[PATCH 1/8] smb: client: compress: fix buffer overrun in lz77_compress()

[PATCH 1/8] smb: client: compress: fix buffer overrun in lz77_compress()

[Enzo Matsumiya]

@dst buffer is allocated with same size as @src, which, for good
compression cases, works fine.

However, when compression goes bad (e.g. random bytes payloads), the
compressed size can increase significantly, and even by stopping the
main loop at 7/8 of @slen, writing leftover literals could write past
the end of @dst because of LZ77 metadata.

To fix this, add lz77_compressed_alloc_size() helper to compute the
correct allocation size for @dst, accounting for metadata and worst
cast scenario (all literals).

While this is overprovisioning memory, it's not only correct, but also
allows lz77_compress() main loop to run without ever checking @dst
limits (i.e. a perf improvement).

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress.c      |  6 +-----
 fs/smb/client/compress/lz77.c | 14 ++++----------
 fs/smb/client/compress/lz77.h | 28 ++++++++++++++++++++++++++++
 3 files changed, 33 insertions(+), 15 deletions(-)

diff --git a/fs/smb/client/compress.c b/fs/smb/client/compress.c
index 3d1e73f5d9af..be9023f841e6 100644
--- a/fs/smb/client/compress.c
+++ b/fs/smb/client/compress.c
@@ -329,11 +329,7 @@ int smb_compress(struct TCP_Server_Info *server, struct smb_rqst *rq, compress_s
 		goto err_free;
 	}
 
-	/*
-	 * This is just overprovisioning, as the algorithm will error out if @dst reaches 7/8
-	 * of @slen.
-	 */
-	dlen = slen;
+	dlen = lz77_compressed_alloc_size(slen);
 	dst = kvzalloc(dlen, GFP_KERNEL);
 	if (!dst) {
 		ret = -ENOMEM;
diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 96e8a8057a77..16c7d8f3ef17 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -137,6 +137,10 @@ noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
 	long flag = 0;
 	u64 *htable;
 
+	/* This is probably a bug, so throw a warning. */
+	if (WARN_ON_ONCE(*dlen < lz77_compressed_alloc_size(slen)))
+		return -EINVAL;
+
 	srcp = src;
 	end = src + slen;
 	dstp = dst;
@@ -180,15 +184,6 @@ noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
 			continue;
 		}
 
-		/*
-		 * Bail out if @dstp reached >= 7/8 of @slen -- already compressed badly, not worth
-		 * going further.
-		 */
-		if (unlikely(dstp - dst >= slen - (slen >> 3))) {
-			*dlen = slen;
-			goto out;
-		}
-
 		dstp = lz77_write_match(dstp, &nib, dist, len);
 		srcp += len;
 
@@ -225,7 +220,6 @@ noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
 	lz77_write32(flag_pos, flag);
 
 	*dlen = dstp - dst;
-out:
 	kvfree(htable);
 
 	if (*dlen < slen)
diff --git a/fs/smb/client/compress/lz77.h b/fs/smb/client/compress/lz77.h
index cdcb191b48a2..2603eab9e071 100644
--- a/fs/smb/client/compress/lz77.h
+++ b/fs/smb/client/compress/lz77.h
@@ -11,5 +11,33 @@
 
 #include <linux/kernel.h>
 
+/**
+ * lz77_compressed_alloc_size() - Compute compressed buffer size.
+ * @size:	uncompressed (src) size
+ *
+ * Compute allocation size for the compressed buffer based on uncompressed size.
+ * Accounts for metadata and overprovision for the worst case scenario.
+ *
+ * LZ77 metadata is a 4-byte flag that is written:
+ * - on dst begin (pos 0)
+ * - every 32 literals or matches
+ * - on end-of-stream (possibly, if last write was another flag)
+ *
+ * Worst case scenario is an all-literal compression, which means:
+ * metadata bytes = 4 + ((@size / 32) * 4) + 4, or, simplified, (@size >> 3) + 8
+ *
+ * The worst case scenario rarely happens, but such overprovisioning also allows lz77_compress()
+ * main loop to run without ever bound checking dst, which is a huge perf improvement, while also
+ * being safe when compression goes bad.
+ *
+ * Return: required (*) allocation size for compressed buffer.
+ *
+ * (*) checked once in the beginning of lz77_compress()
+ */
+static __always_inline u32 lz77_compressed_alloc_size(const u32 size)
+{
+	return size + (size >> 3) + 8;
+}
+
 int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen);
 #endif /* _SMB_COMPRESS_LZ77_H */
-- 
2.53.0

[PATCH 2/8] smb: client: compress: fix bad encoding on last LZ77 flag

[Enzo Matsumiya]

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@suse.de>
---
 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 16c7d8f3ef17..c1e7fada6e61 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -216,7 +216,7 @@ 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

Re: [PATCH 2/8] smb: client: compress: fix bad encoding on last LZ77 flag

[Steve French]

merged into cifs-2.6.git for-next

On Mon, Apr 13, 2026 at 2:07 PM Enzo Matsumiya <ematsumiya@suse.de> wrote:
>
> 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@suse.de>
> ---
>  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 16c7d8f3ef17..c1e7fada6e61 100644
> --- a/fs/smb/client/compress/lz77.c
> +++ b/fs/smb/client/compress/lz77.c
> @@ -216,7 +216,7 @@ 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
>


-- 
Thanks,

Steve

[PATCH 3/8] smb: client: compress: fix counting in LZ77 match finding

[Enzo Matsumiya]

- lz77_match_len() increments @cur before checking for equality,
  leading to off-by-one match len in some cases.

