On Jul 27, 2023, at 5:45 PM, Bobi Jam <bobijam@xxxxxxxxxxx> wrote: > > With LVM it is possible to create an LV with SSD storage at the > beginning of the LV and HDD storage at the end of the LV, and use that > to separate ext4 metadata allocations (that need small random IOs) > from data allocations (that are better suited for large sequential > IOs) depending on the type of underlying storage. Between 0.5-1.0% of > the filesystem capacity would need to be high-IOPS storage in order to > hold all of the internal metadata. > > This would improve performance for inode and other metadata access, > such as ls, find, e2fsck, and in general improve file access latency, > modification, truncate, unlink, transaction commit, etc. > > This patch split largest free order group lists and average fragment > size lists into other two lists for IOPS/fast storage groups, and > cr 0 / cr 1 group scanning for metadata block allocation in following > order: > > cr 0 on largest free order IOPS group list > cr 1 on average fragment size IOPS group list > cr 0 on largest free order non-IOPS group list > cr 1 on average fragment size non-IOPS group list > cr >= 2 perform the linear search as before > > Non-metadata block allocation does not allocate from the IOPS groups. > > Add for mke2fs an option to mark which blocks are in the IOPS region > of storage at format time: > > -E iops=0-1024G,4096-8192G > > so the ext4 mballoc code can then use the EXT4_BG_IOPS flag in the > group descriptors to decide which groups to allocate dynamic filesystem > metadata. Ted, Ritesh, Ojaswin, I would appreciate your review and comments on these two patches. They are a followup to the discussion about hybrid LVM devices a few weeks ago in https://www.spinics.net/lists/linux-ext4/msg90237.html "packed_meta_blocks=1 incompatible with resize2fs?". The 2/2 patch adds an option to mke2fs to mark groups on the SSD/NVMe storage with the "EXT4_BG_IOPS" flag. Together with mke2fs using the existing sparse_super2, flex_bg, and packed_meta_blocks, this would allocate all static metadata to the start of the device. The 1/2 patch changes mballoc to keep two separate allocation lists/trees based on the IOPS flag on each group. This has the dual benefit that all filesystem metadata (typically 4KiB fragmented allocation/read/write) is on fast storage, and these blocks also do not interfere with (usually larger) data allocation/read/write (both in terms of allocation fragmentation and contending IOPS/seeks). This should help both normal IO usage, as well as e2fsck significantly (virtually all e2fsck IO would go to the SSD/NVMe storage). The implementation is relatively simple as you can see. Currently it has a flag in the superblock to indicate that IOPS groups are available during block allocation, but even that is not strictly needed (it could be detected at GDT reading time). Metadata allocations prefer to use the IOPS groups (if available), otherwise fall back to regular groups. For our usage, this is a "soft" feature that does not affect compatibility. It would be mostly harmless if the filesystem was mounted with an older kernel. At worst some performance loss that would disappear again over time, but this would happen rarely I think. This doesn't *directly* address filesystem resize that Roberto was asking about, but having IOPS groups used only for metadata would make it easier to resize later (if only adding HDD capacity). Alternately, because the individual groups are marked with the IOPS flag, a second (third, fourth) flash region could be added at the end of the current filesystem to hold the new bitmaps and inode tables would be relatively straight forward to add on top of this. There might be some work needed for mke2fs to honor the "resize" option with packed_meta_blocks, but maybe not much more. We basically never resize filesystems, so this is not of any interest to implement at this point. Cheers, Andreas > Signed-off-by: Bobi Jam <bobijam@xxxxxxxxxxx> > --- > fs/ext4/balloc.c | 2 +- > fs/ext4/ext4.h | 12 +++++ > fs/ext4/mballoc.c | 154 ++++++++++++++++++++++++++++++++++++++++++------------ > 3 files changed, 134 insertions(+), 34 deletions(-) > > diff --git a/fs/ext4/balloc.c b/fs/ext4/balloc.c > index c1edde8..7b1b3ec 100644 > --- a/fs/ext4/balloc.c > +++ b/fs/ext4/balloc.c > @@ -739,7 +739,7 @@ ext4_fsblk_t ext4_new_meta_blocks(handle_t *handle, struct inode *inode, > ar.inode = inode; > ar.goal = goal; > ar.len = count ? *count : 1; > - ar.flags = flags; > + ar.flags = flags | EXT4_MB_HINT_METADATA; > > ret = ext4_mb_new_blocks(handle, &ar, errp); > if (count) > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h > index 8104a21..3444b6e 100644 > --- a/fs/ext4/ext4.h > +++ b/fs/ext4/ext4.h > @@ -382,6 +382,7 @@ struct flex_groups { > #define EXT4_BG_INODE_UNINIT 0x0001 /* Inode table/bitmap not in use */ > #define EXT4_BG_BLOCK_UNINIT 0x0002 /* Block bitmap not in use */ > #define EXT4_BG_INODE_ZEROED 0x0004 /* On-disk itable initialized to zero */ > +#define EXT4_BG_IOPS 0x0010 /* In IOPS/fast storage */ > > /* > * Macro-instructions used to manage group descriptors > @@ -1112,6 +1113,8 @@ struct ext4_inode_info { > #define EXT2_FLAGS_UNSIGNED_HASH 0x0002 /* Unsigned dirhash in use */ > #define EXT2_FLAGS_TEST_FILESYS 0x0004 /* to test development code */ > > +#define EXT2_FLAGS_HAS_IOPS 0x0080 /* has IOPS storage */ > + > /* > * Mount flags set via mount options or defaults > */ > @@ -1514,8 +1517,12 @@ struct ext4_sb_info { > atomic_t s_retry_alloc_pending; > struct list_head *s_mb_avg_fragment_size; > rwlock_t *s_mb_avg_fragment_size_locks; > + struct list_head *s_avg_fragment_size_list_iops; /* avg_frament_size for IOPS groups */ > + rwlock_t *s_avg_fragment_size_locks_iops; > struct list_head *s_mb_largest_free_orders; > rwlock_t *s_mb_largest_free_orders_locks; > + struct list_head *s_largest_free_orders_list_iops; /* largest_free_orders for IOPS grps */ > + rwlock_t *s_largest_free_orders_locks_iops; > > /* tunables */ > unsigned long s_stripe; > @@ -3366,6 +3373,7 @@ struct ext4_group_info { > #define EXT4_GROUP_INFO_IBITMAP_CORRUPT \ > (1 << EXT4_GROUP_INFO_IBITMAP_CORRUPT_BIT) > #define EXT4_GROUP_INFO_BBITMAP_READ_BIT 4 > +#define EXT4_GROUP_INFO_IOPS_BIT 5 > > #define EXT4_MB_GRP_NEED_INIT(grp) \ > (test_bit(EXT4_GROUP_INFO_NEED_INIT_BIT, &((grp)->bb_state))) > @@ -3382,6 +3390,10 @@ struct ext4_group_info { > (clear_bit(EXT4_GROUP_INFO_WAS_TRIMMED_BIT, &((grp)->bb_state))) > #define EXT4_MB_GRP_TEST_AND_SET_READ(grp) \ > (test_and_set_bit(EXT4_GROUP_INFO_BBITMAP_READ_BIT, &((grp)->bb_state))) > +#define EXT4_MB_GRP_TEST_IOPS(grp) \ > + (test_bit(EXT4_GROUP_INFO_IOPS_BIT, &((grp)->bb_state))) > +#define EXT4_MB_GRP_SET_IOPS(grp) \ > + (set_bit(EXT4_GROUP_INFO_IOPS_BIT, &((grp)->bb_state))) > > #define EXT4_MAX_CONTENTION 8 > #define EXT4_CONTENTION_THRESHOLD 2 > diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c > index 20f67a2..6d218af 100644 > --- a/fs/ext4/mballoc.c > +++ b/fs/ext4/mballoc.c > @@ -828,6 +828,8 @@ static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) > mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_info *grp) > { > struct ext4_sb_info *sbi = EXT4_SB(sb); > + rwlock_t *afs_locks; > + struct list_head *afs_list; > int new_order; > > if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free == 0) > @@ -838,20 +840,23 @@ static int mb_avg_fragment_size_order(struct super_block *sb, ext4_grpblk_t len) > if (new_order == grp->bb_avg_fragment_size_order) > return; > > + if (EXT4_MB_GRP_TEST_IOPS(grp)) { > + afs_locks = sbi->s_avg_fragment_size_locks_iops; > + afs_list = sbi->s_avg_fragment_size_list_iops; > + } else { > + afs_locks = sbi->s_mb_avg_fragment_size_locks; > + afs_list = sbi->s_mb_avg_fragment_size; > + } > + > if (grp->bb_avg_fragment_size_order != -1) { > - write_lock(&sbi->s_mb_avg_fragment_size_locks[ > - grp->bb_avg_fragment_size_order]); > + write_lock(&afs_locks[grp->bb_avg_fragment_size_order]); > list_del(&grp->bb_avg_fragment_size_node); > - write_unlock(&sbi->s_mb_avg_fragment_size_locks[ > - grp->bb_avg_fragment_size_order]); > + write_unlock(&afs_locks[grp->bb_avg_fragment_size_order]); > } > grp->bb_avg_fragment_size_order = new_order; > - write_lock(&sbi->s_mb_avg_fragment_size_locks[ > - grp->bb_avg_fragment_size_order]); > - list_add_tail(&grp->bb_avg_fragment_size_node, > - &sbi->s_mb_avg_fragment_size[grp->bb_avg_fragment_size_order]); > - write_unlock(&sbi->s_mb_avg_fragment_size_locks[ > - grp->bb_avg_fragment_size_order]); > + write_lock(&afs_locks[new_order]); > + list_add_tail(&grp->bb_avg_fragment_size_node, &afs_list[new_order]); > + write_unlock(&afs_locks[new_order]); > } > > /* > @@ -863,6 +868,10 @@ static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, > { > struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); > struct ext4_group_info *iter, *grp; > + bool iops = ac->ac_flags & EXT4_MB_HINT_METADATA && > + ac->ac_sb->s_flags & EXT2_FLAGS_HAS_IOPS; > + rwlock_t *lfo_locks; > + struct list_head *lfo_list; > int i; > > if (ac->ac_status == AC_STATUS_FOUND) > @@ -871,17 +880,25 @@ static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, > if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMIZED)) > atomic_inc(&sbi->s_bal_cr0_bad_suggestions); > > + if (iops) { > + lfo_locks = sbi->s_largest_free_orders_locks_iops; > + lfo_list = sbi->s_largest_free_orders_list_iops; > + } else { > + lfo_locks = sbi->s_mb_largest_free_orders_locks; > + lfo_list = sbi->s_mb_largest_free_orders; > + } > + > grp = NULL; > for (i = ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) { > - if (list_empty(&sbi->s_mb_largest_free_orders[i])) > + if (list_empty(&lfo_list[i])) > continue; > - read_lock(&sbi->s_mb_largest_free_orders_locks[i]); > - if (list_empty(&sbi->s_mb_largest_free_orders[i])) { > - read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); > + read_lock(&lfo_locks[i]); > + if (list_empty(&lfo_list[i])) { > + read_unlock(&lfo_locks[i]); > continue; > } > grp = NULL; > - list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[i], > + list_for_each_entry(iter, &lfo_list[i], > bb_largest_free_order_node) { > if (sbi->s_mb_stats) > atomic64_inc(&sbi->s_bal_cX_groups_considered[0]); > @@ -890,7 +907,7 @@ static void ext4_mb_choose_next_group_cr0(struct ext4_allocation_context *ac, > break; > } > } > - read_unlock(&sbi->s_mb_largest_free_orders_locks[i]); > + read_unlock(&lfo_locks[i]); > if (grp) > break; > } > @@ -913,6 +930,10 @@ static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, > { > struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb); > struct