[PATCH AUTOSEL 7.0-6.12] smb: client: compress: fix bad encoding on last LZ77 flag
From: Sasha Levin
Date: Tue Apr 28 2026 - 06:47:30 EST
From: Enzo Matsumiya <ematsumiya@xxxxxxx>
[ Upstream commit a13e942a03feea211c67a97bc6a57f82aa56e4b6 ]
End-of-stream flag could lead to UB because of int promotion
(overwriting signed bit).
Fix it by changing operand from '1' to '1UL'.
Signed-off-by: Enzo Matsumiya <ematsumiya@xxxxxxx>
Signed-off-by: Steve French <stfrench@xxxxxxxxxxxxx>
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
---
LLM Generated explanations, may be completely bogus:
Based on my investigation, here's the complete analysis:
## Phase 1: Commit Message Forensics
**Step 1.1 - Subject line parsing:**
Record: `smb: client: compress:` prefix, action verb is "fix", one-line
summary: fix bad encoding on last LZ77 end-of-stream flag due to integer
promotion UB.
**Step 1.2 - Tags:**
Record:
- Signed-off-by: Enzo Matsumiya (author, SUSE)
- Signed-off-by: Steve French (SMB client maintainer)
- No Fixes:, Reported-by:, Tested-by:, Reviewed-by:, Link:, or Cc:
stable tags
**Step 1.3 - Commit body:**
Record: Bug described as "End-of-stream flag could lead to UB because of
int promotion (overwriting signed bit)". Fix approach: change operand
from `1` (int) to `1UL` (unsigned long). No crash/stack trace mentioned.
No specific kernel version mentioned.
**Step 1.4 - Hidden bug fix detection:**
Record: Not hidden - explicitly labeled "fix". Author is the original
creator of the LZ77 code, so has clear understanding of intent.
## Phase 2: Diff Analysis
**Step 2.1 - Inventory:**
Record: Single file `fs/smb/client/compress/lz77.c`, 1 line changed
(+1/-1). Function: `lz77_compress()`. Classification: single-file
surgical fix.
**Step 2.2 - Code flow:**
Record: In end-of-stream flag handling:
- Before: `flag |= (1 << (32 - flag_count)) - 1;` — `1` is `int` (signed
32-bit)
- After: `flag |= (1UL << (32 - flag_count)) - 1;` — `1UL` is `unsigned
long` (64-bit on 64-bit arches)
**Step 2.3 - Bug mechanism:**
Record: Category (g) Logic / correctness / UB fix. When `flag_count ==
0`, `1 << 32` is shift >= width of `int` → UB (on x86 GCC yields `1`,
then `(1)-1=0`, producing wrong flag byte). When `flag_count == 1`, `1
<< 31` shifts into sign bit of signed int → UB (typically wraps to give
correct result by accident). `1UL` fits these shifts within a 64-bit
unsigned long without UB.
**Step 2.4 - Fix quality:**
Record: Obviously correct. Minimal (1-line). Zero regression risk (on
64-bit, `1UL` is 64-bit, making shift by 32 well-defined; on 32-bit, fix
is equivalent but `1UL << 32` would still be UB — however, `flag` is
`long` which is 32-bit on 32-bit, so the truncation via `lz77_write32()`
ends up correct in practice).
## Phase 3: Git History Investigation
**Step 3.1 - Blame:**
Record: `git log v6.12:fs/smb/client/compress/lz77.c` confirms `flag |=
(1 << (32 - flag_count)) - 1;` exists at v6.12. Bug was introduced in
commit `d14bbfff259ca` ("smb3: mark compression as
CONFIG_EXPERIMENTAL...", July 2024, v6.12).
**Step 3.2 - Fixes: tag:**
Record: No Fixes: tag. Buggy code was introduced in `d14bbfff259ca` per
blame. That commit is present in v6.12 and all later tags (v6.12 through
v6.19).
**Step 3.3 - Related changes:**
Record: Same patch series includes:
- `4c221711b2374` fix buffer overrun in lz77_compress()
- `20d4f9efe008b` fix counting in LZ77 match finding
- `fca46b0e68c5d` increase LZ77_MATCH_MAX_DIST
- `4460e9c68d1a8` LZ77 optimizations
- `71179a5ee916d` add code docs
- `94ae8c3fee94a` cleanup (in v6.12 already)
**Step 3.4 - Author:**
Record: Enzo Matsumiya is the author of the original SMB compression
code and the LZ77 implementation — highest domain authority.
**Step 3.5 - Dependencies:**
Record: Standalone fix. Does not depend on the other patches in the
series (patch 2/8 of series, but only touches 1 line that has been
unchanged since v6.12).