  Fix by moving pointers increment to inside the loop.
  Also rename @wnd arg to @match (more accurate name).
- both lz77_match_len() and lz77_compress() checked for
  "buf + step < end" when the correct is "<=" for such cases.

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress/lz77.c | 17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index c1e7fada6e61..61cdf1c14612 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -48,17 +48,17 @@ static __always_inline void lz77_write32(u32 *ptr, u32 v)
 	put_unaligned_le32(v, ptr);
 }
 
-static __always_inline u32 lz77_match_len(const void *wnd, const void *cur, const void *end)
+static __always_inline u32 lz77_match_len(const void *match, const void *cur, const void *end)
 {
 	const void *start = cur;
 	u64 diff;
 
 	/* Safe for a do/while because otherwise we wouldn't reach here from the main loop. */
 	do {
-		diff = lz77_read64(cur) ^ lz77_read64(wnd);
+		diff = lz77_read64(cur) ^ lz77_read64(match);
 		if (!diff) {
 			cur += LZ77_STEP_SIZE;
-			wnd += LZ77_STEP_SIZE;
+			match += LZ77_STEP_SIZE;
 
 			continue;
 		}
@@ -67,10 +67,13 @@ static __always_inline u32 lz77_match_len(const void *wnd, const void *cur, cons
 		cur += count_trailing_zeros(diff) >> 3;
 
 		return (cur - start);
-	} while (likely(cur + LZ77_STEP_SIZE < end));
+	} while (likely(cur + LZ77_STEP_SIZE <= end));
 
-	while (cur < end && lz77_read8(cur++) == lz77_read8(wnd++))
-		;
+	/* Fallback to byte-by-byte comparison for last <8 bytes. */
+	while (cur < end && lz77_read8(cur) == lz77_read8(match)) {
+		cur++;
+		match++;
+	}
 
 	return (cur - start);
 }
@@ -195,7 +198,7 @@ noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
 			flag_pos = dstp;
 			dstp += 4;
 		}
-	} while (likely(srcp + LZ77_STEP_SIZE < end));
+	} while (likely(srcp + LZ77_STEP_SIZE <= end));
 
 	while (srcp < end) {
 		u32 c = umin(end - srcp, 32 - flag_count);
-- 
2.53.0

[PATCH 4/8] smb: client: compress: increase LZ77_MATCH_MAX_DIST

[Enzo Matsumiya]

Increase max distance (i.e. window size) from 1k to 8k.
This allows better compression and is just as fast.

Other:
- drop LZ77_MATCH_MIN_DIST as it's nused -- main loop
  already checks if dist > 0

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress/lz77.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 61cdf1c14612..480927dcd4c6 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -17,8 +17,7 @@
  * Compression parameters.
  */
 #define LZ77_MATCH_MIN_LEN	4
-#define LZ77_MATCH_MIN_DIST	1
-#define LZ77_MATCH_MAX_DIST	SZ_1K
+#define LZ77_MATCH_MAX_DIST	SZ_8K
 #define LZ77_HASH_LOG		15
 #define LZ77_HASH_SIZE		(1 << LZ77_HASH_LOG)
 #define LZ77_STEP_SIZE		sizeof(u64)
-- 
2.53.0

[PATCH 5/8] smb: client: compress: LZ77 optimizations

[Enzo Matsumiya]

This patch implements several micro-optimizations on lz77_compress()
with the goal of reducing the number of instructions per [input]
byte (a.k.a. IPB).

Changes:
- change hashtable to be u32 (instead of u64) -- change the hash
  function to reflect that (adds lz77_hash() and lz77_read32() helpers)
- batch-write literals instead of 1 by 1 -- now that we have a well
  defined hot path (match finding) and a cold path (encode literals +
  match), batch writing makes a significant difference
- implement adaptive skipping of input bytes -- skip input bytes more
  aggressively if too few matches are being found
- name some constants for more meaningful context

