linux-stable/fs/btrfs/block-group.h

371 lines
12 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef BTRFS_BLOCK_GROUP_H
#define BTRFS_BLOCK_GROUP_H
#include "free-space-cache.h"
enum btrfs_disk_cache_state {
BTRFS_DC_WRITTEN,
BTRFS_DC_ERROR,
BTRFS_DC_CLEAR,
BTRFS_DC_SETUP,
};
btrfs: introduce size class to block group allocator The aim of this patch is to reduce the fragmentation of block groups under certain unhappy workloads. It is particularly effective when the size of extents correlates with their lifetime, which is something we have observed causing fragmentation in the fleet at Meta. This patch categorizes extents into size classes: - x < 128KiB: "small" - 128KiB < x < 8MiB: "medium" - x > 8MiB: "large" and as much as possible reduces allocations of extents into block groups that don't match the size class. This takes advantage of any (possible) correlation between size and lifetime and also leaves behind predictable re-usable gaps when extents are freed; small writes don't gum up bigger holes. Size classes are implemented in the following way: - Mark each new block group with a size class of the first allocation that goes into it. - Add two new passes to ffe: "unset size class" and "wrong size class". First, try only matching block groups, then try unset ones, then allow allocation of new ones, and finally allow mismatched block groups. - Filtering is done just by skipping inappropriate ones, there is no special size class indexing. Other solutions I considered were: - A best fit allocator with an rb-tree. This worked well, as small writes didn't leak big holes from large freed extents, but led to regressions in ffe and write performance due to lock contention on the rb-tree with every allocation possibly updating it in parallel. Perhaps something clever could be done to do the updates in the background while being "right enough". - A fixed size "working set". This prevents freeing an extent drastically changing where writes currently land, and seems like a good option too. Doesn't take advantage of size in any way. - The same size class idea, but implemented with xarray marks. This turned out to be slower than looping the linked list and skipping wrong block groups, and is also less flexible since we must have only 3 size classes (max #marks). With the current approach we can have as many as we like. Performance testing was done via: https://github.com/josefbacik/fsperf Of particular relevance are the new fragmentation specific tests. A brief summary of the testing results: - Neutral results on existing tests. There are some minor regressions and improvements here and there, but nothing that truly stands out as notable. - Improvement on new tests where size class and extent lifetime are correlated. Fragmentation in these cases is completely eliminated and write performance is generally a little better. There is also significant improvement where extent sizes are just a bit larger than the size class boundaries. - Regression on one new tests: where the allocations are sized intentionally a hair under the borders of the size classes. Results are neutral on the test that intentionally attacks this new scheme by mixing extent size and lifetime. The full dump of the performance results can be found here: https://bur.io/fsperf/size-class-2022-11-15.txt (there are ANSI escape codes, so best to curl and view in terminal) Here is a snippet from the full results for a new test which mixes buffered writes appending to a long lived set of files and large short lived fallocates: bufferedappendvsfallocate results metric baseline current stdev diff ====================================================================================== avg_commit_ms 31.13 29.20 2.67 -6.22% bg_count 14 15.60 0 11.43% commits 11.10 12.20 0.32 9.91% elapsed 27.30 26.40 2.98 -3.30% end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38% end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07% find_free_extent_calls 116244.30 114354.30 964.56 -1.63% find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67% find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80% find_free_extent_ns_min 500 512 6.67 2.40% find_free_extent_ns_p50 2848 2876 37.65 0.98% find_free_extent_ns_p95 4916 5000 75.45 1.71% find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90% frag_pct_max 61.67 0 8.05 -100.00% frag_pct_mean 43.59 0 6.10 -100.00% frag_pct_min 25.91 0 16.60 -100.00% frag_pct_p50 42.53 0 7.25 -100.00% frag_pct_p95 61.67 0 8.05 -100.00% frag_pct_p99 61.67 0 8.05 -100.00% fragmented_bg_count 6.10 0 1.45 -100.00% max_commit_ms 49.80 46 5.37 -7.63% sys_cpu 2.59 2.62 0.29 1.39% write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23% write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14% write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62% write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90% write_io_kbytes 4194304 4194304 0 0.00% write_iops 2476.15 2556.10 274.29 3.23% write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11% write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22% write_lat_ns_min 17353.10 16250 1646.08 -6.36% There are some mixed improvements/regressions in most metrics along with an elimination of fragmentation in this workload. On the balance, the drastic 1->0 improvement in the happy cases seems worth the mix of regressions and improvements we do observe. Some considerations for future work: - Experimenting with more size classes - More hinting/search ordering work to approximate a best-fit allocator Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2022-12-16 00:06:33 +00:00
enum btrfs_block_group_size_class {
/* Unset */
BTRFS_BG_SZ_NONE,
/* 0 < size <= 128K */
BTRFS_BG_SZ_SMALL,
/* 128K < size <= 8M */
BTRFS_BG_SZ_MEDIUM,
/* 8M < size < BG_LENGTH */
BTRFS_BG_SZ_LARGE,
};
/*
* This describes the state of the block_group for async discard. This is due
* to the two pass nature of it where extent discarding is prioritized over
* bitmap discarding. BTRFS_DISCARD_RESET_CURSOR is set when we are resetting
* between lists to prevent contention for discard state variables
* (eg. discard_cursor).
*/
enum btrfs_discard_state {
BTRFS_DISCARD_EXTENTS,
BTRFS_DISCARD_BITMAPS,
BTRFS_DISCARD_RESET_CURSOR,
};
/*
* Control flags for do_chunk_alloc's force field CHUNK_ALLOC_NO_FORCE means to
* only allocate a chunk if we really need one.
*
* CHUNK_ALLOC_LIMITED means to only try and allocate one if we have very few
* chunks already allocated. This is used as part of the clustering code to
* help make sure we have a good pool of storage to cluster in, without filling
* the FS with empty chunks
*
* CHUNK_ALLOC_FORCE means it must try to allocate one
*
* CHUNK_ALLOC_FORCE_FOR_EXTENT like CHUNK_ALLOC_FORCE but called from
* find_free_extent() that also activaes the zone
*/
enum btrfs_chunk_alloc_enum {
CHUNK_ALLOC_NO_FORCE,
CHUNK_ALLOC_LIMITED,
CHUNK_ALLOC_FORCE,
CHUNK_ALLOC_FORCE_FOR_EXTENT,
};
/* Block group flags set at runtime */
enum btrfs_block_group_flags {
BLOCK_GROUP_FLAG_IREF,
BLOCK_GROUP_FLAG_REMOVED,
BLOCK_GROUP_FLAG_TO_COPY,
BLOCK_GROUP_FLAG_RELOCATING_REPAIR,
BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
BLOCK_GROUP_FLAG_ZONED_DATA_RELOC,
/* Does the block group need to be added to the free space tree? */
BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE,
/* Indicate that the block group is placed on a sequential zone */
BLOCK_GROUP_FLAG_SEQUENTIAL_ZONE,
btrfs: fix use-after-free of new block group that became unused If a task creates a new block group and that block group becomes unused before we finish its creation, at btrfs_create_pending_block_groups(), then when btrfs_mark_bg_unused() is called against the block group, we assume that the block group is currently in the list of block groups to reclaim, and we move it out of the list of new block groups and into the list of unused block groups. This has two consequences: 1) We move it out of the list of new block groups associated to the current transaction. So the block group creation is not finished and if we attempt to delete the bg because it's unused, we will not find the block group item in the extent tree (or the new block group tree), its device extent items in the device tree etc, resulting in the deletion to fail due to the missing items; 2) We don't increment the reference count on the block group when we move it to the list of unused block groups, because we assumed the block group was on the list of block groups to reclaim, and in that case it already has the correct reference count. However the block group was on the list of new block groups, in which case no extra reference was taken because it's local to the current task. This later results in doing an extra reference count decrement when removing the block group from the unused list, eventually leading the reference count to 0. This second case was caught when running generic/297 from fstests, which produced the following assertion failure and stack trace: [589.559] assertion failed: refcount_read(&block_group->refs) == 1, in fs/btrfs/block-group.c:4299 [589.559] ------------[ cut here ]------------ [589.559] kernel BUG at fs/btrfs/block-group.c:4299! [589.560] invalid opcode: 0000 [#1] PREEMPT SMP PTI [589.560] CPU: 8 PID: 2819134 Comm: umount Tainted: G W 6.4.0-rc6-btrfs-next-134+ #1 [589.560] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.2-0-gea1b7a073390-prebuilt.qemu.org 04/01/2014 [589.560] RIP: 0010:btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.561] Code: 68 62 da c0 (...) [589.561] RSP: 0018:ffffa55a8c3b3d98 EFLAGS: 00010246 [589.561] RAX: 0000000000000058 RBX: ffff8f030d7f2000 RCX: 0000000000000000 [589.562] RDX: 0000000000000000 RSI: ffffffff953f0878 RDI: 00000000ffffffff [589.562] RBP: ffff8f030d7f2088 R08: 0000000000000000 R09: ffffa55a8c3b3c50 [589.562] R10: 0000000000000001 R11: 0000000000000001 R12: ffff8f05850b4c00 [589.562] R13: ffff8f030d7f2090 R14: ffff8f05850b4cd8 R15: dead000000000100 [589.563] FS: 00007f497fd2e840(0000) GS:ffff8f09dfc00000(0000) knlGS:0000000000000000 [589.563] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [589.563] CR2: 00007f497ff8ec10 CR3: 0000000271472006 CR4: 0000000000370ee0 [589.563] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [589.564] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [589.564] Call Trace: [589.564] <TASK> [589.565] ? __die_body+0x1b/0x60 [589.565] ? die+0x39/0x60 [589.565] ? do_trap+0xeb/0x110 [589.565] ? btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.566] ? do_error_trap+0x6a/0x90 [589.566] ? btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.566] ? exc_invalid_op+0x4e/0x70 [589.566] ? btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.567] ? asm_exc_invalid_op+0x16/0x20 [589.567] ? btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.567] ? btrfs_free_block_groups+0x449/0x4a0 [btrfs] [589.567] close_ctree+0x35d/0x560 [btrfs] [589.568] ? fsnotify_sb_delete+0x13e/0x1d0 [589.568] ? dispose_list+0x3a/0x50 [589.568] ? evict_inodes+0x151/0x1a0 [589.568] generic_shutdown_super+0x73/0x1a0 [589.569] kill_anon_super+0x14/0x30 [589.569] btrfs_kill_super+0x12/0x20 [btrfs] [589.569] deactivate_locked_super+0x2e/0x70 [589.569] cleanup_mnt+0x104/0x160 [589.570] task_work_run+0x56/0x90 [589.570] exit_to_user_mode_prepare+0x160/0x170 [589.570] syscall_exit_to_user_mode+0x22/0x50 [589.570] ? __x64_sys_umount+0x12/0x20 [589.571] do_syscall_64+0x48/0x90 [589.571] entry_SYSCALL_64_after_hwframe+0x72/0xdc [589.571] RIP: 0033:0x7f497ff0a567 [589.571] Code: af 98 0e (...) [589.572] RSP: 002b:00007ffc98347358 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6 [589.572] RAX: 0000000000000000 RBX: 00007f49800b8264 RCX: 00007f497ff0a567 [589.572] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000557f558abfa0 [589.573] RBP: 0000557f558a6ba0 R08: 0000000000000000 R09: 00007ffc98346100 [589.573] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000 [589.573] R13: 0000557f558abfa0 R14: 0000557f558a6cb0 R15: 0000557f558a6dd0 [589.573] </TASK> [589.574] Modules linked in: dm_snapshot dm_thin_pool (...) [589.576] ---[ end trace 0000000000000000 ]--- Fix this by adding a runtime flag to the block group to tell that the block group is still in the list of new block groups, and therefore it should not be moved to the list of unused block groups, at btrfs_mark_bg_unused(), until the flag is cleared, when we finish the creation of the block group at btrfs_create_pending_block_groups(). Fixes: a9f189716cf1 ("btrfs: move out now unused BG from the reclaim list") CC: stable@vger.kernel.org # 5.15+ Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2023-06-28 16:13:37 +00:00
/*
* Indicate that block group is in the list of new block groups of a
* transaction.
*/
BLOCK_GROUP_FLAG_NEW,
};
enum btrfs_caching_type {
BTRFS_CACHE_NO,
BTRFS_CACHE_STARTED,
BTRFS_CACHE_FINISHED,
BTRFS_CACHE_ERROR,
};
struct btrfs_caching_control {
struct list_head list;
struct mutex mutex;
wait_queue_head_t wait;
struct btrfs_work work;
struct btrfs_block_group *block_group;
btrfs: wait for actual caching progress during allocation Recently we've been having mysterious hangs while running generic/475 on the CI system. This turned out to be something like this: Task 1 dmsetup suspend --nolockfs -> __dm_suspend -> dm_wait_for_completion -> dm_wait_for_bios_completion -> Unable to complete because of IO's on a plug in Task 2 Task 2 wb_workfn -> wb_writeback -> blk_start_plug -> writeback_sb_inodes -> Infinite loop unable to make an allocation Task 3 cache_block_group ->read_extent_buffer_pages ->Waiting for IO to complete that can't be submitted because Task 1 suspended the DM device The problem here is that we need Task 2 to be scheduled completely for the blk plug to flush. Normally this would happen, we normally wait for the block group caching to finish (Task 3), and this schedule would result in the block plug flushing. However if there's enough free space available from the current caching to satisfy the allocation we won't actually wait for the caching to complete. This check however just checks that we have enough space, not that we can make the allocation. In this particular case we were trying to allocate 9MiB, and we had 10MiB of free space, but we didn't have 9MiB of contiguous space to allocate, and thus the allocation failed and we looped. We specifically don't cycle through the FFE loop until we stop finding cached block groups because we don't want to allocate new block groups just because we're caching, so we short circuit the normal loop once we hit LOOP_CACHING_WAIT and we found a caching block group. This is normally fine, except in this particular case where the caching thread can't make progress because the DM device has been suspended. Fix this by not only waiting for free space to >= the amount of space we want to allocate, but also that we make some progress in caching from the time we start waiting. This will keep us from busy looping when the caching is taking a while but still theoretically has enough space for us to allocate from, and fixes this particular case by forcing us to actually sleep and wait for forward progress, which will flush the plug. With this fix we're no longer hanging with generic/475. CC: stable@vger.kernel.org # 6.1+ Reviewed-by: Boris Burkov <boris@bur.io> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: David Sterba <dsterba@suse.com>
2023-07-21 20:09:43 +00:00
/* Track progress of caching during allocation. */
atomic_t progress;
refcount_t count;
};
/* Once caching_thread() finds this much free space, it will wake up waiters. */
#define CACHING_CTL_WAKE_UP SZ_2M
struct btrfs_block_group {
struct btrfs_fs_info *fs_info;
struct inode *inode;
spinlock_t lock;
u64 start;
u64 length;
u64 pinned;
u64 reserved;
u64 used;
u64 delalloc_bytes;
u64 bytes_super;
u64 flags;
u64 cache_generation;
u64 global_root_id;
btrfs: skip update of block group item if used bytes are the same [BACKGROUND] When committing a transaction, we will update block group items for all dirty block groups. But in fact, dirty block groups don't always need to update their block group items. It's pretty common to have a metadata block group which experienced several COW operations, but still have the same amount of used bytes. In that case, we may unnecessarily COW a tree block doing nothing. [ENHANCEMENT] This patch will introduce btrfs_block_group::commit_used member to remember the last used bytes, and use that new member to skip unnecessary block group item update. This would be more common for large filesystems, where metadata block group can be as large as 1GiB, containing at most 64K metadata items. In that case, if COW added and then deleted one metadata item near the end of the block group, then it's completely possible we don't need to touch the block group item at all. [BENCHMARK] The change itself can have quite a high chance (20~80%) to skip block group item updates in lot of workloads. As a result, it would result shorter time spent on btrfs_write_dirty_block_groups(), and overall reduce the execution time of the critical section of btrfs_commit_transaction(). Here comes a fio command, which will do random writes in 4K block size, causing a very heavy metadata updates. fio --filename=$mnt/file --size=512M --rw=randwrite --direct=1 --bs=4k \ --ioengine=libaio --iodepth=64 --runtime=300 --numjobs=4 \ --name=random_write --fallocate=none --time_based --fsync_on_close=1 The file size (512M) and number of threads (4) means 2GiB file size in total, but during the full 300s run time, my dedicated SATA SSD is able to write around 20~25GiB, which is over 10 times the file size. Thus after we fill the initial 2G, we should not cause much block group item updates. Please note, the fio numbers by themselves don't have much change, but if we look deeper, there is some reduced execution time, especially for the critical section of btrfs_commit_transaction(). I added extra trace_printk() to measure the following per-transaction execution time: - Critical section of btrfs_commit_transaction() By re-using the existing update_commit_stats() function, which has already calculated the interval correctly. - The while() loop for btrfs_write_dirty_block_groups() Although this includes the execution time of btrfs_run_delayed_refs(), it should still be representative overall. Both result involves transid 7~30, the same amount of transaction committed. The result looks like this: | Before | After | Diff ----------------------+-------------------+----------------+-------- Transaction interval | 229247198.5 | 215016933.6 | -6.2% Block group interval | 23133.33333 | 18970.83333 | -18.0% The change in block group item updates is more obvious, as skipped block group item updates also mean less delayed refs. And the overall execution time for that block group update loop is pretty small, thus we can assume the extent tree is already mostly cached. If we can skip an uncached tree block, it would cause more obvious change. Unfortunately the overall reduction in commit transaction critical section is much smaller, as the block group item updates loop is not really the major part, at least not for the above fio script. But still we have a observable reduction in the critical section. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-09-09 06:45:22 +00:00
/*
* The last committed used bytes of this block group, if the above @used
* is still the same as @commit_used, we don't need to update block
* group item of this block group.
*/
u64 commit_used;
/*
* If the free space extent count exceeds this number, convert the block
* group to bitmaps.
