On 2020/8/5 07:58, Ming Lei wrote: > On Tue, Aug 04, 2020 at 10:23:32PM +0800, Coly Li wrote: >> When some buggy driver doesn't set its queue->limits.discard_granularity >> (e.g. current loop device driver), discard at LBA 0 on such device will >> trigger a kernel BUG() panic from block/blk-mq.c:563. >> >> [ 955.565006][ C39] ------------[ cut here ]------------ >> [ 955.559660][ C39] invalid opcode: 0000 [#1] SMP NOPTI >> [ 955.622171][ C39] CPU: 39 PID: 248 Comm: ksoftirqd/39 Tainted: G E 5.8.0-default+ #40 >> [ 955.622171][ C39] Hardware name: Lenovo ThinkSystem SR650 -[7X05CTO1WW]-/-[7X05CTO1WW]-, BIOS -[IVE160M-2.70]- 07/17/2020 >> [ 955.622175][ C39] RIP: 0010:blk_mq_end_request+0x107/0x110 >> [ 955.622177][ C39] Code: 48 8b 03 e9 59 ff ff ff 48 89 df 5b 5d 41 5c e9 9f ed ff ff 48 8b 35 98 3c f4 00 48 83 c7 10 48 83 c6 19 e8 cb 56 c9 ff eb cb <0f> 0b 0f 1f 80 00 00 00 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 54 >> [ 955.622179][ C39] RSP: 0018:ffffb1288701fe28 EFLAGS: 00010202 >> [ 955.749277][ C39] RAX: 0000000000000001 RBX: ffff956fffba5080 RCX: 0000000000004003 >> [ 955.749278][ C39] RDX: 0000000000000003 RSI: 0000000000000000 RDI: 0000000000000000 >> [ 955.749279][ C39] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000000 >> [ 955.749279][ C39] R10: ffffb1288701fd28 R11: 0000000000000001 R12: ffffffffa8e05160 >> [ 955.749280][ C39] R13: 0000000000000004 R14: 0000000000000004 R15: ffffffffa7ad3a1e >> [ 955.749281][ C39] FS: 0000000000000000(0000) GS:ffff95bfbda00000(0000) knlGS:0000000000000000 >> [ 955.749282][ C39] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 >> [ 955.749282][ C39] CR2: 00007f6f0ef766a8 CR3: 0000005a37012002 CR4: 00000000007606e0 >> [ 955.749283][ C39] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 >> [ 955.749284][ C39] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 >> [ 955.749284][ C39] PKRU: 55555554 >> [ 955.749285][ C39] Call Trace: >> [ 955.749290][ C39] blk_done_softirq+0x99/0xc0 >> [ 957.550669][ C39] __do_softirq+0xd3/0x45f >> [ 957.550677][ C39] ? smpboot_thread_fn+0x2f/0x1e0 >> [ 957.550679][ C39] ? smpboot_thread_fn+0x74/0x1e0 >> [ 957.550680][ C39] ? smpboot_thread_fn+0x14e/0x1e0 >> [ 957.550684][ C39] run_ksoftirqd+0x30/0x60 >> [ 957.550687][ C39] smpboot_thread_fn+0x149/0x1e0 >> [ 957.886225][ C39] ? sort_range+0x20/0x20 >> [ 957.886226][ C39] kthread+0x137/0x160 >> [ 957.886228][ C39] ? kthread_park+0x90/0x90 >> [ 957.886231][ C39] ret_from_fork+0x22/0x30 >> [ 959.117120][ C39] ---[ end trace 3dacdac97e2ed164 ]--- >> >> This is the procedure to reproduce the panic, >> # modprobe scsi_debug delay=0 dev_size_mb=2048 max_queue=1 >> # losetup -f /dev/nvme0n1 --direct-io=on >> # blkdiscard /dev/loop0 -o 0 -l 0x200 >> >> This is how the BUG() panic triggered by __blkdev_issue_discard(), >> - For a NVMe SSD backing loop device, the driver does not initialize >> its queue->limits.discard_granularity and leaves it to 0. >> - When discard on LBA 0 of the loop device, __blkdev_issue_discard() >> is called before loop device driver code. >> - Inside __blkdev_issue_discard(), when calculating value of >> granularity_aligned_lba by >> granularity_aligned_lba = round_up(sector_mapped, >> q->limits.discard_granularity >> SECTOR_SHIFT); >> because sector_mapped is 0 (at LBA 0 and no partition offset), and >> q->limits.discard_granularity is 0 (by the buggy loop driver), the >> calculated granularity_aligned_lba is 0. >> - The inline function bio_aligned_discard_max_sectors() is defined as >> return round_down(UINT_MAX, q->limits.discard_granularity) >> >> SECTOR_SHIFT; >> when q->limits.discard_granularity is 0 from loop device driver, the >> above calculation returns value 0. >> - Now granularity_aligned_lba and sctor_mapped are 0, req_sectors is >> calculated by the following lines in __blkdev_issue_discard(), >> if (granularity_aligned_lba == sector_mapped) >> req_sects = min_t(sector_t, nr_sects, >> bio_aligned_discard_max_sectors(q)); >> because bio_aligned_discard_max_sectors(q) returns 0, req_sects is >> calculated as 