ext4_group_info *grp = NULL, *iter; > + bool iops = ac->ac_flags & EXT4_MB_HINT_METADATA && > + ac->ac_sb->s_flags & EXT2_FLAGS_HAS_IOPS; > + rwlock_t *afs_locks; > + struct list_head *afs_list; > int i; > > if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) { > @@ -920,16 +941,24 @@ static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, > atomic_inc(&sbi->s_bal_cr1_bad_suggestions); > } > > + if (iops) { > + afs_locks = sbi->s_avg_fragment_size_locks_iops; > + afs_list = sbi->s_avg_fragment_size_list_iops; > + } else { > + afs_locks = sbi->s_mb_avg_fragment_size_locks; > + afs_list = sbi->s_mb_avg_fragment_size; > + } > + > for (i = mb_avg_fragment_size_order(ac->ac_sb, ac->ac_g_ex.fe_len); > i < MB_NUM_ORDERS(ac->ac_sb); i++) { > - if (list_empty(&sbi->s_mb_avg_fragment_size[i])) > + if (list_empty(&afs_list[i])) > continue; > - read_lock(&sbi->s_mb_avg_fragment_size_locks[i]); > - if (list_empty(&sbi->s_mb_avg_fragment_size[i])) { > - read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); > + read_lock(&afs_locks[i]); > + if (list_empty(&afs_list[i])) { > + read_unlock(&afs_locks[i]); > continue; > } > - list_for_each_entry(iter, &sbi->s_mb_avg_fragment_size[i], > + list_for_each_entry(iter, &afs_list[i], > bb_avg_fragment_size_node) { > if (sbi->s_mb_stats) > atomic64_inc(&sbi->s_bal_cX_groups_considered[1]); > @@ -938,7 +967,7 @@ static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, > break; > } > } > - read_unlock(&sbi->s_mb_avg_fragment_size_locks[i]); > + read_unlock(&afs_locks[i]); > if (grp) > break; > } > @@ -947,7 +976,15 @@ static void ext4_mb_choose_next_group_cr1(struct ext4_allocation_context *ac, > *group = grp->bb_group; > ac->ac_flags |= EXT4_MB_CR1_OPTIMIZED; > } else { > - *new_cr = 2; > + if (iops) { > + /* cannot find proper group in IOPS storage, > + * fall back to cr0 for non-IOPS groups. > + */ > + ac->ac_flags &= ~EXT4_MB_HINT_METADATA; > + *new_cr = 0; > + } else { > + *new_cr = 2; > + } > } > } > > @@ -1030,6 +1067,8 @@ static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, > mb_set_largest_free_order(struct super_block *sb, struct ext4_group_info *grp) > { > struct ext4_sb_info *sbi = EXT4_SB(sb); > + rwlock_t *lfo_locks; > + struct list_head *lfo_list; > int i; > > for (i = MB_NUM_ORDERS(sb) - 1; i >= 0; i--) > @@ -1042,21 +1081,24 @@ static void ext4_mb_choose_next_group(struct ext4_allocation_context *ac, > return; > } > > + if (EXT4_MB_GRP_TEST_IOPS(grp)) { > + lfo_locks = sbi->s_largest_free_orders_locks_iops; > + lfo_list = sbi->s_largest_free_orders_list_iops; > + } else { > + lfo_locks = sbi->s_mb_largest_free_orders_locks; > + lfo_list = sbi->s_mb_largest_free_orders; > + } > + > if (grp->bb_largest_free_order >= 0) { > - write_lock(&sbi->s_mb_largest_free_orders_locks[ > - grp->bb_largest_free_order]); > + write_lock(&lfo_locks[grp->bb_largest_free_order]); > list_del_init(&grp->bb_largest_free_order_node); > - write_unlock(&sbi->s_mb_largest_free_orders_locks[ > - grp->bb_largest_free_order]); > + write_unlock(&lfo_locks[grp->bb_largest_free_order]); > } > grp->bb_largest_free_order = i; > if (grp->bb_largest_free_order >= 0 && grp->bb_free) { > - write_lock(&sbi->s_mb_largest_free_orders_locks[ > - grp->bb_largest_free_order]); > - list_add_tail(&grp->bb_largest_free_order_node, > - &sbi->s_mb_largest_free_orders[grp->bb_largest_free_order]); > - write_unlock(&sbi->s_mb_largest_free_orders_locks[ > - grp->bb_largest_free_order]); > + write_lock(&lfo_locks[i]); > + list_add_tail(&grp->bb_largest_free_order_node, &lfo_list[i]); > + write_unlock(&lfo_locks[i]); > } > } > > @@ -3150,6 +3192,8 @@ int ext4_mb_add_groupinfo(struct super_block *sb, ext4_group_t group, > INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list); > init_rwsem(&meta_group_info[i]->alloc_sem); > meta_group_info[i]->bb_free_root = RB_ROOT; > + if (desc->bg_flags & EXT4_BG_IOPS) > + EXT4_MB_GRP_SET_IOPS(meta_group_info[i]); > INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node); > INIT_LIST_HEAD(&meta_group_info[i]->bb_avg_fragment_size_node); > meta_group_info[i]->bb_largest_free_order = -1; /* uninit */ > @@ -3423,6 +3467,24 @@ int ext4_mb_init(struct super_block *sb) > INIT_LIST_HEAD(&sbi->s_mb_avg_fragment_size[i]); > rwlock_init(&sbi->s_mb_avg_fragment_size_locks[i]); > } > + sbi->s_avg_fragment_size_list_iops = > + kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), > + GFP_KERNEL); > + if (!sbi->s_avg_fragment_size_list_iops) { > + ret = -ENOMEM; > + goto out; > + } > + sbi->s_avg_fragment_size_locks_iops = > + kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), > + GFP_KERNEL); > + if (!sbi->s_avg_fragment_size_locks_iops) { > + ret = -ENOMEM; > + goto out; > + } > + for (i = 0; i < MB_NUM_ORDERS(sb); i++) { > + INIT_LIST_HEAD(&sbi->s_avg_fragment_size_list_iops[i]); > + rwlock_init(&sbi->s_avg_fragment_size_locks_iops[i]); > + } > sbi->s_mb_largest_free_orders = > kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), > GFP_KERNEL); > @@ -3441,6 +3503,24 @@ int ext4_mb_init(struct super_block *sb) > INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]); > rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]); > } > + sbi->s_largest_free_orders_list_iops = > + kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head), > + GFP_KERNEL); > + if (!sbi->s_largest_free_orders_list_iops) { > + ret = -ENOMEM; > + goto out; > + } > + sbi->s_largest_free_orders_locks_iops = > + kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t), > + GFP_KERNEL); > + if (!sbi->s_largest_free_orders_locks_iops) { > + ret = -ENOMEM; > + goto out; > + } > + for (i = 0; i < MB_NUM_ORDERS(sb); i++) { > + INIT_LIST_HEAD(&sbi->s_largest_free_orders_list_iops[i]); > + rwlock_init(&sbi->s_largest_free_orders_locks_iops[i]); > + } > > spin_lock_init(&sbi->s_md_lock); > sbi->s_mb_free_pending = 0; > @@ -3512,8 +3592,12 @@ int ext4_mb_init(struct super_block *sb) > out: > kfree(sbi->s_mb_avg_fragment_size); > kfree(sbi->s_mb_avg_fragment_size_locks); > + kfree(sbi->s_avg_fragment_size_list_iops); > + kfree(sbi->s_avg_fragment_size_locks_iops); > kfree(sbi->s_mb_largest_free_orders); > kfree(sbi->s_mb_largest_free_orders_locks); > + kfree(sbi->s_largest_free_orders_list_iops); > + kfree(sbi->s_largest_free_orders_locks_iops); > kfree(sbi->s_mb_offsets); > sbi->s_mb_offsets = NULL; > kfree(sbi->s_mb_maxs); > @@ -3582,8 +3666,12 @@ int ext4_mb_release(struct super_block *sb) > } > kfree(sbi->s_mb_avg_fragment_size); > kfree(sbi->s_mb_avg_fragment_size_locks); > + kfree(sbi->s_avg_fragment_size_list_iops); > + kfree(sbi->s_avg_fragment_size_locks_iops); > kfree(sbi->s_mb_largest_free_orders); > kfree(sbi->s_mb_largest_free_orders_locks); > + kfree(sbi->s_largest_free_orders_list_iops); > + kfree(sbi->s_largest_free_orders_locks_iops); > kfree(sbi->s_mb_offsets); > kfree(sbi->s_mb_maxs); > iput(sbi->s_buddy_cache); > -- > 1.8.3.1 > Cheers, Andreas
Attachment:
signature.asc
Description: Message signed with OpenPGP