*/
u32 bitmap_high_thresh;
/*
* If the free space extent count drops below this number, convert the
* block group back to extents.
*/
u32 bitmap_low_thresh;
/*
* It is just used for the delayed data space allocation because
* only the data space allocation and the relative metadata update
* can be done cross the transaction.
*/
struct rw_semaphore data_rwsem;
/* For raid56, this is a full stripe, without parity */
unsigned long full_stripe_len;
unsigned long runtime_flags;
unsigned int ro;
int disk_cache_state;
/* Cache tracking stuff */
int cached;
struct btrfs_caching_control *caching_ctl;
struct btrfs_space_info *space_info;
/* Free space cache stuff */
struct btrfs_free_space_ctl *free_space_ctl;
/* Block group cache stuff */
struct rb_node cache_node;
/* For block groups in the same raid type */
struct list_head list;
refcount_t refs;
/*
* List of struct btrfs_free_clusters for this block group.
* Today it will only have one thing on it, but that may change
*/
struct list_head cluster_list;
/*
* Used for several lists:
*
* 1) struct btrfs_fs_info::unused_bgs
* 2) struct btrfs_fs_info::reclaim_bgs
* 3) struct btrfs_transaction::deleted_bgs
* 4) struct btrfs_trans_handle::new_bgs
*/
struct list_head bg_list;
/* For read-only block groups */
struct list_head ro_list;
/*
* When non-zero it means the block group's logical address and its
* device extents can not be reused for future block group allocations
* until the counter goes down to 0. This is to prevent them from being
* reused while some task is still using the block group after it was
* deleted - we want to make sure they can only be reused for new block
* groups after that task is done with the deleted block group.
*/
atomic_t frozen;
btrfs: add the beginning of async discard, discard workqueue When discard is enabled, everytime a pinned extent is released back to the block_group's free space cache, a discard is issued for the extent. This is an overeager approach when it comes to discarding and helping the SSD maintain enough free space to prevent severe garbage collection situations. This adds the beginning of async discard. Instead of issuing a discard prior to returning it to the free space, it is just marked as untrimmed. The block_group is then added to a LRU which then feeds into a workqueue to issue discards at a much slower rate. Full discarding of unused block groups is still done and will be addressed in a future patch of the series. For now, we don't persist the discard state of extents and bitmaps. Therefore, our failure recovery mode will be to consider extents untrimmed. This lets us handle failure and unmounting as one in the same. On a number of Facebook webservers, I collected data every minute accounting the time we spent in btrfs_finish_extent_commit() (col. 1) and in btrfs_commit_transaction() (col. 2). btrfs_finish_extent_commit() is where we discard extents synchronously before returning them to the free space cache. discard=sync: p99 total per minute p99 total per minute Drive | extent_commit() (ms) | commit_trans() (ms) --------------------------------------------------------------- Drive A | 434 | 1170 Drive B | 880 | 2330 Drive C | 2943 | 3920 Drive D | 4763 | 5701 discard=async: p99 total per minute p99 total per minute Drive | extent_commit() (ms) | commit_trans() (ms) -------------------------------------------------------------- Drive A | 134 | 956 Drive B | 64 | 1972 Drive C | 59 | 1032 Drive D | 62 | 1200 While it's not great that the stats are cumulative over 1m, all of these servers are running the same workload and and the delta between the two are substantial. We are spending significantly less time in btrfs_finish_extent_commit() which is responsible for discarding. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Dennis Zhou <dennis@kernel.org> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-14 00:22:14 +00:00
/* For discard operations */
struct list_head discard_list;
int discard_index;
u64 discard_eligible_time;
u64 discard_cursor;
enum btrfs_discard_state discard_state;
/* For dirty block groups */
struct list_head dirty_list;
struct list_head io_list;
struct btrfs_io_ctl io_ctl;
/*
* Incremented when doing extent allocations and holding a read lock
* on the space_info's groups_sem semaphore.
* Decremented when an ordered extent that represents an IO against this
* block group's range is created (after it's added to its inode's
* root's list of ordered extents) or immediately after the allocation
* if it's a metadata extent or fallocate extent (for these cases we
* don't create ordered extents).
*/
atomic_t reservations;
/*
* Incremented while holding the spinlock *lock* by a task checking if
* it can perform a nocow write (incremented if the value for the *ro*
* field is 0). Decremented by such tasks once they create an ordered
* extent or before that if some error happens before reaching that step.
* This is to prevent races between block group relocation and nocow
* writes through direct IO.
*/
atomic_t nocow_writers;
/* Lock for free space tree operations. */
struct mutex free_space_lock;
btrfs: fix race between writes to swap files and scrub When we active a swap file, at btrfs_swap_activate(), we acquire the exclusive operation lock to prevent the physical location of the swap file extents to be changed by operations such as balance and device replace/resize/remove. We also call there can_nocow_extent() which, among other things, checks if the block group of a swap file extent is currently RO, and if it is we can not use the extent, since a write into it would result in COWing the extent. However we have no protection against a scrub operation running after we activate the swap file, which can result in the swap file extents to be COWed while the scrub is running and operating on the respective block group, because scrub turns a block group into RO before it processes it and then back again to RW mode after processing it. That means an attempt to write into a swap file extent while scrub is processing the respective block group, will result in COWing the extent, changing its physical location on disk. Fix this by making sure that block groups that have extents that are used by active swap files can not be turned into RO mode, therefore making it not possible for a scrub to turn them into RO mode. When a scrub finds a block group that can not be turned to RO due to the existence of extents used by swap files, it proceeds to the next block group and logs a warning message that mentions the block group was skipped due to active swap files - this is the same approach we currently use for balance. Fixes: ed46ff3d42378 ("Btrfs: support swap files") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Anand Jain <anand.jain@oracle.com> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-05 12:55:37 +00:00
/*
* Number of extents in this block group used for swap files.
* All accesses protected by the spinlock 'lock'.
*/
int swap_extents;
/*
* Allocation offset for the block group to implement sequential
* allocation. This is used only on a zoned filesystem.
*/
u64 alloc_offset;
u64 zone_unusable;
u64 zone_capacity;
u64 meta_write_pointer;
struct map_lookup *physical_map;
struct list_head active_bg_list;
struct work_struct zone_finish_work;
struct extent_buffer *last_eb;
btrfs: introduce size class to block group allocator The aim of this patch is to reduce the fragmentation of block groups under certain unhappy workloads. It is particularly effective when the size of extents correlates with their lifetime, which is something we have observed causing fragmentation in the fleet at Meta. This patch categorizes extents into size classes: - x < 128KiB: "small" - 128KiB < x < 8MiB: "medium" - x > 8MiB: "large" and as much as possible reduces allocations of extents into block groups that don't match the size class. This takes advantage of any (possible) correlation between size and lifetime and also leaves behind predictable re-usable gaps when extents are freed; small writes don't gum up bigger holes. Size classes are implemented in the following way: - Mark each new block group with a size class of the first allocation that goes into it. - Add two new passes to ffe: "unset size class" and "wrong size class". First, try only matching block groups, then try unset ones, then allow allocation of new ones, and finally allow mismatched block groups. - Filtering is done just by skipping inappropriate ones, there is no special size class indexing. Other solutions I considered were: - A best fit allocator with an rb-tree. This worked well, as small writes didn't leak big holes from large freed extents, but led to regressions in ffe and write performance due to lock contention on the rb-tree with every allocation possibly updating it in parallel. Perhaps something clever could be done to do the updates in the background while being "right enough". - A fixed size "working set". This prevents freeing an extent drastically changing where writes currently land, and seems like a good option too. Doesn't take advantage of size in any way. - The same size class idea, but implemented with xarray marks. This turned out to be slower than looping the linked list and skipping wrong block groups, and is also less flexible since we must have only 3 size classes (max #marks). With the current approach we can have as many as we like. Performance testing was done via: https://github.com/josefbacik/fsperf Of particular relevance are the new fragmentation specific tests. A brief summary of the testing results: - Neutral results on existing tests. There are some minor regressions and improvements here and there, but nothing that truly stands out as notable. - Improvement on new tests where size class and extent lifetime are correlated. Fragmentation in these cases is completely eliminated and write performance is generally a little better. There is also significant improvement where extent sizes are just a bit larger than the size class boundaries. - Regression on one new tests: where the allocations are sized intentionally a hair under the borders of the size classes. Results are neutral on the test that intentionally attacks this new scheme by mixing extent size and lifetime. The full dump of the performance results can be found here: https://bur.io/fsperf/size-class-2022-11-15.txt (there are ANSI escape codes, so best to curl and view in terminal) Here is a snippet from the full results for a new test which mixes buffered writes appending to a long lived set of files and large short lived fallocates: bufferedappendvsfallocate results metric baseline current stdev diff ====================================================================================== avg_commit_ms 31.13 29.20 2.67 -6.22% bg_count 14 15.60 0 11.43% commits 11.10 12.20 0.32 9.91% elapsed 27.30 26.40 2.98 -3.30% end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38% end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07% find_free_extent_calls 116244.30 114354.30 964.56 -1.63% find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67% find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80% find_free_extent_ns_min 500 512 6.67 2.40% find_free_extent_ns_p50 2848 2876 37.65 0.98% find_free_extent_ns_p95 4916 5000 75.45 1.71% find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90% frag_pct_max 61.67 0 8.05 -100.00% frag_pct_mean 43.59 0 6.10 -100.00% frag_pct_min 25.91 0 16.60 -100.00% frag_pct_p50 42.53 0 7.25 -100.00% frag_pct_p95 61.67 0 8.05 -100.00% frag_pct_p99 61.67 0 8.05 -100.00% fragmented_bg_count 6.10 0 1.45 -100.00% max_commit_ms 49.80 46 5.37 -7.63% sys_cpu 2.59 2.62 0.29 1.39% write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23% write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14% write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62% write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90% write_io_kbytes 4194304 4194304 0 0.00% write_iops 2476.15 2556.10 274.29 3.23% write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11% write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22% write_lat_ns_min 17353.10 16250 1646.08 -6.36% There are some mixed improvements/regressions in most metrics along with an elimination of fragmentation in this workload. On the balance, the drastic 1->0 improvement in the happy cases seems worth the mix of regressions and improvements we do observe. Some considerations for future work: - Experimenting with more size classes - More hinting/search ordering work to approximate a best-fit allocator Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2022-12-16 00:06:33 +00:00
enum btrfs_block_group_size_class size_class;
};
btrfs: add the beginning of async discard, discard workqueue When discard is enabled, everytime a pinned extent is released back to the block_group's free space cache, a discard is issued for the extent. This is an overeager approach when it comes to discarding and helping the SSD maintain enough free space to prevent severe garbage collection situations. This adds the beginning of async discard. Instead of issuing a discard prior to returning it to the free space, it is just marked as untrimmed. The block_group is then added to a LRU which then feeds into a workqueue to issue discards at a much slower rate. Full discarding of unused block groups is still done and will be addressed in a future patch of the series. For now, we don't persist the discard state of extents and bitmaps. Therefore, our failure recovery mode will be to consider extents untrimmed. This lets us handle failure and unmounting as one in the same. On a number of Facebook webservers, I collected data every minute accounting the time we spent in btrfs_finish_extent_commit() (col. 1) and in btrfs_commit_transaction() (col. 2). btrfs_finish_extent_commit() is where we discard extents synchronously before returning them to the free space cache. discard=sync: p99 total per minute p99 total per minute Drive | extent_commit() (ms) | commit_trans() (ms) --------------------------------------------------------------- Drive A | 434 | 1170 Drive B | 880 | 2330 Drive C | 2943 | 3920 Drive D | 4763 | 5701 discard=async: p99 total per minute p99 total per minute Drive | extent_commit() (ms) | commit_trans() (ms) -------------------------------------------------------------- Drive A | 134 | 956 Drive B | 64 | 1972 Drive C | 59 | 1032 Drive D | 62 | 1200 While it's not great that the stats are cumulative over 1m, all of these servers are running the same workload and and the delta between the two are substantial. We are spending significantly less time in btrfs_finish_extent_commit() which is responsible for discarding. Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Dennis Zhou <dennis@kernel.org> Reviewed-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-12-14 00:22:14 +00:00
static inline u64 btrfs_block_group_end(struct btrfs_block_group *block_group)
{
return (block_group->start + block_group->length);
}
static inline bool btrfs_is_block_group_data_only(
struct btrfs_block_group *block_group)
{
/*
* In mixed mode the fragmentation is expected to be high, lowering the
* efficiency, so only proper data block groups are considered.
*/
return (block_group->flags & BTRFS_BLOCK_GROUP_DATA) &&
!(block_group->flags & BTRFS_BLOCK_GROUP_METADATA);
}
#ifdef CONFIG_BTRFS_DEBUG
int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group);
#endif
struct btrfs_block_group *btrfs_lookup_first_block_group(
struct btrfs_fs_info *info, u64 bytenr);
struct btrfs_block_group *btrfs_lookup_block_group(
struct btrfs_fs_info *info, u64 bytenr);
struct btrfs_block_group *btrfs_next_block_group(
struct btrfs_block_group *cache);
void btrfs_get_block_group(struct btrfs_block_group *cache);
void btrfs_put_block_group(struct btrfs_block_group *cache);
void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
const u64 start);
void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg);
btrfs: avoid double search for block group during NOCOW writes When doing a NOCOW write, either through direct IO or buffered IO, we do two lookups for the block group that contains the target extent: once when we call btrfs_inc_nocow_writers() and then later again when we call btrfs_dec_nocow_writers() after creating the ordered extent. The lookups require taking a lock and navigating the red black tree used to track all block groups, which can take a non-negligible amount of time for a large filesystem with thousands of block groups, as well as lock contention and cache line bouncing. Improve on this by having a single block group search: making btrfs_inc_nocow_writers() return the block group to its caller and then have the caller pass that block group to btrfs_dec_nocow_writers(). This is part of a patchset comprised of the following patches: btrfs: remove search start argument from first_logical_byte() btrfs: use rbtree with leftmost node cached for tracking lowest block group btrfs: use a read/write lock for protecting the block groups tree btrfs: return block group directly at btrfs_next_block_group() btrfs: avoid double search for block group during NOCOW writes The following test was used to test these changes from a performance perspective: $ cat test.sh #!/bin/bash modprobe null_blk nr_devices=0 NULL_DEV_PATH=/sys/kernel/config/nullb/nullb0 mkdir $NULL_DEV_PATH if [ $? -ne 0 ]; then echo "Failed to create nullb0 directory." exit 1 fi echo 2 > $NULL_DEV_PATH/submit_queues echo 16384 > $NULL_DEV_PATH/size # 16G echo 1 > $NULL_DEV_PATH/memory_backed echo 1 > $NULL_DEV_PATH/power DEV=/dev/nullb0 MNT=/mnt/nullb0 LOOP_MNT="$MNT/loop" MOUNT_OPTIONS="-o ssd -o nodatacow" MKFS_OPTIONS="-R free-space-tree -O no-holes" cat <<EOF > /tmp/fio-job.ini [io_uring_writes] rw=randwrite fsync=0 fallocate=posix group_reporting=1 direct=1 ioengine=io_uring iodepth=64 bs=64k filesize=1g runtime=300 time_based directory=$LOOP_MNT numjobs=8 thread EOF echo performance | \ tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor echo echo "Using config:" echo cat /tmp/fio-job.ini echo umount $MNT &> /dev/null mkfs.btrfs -f $MKFS_OPTIONS $DEV &> /dev/null mount $MOUNT_OPTIONS $DEV $MNT mkdir $LOOP_MNT truncate -s 4T $MNT/loopfile mkfs.btrfs -f $MKFS_OPTIONS $MNT/loopfile &> /dev/null mount $MOUNT_OPTIONS $MNT/loopfile $LOOP_MNT # Trigger the allocation of about 3500 data block groups, without # actually consuming space on underlying filesystem, just to make # the tree of block group large. fallocate -l 3500G $LOOP_MNT/filler fio /tmp/fio-job.ini umount $LOOP_MNT umount $MNT echo 0 > $NULL_DEV_PATH/power rmdir $NULL_DEV_PATH The test was run on a non-debug kernel (Debian's default kernel config), the result were the following. Before patchset: WRITE: bw=1455MiB/s (1526MB/s), 1455MiB/s-1455MiB/s (1526MB/s-1526MB/s), io=426GiB (458GB), run=300006-300006msec After patchset: WRITE: bw=1503MiB/s (1577MB/s), 1503MiB/s-1503MiB/s (1577MB/s-1577MB/s), io=440GiB (473GB), run=300006-300006msec +3.3% write throughput and +3.3% IO done in the same time period. The test has somewhat limited coverage scope, as with only NOCOW writes we get less contention on the red black tree of block groups, since we don't have the extra contention caused by COW writes, namely when allocating data extents, pinning and unpinning data extents, but on the hand there's access to tree in the NOCOW path, when incrementing a block group's number of NOCOW writers. Reviewed-by: Nikolay Borisov <nborisov@suse.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-04-13 15:20:43 +00:00
struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
u64 bytenr);
void btrfs_dec_nocow_writers(struct btrfs_block_group *bg);
void btrfs_wait_nocow_writers(struct btrfs_block_group *bg);
void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
u64 num_bytes);