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress/lz77.c | 173 +++++++++++++++++++++-------------
 fs/smb/client/compress/lz77.h |   4 +-
 2 files changed, 108 insertions(+), 69 deletions(-)

diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 480927dcd4c6..96744f52e364 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: GPL-2.0-only
 /*
- * Copyright (C) 2024, SUSE LLC
+ * Copyright (C) 2024-2026, SUSE LLC
  *
  * Authors: Enzo Matsumiya <ematsumiya@suse.de>
  *
@@ -16,17 +16,26 @@
 /*
  * Compression parameters.
  */
-#define LZ77_MATCH_MIN_LEN	4
 #define LZ77_MATCH_MAX_DIST	SZ_8K
 #define LZ77_HASH_LOG		15
 #define LZ77_HASH_SIZE		(1 << LZ77_HASH_LOG)
-#define LZ77_STEP_SIZE		sizeof(u64)
+#define LZ77_RSTEP_SIZE		sizeof(u32)
+#define LZ77_MSTEP_SIZE		sizeof(u64)
+#define LZ77_SKIP_TRIGGER	4
+
+#define LZ77_PREFETCH(ptr)	__builtin_prefetch((ptr), 0, 3)
+#define LZ77_FLAG_MAX		32
 
 static __always_inline u8 lz77_read8(const u8 *ptr)
 {
 	return get_unaligned(ptr);
 }
 
+static __always_inline u32 lz77_read32(const u32 *ptr)
+{
+	return get_unaligned(ptr);
+}
+
 static __always_inline u64 lz77_read64(const u64 *ptr)
 {
 	return get_unaligned(ptr);
@@ -50,14 +59,14 @@ static __always_inline void lz77_write32(u32 *ptr, u32 v)
 static __always_inline u32 lz77_match_len(const void *match, const void *cur, const void *end)
 {
 	const void *start = cur;
-	u64 diff;
 
 	/* Safe for a do/while because otherwise we wouldn't reach here from the main loop. */
 	do {
-		diff = lz77_read64(cur) ^ lz77_read64(match);
+		const u64 diff = lz77_read64(cur) ^ lz77_read64(match);
+
 		if (!diff) {
-			cur += LZ77_STEP_SIZE;
-			match += LZ77_STEP_SIZE;
+			cur += LZ77_MSTEP_SIZE;
+			match += LZ77_MSTEP_SIZE;
 
 			continue;
 		}
@@ -66,7 +75,7 @@ static __always_inline u32 lz77_match_len(const void *match, const void *cur, co
 		cur += count_trailing_zeros(diff) >> 3;
 
 		return (cur - start);
-	} while (likely(cur + LZ77_STEP_SIZE <= end));
+	} while (likely(cur + LZ77_MSTEP_SIZE <= end));
 
 	/* Fallback to byte-by-byte comparison for last <8 bytes. */
 	while (cur < end && lz77_read8(cur) == lz77_read8(match)) {
@@ -77,7 +86,7 @@ static __always_inline u32 lz77_match_len(const void *match, const void *cur, co
 	return (cur - start);
 }
 
-static __always_inline void *lz77_write_match(void *dst, void **nib, u32 dist, u32 len)
+static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist, u32 len)
 {
 	len -= 3;
 	dist--;
@@ -131,94 +140,124 @@ static __always_inline void *lz77_write_match(void *dst, void **nib, u32 dist, u
 	return dst + 4;
 }
 
-noinline int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen)
+static __always_inline void *lz77_encode_literals(const void *start, const void *end, void *dst,
+						  long *f, u32 *fc, void **fp)
+{
+	if (start >= end)
+		return dst;
+
+	do {
+		const u32 len = umin(end - start, LZ77_FLAG_MAX - *fc);
+
+		memcpy(dst, start, len);
+
+		dst += len;
+		start += len;
+
+		*f <<= len;
+		*fc += len;
+		if (*fc == LZ77_FLAG_MAX) {
+			lz77_write32(*fp, *f);
+			*fc = 0;
+			*fp = dst;
+			dst += 4;
+		}
+	} while (start < end);
+
+	return dst;
+}
+
+static __always_inline u32 lz77_hash(const u32 v)
+{
+	return ((v ^ 0x9E3779B9) * 0x85EBCA6B) >> (32 - LZ77_HASH_LOG);
+}
+
+noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen)
 {
-	const void *srcp, *end;
+	const void *srcp, *rlim, *end, *anchor;
+	u32 *htable, hash, flag_count = 0;
 	void *dstp, *nib, *flag_pos;
-	u32 flag_count = 0;
 	long flag = 0;
-	u64 *htable;
 
 	/* This is probably a bug, so throw a warning. */
 	if (WARN_ON_ONCE(*dlen < lz77_compressed_alloc_size(slen)))
 		return -EINVAL;
 
-	srcp = src;
-	end = src + slen;
+	srcp = anchor = src;
+	end = srcp + slen; /* absolute end */
+	rlim = end - LZ77_MSTEP_SIZE; /* read limit (for lz77_match_len()) */
 	dstp = dst;
-	nib = NULL;
 	flag_pos = dstp;
 	dstp += 4;
+	nib = NULL;
 
 	htable = kvcalloc(LZ77_HASH_SIZE, sizeof(*htable), GFP_KERNEL);
 	if (!htable)
 		return -ENOMEM;
 
-	/* Main loop. */
-	do {
-		u32 dist, len = 0;
-		const void *wnd;
-		u64 hash;
-
-		hash = ((lz77_read64(srcp) << 24) * 889523592379ULL) >> (64 - LZ77_HASH_LOG);
-		wnd = src + htable[hash];
-		htable[hash] = srcp - src;
-		dist = srcp - wnd;
-
-		if (dist && dist < LZ77_MATCH_MAX_DIST)
-			len = lz77_match_len(wnd, srcp, end);
+	LZ77_PREFETCH(srcp + LZ77_RSTEP_SIZE);
 
