSZ == (wp & Z_WP_VALUE_MASK)) { + wpg->wp_alloc[gzoff] = cpu_to_le32(wp + | Z_WP_GC_READY); + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + } else { + Z_ERR(znd, "Zone %u seems bogus.", zone); + } + } + spin_unlock(&wpg->wplck); + + dmz_close_zone(znd, zone); + update_stale_ratio(znd, zone); + } +} + +/** + * z_acquire() - Allocate blocks for writing + * @znd: ZDM Instance + * @flags: Alloc strategy and stream id. + * @nblks: Number of blocks desired. + * @nfound: Number of blocks available. + * + * Return: Lba for writing. + */ +static u64 z_acquire(struct zoned *znd, u32 flags, u32 nblks, u32 *nfound) +{ + sector_t found = 0; + u32 z_pref = znd->z_current; + u32 stream_id = 0; + u32 z_find; + const int wait = 1; + gfp_t gfp = (flags & Z_AQ_NORMAL) ? GFP_ATOMIC : GFP_KERNEL; + + zone_filled_cleanup(znd); + + if (flags & Z_AQ_STREAM_ID) { + stream_id = flags & Z_AQ_STREAM_MASK; + z_pref = le32_to_cpu(znd->bmkeys->stream[stream_id]); + } + if (z_pref >= znd->data_zones) { + z_pref = next_open_zone(znd, znd->z_current); + if (z_pref < znd->data_zones) + set_current(znd, flags, z_pref); + } + + if (z_pref < znd->data_zones) { + found = _blkalloc(znd, z_pref, flags, nblks, nfound); + if (found || *nfound) + goto out; + } + + if (znd->z_gc_free < 5) { + Z_DBG(znd, "... alloc - gc low on free space."); + gc_immediate(znd, !wait, gfp); + } + +retry: + z_find = next_open_zone(znd, znd->z_current); + if (z_find < znd->data_zones) { + found = _blkalloc(znd, z_find, flags, nblks, nfound); + if (found || *nfound) { + set_current(znd, flags, z_find); + goto out; + } + } + + if (flags & Z_AQ_GC) { + u32 gresv = znd->z_gc_resv & Z_WP_VALUE_MASK; + + Z_ERR(znd, "Using GC Reserve (%u)", gresv); + found = _blkalloc(znd, gresv, flags, nblks, nfound); + znd->z_gc_resv |= Z_WP_GC_ACTIVE; + } + + if (flags & Z_AQ_META) { + int can_retry = gc_immediate(znd, wait, gfp); + u32 mresv = znd->z_meta_resv & Z_WP_VALUE_MASK; + + Z_DBG(znd, "GC: Need META."); + if (can_retry) + goto retry; + + Z_ERR(znd, "Using META Reserve (%u)", znd->z_meta_resv); + found = _blkalloc(znd, mresv, flags, nblks, nfound); + } + +out: + if (!found && (*nfound == 0)) { + Z_DBG(znd, "... alloc - out of space?"); + if (gc_immediate(znd, wait, gfp)) + goto retry; + + Z_ERR(znd, "%s: -> Out of space.", __func__); + } + return found; +} + +/** + * compare_lba() - Compare map page on lba. + * @x1: map page + * @x2: map page + * + * Return -1, 0, or 1 if x1 < x2, equal, or >, respectivly. + */ +static int compare_lba(const void *x1, const void *x2) +{ + const struct map_pg *v1 = *(const struct map_pg **)x1; + const struct map_pg *v2 = *(const struct map_pg **)x2; + int cmp = (v1->lba < v2->lba) ? -1 : ((v1->lba > v2->lba) ? 1 : 0); + + return cmp; +} + +/** + * is_dirty() - Test map page if is of bit_type and dirty. + * @expg: map page + * @bit_type: map page flag to test for... + * + * Return 1 if page had dirty and bit_type flags set. + * + * Note: bit_type is IS_CRC and IS_LUT most typically. + */ +static __always_inline int is_dirty(struct map_pg *expg, int bit_type) +{ + return (test_bit(bit_type, &expg->flags) && + test_bit(IS_DIRTY, &expg->flags)); +} + + +/** + * is_old_and_clean() - Test map page if it expired and can be dropped. + * @expg: map page + * @bit_type: map page flag to test for... + * + * Return 1 if page is clean and old. + */ +static __always_inline int is_old_and_clean(struct map_pg *expg, int bit_type) +{ + int is_match = 0; + + if (!test_bit(IS_DIRTY, &expg->flags) && is_expired(expg->age)) { + if (getref_pg(expg) == 1) + is_match = 