btrfs: fix space cache corruption and potential double allocations When testing space_cache v2 on a large set of machines, we encountered a few symptoms: 1. "unable to add free space :-17" (EEXIST) errors. 2. Missing free space info items, sometimes caught with a "missing free space info for X" error. 3. Double-accounted space: ranges that were allocated in the extent tree and also marked as free in the free space tree, ranges that were marked as allocated twice in the extent tree, or ranges that were marked as free twice in the free space tree. If the latter made it onto disk, the next reboot would hit the BUG_ON() in add_new_free_space(). 4. On some hosts with no on-disk corruption or error messages, the in-memory space cache (dumped with drgn) disagreed with the free space tree. All of these symptoms have the same underlying cause: a race between caching the free space for a block group and returning free space to the in-memory space cache for pinned extents causes us to double-add a free range to the space cache. This race exists when free space is cached from the free space tree (space_cache=v2) or the extent tree (nospace_cache, or space_cache=v1 if the cache needs to be regenerated). struct btrfs_block_group::last_byte_to_unpin and struct btrfs_block_group::progress are supposed to protect against this race, but commit d0c2f4fa555e ("btrfs: make concurrent fsyncs wait less when waiting for a transaction commit") subtly broke this by allowing multiple transactions to be unpinning extents at the same time. Specifically, the race is as follows: 1. An extent is deleted from an uncached block group in transaction A. 2. btrfs_commit_transaction() is called for transaction A. 3. btrfs_run_delayed_refs() -> __btrfs_free_extent() runs the delayed ref for the deleted extent. 4. __btrfs_free_extent() -> do_free_extent_accounting() -> add_to_free_space_tree() adds the deleted extent back to the free space tree. 5. do_free_extent_accounting() -> btrfs_update_block_group() -> btrfs_cache_block_group() queues up the block group to get cached. block_group->progress is set to block_group->start. 6. btrfs_commit_transaction() for transaction A calls switch_commit_roots(). It sets block_group->last_byte_to_unpin to block_group->progress, which is block_group->start because the block group hasn't been cached yet. 7. The caching thread gets to our block group. Since the commit roots were already switched, load_free_space_tree() sees the deleted extent as free and adds it to the space cache. It finishes caching and sets block_group->progress to U64_MAX. 8. btrfs_commit_transaction() advances transaction A to TRANS_STATE_SUPER_COMMITTED. 9. fsync calls btrfs_commit_transaction() for transaction B. Since transaction A is already in TRANS_STATE_SUPER_COMMITTED and the commit is for fsync, it advances. 10. btrfs_commit_transaction() for transaction B calls switch_commit_roots(). This time, the block group has already been cached, so it sets block_group->last_byte_to_unpin to U64_MAX. 11. btrfs_commit_transaction() for transaction A calls btrfs_finish_extent_commit(), which calls unpin_extent_range() for the deleted extent. It sees last_byte_to_unpin set to U64_MAX (by transaction B!), so it adds the deleted extent to the space cache again! This explains all of our symptoms above: * If the sequence of events is exactly as described above, when the free space is re-added in step 11, it will fail with EEXIST. * If another thread reallocates the deleted extent in between steps 7 and 11, then step 11 will silently re-add that space to the space cache as free even though it is actually allocated. Then, if that space is allocated *again*, the free space tree will be corrupted (namely, the wrong item will be deleted). * If we don't catch this free space tree corruption, it will continue to get worse as extents are deleted and reallocated. The v1 space_cache is synchronously loaded when an extent is deleted (btrfs_update_block_group() with alloc=0 calls btrfs_cache_block_group() with load_cache_only=1), so it is not normally affected by this bug. However, as noted above, if we fail to load the space cache, we will fall back to caching from the extent tree and may hit this bug. The easiest fix for this race is to also make caching from the free space tree or extent tree synchronous. Josef tested this and found no performance regressions. A few extra changes fall out of this change. Namely, this fix does the following, with step 2 being the crucial fix: 1. Factor btrfs_caching_ctl_wait_done() out of btrfs_wait_block_group_cache_done() to allow waiting on a caching_ctl that we already hold a reference to. 2. Change the call in btrfs_cache_block_group() of btrfs_wait_space_cache_v1_finished() to btrfs_caching_ctl_wait_done(), which makes us wait regardless of the space_cache option. 3. Delete the now unused btrfs_wait_space_cache_v1_finished() and space_cache_v1_done(). 4. Change btrfs_cache_block_group()'s `int load_cache_only` parameter to `bool wait` to more accurately describe its new meaning. 5. Change a few callers which had a separate call to btrfs_wait_block_group_cache_done() to use wait = true instead. 6. Make btrfs_wait_block_group_cache_done() static now that it's not used outside of block-group.c anymore. Fixes: d0c2f4fa555e ("btrfs: make concurrent fsyncs wait less when waiting for a transaction commit") CC: stable@vger.kernel.org # 5.12+ Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Omar Sandoval <osandov@fb.com> Signed-off-by: David Sterba <dsterba@suse.com>
2022-08-23 18:28:13 +00:00
int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait);
void btrfs_put_caching_control(struct btrfs_caching_control *ctl);
struct btrfs_caching_control *btrfs_get_caching_control(
struct btrfs_block_group *cache);
int btrfs_add_new_free_space(struct btrfs_block_group *block_group,
u64 start, u64 end, u64 *total_added_ret);
struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
struct btrfs_fs_info *fs_info,
const u64 chunk_offset);
int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
u64 group_start, struct extent_map *em);
void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info);
void btrfs_mark_bg_unused(struct btrfs_block_group *bg);
void btrfs_reclaim_bgs_work(struct work_struct *work);
void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info);
void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg);
int btrfs_read_block_groups(struct btrfs_fs_info *info);
btrfs: rework chunk allocation to avoid exhaustion of the system chunk array Commit eafa4fd0ad0607 ("btrfs: fix exhaustion of the system chunk array due to concurrent allocations") fixed a problem that resulted in exhausting the system chunk array in the superblock when there are many tasks allocating chunks in parallel. Basically too many tasks enter the first phase of chunk allocation without previous tasks having finished their second phase of allocation, resulting in too many system chunks being allocated. That was originally observed when running the fallocate tests of stress-ng on a PowerPC machine, using a node size of 64K. However that commit also introduced a deadlock where a task in phase 1 of the chunk allocation waited for another task that had allocated a system chunk to finish its phase 2, but that other task was waiting on an extent buffer lock held by the first task, therefore resulting in both tasks not making any progress. That change was later reverted by a patch with the subject "btrfs: fix deadlock with concurrent chunk allocations involving system chunks", since there is no simple and short solution to address it and the deadlock is relatively easy to trigger on zoned filesystems, while the system chunk array exhaustion is not so common. This change reworks the chunk allocation to avoid the system chunk array exhaustion. It accomplishes that by making the first phase of chunk allocation do the updates of the device items in the chunk btree and the insertion of the new chunk item in the chunk btree. This is done while under the protection of the chunk mutex (fs_info->chunk_mutex), in the same critical section that checks for available system space, allocates a new system chunk if needed and reserves system chunk space. This way we do not have chunk space reserved until the second phase completes. The same logic is applied to chunk removal as well, since it keeps reserved system space long after it is done updating the chunk btree. For direct allocation of system chunks, the previous behaviour remains, because otherwise we would deadlock on extent buffers of the chunk btree. Changes to the chunk btree are by large done by chunk allocation and chunk removal, which first reserve chunk system space and then later do changes to the chunk btree. The other remaining cases are uncommon and correspond to adding a device, removing a device and resizing a device. All these other cases do not pre-reserve system space, they modify the chunk btree right away, so they don't hold reserved space for a long period like chunk allocation and chunk removal do. The diff of this change is huge, but more than half of it is just addition of comments describing both how things work regarding chunk allocation and removal, including both the new behavior and the parts of the old behavior that did not change. CC: stable@vger.kernel.org # 5.12+ Tested-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com> Tested-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-29 13:43:06 +00:00
struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
u64 type,