-		if (len < LZ77_MATCH_MIN_LEN) {
-			lz77_write8(dstp, lz77_read8(srcp));
-
-			dstp++;
-			srcp++;
-
-			flag <<= 1;
-			flag_count++;
-			if (flag_count == 32) {
-				lz77_write32(flag_pos, flag);
-				flag_count = 0;
-				flag_pos = dstp;
-				dstp += 4;
-			}
-
-			continue;
-		}
+	hash = lz77_hash(lz77_read32(srcp++));
+	htable[hash] = 0;
+	hash = lz77_hash(lz77_read32(srcp));
 
-		dstp = lz77_write_match(dstp, &nib, dist, len);
+	/*
+	 * Main loop.
+	 *
+	 * @dlen is >= lz77_compressed_alloc_size(), so run without bound-checking @dstp.
+	 *
+	 * This code was crafted in a way to best utilise fetch-decode-execute CPU flow.
+	 * Any attempt to optimize it, or even organize it, can lead to huge performance loss.
+	 */
+	do {
+		const void *match, *next = srcp;
+		u32 len, step = 1, skip = 1U << LZ77_SKIP_TRIGGER;
+
+		/* Match finding (hot path -- don't change the read/check/write order). */
+		do {
+			const u32 cur_hash = hash;
+
+			srcp = next;
+			next += step;
+			step = (skip++ >> LZ77_SKIP_TRIGGER);
+			if (unlikely(next > rlim))
+				goto out;
+
+			hash = lz77_hash(lz77_read32(next));
+			match = src + htable[cur_hash];
+			htable[cur_hash] = srcp - src;
+		} while (likely(match + LZ77_MATCH_MAX_DIST < srcp) ||
+			 lz77_read32(match) != lz77_read32(srcp));
+
+		dstp = lz77_encode_literals(anchor, srcp, dstp, &flag, &flag_count, &flag_pos);
+		len = lz77_match_len(match, srcp, end);
+		dstp = lz77_encode_match(dstp, &nib, srcp - match, len);
 		srcp += len;
+		anchor = srcp;
+
+		LZ77_PREFETCH(srcp);
 
 		flag = (flag << 1) | 1;
 		flag_count++;
-		if (flag_count == 32) {
+		if (flag_count == LZ77_FLAG_MAX) {
 			lz77_write32(flag_pos, flag);
 			flag_count = 0;
 			flag_pos = dstp;
 			dstp += 4;
 		}
-	} while (likely(srcp + LZ77_STEP_SIZE <= end));
-
-	while (srcp < end) {
-		u32 c = umin(end - srcp, 32 - flag_count);
 
-		memcpy(dstp, srcp, c);
+		if (unlikely(srcp > rlim))
+			break;
 
-		dstp += c;
-		srcp += c;
-
-		flag <<= c;
-		flag_count += c;
-		if (flag_count == 32) {
-			lz77_write32(flag_pos, flag);
-			flag_count = 0;
-			flag_pos = dstp;
-			dstp += 4;
-		}
-	}
+		/* Prepare for next loop. */
+		hash = lz77_hash(lz77_read32(srcp));
+	} while (srcp < end);
+out:
+	dstp = lz77_encode_literals(anchor, end, dstp, &flag, &flag_count, &flag_pos);
 
-	flag <<= (32 - flag_count);
-	flag |= (1UL << (32 - flag_count)) - 1;
+	flag_count = LZ77_FLAG_MAX - flag_count;
+	flag <<= flag_count;
+	flag |= (1UL << flag_count) - 1;
 	lz77_write32(flag_pos, flag);
 
 	*dlen = dstp - dst;
diff --git a/fs/smb/client/compress/lz77.h b/fs/smb/client/compress/lz77.h
index 2603eab9e071..4e570846aefa 100644
--- a/fs/smb/client/compress/lz77.h
+++ b/fs/smb/client/compress/lz77.h
@@ -1,6 +1,6 @@
 /* SPDX-License-Identifier: GPL-2.0-only */
 /*
- * Copyright (C) 2024, SUSE LLC
+ * Copyright (C) 2024-2026, SUSE LLC
  *
  * Authors: Enzo Matsumiya <ematsumiya@suse.de>
  *
@@ -39,5 +39,5 @@ static __always_inline u32 lz77_compressed_alloc_size(const u32 size)
 	return size + (size >> 3) + 8;
 }
 
-int lz77_compress(const void *src, u32 slen, void *dst, u32 *dlen);
+int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen);
 #endif /* _SMB_COMPRESS_LZ77_H */
-- 
2.53.0

[PATCH 6/8] smb: client: compress: add code docs to lz77.c

[Enzo Matsumiya]

Document parts of the code, especially the apparently
non-sense parts.