## Phase 4: Mailing List Research
**Step 4.1 - b4 dig:**
Record: `b4 dig -c a13e942a03fee` →
https://lore.kernel.org/all/20260413190713.283939-2-ematsumiya@xxxxxxx/
Patch is 2/8 in series "[PATCH 0-8]" of SMB compression fixes.
**Step 4.1 continued - b4 dig -a:**
Record: Only v1 exists. Applied version is the only version sent.
**Step 4.2 - Reviewers:**
Record: Recipients included smfrench@xxxxxxxxx (maintainer),
pc@xxxxxxxxxxxxx, ronniesahlberg@xxxxxxxxx, sprasad@xxxxxxxxxxxxx,
tom@xxxxxxxxxx, bharathsm@xxxxxxxxxxxxx, henrique.carvalho@xxxxxxxx —
appropriate SMB reviewers.
**Step 4.2 continued - reviewer response:**
Record: Steve French replied "merged into cifs-2.6.git for-next"
confirming acceptance by maintainer. No NAKs or concerns. No explicit
stable nomination.
**Step 4.3 - Bug report:**
Record: No Reported-by or Link tag. Appears to be developer-found
(likely via code inspection or UBSAN).
**Step 4.4 - Series context:**
Record: Part of larger SMB client fixes series. Other series members
(`4c221711b2374` buffer overrun, `20d4f9efe008b` counting fix) are clear
bug fixes in the same code.
**Step 4.5 - Stable ML:**
Record: No explicit stable discussion found.
## Phase 5: Code Semantic Analysis
**Step 5.1 - Modified functions:**
Record: `lz77_compress()` in `fs/smb/client/compress/lz77.c`.
**Step 5.2 - Callers:**
Record: Called from `smb_compress()` in `fs/smb/client/compress.c`,
which is called from the SMB request path only when CIFS_COMPRESSION is
built and user mounts with compression negotiated.
**Step 5.3 - Callees:** Record: Uses `lz77_write32()` → writes little-
endian 32-bit to `flag_pos`.
**Step 5.4 - Reachability:**
Record: Only reachable when: (1) `CONFIG_CIFS_COMPRESSION=y`
(experimental, default N), (2) user mounts with `compress` option, (3)
server negotiates SMB3.1.1 compression, (4) request eligible for
compression (>= PAGE_SIZE write). Not reachable by default kernel
builds.
**Step 5.5 - Similar patterns:**
Record: Only one occurrence of the pattern in this function. Other
shifts in the file (e.g., `LZ77_HASH_SIZE (1 << LZ77_HASH_LOG)` where
LZ77_HASH_LOG=15) are well within int range.
## Phase 6: Cross-Referencing Stable Trees
**Step 6.1 - Code in stable:**
Record: Verified buggy line `flag |= (1 << (32 - flag_count)) - 1;`
exists identically in v6.12, v6.15, v6.17, v6.18, v6.19. Code introduced
in v6.12 (d14bbfff259ca), moved to current form in v6.12 (94ae8c3fee94a)
- both before v6.12 release.
**Step 6.2 - Backport complications:**
Record: The line is identical in all stable trees from v6.12+. Clean
backport expected.
**Step 6.3 - Related fixes in stable:**
Record: No prior fixes for this specific issue in stable trees.
## Phase 7: Subsystem Context
**Step 7.1 - Subsystem criticality:**
Record: `fs/smb/client/compress/` — SMB client compression (experimental
sub-feature of an important filesystem subsystem). Classification:
PERIPHERAL (experimental, opt-in, default off).
**Step 7.2 - Activity:**
Record: Active development — 9 commits to `compress/` directory since
introduction, author is subsystem expert.
## Phase 8: Impact and Risk
**Step 8.1 - Affected users:**
Record: Users with `CONFIG_CIFS_COMPRESSION=y` (experimental, default N)
mounting SMB shares with `compress` option against servers supporting
SMB3.1.1 compression. Very limited population.
**Step 8.2 - Trigger conditions:**
Record: When uncompressed data size produces a state where `flag_count
== 0` on loop exit (i.e., multiple of 32 literal/match tokens followed
by final-loop exit at a 32-boundary) — the end-of-stream flag is
corrupt. Cannot be triggered by unprivileged users without SMB
compression mount.
**Step 8.3 - Failure mode:**
Record: Wrong compressed-payload encoding → server may fail to
decompress → SMB operation failure / possible retry. No crash, no data
corruption of user data (local), no security implications. Severity:
MEDIUM (only affects an experimental feature's correctness).