btrfs: rework chunk allocation to avoid exhaustion of the system chunk array Commit eafa4fd0ad0607 ("btrfs: fix exhaustion of the system chunk array due to concurrent allocations") fixed a problem that resulted in exhausting the system chunk array in the superblock when there are many tasks allocating chunks in parallel. Basically too many tasks enter the first phase of chunk allocation without previous tasks having finished their second phase of allocation, resulting in too many system chunks being allocated. That was originally observed when running the fallocate tests of stress-ng on a PowerPC machine, using a node size of 64K. However that commit also introduced a deadlock where a task in phase 1 of the chunk allocation waited for another task that had allocated a system chunk to finish its phase 2, but that other task was waiting on an extent buffer lock held by the first task, therefore resulting in both tasks not making any progress. That change was later reverted by a patch with the subject "btrfs: fix deadlock with concurrent chunk allocations involving system chunks", since there is no simple and short solution to address it and the deadlock is relatively easy to trigger on zoned filesystems, while the system chunk array exhaustion is not so common. This change reworks the chunk allocation to avoid the system chunk array exhaustion. It accomplishes that by making the first phase of chunk allocation do the updates of the device items in the chunk btree and the insertion of the new chunk item in the chunk btree. This is done while under the protection of the chunk mutex (fs_info->chunk_mutex), in the same critical section that checks for available system space, allocates a new system chunk if needed and reserves system chunk space. This way we do not have chunk space reserved until the second phase completes. The same logic is applied to chunk removal as well, since it keeps reserved system space long after it is done updating the chunk btree. For direct allocation of system chunks, the previous behaviour remains, because otherwise we would deadlock on extent buffers of the chunk btree. Changes to the chunk btree are by large done by chunk allocation and chunk removal, which first reserve chunk system space and then later do changes to the chunk btree. The other remaining cases are uncommon and correspond to adding a device, removing a device and resizing a device. All these other cases do not pre-reserve system space, they modify the chunk btree right away, so they don't hold reserved space for a long period like chunk allocation and chunk removal do. The diff of this change is huge, but more than half of it is just addition of comments describing both how things work regarding chunk allocation and removal, including both the new behavior and the parts of the old behavior that did not change. CC: stable@vger.kernel.org # 5.12+ Tested-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com> Tested-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Tested-by: David Sterba <dsterba@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-06-29 13:43:06 +00:00
u64 chunk_offset, u64 size);
void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans);
btrfs: scrub: Don't check free space before marking a block group RO [BUG] When running btrfs/072 with only one online CPU, it has a pretty high chance to fail: btrfs/072 12s ... _check_dmesg: something found in dmesg (see xfstests-dev/results//btrfs/072.dmesg) - output mismatch (see xfstests-dev/results//btrfs/072.out.bad) --- tests/btrfs/072.out 2019-10-22 15:18:14.008965340 +0800 +++ /xfstests-dev/results//btrfs/072.out.bad 2019-11-14 15:56:45.877152240 +0800 @@ -1,2 +1,3 @@ QA output created by 072 Silence is golden +Scrub find errors in "-m dup -d single" test ... And with the following call trace: BTRFS info (device dm-5): scrub: started on devid 1 ------------[ cut here ]------------ BTRFS: Transaction aborted (error -27) WARNING: CPU: 0 PID: 55087 at fs/btrfs/block-group.c:1890 btrfs_create_pending_block_groups+0x3e6/0x470 [btrfs] CPU: 0 PID: 55087 Comm: btrfs Tainted: G W O 5.4.0-rc1-custom+ #13 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:btrfs_create_pending_block_groups+0x3e6/0x470 [btrfs] Call Trace: __btrfs_end_transaction+0xdb/0x310 [btrfs] btrfs_end_transaction+0x10/0x20 [btrfs] btrfs_inc_block_group_ro+0x1c9/0x210 [btrfs] scrub_enumerate_chunks+0x264/0x940 [btrfs] btrfs_scrub_dev+0x45c/0x8f0 [btrfs] btrfs_ioctl+0x31a1/0x3fb0 [btrfs] do_vfs_ioctl+0x636/0xaa0 ksys_ioctl+0x67/0x90 __x64_sys_ioctl+0x43/0x50 do_syscall_64+0x79/0xe0 entry_SYSCALL_64_after_hwframe+0x49/0xbe ---[ end trace 166c865cec7688e7 ]--- [CAUSE] The error number -27 is -EFBIG, returned from the following call chain: btrfs_end_transaction() |- __btrfs_end_transaction() |- btrfs_create_pending_block_groups() |- btrfs_finish_chunk_alloc() |- btrfs_add_system_chunk() This happens because we have used up all space of btrfs_super_block::sys_chunk_array. The root cause is, we have the following bad loop of creating tons of system chunks: 1. The only SYSTEM chunk is being scrubbed It's very common to have only one SYSTEM chunk. 2. New SYSTEM bg will be allocated As btrfs_inc_block_group_ro() will check if we have enough space after marking current bg RO. If not, then allocate a new chunk. 3. New SYSTEM bg is still empty, will be reclaimed During the reclaim, we will mark it RO again. 4. That newly allocated empty SYSTEM bg get scrubbed We go back to step 2, as the bg is already mark RO but still not cleaned up yet. If the cleaner kthread doesn't get executed fast enough (e.g. only one CPU), then we will get more and more empty SYSTEM chunks, using up all the space of btrfs_super_block::sys_chunk_array. [FIX] Since scrub/dev-replace doesn't always need to allocate new extent, especially chunk tree extent, so we don't really need to do chunk pre-allocation. To break above spiral, here we introduce a new parameter to btrfs_inc_block_group(), @do_chunk_alloc, which indicates whether we need extra chunk pre-allocation. For relocation, we pass @do_chunk_alloc=true, while for scrub, we pass @do_chunk_alloc=false. This should keep unnecessary empty chunks from popping up for scrub. Also, since there are two parameters for btrfs_inc_block_group_ro(), add more comment for it. Reviewed-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: Qu Wenruo <wqu@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2019-11-15 02:09:00 +00:00
int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
bool do_chunk_alloc);
void btrfs_dec_block_group_ro(struct btrfs_block_group *cache);
int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans);
int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans);
int btrfs_setup_space_cache(struct btrfs_trans_handle *trans);
int btrfs_update_block_group(struct btrfs_trans_handle *trans,
u64 bytenr, u64 num_bytes, bool alloc);
int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
btrfs: introduce size class to block group allocator The aim of this patch is to reduce the fragmentation of block groups under certain unhappy workloads. It is particularly effective when the size of extents correlates with their lifetime, which is something we have observed causing fragmentation in the fleet at Meta. This patch categorizes extents into size classes: - x < 128KiB: "small" - 128KiB < x < 8MiB: "medium" - x > 8MiB: "large" and as much as possible reduces allocations of extents into block groups that don't match the size class. This takes advantage of any (possible) correlation between size and lifetime and also leaves behind predictable re-usable gaps when extents are freed; small writes don't gum up bigger holes. Size classes are implemented in the following way: - Mark each new block group with a size class of the first allocation that goes into it. - Add two new passes to ffe: "unset size class" and "wrong size class". First, try only matching block groups, then try unset ones, then allow allocation of new ones, and finally allow mismatched block groups. - Filtering is done just by skipping inappropriate ones, there is no special size class indexing. Other solutions I considered were: - A best fit allocator with an rb-tree. This worked well, as small writes didn't leak big holes from large freed extents, but led to regressions in ffe and write performance due to lock contention on the rb-tree with every allocation possibly updating it in parallel. Perhaps something clever could be done to do the updates in the background while being "right enough". - A fixed size "working set". This prevents freeing an extent drastically changing where writes currently land, and seems like a good option too. Doesn't take advantage of size in any way. - The same size class idea, but implemented with xarray marks. This turned out to be slower than looping the linked list and skipping wrong block groups, and is also less flexible since we must have only 3 size classes (max #marks). With the current approach we can have as many as we like. Performance testing was done via: https://github.com/josefbacik/fsperf Of particular relevance are the new fragmentation specific tests. A brief summary of the testing results: - Neutral results on existing tests. There are some minor regressions and improvements here and there, but nothing that truly stands out as notable. - Improvement on new tests where size class and extent lifetime are correlated. Fragmentation in these cases is completely eliminated and write performance is generally a little better. There is also significant improvement where extent sizes are just a bit larger than the size class boundaries. - Regression on one new tests: where the allocations are sized intentionally a hair under the borders of the size classes. Results are neutral on the test that intentionally attacks this new scheme by mixing extent size and lifetime. The full dump of the performance results can be found here: https://bur.io/fsperf/size-class-2022-11-15.txt (there are ANSI escape codes, so best to curl and view in terminal) Here is a snippet from the full results for a new test which mixes buffered writes appending to a long lived set of files and large short lived fallocates: bufferedappendvsfallocate results metric baseline current stdev diff ====================================================================================== avg_commit_ms 31.13 29.20 2.67 -6.22% bg_count 14 15.60 0 11.43% commits 11.10 12.20 0.32 9.91% elapsed 27.30 26.40 2.98 -3.30% end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38% end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07% find_free_extent_calls 116244.30 114354.30 964.56 -1.63% find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67% find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80% find_free_extent_ns_min 500 512 6.67 2.40% find_free_extent_ns_p50 2848 2876 37.65 0.98% find_free_extent_ns_p95 4916 5000 75.45 1.71% find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90% frag_pct_max 61.67 0 8.05 -100.00% frag_pct_mean 43.59 0 6.10 -100.00% frag_pct_min 25.91 0 16.60 -100.00% frag_pct_p50 42.53 0 7.25 -100.00% frag_pct_p95 61.67 0 8.05 -100.00% frag_pct_p99 61.67 0 8.05 -100.00% fragmented_bg_count 6.10 0 1.45 -100.00% max_commit_ms 49.80 46 5.37 -7.63% sys_cpu 2.59 2.62 0.29 1.39% write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23% write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14% write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62% write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90% write_io_kbytes 4194304 4194304 0 0.00% write_iops 2476.15 2556.10 274.29 3.23% write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11% write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22% write_lat_ns_min 17353.10 16250 1646.08 -6.36% There are some mixed improvements/regressions in most metrics along with an elimination of fragmentation in this workload. On the balance, the drastic 1->0 improvement in the happy cases seems worth the mix of regressions and improvements we do observe. Some considerations for future work: - Experimenting with more size classes - More hinting/search ordering work to approximate a best-fit allocator Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2022-12-16 00:06:33 +00:00