Other:
- change pointer increment constants to sizeof() values

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress/lz77.c | 81 +++++++++++++++++++++++++++++++----
 1 file changed, 73 insertions(+), 8 deletions(-)

diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 96744f52e364..7365d0f97396 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -15,6 +15,20 @@
 
 /*
  * Compression parameters.
+ *
+ * LZ77_MATCH_MAX_DIST:		Farthest back a match can be from current position (can be 1 - 8K).
+ * LZ77_HASH_LOG:
+ * LZ77_HASH_SIZE:		ilog2 hash size (recommended to be 13 - 18, default 15 (hash size
+ *				32k)).
+ * LZ77_RSTEP_SIZE:		Number of bytes to read from input buffer for hashing and initial
+ *				match check (default 4 bytes, this effectivelly makes this the min
+ *				match len).
+ * LZ77_MSTEP_SIZE:		Number of bytes to extend-compare a found match (default 8 bytes).
+ * LZ77_SKIP_TRIGGER:		ilog2 value for adaptive skipping, i.e. to progressively skip input
+ *				bytes when we can't find matches.  Default is 4.
+ *				Higher values (>0) will decrease compression time, but will result
+ *				in worse compression ratio.  Lower values will give better
+ *				compression ratio (more matches found), but will increase time.
  */
 #define LZ77_MATCH_MAX_DIST	SZ_8K
 #define LZ77_HASH_LOG		15
@@ -86,6 +100,19 @@ static __always_inline u32 lz77_match_len(const void *match, const void *cur, co
 	return (cur - start);
 }
 
+/**
+ * lz77_encode_match() - Match encoding.
+ * @dst:	compressed buffer
+ * @nib:	pointer to an address in @dst
+ * @dist:	match distance
+ * @len:	match length
+ *
+ * Assumes all args were previously checked.
+ *
+ * Return: @dst advanced to new position
+ *
+ * Ref: MS-XCA 2.3.4 "Plain LZ77 Compression Algorithm Details" - "Processing"
+ */
 static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist, u32 len)
 {
 	len -= 3;
@@ -95,12 +122,12 @@ static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist,
 	if (len < 7) {
 		lz77_write16(dst, dist + len);
 
-		return dst + 2;
+		return dst + sizeof(u16);
 	}
 
 	dist |= 7;
 	lz77_write16(dst, dist);
-	dst += 2;
+	dst += sizeof(u16);
 	len -= 7;
 
 	if (!*nib) {
@@ -130,16 +157,32 @@ static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist,
 	if (len <= 0xffff) {
 		lz77_write16(dst, len);
 
-		return dst + 2;
+		return dst + sizeof(u16);
 	}
 
 	lz77_write16(dst, 0);
-	dst += 2;
+	dst += sizeof(u16);
 	lz77_write32(dst, len);
 
-	return dst + 4;
+	return dst + sizeof(u32);
 }
 
+/**
+ * lz77_encode_literals() - Literals encoding.
+ * @start:	where to start copying literals (uncompressed buffer)
+ * @end:	when to stop copying (uncompressed buffer)
+ * @dst:	compressed buffer
+ * @f:		pointer to current flag value
+ * @fc:		pointer to current flag count
+ * @fp:		pointer to current flag address
+ *
+ * Batch copy literals from @start to @dst, updating flag values accordingly.
+ * Assumes all args were previously checked.
+ *
+ * Return: @dst advanced to new position
+ *
+ * MS-XCA 2.3.4 "Plain LZ77 Compression Algorithm Details" - "Processing"
+ */
 static __always_inline void *lz77_encode_literals(const void *start, const void *end, void *dst,
 						  long *f, u32 *fc, void **fp)
 {
@@ -160,7 +203,7 @@ static __always_inline void *lz77_encode_literals(const void *start, const void
 			lz77_write32(*fp, *f);
 			*fc = 0;
 			*fp = dst;
-			dst += 4;
+			dst += sizeof(u32);
 		}
 	} while (start < end);
 
@@ -188,7 +231,7 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 	rlim = end - LZ77_MSTEP_SIZE; /* read limit (for lz77_match_len()) */
 	dstp = dst;
 	flag_pos = dstp;
-	dstp += 4;
+	dstp += sizeof(u32);
 	nib = NULL;
 
 	htable = kvcalloc(LZ77_HASH_SIZE, sizeof(*htable), GFP_KERNEL);
@@ -197,6 +240,10 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 
 	LZ77_PREFETCH(srcp + LZ77_RSTEP_SIZE);
 
+	/*
+	 * Adjust @srcp so we don't get a false positive match on first iteration.
+	 * Then prepare hash for first loop iteration (don't advance @srcp again).
+	 */
 	hash = lz77_hash(lz77_read32(srcp++));
 	htable[hash] = 0;
 	hash = lz77_hash(lz77_read32(srcp));
@@ -219,6 +266,14 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 
 			srcp = next;
 			next += step;
+
+			/*
+			 * Adaptive skipping.
+			 *
+			 * Increment @step every (1 << LZ77_SKIP_TRIGGER, 16 in our case) bytes
+			 * without a match.
+			 * Reset to 1 when a match is found.
+			 */
 			step = (skip++ >> LZ77_SKIP_TRIGGER);
 			if (unlikely(next > rlim))
 				goto out;
@@ -229,6 +284,16 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 		} while (likely(match + LZ77_MATCH_MAX_DIST < srcp) ||
 			 lz77_read32(match) != lz77_read32(srcp));
 
