[PATCH 0/2] mm/zswap: distinguish incompressible pages from compressor failures

[Suseendran Bhaskaran <suseendranb@google.com>]

Commit dca4437a5861 ("mm/zswap: store <PAGE_SIZE compression failed page
as-is") and commit 0b1bf60c324a ("mm/zswap: reduce the size of the
compression buffer to a single page") updated zswap to pass a
single-page (PAGE_SIZE) destination buffer to crypto_acomp_compress()
(so incompressible pages typically overflow the PAGE_SIZE buffer due to
compressor metadata overhead). Because the single-buffer software
compressors return -EINVAL on destination buffer overflow and
zswap_compress() checks 'if (comp_ret || !dlen || dlen >= PAGE_SIZE)',
destination buffer overflow (incompressible page) is conflated with
unexpected compressor failures, causing two user-visible bugs:

1. Broken debugfs telemetry when writeback is disabled
   (memory.zswap.writeback = 0):
   Single-buffer software compressors (zstd_compress_one(), lz4, lz4hc,
   lzo, and lzo-rle) return -EINVAL on destination buffer overflow
   (whereas the multi-segment streaming path in zstd_compress(),
   deflate, and 842 return -ENOSPC). Furthermore, in zswap_compress(),
   'comp_ret = comp_ret ? comp_ret : -EINVAL' forces comp_ret to -EINVAL
   even when comp_ret == 0 and dlen == PAGE_SIZE. As a result,
   incompressible pages are misaccounted in
   /sys/kernel/debug/zswap/reject_compress_fail instead of
   reject_compress_poor, leaving reject_compress_poor at 0 and causing
   false-positive compressor failure alerts in monitoring.

2. Silent masking of real compressor failures when writeback is enabled
   (memory.zswap.writeback = 1):
   Because zswap_compress() checks 'if (comp_ret || !dlen || dlen >=
   PAGE_SIZE)' and unconditionally clears comp_ret = 0 to store the raw
   page in zsmalloc, genuine compressor errors (such as -ENOMEM,
   -EINVAL from an invalid context, or hardware accelerator errors) are
   silently masked as incompressible pages. Users see 0
   reject_compress_fail events, and pages that encounter an unexpected
   compressor error are miscounted under stored_incompressible_pages.

Patch 1 updates zstd_compress_one(), lz4, lz4hc, lzo, and lzo-rle to
return -ENOSPC on destination buffer overflow, matching the streaming
path in zstd_compress(), deflate, and 842.

Patch 2 updates zswap_compress() to treat only comp_ret == -ENOSPC or
(!comp_ret && dlen >= PAGE_SIZE) as an incompressible page (storing it
uncompressed in zswap when writeback is enabled to preserve LRU order,
or rejecting it with comp_ret = -ENOSPC when writeback is disabled so
zswap_reject_compress_poor is incremented), while jumping to unlock on
unexpected compressor errors (comp_ret || !dlen) so
zswap_reject_compress_fail is properly incremented.

Suseendran Bhaskaran (2):
  crypto: return -ENOSPC on compression destination overflow
  mm/zswap: classify compression dst overflow as poor compression

 crypto/lz4.c     | 2 +-
 crypto/lz4hc.c   | 2 +-
 crypto/lzo-rle.c | 2 +-
 crypto/lzo.c     | 2 +-
 crypto/zstd.c    | 5 ++++-
 mm/zswap.c       | 7 +++++--
 6 files changed, 13 insertions(+), 7 deletions(-)

-- 
2.56.0.218.g1ec904b173-goog