u64 ram_bytes, u64 num_bytes, int delalloc,
bool force_wrong_size_class);
void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
u64 num_bytes, int delalloc);
int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
enum btrfs_chunk_alloc_enum force);
int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type);
void check_system_chunk(struct btrfs_trans_handle *trans, const u64 type);
btrfs: fix deadlock between chunk allocation and chunk btree modifications When a task is doing some modification to the chunk btree and it is not in the context of a chunk allocation or a chunk removal, it can deadlock with another task that is currently allocating a new data or metadata chunk. These contexts are the following: * When relocating a system chunk, when we need to COW the extent buffers that belong to the chunk btree; * When adding a new device (ioctl), where we need to add a new device item to the chunk btree; * When removing a device (ioctl), where we need to remove a device item from the chunk btree; * When resizing a device (ioctl), where we need to update a device item in the chunk btree and may need to relocate a system chunk that lies beyond the new device size when shrinking a device. The problem happens due to a sequence of steps like the following: 1) Task A starts a data or metadata chunk allocation and it locks the chunk mutex; 2) Task B is relocating a system chunk, and when it needs to COW an extent buffer of the chunk btree, it has locked both that extent buffer as well as its parent extent buffer; 3) Since there is not enough available system space, either because none of the existing system block groups have enough free space or because the only one with enough free space is in RO mode due to the relocation, task B triggers a new system chunk allocation. It blocks when trying to acquire the chunk mutex, currently held by task A; 4) Task A enters btrfs_chunk_alloc_add_chunk_item(), in order to insert the new chunk item into the chunk btree and update the existing device items there. But in order to do that, it has to lock the extent buffer that task B locked at step 2, or its parent extent buffer, but task B is waiting on the chunk mutex, which is currently locked by task A, therefore resulting in a deadlock. One example report when the deadlock happens with system chunk relocation: INFO: task kworker/u9:5:546 blocked for more than 143 seconds. Not tainted 5.15.0-rc3+ #1 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:kworker/u9:5 state:D stack:25936 pid: 546 ppid: 2 flags:0x00004000 Workqueue: events_unbound btrfs_async_reclaim_metadata_space Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xcd9/0x2530 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 rwsem_down_read_slowpath+0x4ee/0x9d0 kernel/locking/rwsem.c:993 __down_read_common kernel/locking/rwsem.c:1214 [inline] __down_read kernel/locking/rwsem.c:1223 [inline] down_read_nested+0xe6/0x440 kernel/locking/rwsem.c:1590 __btrfs_tree_read_lock+0x31/0x350 fs/btrfs/locking.c:47 btrfs_tree_read_lock fs/btrfs/locking.c:54 [inline] btrfs_read_lock_root_node+0x8a/0x320 fs/btrfs/locking.c:191 btrfs_search_slot_get_root fs/btrfs/ctree.c:1623 [inline] btrfs_search_slot+0x13b4/0x2140 fs/btrfs/ctree.c:1728 btrfs_update_device+0x11f/0x500 fs/btrfs/volumes.c:2794 btrfs_chunk_alloc_add_chunk_item+0x34d/0xea0 fs/btrfs/volumes.c:5504 do_chunk_alloc fs/btrfs/block-group.c:3408 [inline] btrfs_chunk_alloc+0x84d/0xf50 fs/btrfs/block-group.c:3653 flush_space+0x54e/0xd80 fs/btrfs/space-info.c:670 btrfs_async_reclaim_metadata_space+0x396/0xa90 fs/btrfs/space-info.c:953 process_one_work+0x9df/0x16d0 kernel/workqueue.c:2297 worker_thread+0x90/0xed0 kernel/workqueue.c:2444 kthread+0x3e5/0x4d0 kernel/kthread.c:319 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295 INFO: task syz-executor:9107 blocked for more than 143 seconds. Not tainted 5.15.0-rc3+ #1 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. task:syz-executor state:D stack:23200 pid: 9107 ppid: 7792 flags:0x00004004 Call Trace: context_switch kernel/sched/core.c:4940 [inline] __schedule+0xcd9/0x2530 kernel/sched/core.c:6287 schedule+0xd3/0x270 kernel/sched/core.c:6366 schedule_preempt_disabled+0xf/0x20 kernel/sched/core.c:6425 __mutex_lock_common kernel/locking/mutex.c:669 [inline] __mutex_lock+0xc96/0x1680 kernel/locking/mutex.c:729 btrfs_chunk_alloc+0x31a/0xf50 fs/btrfs/block-group.c:3631 find_free_extent_update_loop fs/btrfs/extent-tree.c:3986 [inline] find_free_extent+0x25cb/0x3a30 fs/btrfs/extent-tree.c:4335 btrfs_reserve_extent+0x1f1/0x500 fs/btrfs/extent-tree.c:4415 btrfs_alloc_tree_block+0x203/0x1120 fs/btrfs/extent-tree.c:4813 __btrfs_cow_block+0x412/0x1620 fs/btrfs/ctree.c:415 btrfs_cow_block+0x2f6/0x8c0 fs/btrfs/ctree.c:570 btrfs_search_slot+0x1094/0x2140 fs/btrfs/ctree.c:1768 relocate_tree_block fs/btrfs/relocation.c:2694 [inline] relocate_tree_blocks+0xf73/0x1770 fs/btrfs/relocation.c:2757 relocate_block_group+0x47e/0xc70 fs/btrfs/relocation.c:3673 btrfs_relocate_block_group+0x48a/0xc60 fs/btrfs/relocation.c:4070 btrfs_relocate_chunk+0x96/0x280 fs/btrfs/volumes.c:3181 __btrfs_balance fs/btrfs/volumes.c:3911 [inline] btrfs_balance+0x1f03/0x3cd0 fs/btrfs/volumes.c:4301 btrfs_ioctl_balance+0x61e/0x800 fs/btrfs/ioctl.c:4137 btrfs_ioctl+0x39ea/0x7b70 fs/btrfs/ioctl.c:4949 vfs_ioctl fs/ioctl.c:51 [inline] __do_sys_ioctl fs/ioctl.c:874 [inline] __se_sys_ioctl fs/ioctl.c:860 [inline] __x64_sys_ioctl+0x193/0x200 fs/ioctl.c:860 do_syscall_x64 arch/x86/entry/common.c:50 [inline] do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80 entry_SYSCALL_64_after_hwframe+0x44/0xae So fix this by making sure that whenever we try to modify the chunk btree and we are neither in a chunk allocation context nor in a chunk remove context, we reserve system space before modifying the chunk btree. Reported-by: Hao Sun <sunhao.th@gmail.com> Link: https://lore.kernel.org/linux-btrfs/CACkBjsax51i4mu6C0C3vJqQN3NR_iVuucoeG3U1HXjrgzn5FFQ@mail.gmail.com/ Fixes: 79bd37120b1495 ("btrfs: rework chunk allocation to avoid exhaustion of the system chunk array") CC: stable@vger.kernel.org # 5.14+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-10-13 09:12:49 +00:00
void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
bool is_item_insertion);
u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags);
void btrfs_put_block_group_cache(struct btrfs_fs_info *info);
int btrfs_free_block_groups(struct btrfs_fs_info *info);
int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
u64 physical, u64 **logical, int *naddrs, int *stripe_len);
static inline u64 btrfs_data_alloc_profile(struct btrfs_fs_info *fs_info)
{
return btrfs_get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_DATA);
}
static inline u64 btrfs_metadata_alloc_profile(struct btrfs_fs_info *fs_info)
{
return btrfs_get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_METADATA);
}
static inline u64 btrfs_system_alloc_profile(struct btrfs_fs_info *fs_info)
{
return btrfs_get_alloc_profile(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
}
static inline int btrfs_block_group_done(struct btrfs_block_group *cache)
{
smp_mb();
return cache->cached == BTRFS_CACHE_FINISHED ||
cache->cached == BTRFS_CACHE_ERROR;
}
void btrfs_freeze_block_group(struct btrfs_block_group *cache);
void btrfs_unfreeze_block_group(struct btrfs_block_group *cache);
btrfs: fix race between writes to swap files and scrub When we active a swap file, at btrfs_swap_activate(), we acquire the exclusive operation lock to prevent the physical location of the swap file extents to be changed by operations such as balance and device replace/resize/remove. We also call there can_nocow_extent() which, among other things, checks if the block group of a swap file extent is currently RO, and if it is we can not use the extent, since a write into it would result in COWing the extent. However we have no protection against a scrub operation running after we activate the swap file, which can result in the swap file extents to be COWed while the scrub is running and operating on the respective block group, because scrub turns a block group into RO before it processes it and then back again to RW mode after processing it. That means an attempt to write into a swap file extent while scrub is processing the respective block group, will result in COWing the extent, changing its physical location on disk. Fix this by making sure that block groups that have extents that are used by active swap files can not be turned into RO mode, therefore making it not possible for a scrub to turn them into RO mode. When a scrub finds a block group that can not be turned to RO due to the existence of extents used by swap files, it proceeds to the next block group and logs a warning message that mentions the block group was skipped due to active swap files - this is the same approach we currently use for balance. Fixes: ed46ff3d42378 ("Btrfs: support swap files") CC: stable@vger.kernel.org # 5.4+ Reviewed-by: Anand Jain <anand.jain@oracle.com> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
2021-02-05 12:55:37 +00:00
bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg);
void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount);