+		/*
+		 * Match found.  Warm/cold path; begin parsing @srcp and writing to @dstp:
+		 * - flush literals
+		 * - compute match length (*)
+		 * - encode match
+		 *
+		 * (*) Current minimum match length is defined by the memory read size above, so
+		 * here we already know that we have 4 matching bytes, but it's just faster to
+		 * redundantly compute it again in lz77_match_len() than to adjust pointers/len.
+		 */
 		dstp = lz77_encode_literals(anchor, srcp, dstp, &flag, &flag_count, &flag_pos);
 		len = lz77_match_len(match, srcp, end);
 		dstp = lz77_encode_match(dstp, &nib, srcp - match, len);
@@ -243,7 +308,7 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 			lz77_write32(flag_pos, flag);
 			flag_count = 0;
 			flag_pos = dstp;
-			dstp += 4;
+			dstp += sizeof(u32);
 		}
 
 		if (unlikely(srcp > rlim))
-- 
2.53.0

[PATCH 7/8] smb: client: compress: add compress/common.h

[Enzo Matsumiya]

Add compress/common.h to aggregate helpers and definitions that will be
shared with other compression algorithms.

Also add a few build time checks for proper support.

Changes:
- update affected call paths in compress/lz77.c

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/client/compress/common.h | 152 ++++++++++++++++++++++++++++++++
 fs/smb/client/compress/lz77.c   | 122 +++++--------------------
 2 files changed, 175 insertions(+), 99 deletions(-)
 create mode 100644 fs/smb/client/compress/common.h

