Skip to content

btrfs: zoned: fix active zone tracking issues

Naohiro Aota linux-block linux-btrfs 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest13 patches 9 repliesIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/13block: add bdev_max_segments() helperinclude/linux/blkdev.h+5 -02 replies
2/13btrfs: zoned: revive max_zone_append_bytesfs/btrfs/ctree.h fs/btrfs/zoned.c fs/btrfs/zoned.h+20 -01 reply
3/13btrfs: replace BTRFS_MAX_EXTENT_SIZE with fs_info->max_extent_sizefs/btrfs/ctree.h fs/btrfs/disk-io.c fs/btrfs/extent_io.c +2 more+18 -32 replies
4/13btrfs: convert count_max_extents() to use fs_info->max_extent_sizefs/btrfs/ctree.h fs/btrfs/delalloc-space.c fs/btrfs/inode.c+24 -191 reply
5/13btrfs: use fs_info->max_extent_size in get_extent_max_capacity()fs/btrfs/ioctl.c+7 -40 replies
6/13btrfs: let can_allocate_chunk return intfs/btrfs/extent-tree.c+8 -70 replies
7/13btrfs: zoned: finish least available block group on data BG allocationfs/btrfs/extent-tree.c fs/btrfs/zoned.c fs/btrfs/zoned.h+86 -101 reply
8/13btrfs: zoned: introduce space_info->active_total_bytesfs/btrfs/block-group.c fs/btrfs/space-info.c fs/btrfs/space-info.h +1 more+50 -130 replies
9/13btrfs: zoned: disable metadata overcommit for zonedfs/btrfs/space-info.c+4 -10 replies
10/13btrfs: zoned: activate metadata BG on flush_spacefs/btrfs/space-info.c fs/btrfs/zoned.c fs/btrfs/zoned.h+93 -00 replies
11/13btrfs: zoned: activate necessary block groupfs/btrfs/block-group.c+16 -00 replies
12/13btrfs: zoned: write out partially allocated regionfs/btrfs/extent-tree.c fs/btrfs/inode.c+59 -140 replies
13/13btrfs: zoned: wait until zone is finished when allocation didn't progressfs/btrfs/ctree.h fs/btrfs/disk-io.c fs/btrfs/inode.c +1 more+18 -20 replies

Reviews trailers per patch

7 of 13 patches have a Reviewed-by or Acked-by.

PatchJohannes ThumshirnJens AxboeChristoph Hellwig
1/13R-b R-b R-b
2/13R-b
3/13R-b
4/13
5/13R-b
6/13R-b
7/13R-b
8/13
9/13R-b
10/13
11/13
12/13
13/13

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes. Scroll the table sideways for all 3 reviewers.

Read the v1 thread

Cover letter [PATCH 00/13] btrfs: zoned: fix active zone tracking issues

This series addresses mainly two issues on zoned btrfs' active zone tracking and one issue which is a dependency of the main issue.

* ChangeLog
- v2
- Support sanity tests (Johannes)
- fs_info can be NULL while it is running sanity tests. Consider that case in CONFIG_FS_BTRFS_RUN_SANITY_TESTS.
- Propagete an error of btrfs_zone_finish() (Johannes)
- Add a comment to max_segments limitation (Christoph)
- Rename btrfs_finish_one_bg() to btrfs_zone_finish_one_bg() to make the it clear it is related to zoned code.
- Do not reduce active_total_bytes when finishing a block group.
- While it's no longer active, but it still can have "used" bytes. So, it should be counted to host "total_bytes". Or, it breaks free space calculation.
- Do not try to activate a fully allocated block group.

* Background

A ZNS drive has an upper limit of zones that simultaneously can be written out. We call the limit max_active_zones. An active zone is deactivated when we write fully to the zone, or when we explicitly send a REQ_OP_ZONE_FINISH command to make it full.

The zoned btrfs must be aware of max_active_zones to use a ZNS drive. So, we have an active zone tracking system that considers a block group as active iff the underlying zone is active. In fact, we consider a block group (and its underlying zones) as active when we start allocating from it. Then, when the last region which can be allocated in the block group is written, we send a REQ_OP_ZONE_FINISH command to each zone and consider the block group as inactive.

So, in short, we currently depend on writing fully to a zone to finish a block group.

* Issues
** Issue A

In a certain situation, the current zoned btrfs's extent allocation fails with an early -ENOSPC on a ZNS drive. When all the block groups do not have enough space left for the allocation, it tries to allocate a new block group if we can activate a new zone. If not, it returns -ENOSPC while the device still has free space left.

** Issue B

When doing a buffered write, we call cow_file_range() to allocate the data extent. The cow_file_range() works like an all-or-nothing manner: if it can allocate for all the range it returns 0, or -ENOSPC if not. Thus, when all the block group have small free space left, and btrfs cannot finish any block group, the allocation partly succeed but fails in the end. This also results in an early -ENOSPC.