1; + else + pr_err("%llu: dirty, exp, and elev: %d\n", + expg->lba, getref_pg(expg)); + } + + (void)bit_type; + + return is_match; +} + +/** + * _pool_write() - Sort/Write and array of ZLT pages. + * @znd: ZDM Instance + * @wset: Array of pages to be written. + * @count: Number of entries. + * + * NOTE: On entry all map_pg entries have elevated refcount from _pool_fill(). + * write_if_dirty() will dec the refcount when the block hits disk. + */ +static int _pool_write(struct zoned *znd, struct map_pg **wset, int count) +{ + const int use_wq = 0; + int iter; + struct map_pg *expg; + int err = 0; + + /* write dirty table pages */ + if (count <= 0) + goto out; + + sort(wset, count, sizeof(*wset), compare_lba, NULL); + + for (iter = 0; iter < count; iter++) { + expg = wset[iter]; + if (expg) { + if (iter && expg->lba == wset[iter-1]->lba) { + wset[iter] = NULL; + Z_ERR(znd, "** DUPE ** ignored."); + } else + cache_if_dirty(znd, expg, use_wq); + } + } + + for (iter = 0; iter < count; iter++) { + const int sync = 1; + /* REDUX: For Async WB: int sync = (iter == last) ? 1 : 0; */ + + expg = wset[iter]; + if (expg) { + err = write_if_dirty(znd, expg, use_wq, sync); + if (err) { + Z_ERR(znd, "Write failed: %d", err); + goto out; + } + } + } + err = count; + +out: + return err; +} + +/** + * z_metadata_lba() - Alloc a block of metadata. + * @znd: ZDM Instance + * @map: Block of metadata. + * @num: Number of blocks allocated. + * + * Return: lba or 0 on failure. + * + * When map->lba is less than data_lba the metadata is pinned to it's logical + * location. + * When map->lba lands in data space it is dynmaically allocated and intermixed + * within the datapool. + * + * [FUTURE: Pick a stream_id for floating metadata] + */ +static u64 z_metadata_lba(struct zoned *znd, struct map_pg *map, u32 *num) +{ + u32 nblks = 1; + u64 lba = map->lba; + + if (lba < znd->data_lba) { + *num = 1; + return map->lba; + } + return z_acquire(znd, Z_AQ_META_STREAM, nblks, num); +} + +/** + * pg_crc() - + * @znd: ZDM Instance + * @pg: + * @md_crc: 16 bit crc of page. + * + * callback from dm_io notify.. cannot hold mutex here + */ +static void pg_crc(struct zoned *znd, struct map_pg *pg, __le16 md_crc) +{ + struct mpinfo mpi; + + to_table_entry(znd, pg->lba, &mpi); + if (pg->crc_pg) { + struct map_pg *crc_pg = pg->crc_pg; + int entry = mpi.crc.pg_idx; + + if (crc_pg && crc_pg->data.crc) { + ref_pg(crc_pg); + crc_pg->data.crc[entry] = md_crc; + clear_bit(IS_READA, &crc_pg->flags); + set_bit(IS_DIRTY, &crc_pg->flags); + clear_bit(IS_FLUSH, &crc_pg->flags); + crc_pg->age = jiffies_64 + + msecs_to_jiffies(crc_pg->hotness); + deref_pg(crc_pg); + } + put_map_entry(crc_pg); + if (getref_pg(crc_pg) == 0) + pg->crc_pg = NULL; + } else if (!test_bit(IS_LUT, &pg->flags)) { + znd->md_crcs[mpi.crc.pg_idx] = md_crc; + } else { + Z_ERR(znd, "unexpected state."); + dump_stack(); + } +} + + +/** + * pg_written() - Handle accouting related to lookup table page writes + * @pg: The page of lookup table [or CRC] that was written. + * @error: non-zero if an error occurred. + * + * callback from dm_io notify.. cannot hold mutex here, cannot sleep. + * + */ +static int pg_written(struct map_pg *pg, unsigned long error) +{ + int rcode = 0; + struct zoned *znd = pg->znd; + __le16 md_crc; + + if (error) { + Z_ERR(znd, "write_page: %" PRIx64 " -> %" PRIx64 + " ERR: %ld", pg->lba, pg->last_write, error); + rcode = -EIO; + goto out; + } + + /* + * Re-calculate CRC on current memory page. If unchanged then on-disk + * is stable and in-memory is not dirty. Otherwise in memory