diff --git a/fs/smb/client/compress/common.h b/fs/smb/client/compress/common.h
new file mode 100644
index 000000000000..b5ccf5debd22
--- /dev/null
+++ b/fs/smb/client/compress/common.h
@@ -0,0 +1,152 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2026, SUSE LLC
+ *
+ * Authors: Enzo Matsumiya <ematsumiya@suse.de>
+ *
+ * Common helpers and definitions for compression/decompression.
+ *
+ * Naming convention:
+ * - smb_compress_*:	MS-SMB2 related API
+ * - compress_*:	Compression internals
+ * - <alg>_*:		Compression algorithm specifics
+ *
+ * Upper case for macros and constants, lower case for everything else.
+ */
+#ifndef _COMPRESS_COMMON_H
+#define _COMPRESS_COMMON_H
+
+#include <linux/kernel.h>
+#include <linux/count_zeros.h>
+#include <linux/string.h>
+#include <linux/sizes.h>
+#include <linux/slab.h>
+
+/*
+ * Build time checks/asserts.
+ * These are assumptions/expectations for all algorithms implemented.
+ */
+#ifndef __LITTLE_ENDIAN /* TODO */
+# error "SMB3 compression is only supported on little endian architectures"
+#endif /* !__LITTLE_ENDIAN */
+
+#if BITS_PER_LONG < 64 /* TODO */
+# error "SMB3 compression is only supported on 64-bit architectures"
+#endif /* BITS_PER_LONG < 64 */
+
+/* Build time double checks (probably unnecessary) */
+static_assert(sizeof(u16) == 2);
+static_assert(sizeof(u32) == 4);
+static_assert(sizeof(u64) == 8);
+static_assert(sizeof(size_t) == 8);
+
+#ifdef CONFIG_CIFS_DEBUG2
+# define COMPRESS_DEBUG
+#endif /* CONFIG_CIFS_DEBUG2 */
+
+#define COMPRESS_LOG_FMT	"CIFS: %s:%d %s(): "
+
+#ifdef COMPRESS_DEBUG
+# define compress_log(fmt, ...) \
+	pr_err(COMPRESS_LOG_FMT fmt, __FILE__, __LINE__, __func__, ## __VA_ARGS__)
+
+/* Make sure to handle assertion failures! */
+# define COMPRESS_ASSERT(cond) \
+	!WARN_ONCE(!(cond), COMPRESS_LOG_FMT "assertion failed: %s\n", \
+		   __FILE__, __LINE__, __func__, #cond)
+#else /* COMPRESS_DEBUG */
+# define compress_log(...)
+/* XXX: should not fail silently on non-debug? */
+# define COMPRESS_ASSERT(cond)	likely(cond)
+#endif /* !COMPRESS_DEBUG */
+
+/*
+ * Memory ops helpers.
+ */
+#undef MEM_UNALIGNED_READ
+#undef MEM_UNALIGNED_WRITE
+
+/* Read prefetch */
+#define MEM_PREFETCH(ptr)		__builtin_prefetch((ptr), 0, 3)
+
+#ifdef CONFIG_X86
+# define MEM_UNALIGNED_READ(ptr, t)	(*(const t *)(ptr))
+# define MEM_UNALIGNED_WRITE(ptr, v, t)	(*(t *)(ptr) = (t)(v))
+#else
+# include <linux/unaligned.h>
+# define MEM_UNALIGNED_READ(ptr, t)	get_unaligned((const t *)(ptr))
+# define MEM_UNALIGNED_WRITE(ptr, v, t)	put_unaligned((v), (t *)(ptr))
+#endif /* !CONFIG_X86 */
+
+#define mem_read8(ptr)			MEM_UNALIGNED_READ(ptr, u8)
+#define mem_read16(ptr)			MEM_UNALIGNED_READ(ptr, u16)
+#define mem_read32(ptr)			MEM_UNALIGNED_READ(ptr, u32)
+#define mem_read64(ptr)			MEM_UNALIGNED_READ(ptr, u64)
+#define mem_write8(ptr, v)		MEM_UNALIGNED_WRITE(ptr, v, u8)
+#define mem_write16(ptr, v)		MEM_UNALIGNED_WRITE(ptr, v, u16)
+#define mem_write32(ptr, v)		MEM_UNALIGNED_WRITE(ptr, v, u32)
+/* mem_write64() not implemented because it's not used anywhere. */
+
+/*
+ * COMPRESS_RSTEP_SIZE:		Number of bytes to read from input buffer for hashing and initial
+ *				match check (default 4 bytes).
+ * COMPRESS_MSTEP_SIZE:		Number of bytes to extend-compare a found match (default 8 bytes).
+ */
+#define COMPRESS_RSTEP_SIZE	sizeof(u32)
+#define COMPRESS_MSTEP_SIZE	sizeof(u64)
+
+static __always_inline size_t mem_match_len(const void *match, const void *cur, const void *end)
+{
+	const void *start = cur;
+
+	/* Callers must ensure @cur + COMPRESS_MSTEP_SIZE < @end. */
+	do {
+		const u64 diff = mem_read64(cur) ^ mem_read64(match);
+
+		if (!diff) {
+			cur += COMPRESS_MSTEP_SIZE;
+			match += COMPRESS_MSTEP_SIZE;
+
+			continue;
+		}
+
+		/* This computes the number of common bytes in @diff. */
+		cur += count_trailing_zeros(diff) >> 3;
+
+		return (cur - start);
+	} while (likely(cur + COMPRESS_MSTEP_SIZE <= end));
+
+	/* Fallback to byte-by-byte comparison for last bytes (< COMPRESS_MSTEP_SIZE). */
+	while (cur < end && mem_read8(match) == mem_read8(cur)) {
+		cur++;
+		match++;
+	}
+
+	return (cur - start);
+}
+
+/*
+ * Hashing
+ *
+ * Same for all algorithms.
+ *
+ * XXX: these are fixed for now, might make them tunables in the future.
+ */
+
+/*
+ * COMPRESS_HASH_LOG:		ilog2 hash size (recommended to be 13 - 18, default 15)
+ * COMPRESS_HASH_SIZE:		Hashtable size (default is 32k (1 << COMPRESS_HASH_LOG))).
+ */
+#define COMPRESS_HASH_LOG	15
+#define COMPRESS_HASH_SIZE	(1U << COMPRESS_HASH_LOG)
+
+static __always_inline u32 compress_hash(const u32 v)
+{
+	return ((v ^ 0x9E3779B9U) * 0x85EBCA6BU) >> (32 - COMPRESS_HASH_LOG);
+}
+
+static __always_inline u32 compress_hash_ptr(const void *ptr)
+{
+	return compress_hash(mem_read32(ptr));
+}
+#endif /* _COMPRESS_COMMON_H */
diff --git a/fs/smb/client/compress/lz77.c b/fs/smb/client/compress/lz77.c
index 7365d0f97396..d8c73f7f7483 100644
--- a/fs/smb/client/compress/lz77.c
+++ b/fs/smb/client/compress/lz77.c
@@ -11,19 +11,13 @@
 #include <linux/count_zeros.h>
 #include <linux/unaligned.h>
 
+#include "common.h"
 #include "lz77.h"
 
 /*
  * Compression parameters.
  *
  * LZ77_MATCH_MAX_DIST:		Farthest back a match can be from current position (can be 1 - 8K).
- * LZ77_HASH_LOG:
- * LZ77_HASH_SIZE:		ilog2 hash size (recommended to be 13 - 18, default 15 (hash size
- *				32k)).
- * LZ77_RSTEP_SIZE:		Number of bytes to read from input buffer for hashing and initial
- *				match check (default 4 bytes, this effectivelly makes this the min
- *				match len).
- * LZ77_MSTEP_SIZE:		Number of bytes to extend-compare a found match (default 8 bytes).
  * LZ77_SKIP_TRIGGER:		ilog2 value for adaptive skipping, i.e. to progressively skip input
  *				bytes when we can't find matches.  Default is 4.
  *				Higher values (>0) will decrease compression time, but will result
@@ -31,75 +25,10 @@
  *				compression ratio (more matches found), but will increase time.
  */
 #define LZ77_MATCH_MAX_DIST	SZ_8K
-#define LZ77_HASH_LOG		15
-#define LZ77_HASH_SIZE		(1 << LZ77_HASH_LOG)
-#define LZ77_RSTEP_SIZE		sizeof(u32)
-#define LZ77_MSTEP_SIZE		sizeof(u64)
 #define LZ77_SKIP_TRIGGER	4
 
