Re: [PATCH v2 0/4] btrfs: prepare compression for bs > ps support

[Qu Wenruo <wqu@suse.com>]

在 2025/9/10 14:48, Qu Wenruo 写道:
> [CHANGELOG]
> v2:
> - Fix a missing callsite inside btrfs_compress_file_range() which only
>    zeros the range inside the first page
>    The folio_zero_range() of the last compressed folio should cover the
>    full folio, not only the first page.

And there is a missing call site in btrfs_decompress() which uses 
ASSERT() to check against PAGE_SIZE, not folio size and can crash 
btrfs/056 during tests.

(Yep, I'm already testing bs > ps with fstests now, and it can reach 
btrfs/056 except a weird crash in btrfs/004 that I'm still debugging)

Will update the series when no more compression bugs exposed by default 
fstests runs.

Thanks,
Qu

> 
> This is the compression part support for bs > ps cases.
> 
> The main trick involved is the handling of compr folios, the main
> changes are:
> 
> - Compressed folios now need to follow the minimal order
>    This is the requirement for the recently added btrfs_for_each_block*()
>    helpers, and this keeps our code from handling sub-block sized ranges.
> 
> - No cached compression folios for bs > ps cases
>    Those folios are large and are not sharable between other fses, and
>    most of btrfs will use 4K (until storage with 16K block size got
>    popular).
> 
> - Extra rejection of HIGHMEM systems with bs > ps support
>    Unfortunately HIGHMEM large folios need us to map them page by page,
>    this breaks our principle of no sub-block handling.
> 
>    Considering HIGHMEM is always a pain in the backend and is already
>    planned for deprecation, it's best for everyone to just reject bs > ps
>    btrfses on HIGHMEM systems.
> 
> Please still keep in mind that, raid56, scrub, encoded write are not yet
> supporting bs > ps cases.
> 
> For now I have only done basic read/write/balance/offline data check
> tests on bs > ps cases with all 4 compression algorithms (none, lzo, zlib,
> zstd), so far so good.
> 
> If some one wants to play with the incomplete bs > ps cases, the
> following simple diff will enable the work:
> 
>   --- a/fs/btrfs/fs.c
>   +++ b/fs/btrfs/fs.c
>   @@ -96,8 +96,7 @@ bool __attribute_const__ btrfs_supported_blocksize(u32 blocksize)
>            */
>           if (IS_ENABLED(CONFIG_HIGHMEM) && blocksize > PAGE_SIZE)
>                   return false;
>   -       if (blocksize <= PAGE_SIZE)
>   -               return true;
>   +       return true;
>    #endif
>           return false;
>    }
> 
> The remaining features and their road maps are:
> 
> - Encoded writes
>    This should be the most simple part.
> 
> - RAID56
>    Needs to convert the page usage into folio one first.
> 
> - Scrub
>    This relies on some RAID56 interfaces for parity handling.
>    Otherwise pretty like RAID56, we need to convert the page usage to
>    folios first.
> 
> Qu Wenruo (4):
>    btrfs: prepare compression folio alloc/free for bs > ps cases
>    btrfs: prepare zstd to support bs > ps cases
>    btrfs: prepare lzo to support bs > ps cases
>    btrfs: prepare zlib to support bs > ps cases
> 
>   fs/btrfs/compression.c | 38 +++++++++++++++++++-------
>   fs/btrfs/compression.h |  2 +-
>   fs/btrfs/extent_io.c   |  7 +++--
>   fs/btrfs/extent_io.h   |  3 ++-
>   fs/btrfs/fs.c          | 17 ++++++++++++
>   fs/btrfs/fs.h          |  6 +++++
>   fs/btrfs/inode.c       | 16 ++++++-----
>   fs/btrfs/lzo.c         | 59 ++++++++++++++++++++++-------------------
>   fs/btrfs/zlib.c        | 60 +++++++++++++++++++++++++++---------------
>   fs/btrfs/zstd.c        | 44 +++++++++++++++++--------------
>   10 files changed, 163 insertions(+), 89 deletions(-)
>