changed + * during write back so leave the dirty flag set. For the purpose of + * the CRC table we assume that in-memory == on-disk although this + * is not strictly true as the page could have updated post disk write. + */ + + md_crc = crc_md_le16(pg->data.addr, Z_CRC_4K); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + if (md_crc == pg->md_crc) + clear_bit(IS_DIRTY, &pg->flags); + + clear_bit(W_IN_FLIGHT, &pg->flags); + pg_crc(znd, pg, md_crc); + + if (pg->last_write < znd->data_lba) + goto out; + +/* + * NOTE: If we reach here it's a problem. + * TODO: mutex free path for adding a map entry ... + */ + + /* written lba was allocated from data-pool */ + rcode = z_mapped_addmany(znd, pg->lba, pg->last_write, 1, CRIT); + if (rcode) { + Z_ERR(znd, "%s: Journal MANY failed.", __func__); + goto out; + } + increment_used_blks(znd, pg->last_write, 1); + +out: + return rcode; +} + + +/** + * on_pg_written() - A block of map table was written. + * @error: Any error code that occurred during the I/O. + * @context: The map_pg that was queued/written. + */ +static void on_pg_written(unsigned long error, void *context) +{ + struct map_pg *pg = context; + int rcode; + + rcode = pg_written(pg, error); + deref_pg(pg); + if (rcode < 0) + pg->znd->meta_result = rcode; +} + +/** + * queue_pg() - Queue a map table page for writeback + * @znd: ZDM Instance + * @pg: The target page to ensure the cover CRC blocks is cached. + * @lba: The address to write the block to. + */ +static int queue_pg(struct zoned *znd, struct map_pg *pg, u64 lba) +{ + const int use_wq = 0; + sector_t block = lba << Z_SHFT4K; + unsigned int nDMsect = 1 << Z_SHFT4K; + int rc; + + pg->znd = znd; + MutexLock(&pg->md_lock); + pg->md_crc = crc_md_le16(pg->data.addr, Z_CRC_4K); + pg->last_write = lba; /* presumably */ + mutex_unlock(&pg->md_lock); + + rc = znd_async_io(znd, DM_IO_KMEM, pg->data.addr, block, nDMsect, WRITE, + use_wq, on_pg_written, pg); + if (rc) { + Z_ERR(znd, "queue error: %d Q: %" PRIx64 " [%u dm sect] (Q:%d)", + rc, lba, nDMsect, use_wq); + dump_stack(); + } + + return rc; +} + +/** + * cache_if_dirty() - Load a page of CRC's into memory. + * @znd: ZDM Instance + * @pg: The target page to ensure the cover CRC blocks is cached. + * @wp: If a queue is needed for I/O. + * + * The purpose of loading is to ensure the pages are in memory with the + * async_io (write) completes the CRC accounting doesn't cause a sleep + * and violate the callback() API rules. + */ +static void cache_if_dirty(struct zoned *znd, struct map_pg *pg, int wq) +{ + if (test_bit(IS_DIRTY, &pg->flags) && test_bit(IS_LUT, &pg->flags) && + pg->data.addr) { + u64 base = znd->c_base; + struct map_pg *crc_pg; + struct mpinfo mpi; + + if (test_bit(IS_REV, &pg->flags)) + base = znd->c_mid; + + to_table_entry(znd, pg->lba, &mpi); + base += mpi.crc.pg_no; + crc_pg = get_map_entry(znd, base, NORMAL); + if (!crc_pg) + Z_ERR(znd, "Out of memory. No CRC Pg"); + pg->crc_pg = crc_pg; + } +} + +/** + * write_if_dirty() - Write old pages (flagged as DIRTY) pages of table map. + * @znd: ZDM instance. + * @pg: A page of table map data. + * @wq: Use worker queue for sync writes. + * @snc: Performa a Sync or Async write. + * + * Return: 0 on success or -errno value + */ +static int write_if_dirty(struct zoned *znd, struct map_pg *pg, int wq, int snc) +{ + int rcode = 0; + + if (!pg) + return rcode; + + if (test_bit(IS_DIRTY, &pg->flags) && pg->data.addr) { + u64 dm_s = pg->lba; + u32 nf; + u64 lba = z_metadata_lba(znd, pg, &nf); + + Z_DBG(znd, "Write if dirty: %" PRIx64" -> %" PRIx64, dm_s, lba); + + if (lba && nf) { + int rcwrt; + int count = 1; + __le16 md_crc; + + set_bit(W_IN_FLIGHT, &pg->flags); + + if (!snc) { + rcode = queue_pg(znd, pg, lba); + goto out_queued; + } + + md_crc = crc_md_le16(pg->data.addr, Z_CRC_4K); + rcwrt = write_block(znd->ti, DM_IO_KMEM, + pg->data.addr, lba, count, wq); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + pg->last_write = lba; + pg_crc(znd, pg, md_crc); + + if (rcwrt) { + Z_ERR(znd, "write_page: %" PRIx64 " -> %" PRIx64 + " ERR: %d", pg->lba, lba, rcwrt); + rcode = rcwrt; + goto out; + } + + if (crc_md_le16(pg->data.addr, Z_CRC_4K) == md_crc) + clear_bit(IS_DIRTY, &pg->flags); + clear_bit(W_IN_FLIGHT, &pg->flags); + if (lba < znd->data_lba) + goto out; + + rcwrt = z_mapped_addmany(znd, dm_s, lba, nf, NORMAL); + if (rcwrt) { + Z_ERR(znd, "%s: Journal MANY failed.", + __func__); + rcode = rcwrt; + goto out; + } + increment_used_blks(znd, lba, nf); + } else { + Z_ERR(znd, "%s: Out of space for metadata?", __func__); + rcode = -ENOSPC; + goto out; + } + } + +out: + deref_pg(pg); + +out_queued: + if (rcode < 0) + znd->meta_result = rcode; + + return rcode; +} + +/** + * _sync_dirty() - Write all *dirty* ZLT blocks to disk (journal->SYNC->home) + * @znd: ZDM instance + * @bit_type: MAP blocks then CRC blocks. + * @sync: If true write dirty blocks to disk + * @drop: Number of ZLT blocks to free. + * + * Return: 0 on success or -errno value + */ +static int _sync_dirty(struct zoned *znd, int bit_type, int sync, int drop) +{ + int err = 0; + int entries = 0; + int want_flush = 0; + struct map_pg *expg = NULL; + struct map_pg *_tpg; + struct map_pg **wset = NULL; + int dlstsz = 0; + LIST_HEAD(droplist); + + wset = ZDM_CALLOC(znd, sizeof(*wset), MAX_WSET, KM_19, NORMAL); + + spin_lock(&znd->zlt_lck); + if (list_empty(&znd->zltpool)) + goto writeback; + + expg = list_last_entry(&znd->zltpool, typeof(*expg), zltlst); + if (!expg || &expg->zltlst == (&znd->zltpool)) + goto writeback; + + _tpg = list_prev_entry(expg, zltlst); + while (&expg->zltlst != &znd->zltpool) { + ref_pg(expg); + + if (sync && (entries < MAX_WSET) && is_dirty(expg, bit_type)) { + wset[entries++] = expg; + } else if ((drop > 0) && is_old_and_clean(expg, bit_type)) { + struct map_pg **zlt; + spinlock_t *lock; + int index = expg->index; + int is_lut = test_bit(IS_LUT, &expg->flags); + int is_fwd = test_bit(IS_FWD, &expg->flags); + u32 count = is_lut ? znd->map_count : znd->crc_count; + + if (is_lut) { + zlt = is_fwd ? znd->fwd_tm : znd->rev_tm; + lock = &znd->mapkey_lock; + } else { + zlt = is_fwd ? znd->fwd_crc : znd->rev_crc; + lock = &znd->ct_lock; + } + spin_lock(lock); + if (index > -1 && index < count && + getref_pg(expg) == 1 && expg == zlt[index]) { + zlt[expg->index] = NULL; + list_del(&expg->zltlst); + znd->in_zlt--; + clear_bit(IN_ZLT, &expg->flags); + drop--; + expg->age = jiffies_64; + + msecs_to_jiffies(expg->hotness); + +#if ENABLE_PG_FREE_VIA_LAZY + if (test_bit(IS_LAZY, &expg->flags)) + Z_ERR(znd, "** Pg is lazy && zlt %llx", + expg->lba); + + if (!test_bit(IS_LAZY, &expg->flags)) { + list_add(&expg->lazy, &droplist); + znd->in_lzy++; + set_bit(IS_LAZY, &expg->flags); + set_bit(IS_DROPPED, &expg->flags); + dlstsz++; + } + zlt[expg->index] = expg; /* so undrop works */ + deref_pg(expg); + if (getref_pg(expg) > 0) + Z_ERR(znd, "Moving elv ref: %u", + getref_pg(expg)); +#else + MutexLock(&expg->md_lock); + if (expg->data.addr) { + void *pg = expg->data.addr; + + expg->data.addr = NULL; + ZDM_FREE(znd, pg, Z_C4K, PG_27); + atomic_dec(&znd->incore); + if (!test_bit(IS_FLUSH, &expg->flags)) + want_flush++; + } + mutex_unlock(&expg->md_lock); + deref_pg(expg); + if (getref_pg(expg) > 0) + Z_ERR(znd, "Dropped elv ref: %u", + getref_pg(expg)); + ZDM_FREE(znd, expg, sizeof(*expg), KM_20); +#endif + } + spin_unlock(lock); + } else { + deref_pg(expg); + } + if (entries == MAX_WSET) + break; + + expg = _tpg; + _tpg = list_prev_entry(expg, zltlst); + } + +writeback: + spin_unlock(&znd->zlt_lck); + + if (entries > 0) { + err = _pool_write(znd, wset, entries); + if (err < 0) + goto out; + if (entries == MAX_WSET) + err = -EBUSY; + } + +out: + if (want_flush) + set_bit(DO_FLUSH, &znd->flags); + if (!list_empty(&droplist)) + lazy_pool_splice(znd, &droplist); + + if (wset) + ZDM_FREE(znd, wset, sizeof(*wset) * MAX_WSET, KM_19); + + return err; +} + +/** + * sync_dirty() - Write all *dirty* ZLT blocks to disk (journal->SYNC->home) + * @znd: ZDM instance + * @bit_type: MAP blocks then CRC blocks. + * @sync: Write dirty blocks + * @drop: IN: # of pages to free. + * + * Return: 0 on success or -errno value + */ +static int sync_dirty(struct zoned *znd, int bit_type, int sync, int drop) +{ + int err; + + MutexLock(&znd->pool_mtx); + do { + err = _sync_dirty(znd, bit_type, sync, drop); + drop = 0; + } while (err == -EBUSY); + + if (err > 0) + err = 0; + mutex_unlock(&znd->pool_mtx); + + return err; +} + +/** + * sync_mapped_pages() - Migrate lookup tables and crc pages to disk + * @znd: ZDM instance + * @sync: If dirty blocks need to be written. + * @drop: Number of blocks to drop. + * + * Return: 0 on success or -errno value + */ +static int sync_mapped_pages(struct zoned *znd, int sync, int drop) +{ + int err; + int remove = drop ? 1 : 0; + + err = sync_dirty(znd, IS_LUT, sync, drop); + + /* on error return */ + if (err < 0) + return err; + + /* TBD: purge CRC's on ref-count? */ + err = sync_dirty(znd, IS_CRC, sync, remove); + + return err; +} + +/** + * dm_s is a logical sector that maps 1:1 to the whole disk in 4k blocks + * Here the logical LBA and field are calculated for the lookup table + * where the physical LBA can be read from disk. + */ +static int map_addr_aligned(struct zoned *znd, u64 dm_s, struct map_addr *out) +{ + u64 block = dm_s >> 10; + + out->zone_id = block >> 6; + out->lut_s = block + znd->s_base; + out->lut_r = block + znd->r_base; + out->pg_idx = dm_s & 0x3FF; + + return 0; +} + +/* -------------------------------------------------------------------------- */ +/* -------------------------------------------------------------------------- */ + +/** + * map_addr_calc() + * @znd: ZDM instance + * @origin: address to calc + * @out: address, zone, crc, lut addr + * + * dm_s is a logical sector that maps 1:1 to the whole disk in 4k blocks + * Here the logical LBA and field are calculated for the lookup table + * where the physical LBA can be read from disk. + */ +static int map_addr_calc(struct zoned *znd, u64 origin, struct map_addr *out) +{ + u64 offset = ((origin < znd->md_end) ? znd->md_start : znd->md_end); + + out->dm_s = origin; + return map_addr_aligned(znd, origin - offset, out); +} + +/** + * read_pg() - Read a page of LUT/CRC from disk. + * @znd: ZDM instance + * @pg: Page to fill + * @gfp: Memory allocation rule + * @mpi: Backing page locations. + * + * Load a page of the sector lookup table that maps to pg->lba + * If pg->lba is not on disk return 0 + * + * Return: 1 if page exists, 0 if unmodified, else -errno on error. + */ +static int read_pg(struct zoned *znd, struct map_pg *pg, u64 lba48, gfp_t gfp, + struct mpinfo *mpi) +{ + int rcode = 0; + const int count = 1; + const