btrfs: introduce size class to block group allocator The aim of this patch is to reduce the fragmentation of block groups under certain unhappy workloads. It is particularly effective when the size of extents correlates with their lifetime, which is something we have observed causing fragmentation in the fleet at Meta. This patch categorizes extents into size classes: - x < 128KiB: "small" - 128KiB < x < 8MiB: "medium" - x > 8MiB: "large" and as much as possible reduces allocations of extents into block groups that don't match the size class. This takes advantage of any (possible) correlation between size and lifetime and also leaves behind predictable re-usable gaps when extents are freed; small writes don't gum up bigger holes. Size classes are implemented in the following way: - Mark each new block group with a size class of the first allocation that goes into it. - Add two new passes to ffe: "unset size class" and "wrong size class". First, try only matching block groups, then try unset ones, then allow allocation of new ones, and finally allow mismatched block groups. - Filtering is done just by skipping inappropriate ones, there is no special size class indexing. Other solutions I considered were: - A best fit allocator with an rb-tree. This worked well, as small writes didn't leak big holes from large freed extents, but led to regressions in ffe and write performance due to lock contention on the rb-tree with every allocation possibly updating it in parallel. Perhaps something clever could be done to do the updates in the background while being "right enough". - A fixed size "working set". This prevents freeing an extent drastically changing where writes currently land, and seems like a good option too. Doesn't take advantage of size in any way. - The same size class idea, but implemented with xarray marks. This turned out to be slower than looping the linked list and skipping wrong block groups, and is also less flexible since we must have only 3 size classes (max #marks). With the current approach we can have as many as we like. Performance testing was done via: https://github.com/josefbacik/fsperf Of particular relevance are the new fragmentation specific tests. A brief summary of the testing results: - Neutral results on existing tests. There are some minor regressions and improvements here and there, but nothing that truly stands out as notable. - Improvement on new tests where size class and extent lifetime are correlated. Fragmentation in these cases is completely eliminated and write performance is generally a little better. There is also significant improvement where extent sizes are just a bit larger than the size class boundaries. - Regression on one new tests: where the allocations are sized intentionally a hair under the borders of the size classes. Results are neutral on the test that intentionally attacks this new scheme by mixing extent size and lifetime. The full dump of the performance results can be found here: https://bur.io/fsperf/size-class-2022-11-15.txt (there are ANSI escape codes, so best to curl and view in terminal) Here is a snippet from the full results for a new test which mixes buffered writes appending to a long lived set of files and large short lived fallocates: bufferedappendvsfallocate results metric baseline current stdev diff ====================================================================================== avg_commit_ms 31.13 29.20 2.67 -6.22% bg_count 14 15.60 0 11.43% commits 11.10 12.20 0.32 9.91% elapsed 27.30 26.40 2.98 -3.30% end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38% end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07% find_free_extent_calls 116244.30 114354.30 964.56 -1.63% find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67% find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80% find_free_extent_ns_min 500 512 6.67 2.40% find_free_extent_ns_p50 2848 2876 37.65 0.98% find_free_extent_ns_p95 4916 5000 75.45 1.71% find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90% frag_pct_max 61.67 0 8.05 -100.00% frag_pct_mean 43.59 0 6.10 -100.00% frag_pct_min 25.91 0 16.60 -100.00% frag_pct_p50 42.53 0 7.25 -100.00% frag_pct_p95 61.67 0 8.05 -100.00% frag_pct_p99 61.67 0 8.05 -100.00% fragmented_bg_count 6.10 0 1.45 -100.00% max_commit_ms 49.80 46 5.37 -7.63% sys_cpu 2.59 2.62 0.29 1.39% write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23% write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14% write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62% write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90% write_io_kbytes 4194304 4194304 0 0.00% write_iops 2476.15 2556.10 274.29 3.23% write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11% write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22% write_lat_ns_min 17353.10 16250 1646.08 -6.36% There are some mixed improvements/regressions in most metrics along with an elimination of fragmentation in this workload. On the balance, the drastic 1->0 improvement in the happy cases seems worth the mix of regressions and improvements we do observe. Some considerations for future work: - Experimenting with more size classes - More hinting/search ordering work to approximate a best-fit allocator Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2022-12-16 00:06:33 +00:00
enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size);
int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
enum btrfs_block_group_size_class size_class,
bool force_wrong_size_class);
bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg);
btrfs: introduce size class to block group allocator The aim of this patch is to reduce the fragmentation of block groups under certain unhappy workloads. It is particularly effective when the size of extents correlates with their lifetime, which is something we have observed causing fragmentation in the fleet at Meta. This patch categorizes extents into size classes: - x < 128KiB: "small" - 128KiB < x < 8MiB: "medium" - x > 8MiB: "large" and as much as possible reduces allocations of extents into block groups that don't match the size class. This takes advantage of any (possible) correlation between size and lifetime and also leaves behind predictable re-usable gaps when extents are freed; small writes don't gum up bigger holes. Size classes are implemented in the following way: - Mark each new block group with a size class of the first allocation that goes into it. - Add two new passes to ffe: "unset size class" and "wrong size class". First, try only matching block groups, then try unset ones, then allow allocation of new ones, and finally allow mismatched block groups. - Filtering is done just by skipping inappropriate ones, there is no special size class indexing. Other solutions I considered were: - A best fit allocator with an rb-tree. This worked well, as small writes didn't leak big holes from large freed extents, but led to regressions in ffe and write performance due to lock contention on the rb-tree with every allocation possibly updating it in parallel. Perhaps something clever could be done to do the updates in the background while being "right enough". - A fixed size "working set". This prevents freeing an extent drastically changing where writes currently land, and seems like a good option too. Doesn't take advantage of size in any way. - The same size class idea, but implemented with xarray marks. This turned out to be slower than looping the linked list and skipping wrong block groups, and is also less flexible since we must have only 3 size classes (max #marks). With the current approach we can have as many as we like. Performance testing was done via: https://github.com/josefbacik/fsperf Of particular relevance are the new fragmentation specific tests. A brief summary of the testing results: - Neutral results on existing tests. There are some minor regressions and improvements here and there, but nothing that truly stands out as notable. - Improvement on new tests where size class and extent lifetime are correlated. Fragmentation in these cases is completely eliminated and write performance is generally a little better. There is also significant improvement where extent sizes are just a bit larger than the size class boundaries. - Regression on one new tests: where the allocations are sized intentionally a hair under the borders of the size classes. Results are neutral on the test that intentionally attacks this new scheme by mixing extent size and lifetime. The full dump of the performance results can be found here: https://bur.io/fsperf/size-class-2022-11-15.txt (there are ANSI escape codes, so best to curl and view in terminal) Here is a snippet from the full results for a new test which mixes buffered writes appending to a long lived set of files and large short lived fallocates: bufferedappendvsfallocate results metric baseline current stdev diff ====================================================================================== avg_commit_ms 31.13 29.20 2.67 -6.22% bg_count 14 15.60 0 11.43% commits 11.10 12.20 0.32 9.91% elapsed 27.30 26.40 2.98 -3.30% end_state_mount_ns 11122551.90 10635118.90 851143.04 -4.38% end_state_umount_ns 1.36e+09 1.35e+09 12248056.65 -1.07% find_free_extent_calls 116244.30 114354.30 964.56 -1.63% find_free_extent_ns_max 599507.20 1047168.20 103337.08 74.67% find_free_extent_ns_mean 3607.19 3672.11 101.20 1.80% find_free_extent_ns_min 500 512 6.67 2.40% find_free_extent_ns_p50 2848 2876 37.65 0.98% find_free_extent_ns_p95 4916 5000 75.45 1.71% find_free_extent_ns_p99 20734.49 20920.48 1670.93 0.90% frag_pct_max 61.67 0 8.05 -100.00% frag_pct_mean 43.59 0 6.10 -100.00% frag_pct_min 25.91 0 16.60 -100.00% frag_pct_p50 42.53 0 7.25 -100.00% frag_pct_p95 61.67 0 8.05 -100.00% frag_pct_p99 61.67 0 8.05 -100.00% fragmented_bg_count 6.10 0 1.45 -100.00% max_commit_ms 49.80 46 5.37 -7.63% sys_cpu 2.59 2.62 0.29 1.39% write_bw_bytes 1.62e+08 1.68e+08 17975843.50 3.23% write_clat_ns_mean 57426.39 54475.95 2292.72 -5.14% write_clat_ns_p50 46950.40 42905.60 2101.35 -8.62% write_clat_ns_p99 148070.40 143769.60 2115.17 -2.90% write_io_kbytes 4194304 4194304 0 0.00% write_iops 2476.15 2556.10 274.29 3.23% write_lat_ns_max 2101667.60 2251129.50 370556.59 7.11% write_lat_ns_mean 59374.91 55682.00 2523.09 -6.22% write_lat_ns_min 17353.10 16250 1646.08 -6.36% There are some mixed improvements/regressions in most metrics along with an elimination of fragmentation in this workload. On the balance, the drastic 1->0 improvement in the happy cases seems worth the mix of regressions and improvements we do observe. Some considerations for future work: - Experimenting with more size classes - More hinting/search ordering work to approximate a best-fit allocator Signed-off-by: Boris Burkov <boris@bur.io> Signed-off-by: David Sterba <dsterba@suse.com>
2022-12-16 00:06:33 +00:00
#endif /* BTRFS_BLOCK_GROUP_H */