**Step 8.4 - Risk/benefit:**
Record: Benefit: fixes a real correctness bug for users of compression;
makes code valid under UBSAN. Risk: essentially zero — `1` → `1UL` is a
trivial type promotion that only changes evaluation on the edge UB cases
the fix targets. Very favorable risk/benefit ratio.
## Phase 9: Final Synthesis
**Step 9.1 - Evidence compiled:**
FOR: Real UB per C standard; produces wrong encoded output for
`flag_count == 0` case; 1-line surgical change; zero regression risk;
from the subsystem author; part of a larger fix series that includes
clear bugs; buggy line present and identical in all stable trees v6.12+;
trivial clean apply.
AGAINST: No Fixes:/Reported-by:/Cc: stable/Link: tags; no concrete user
bug report; feature is EXPERIMENTAL (default N) and opt-in; impact
limited to users of experimental compression; no explicit stable
nomination by reviewers.
**Step 9.2 - Stable rules checklist:**
1. Obviously correct and tested? YES — merged by maintainer, trivially
correct.
2. Fixes a real bug? YES — real C-standard UB producing incorrect bytes.
3. Important issue? BORDERLINE — correctness bug but in experimental
feature.
4. Small and contained? YES — 1 line.
5. No new features/APIs? YES.
6. Applies cleanly to stable? YES — identical code in v6.12 through
v6.19.
**Step 9.3 - Exceptions:**
Record: Not a device ID/quirk/DT/build/doc fix. Falls under normal bug-
fix category.
**Step 9.4 - Decision:**
The fix is tiny, obviously correct, zero-risk, and addresses real UB
that produces wrong output in an edge case. Although the feature is
marked EXPERIMENTAL, the fix:
- Is absolutely safe (1 → 1UL)
- Applies cleanly
- Is from the subsystem author
- Makes the code correct under UBSAN
- Complements other bug fixes in the series that should be backported
This is exactly the kind of low-risk correctness fix that stable trees
accept. Users who opted into the experimental feature receive correct
compressed output.
## Verification
- [Phase 1] Parsed tags: no Fixes, no Reported-by, no Cc: stable, no
Link; two Signed-off-by (author + maintainer)
- [Phase 2] Diff analysis: confirmed via `git show` — single line change
`1` → `1UL` in `lz77_compress()`
- [Phase 3] git blame/show d14bbfff259ca: original compression commit
July 2024 → v6.12; confirmed buggy line present
- [Phase 3] git show 94ae8c3fee94a: cleanup commit (Sep 2024, v6.12)
also contains buggy line — line has been unchanged since v6.12
- [Phase 3] `git tag --contains d14bbfff259ca`: bug reaches v6.12 and
all releases through v6.19
- [Phase 4] `b4 dig -c a13e942a03fee`: matched to https://lore.kernel.or
g/all/20260413190713.283939-2-ematsumiya@xxxxxxx/ as patch 2/8
- [Phase 4] `b4 dig -c a13e942a03fee -a`: only v1 exists
- [Phase 4] mbox content: Steve French replied "merged into cifs-2.6.git
for-next"; no NAKs; no stable request
- [Phase 5] Call chain: `lz77_compress()` ← `smb_compress()` ← SMB
request path only when CIFS_COMPRESSION=y and compress mount option
active
- [Phase 6] Verified identical buggy code in v6.12, v6.15, v6.17, v6.18,
v6.19 via `git show $TAG:fs/smb/client/compress/lz77.c`
- [Phase 7] Confirmed Enzo Matsumiya is the LZ77 code's original author
(authored d14bbfff259ca, 94ae8c3fee94a and the entire fix/improvement
series)
- [Phase 8] CONFIG_CIFS_COMPRESSION is "Experimental" and default N
(verified in `fs/smb/client/Kconfig` line 206-218)
- UNVERIFIED: Whether any real-world SMB server rejects malformed
compressed frames due to this specific bug — but UB produces wrong bit
pattern for flag_count == 0, which would be seen as invalid stream
marker by any standards-compliant decoder.
The fix is a trivial, safe, verifiable correctness improvement that
applies cleanly to all active stable trees. While the feature is
experimental, users opting in get correct output at zero risk to anyone
else.
**YES**
fs/smb/client/compress/lz77.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 96e8a8057a772..cdd6b53766b0a 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -221,7 +221,7 @@ noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
}
flag <<= (32 - flag_count);
- flag |= (1 << (32 - flag_count)) - 1;
+ flag |= (1UL << (32 - flag_count)) - 1;
lz77_write32(flag_pos, flag);
*dlen = dstp - dst;
--
2.53.0