int wq = 1; + int rd; + __le16 check; + __le16 expect = 0; + + /** + * This table entry may be on-disk, if so it needs to + * be loaded. + * If not it needs to be initialized to 0xFF + * + * When ! ZF_POOL_FWD and mapped block is FWD && LUT + * the mapped->lba may be pinned in sk_ pool + * if so the sk_ pool entry may need to be pulled + * to resvole the current address of mapped->lba + */ + +/* TODO: Async reads */ + + if (warn_bad_lba(znd, lba48)) + Z_ERR(znd, "Bad PAGE %" PRIx64, lba48); + + rd = read_block(znd->ti, DM_IO_KMEM, pg->data.addr, lba48, count, wq); + if (rd) { + Z_ERR(znd, "%s: read_block: ERROR: %d", __func__, rd); + rcode = -EIO; + goto out; + } + + /* + * Now check block crc + */ + check = crc_md_le16(pg->data.addr, Z_CRC_4K); + if (test_bit(IS_LUT, &pg->flags)) { + struct map_pg *crc_pg; + u64 base = znd->c_base; + + if (test_bit(IS_REV, &pg->flags)) + base = znd->c_mid; + + crc_pg = get_map_entry(znd, mpi->crc.pg_no + base, gfp); + if (crc_pg) { + ref_pg(crc_pg); + if (crc_pg->data.crc) { + MutexLock(&crc_pg->md_lock); + expect = crc_pg->data.crc[mpi->crc.pg_idx]; + mutex_unlock(&crc_pg->md_lock); + crc_pg->age = jiffies_64 + + msecs_to_jiffies(crc_pg->hotness); + } + deref_pg(crc_pg); + } + put_map_entry(crc_pg); + } else { + expect = znd->md_crcs[mpi->crc.pg_idx]; + } + + if (check != expect) { + znd->meta_result = -ENOSPC; + + Z_ERR(znd, + "Corrupt metadata: %" PRIx64 " from %" PRIx64 + " [%04x != %04x]", + pg->lba, lba48, + le16_to_cpu(check), + le16_to_cpu(expect)); + } + rcode = 1; + +out: + return rcode; +} + +/** + * cache_pg() - Load a page of LUT/CRC into memory from disk, or default values. + * @znd: ZDM instance + * @pg: Page to fill + * @gfp: Memory allocation rule + * @mpi: Backing page locations. + * + * Return: 1 if page loaded from disk, 0 if empty, else -errno on error. + */ +static int cache_pg(struct zoned *znd, struct map_pg *pg, gfp_t gfp, + struct mpinfo *mpi) +{ + int rc = 0; + int empty_val = test_bit(IS_LUT, &pg->flags) ? 0xff : 0; + u64 lba48 = current_mapping(znd, pg->lba, gfp); + + ref_pg(pg); + MutexLock(&pg->md_lock); + if (!pg->data.addr) { + pg->data.addr = ZDM_ALLOC(znd, Z_C4K, PG_27, gfp); + if (pg->data.addr) { + if (lba48) + rc = read_pg(znd, pg, lba48, gfp, mpi); + else + memset(pg->data.addr, empty_val, Z_C4K); + + if (rc < 0) { + Z_ERR(znd, "%s: read_pg from %" PRIx64 + " [to? %x] error: %d", __func__, + pg->lba, empty_val, rc); + ZDM_FREE(znd, pg->data.addr, Z_C4K, PG_27); + } + } else { + Z_ERR(znd, "%s: Out of memory.", __func__); + rc = -ENOMEM; + } + if (pg->data.addr) { + set_bit(IS_FLUSH, &pg->flags); + atomic_inc(&znd->incore); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + rc = pool_add(znd, pg); + Z_DBG(znd, "Page loaded: lba: %" PRIx64, pg->lba); + } + } + mutex_unlock(&pg->md_lock); + + deref_pg(pg); + return rc; +} + +/** + * z_lookup_table() - resolve a sector mapping via ZLT mapping + * @znd: ZDM Instance + * @addr: Address to resolve (via FWD map). + * @gfp: Current allocation flags. + */ +static u64 z_lookup_table(struct zoned *znd, u64 addr, gfp_t gfp) +{ + struct map_addr maddr; + struct map_pg *pg; + u64 old_phy = 0; + + map_addr_calc(znd, addr, &maddr); + pg = get_map_entry(znd, maddr.lut_s, gfp); + if (pg) { + ref_pg(pg); + if (pg->data.addr) { + __le32 delta; + + MutexLock(&pg->md_lock); + delta = pg->data.addr[maddr.pg_idx]; + mutex_unlock(&pg->md_lock); + old_phy = map_value(znd, delta); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + clear_bit(IS_READA, &pg->flags); + } + deref_pg(pg); + put_map_entry(pg); + } + return