0. >> - Now a discard bio is mistakenly initialized as a 0 byte bio by, >> bio->bi_iter.bi_size = req_sects << 9; >> and sent to loop device driver. >> - This discard request is handled by loop device driver by following >> code path, >> loop_handle_cmd => do_req_filebacked => lo_fallocate => >> file->f_op->fallocate => blkdev_fallocate => blkdev_issue_zeroout => >> __blkdev_issue_write_zeroes >> - Inside __blkdev_issue_write_zeroes(), a 0 byte length discard bio is >> composed and sent to the backing device of the loop device. >> - In the I/O completion code path, in my case it is, >> blk_done_softirq => nrq->q->mq_ops->complete => nvme_pci_complete_rq >> => nvme_complete_rq => blk_mq_end_request >> inside blk_mq_end_request(), blk_update_request() is called and due to >> req->bio is NULL in previous step, blk_update_request() returns false >> then the BUG() panic in blk_mq_end_request() is triggered. >> >> Although the above panic can be fixed in loop device driver, the generic >> __blkdev_issue_discard() should also be fixed to tolerate the incorrect >> 0 value from queue->limits.discard_granularity, in case some other buggy >> driver makes such mistake again. >> >> This patch checks whether q->limits.discard_granularity is 0 in >> __blkdev_issue_discard() and bio_aligned_discard_max_sectors(). If it is >> 0 from some buggy driver queue, prints a warning oops information and >> set queue_logical_block_size(q) to a local variable discard_granularity. >> This local variable is used in round_up() and round_down() calculation, >> now req_sects won't be 0 and no empty discard request is generated. >> >> Fixes: 9b15d109a6b2 ("block: improve discard bio alignment in __blkdev_issue_discard()") >> Fixes: c52abf563049 ("loop: Better discard support for block devices") >> Reported-by: Ming Lei <ming.lei@xxxxxxxxxx> >> Signed-off-by: Coly Li <colyli@xxxxxxx> >> Cc: Hannes Reinecke <hare@xxxxxxxx> >> Cc: Ming Lei <ming.lei@xxxxxxxxxx> >> Cc: Xiao Ni <xni@xxxxxxxxxx> >> Cc: Bart Van Assche <bvanassche@xxxxxxx> >> Cc: Christoph Hellwig <hch@xxxxxx> >> Cc: Enzo Matsumiya <ematsumiya@xxxxxxxx> >> Cc: Jens Axboe <axboe@xxxxxxxxx> >> Cc: Evan Green <evgreen@xxxxxxxxxxxx> >> --- >> block/blk-lib.c | 8 +++++++- >> block/blk.h | 9 +++++++-- >> 2 files changed, 14 insertions(+), 3 deletions(-) >> >> diff --git a/block/blk-lib.c b/block/blk-lib.c >> index 019e09bb9c0e..3017e4cba923 100644 >> --- a/block/blk-lib.c >> +++ b/block/blk-lib.c >> @@ -30,6 +30,7 @@ int __blkdev_issue_discard(struct block_device *bdev, sector_t sector, >> struct bio *bio = *biop; >> unsigned int op; >> sector_t bs_mask, part_offset = 0; >> + sector_t discard_granularity; >> >> if (!q) >> return -ENXIO; >> @@ -54,6 +55,11 @@ int __blkdev_issue_discard(struct block_device *bdev, sector_t sector, >> if (!nr_sects) >> return -EINVAL; >> >> + discard_granularity = q->limits.discard_granularity; >> + /* In case some buggy driver does not set limits.discard_granularity */ >> + if (WARN_ON_ONCE(discard_granularity == 0)) >> + discard_granularity = queue_logical_block_size(q); > > This code path is supposed to not run in case of zero q->limits.discard_granularity, > and looks it is fine to just warn and return -EINVAL in this case, > see Documentation/ABI/testing/sysfs-block: > > What: /sys/block/<disk>/queue/discard_granularity > Date: May 2011 > Contact: Martin K. Petersen <martin.petersen@xxxxxxxxxx> > Description: > Devices that support discard functionality may > internally allocate space using units that are bigger > than the logical block size. The discard_granularity > parameter indicates the size of the internal allocation > unit in bytes if reported by the device. Otherwise the > discard_granularity will be set to match the device's > physical block size. A discard_granularity of 0 means > that the device does not support discard functionality. > > What we need to fix is loop driver, if it claims to support discard, > q->limits.discard_granularity has to be one valid value. Yes your suggestion is much simpler, let me do it :-) Thanks. Coly Li