We cannot finish any block group in a certain situation. Let's consider that we have 8 active data block groups (forget about metadata/system block groups here) and each of them has 1 MB free space left. Now, we want to do 10 MB buffered write. We can allocate blocks for the 8 of 10 MB. And, we can no longer allocate from any block group. Furthermore, we cannot finish any block group, because all the block groups have 1 MB reserved unwritten space left now. And, since this 1 MB regions are owned by the allocating process itself, simply waiting for the region to be written won't work.

** Issue C

To address issue A, we needed to disable metadata reservation over-commit. That reveals that we under-estimate the number of extents to be written on zoned btrfs. On zoned btrfs, we use a ZONE APPEND command to write data, whose bio size is limited by max_zone_append_sectors and max_segments. So, a data extent is always split at most at the size of the limit. As a result, if BTRFS_MAX_EXTENT_SIZE is larger than the limit, we tend to have more extents than expected from the estimation using BTRFS_MAX_EXTENT_SIZE.

Since the metadata reservation is done before allocation (e.g, at btrfs_buffered_write) and released afterward along with the delalloc process or ordered extent creation. As a result, we can be short of the metadata reservation in a certain situation, and can cause a WARN by that.

* Solutions
** For issue A

Issue A is that we can have early -ENOSPC if we cannot activate another block group and no block group has enough space left.

To avoid the early -ENOSPC, we need to choose one block group and finish it to make rooms for a new block group to be activated. But, that is only possible from the data extent allocation context. From the metadata context, we can cause a deadlock because we might need to wait for a running transaction to make the finishing block group read-only.

So, we use two different methods for data allocation and metadata allocation. For data allocation, we can finish a block group on-demand from btrfs_reserve_extent() context. The finishing block group will be the block group with a least free space left.

For metadata allocation, we use flush_space() to ensure that reserved bytes can be written into active block groups. To do so, we track active block groups' total bytes as active_total_bytes, and activate a block group on-demand from flush_space().

Also, a newly allocated block group from some contexts must be activated

** For issue B

Issue B is about when we cannot allocate space from any block group, and we cannot finish any block group. This issue only occurs when allocating a data extent, because metadata reservation is ensured to be contained in active block groups by solution for issue A.

In this case, writing out the partially allocated region will close the gap between the allocation pointer and the capacity of the block group, make the zone finished, and opens up rooms to activate a new block group. So, this series implements the partial writing out and retrying of the alloction.

In a certain case, we can't allocate anything from the block groups. In that case, we'd expect there is on-going IOs to finish a block group. So, we wait for it and retry the allocation.

** For issue C

Issue C is about that we underestimate the number of extents to be written on zoned btrfs, because we don't expect an ordered extent is split by the size of a bio.

We need to use a proper extent size limit to fix issue C. For that, we revive the fs_info->max_zone_append_size and use it to calculate count_max_extents(). Technically, the bio size is also limited by the max_segments, so the limit is also capped by it.

* Patch structure

The fix for issue C comes first because it is a dependency of the fixes for issue A and B.

Patches 1 to 5 address issue C by reviving fs_info->max_zone_append_bytes and use it to replace BTRFS_MAX_EXTENT_SIZE on zoned btrfs.

Patches 6 to 11 address issue A. In detail, patch 7 fixes the data allocation by finishing a block group when we cannot activate another block group. Patch 10 fixes the metadata allocation by finishing a block group at space reservation time.

Patches 12 and 13 address issue B by writing out a successfully allocated part first and retrying the rest allocation.

Naohiro Aota (13):
block: add bdev_max_segments() helper
btrfs: zoned: revive max_zone_append_bytes
btrfs: replace BTRFS_MAX_EXTENT_SIZE with fs_info->max_extent_size
btrfs: convert count_max_extents() to use fs_info->max_extent_size
btrfs: use fs_info->max_extent_size in get_extent_max_capacity()
btrfs: let can_allocate_chunk return int
btrfs: zoned: finish least available block group on data BG allocation
btrfs: zoned: introduce space_info->active_total_bytes
btrfs: zoned: disable metadata overcommit for zoned
btrfs: zoned: activate metadata BG on flush_space
btrfs: zoned: activate necessary block group
btrfs: zoned: write out partially allocated region
btrfs: zoned: wait until zone is finished when allocation didn't
progress

 fs/btrfs/block-group.c    |  28 ++++++++-
 fs/btrfs/ctree.h          |  30 ++++++---
 fs/btrfs/delalloc-space.c |   6 +-
 fs/btrfs/disk-io.c        |   3 +
 fs/btrfs/extent-tree.c    |  70 ++++++++++++++++-----
 fs/btrfs/extent_io.c      |   8 ++-
 fs/btrfs/inode.c          |  90 +++++++++++++++++++--------
 fs/btrfs/ioctl.c          |  11 ++--
 fs/btrfs/space-info.c     |  76 ++++++++++++++++++++---
 fs/btrfs/space-info.h     |   4 +-
 fs/btrfs/zoned.c          | 124 ++++++++++++++++++++++++++++++++++++++
 fs/btrfs/zoned.h          |  18 ++++++
 include/linux/blkdev.h    |   5 ++
 13 files changed, 404 insertions(+), 69 deletions(-)
Signature · 1 line
2.35.1