old_phy; +} + +/** + * update_map_entry() - Migrate memcache to lookup table map entries. + * @znd: ZDM instance + * @mapped: memcache block. + * @maddr: map_addr + * @to_addr: LBA or sector #. + * @is_fwd: flag forward or reverse lookup table. + * + * when is_fwd is 0: + * - maddr->dm_s is a sector -> lba. + * in this case the old lba is discarded and scheduled for cleanup + * by updating the reverse map lba tables noting that this location + * is now unused. + * when is_fwd is 0: + * - maddr->dm_s is an lba, lba -> dm_s + * + * Return: non-zero on error. + */ +static int update_map_entry(struct zoned *znd, struct map_pg *pg, + struct map_addr *maddr, u64 to_addr, int is_fwd) +{ + int err = -ENOMEM; + + if (pg && pg->data.addr) { + u64 index = maddr->pg_idx; + __le32 delta; + __le32 value; + int was_updated = 0; + + ref_pg(pg); + MutexLock(&pg->md_lock); + delta = pg->data.addr[index]; + err = map_encode(znd, to_addr, &value); + if (!err) { + /* + * if the value is modified update the table and + * place it on top of the active [zltlst] list + * this will keep the chunk of lookup table in + * memory. + */ + if (pg->data.addr[index] != value) { + pg->data.addr[index] = value; + pg->age = jiffies_64 + + msecs_to_jiffies(pg->hotness); + set_bit(IS_DIRTY, &pg->flags); + clear_bit(IS_FLUSH, &pg->flags); + clear_bit(IS_READA, &pg->flags); + was_updated = 1; + } else if (value != MZTEV_UNUSED) { + Z_ERR(znd, "*ERR* %" PRIx64 + " -> data.addr[index] (%x) == (%x)", + maddr->dm_s, pg->data.addr[index], + value); + dump_stack(); + } + } else { + Z_ERR(znd, "*ERR* Mapping: %" PRIx64 " to %" PRIx64, + to_addr, maddr->dm_s); + } + mutex_unlock(&pg->md_lock); + + if (was_updated && is_fwd && (delta != MZTEV_UNUSED)) { + u64 old_phy = map_value(znd, delta); + + /* + * add to discard list of the controlling mzone + * for the 'delta' physical block + */ + Z_DBG(znd, "%s: unused_phy: %" PRIu64 + " (new lba: %" PRIu64 ")", + __func__, old_phy, to_addr); + + WARN_ON(old_phy >= znd->nr_blocks); + + err = unused_phy(znd, old_phy, 0, NORMAL); + if (err) + err = -ENOSPC; + } + deref_pg(pg); + } + return err; +} + +/** + * move_to_map_tables() - Migrate memcache to lookup table map entries. + * @znd: ZDM instance + * @mcache: memcache block. + * + * Return: non-zero on error. + */ +static int move_to_map_tables(struct zoned *znd, struct map_cache *mcache) +{ + struct map_pg *smtbl = NULL; + struct map_pg *rmtbl = NULL; + struct map_addr maddr = { .dm_s = 0ul }; + struct map_addr rev = { .dm_s = 0ul }; + u64 lut_s = BAD_ADDR; + u64 lut_r = BAD_ADDR; + int jentry; + int err = 0; + int is_fwd = 1; + + /* the journal being move must remain stable so sorting + * is disabled. If a sort is desired due to an unsorted + * page the search devolves to a linear lookup. + */ + mcache_busy(mcache); + for (jentry = mcache->jcount; jentry > 0;) { + u64 dm_s = le64_to_lba48(mcache->jdata[jentry].tlba, NULL); + u64 lba = le64_to_lba48(mcache->jdata[jentry].bval, NULL); + + if (dm_s == Z_LOWER48 || lba == Z_LOWER48) { + mcache->jcount = --jentry; + continue; + } + + map_addr_calc(znd, dm_s, &maddr); + if (lut_s != maddr.lut_s) { + put_map_entry(smtbl); + if (smtbl) + deref_pg(smtbl); + smtbl = get_map_entry(znd, maddr.lut_s, NORMAL); + if (!smtbl) { + err = -ENOMEM; + goto out; + } + ref_pg(smtbl); + lut_s = smtbl->lba; + } + + is_fwd = 1; + if (lba == 0ul) + lba = BAD_ADDR; + err = update_map_entry(znd, smtbl, &maddr, lba, is_fwd); + if (err < 0) + goto out; + + /* + * In this case the reverse was handled as part of + * discarding the forward map entry -- if it was in use. + */ + if (lba != BAD_ADDR) { + map_addr_calc(znd, lba, &rev); + if (lut_r != rev.lut_r) { + put_map_entry(rmtbl); + if (rmtbl) + deref_pg(rmtbl); + rmtbl = get_map_entry(znd, rev.lut_r, NORMAL); + if (!rmtbl) { + err = -ENOMEM; + goto out; + } + ref_pg(rmtbl); + lut_r = rmtbl->lba; + } + is_fwd = 0; + err = update_map_entry(znd, rmtbl, &rev, dm_s, is_fwd); + if (err == 1) + err = 0; + } + + if (err < 0) + goto out; + + mcache->jdata[jentry].tlba = MC_INVALID; + mcache->jdata[jentry].bval = MC_INVALID; + if (mcache->jsorted == mcache->jcount) + mcache->jsorted--; + mcache->jcount = --jentry; + } +out: + if (smtbl) + deref_pg(smtbl); + if (rmtbl) + deref_pg(rmtbl); + put_map_entry(smtbl); + put_map_entry(rmtbl); + set_bit(DO_MEMPOOL, &znd->flags); + mcache_unbusy(mcache); + + return err; +} + +/** + * unused_phy() - Mark a block as unused. + * @znd: ZDM instance + * @lba: Logical LBA of block. + * @orig_s: Sector being marked. + * @gfp: Memory allocation rule + * + * Return: non-zero on error. + * + * Add an unused block to the list of blocks to be discarded during + * garbage collection. + */ +static int unused_phy(struct zoned *znd, u64 lba, u64 orig_s, gfp_t gfp) +{ + int err = 0; + struct map_pg *pg; + struct map_addr reverse; + int z_off; + + if (lba < znd->data_lba) + return 0; + + map_addr_calc(znd, lba, &reverse); + z_off = reverse.zone_id % 1024; + pg = get_map_entry(znd, reverse.lut_r, gfp); + if (!pg) { + err = -EIO; + Z_ERR(znd, "unused_phy: Reverse Map Entry not found."); + goto out; + } + + if (!pg->data.addr) { + Z_ERR(znd, "Catastrophic missing LUT page."); + dump_stack(); + err = -EIO; + goto out; + } + ref_pg(pg); + + /* + * if the value is modified update the table and + * place it on top of the active [zltlst] list + */ + if (pg->data.addr[reverse.pg_idx] != MZTEV_UNUSED) { + u32 gzno = reverse.zone_id >> GZ_BITS; + u32 gzoff = reverse.zone_id & GZ_MMSK; + struct meta_pg *wpg = &znd->wp[gzno]; + u32 wp; + u32 zf; + u32 stream_id; + + if (orig_s) { + __le32 enc = pg->data.addr[reverse.pg_idx]; + u64 dm_s = map_value(znd, enc); + int drop_discard = 0; + + if (dm_s < znd->data_lba) { + drop_discard = 1; + Z_ERR(znd, "Discard invalid target %" + PRIx64" - Is ZDM Meta %"PRIx64" vs %" + PRIx64, lba, orig_s, dm_s); + } + if (orig_s != dm_s) { + drop_discard = 1; + Z_ERR(znd, + "Discard %" PRIx64 + " mismatched src: %"PRIx64 " vs %" PRIx64, + lba, orig_s, dm_s); + } + if (drop_discard) + goto out_unlock; + } + MutexLock(&pg->md_lock); + pg->data.addr[reverse.pg_idx] = MZTEV_UNUSED; + mutex_unlock(&pg->md_lock); + pg->age = jiffies_64 + msecs_to_jiffies(pg->hotness); + set_bit(IS_DIRTY, &pg->flags); + clear_bit(IS_READA, &pg->flags); + clear_bit(IS_FLUSH, &pg->flags); + + SpinLock(&wpg->wplck); + wp = le32_to_cpu(wpg->wp_alloc[gzoff]); + zf = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_VALUE_MASK; + stream_id = le32_to_cpu(wpg->zf_est[gzoff]) & Z_WP_STREAM_MASK; + if (wp > 0 && zf < Z_BLKSZ) { + zf++; + wpg->zf_est[gzoff] = cpu_to_le32(zf | stream_id); + wpg->wp_alloc[gzoff] = cpu_to_le32(wp | Z_WP_RRECALC); + set_bit(IS_DIRTY, &wpg->flags); + clear_bit(IS_FLUSH, &wpg->flags); + } + spin_unlock(&wpg->wplck); + if ((wp & Z_WP_VALUE_MASK) == Z_BLKSZ) + znd->discard_count++; + } else { + Z_DBG(znd, "lba: %" PRIx64 " alread reported as free?", lba); + } + +out_unlock: + deref_pg(pg); + +out: + put_map_entry(pg); + + return err; +} -- 1.9.1 -- To unsubscribe from this list: send the line "unsubscribe linux-ide" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html