-#define LZ77_PREFETCH(ptr)	__builtin_prefetch((ptr), 0, 3)
 #define LZ77_FLAG_MAX		32
 
-static __always_inline u8 lz77_read8(const u8 *ptr)
-{
-	return get_unaligned(ptr);
-}
-
-static __always_inline u32 lz77_read32(const u32 *ptr)
-{
-	return get_unaligned(ptr);
-}
-
-static __always_inline u64 lz77_read64(const u64 *ptr)
-{
-	return get_unaligned(ptr);
-}
-
-static __always_inline void lz77_write8(u8 *ptr, u8 v)
-{
-	put_unaligned(v, ptr);
-}
-
-static __always_inline void lz77_write16(u16 *ptr, u16 v)
-{
-	put_unaligned_le16(v, ptr);
-}
-
-static __always_inline void lz77_write32(u32 *ptr, u32 v)
-{
-	put_unaligned_le32(v, ptr);
-}
-
-static __always_inline u32 lz77_match_len(const void *match, const void *cur, const void *end)
-{
-	const void *start = cur;
-
-	/* Safe for a do/while because otherwise we wouldn't reach here from the main loop. */
-	do {
-		const u64 diff = lz77_read64(cur) ^ lz77_read64(match);
-
-		if (!diff) {
-			cur += LZ77_MSTEP_SIZE;
-			match += LZ77_MSTEP_SIZE;
-
-			continue;
-		}
-
-		/* This computes the number of common bytes in @diff. */
-		cur += count_trailing_zeros(diff) >> 3;
-
-		return (cur - start);
-	} while (likely(cur + LZ77_MSTEP_SIZE <= end));
-
-	/* Fallback to byte-by-byte comparison for last <8 bytes. */
-	while (cur < end && lz77_read8(cur) == lz77_read8(match)) {
-		cur++;
-		match++;
-	}
-
-	return (cur - start);
-}
-
 /**
  * lz77_encode_match() - Match encoding.
  * @dst:	compressed buffer
@@ -120,24 +49,24 @@ static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist,
 	dist <<= 3;
 
 	if (len < 7) {
-		lz77_write16(dst, dist + len);
+		mem_write16(dst, dist + len);
 
 		return dst + sizeof(u16);
 	}
 
 	dist |= 7;
-	lz77_write16(dst, dist);
+	mem_write16(dst, dist);
 	dst += sizeof(u16);
 	len -= 7;
 
 	if (!*nib) {
-		lz77_write8(dst, umin(len, 15));
+		mem_write8(dst, umin(len, 15));
 		*nib = dst;
 		dst++;
 	} else {
 		u8 *b = *nib;
 
-		lz77_write8(b, *b | umin(len, 15) << 4);
+		mem_write8(b, *b | umin(len, 15) << 4);
 		*nib = NULL;
 	}
 
@@ -146,23 +75,23 @@ static __always_inline void *lz77_encode_match(void *dst, void **nib, u16 dist,
 
 	len -= 15;
 	if (len < 255) {
-		lz77_write8(dst, len);
+		mem_write8(dst, len);
 
 		return dst + 1;
 	}
 
-	lz77_write8(dst, 0xff);
+	mem_write8(dst, 0xff);
 	dst++;
 	len += 7 + 15;
 	if (len <= 0xffff) {
-		lz77_write16(dst, len);
+		mem_write16(dst, len);
 
 		return dst + sizeof(u16);
 	}
 
-	lz77_write16(dst, 0);
+	mem_write16(dst, 0);
 	dst += sizeof(u16);
-	lz77_write32(dst, len);
+	mem_write32(dst, len);
 
 	return dst + sizeof(u32);
 }
@@ -200,7 +129,7 @@ static __always_inline void *lz77_encode_literals(const void *start, const void
 		*f <<= len;
 		*fc += len;
 		if (*fc == LZ77_FLAG_MAX) {
-			lz77_write32(*fp, *f);
+			mem_write32(*fp, *f);
 			*fc = 0;
 			*fp = dst;
 			dst += sizeof(u32);
@@ -210,11 +139,6 @@ static __always_inline void *lz77_encode_literals(const void *start, const void
 	return dst;
 }
 
-static __always_inline u32 lz77_hash(const u32 v)
-{
-	return ((v ^ 0x9E3779B9) * 0x85EBCA6B) >> (32 - LZ77_HASH_LOG);
-}
-
 noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen)
 {
 	const void *srcp, *rlim, *end, *anchor;
@@ -228,25 +152,25 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 
 	srcp = anchor = src;
 	end = srcp + slen; /* absolute end */
-	rlim = end - LZ77_MSTEP_SIZE; /* read limit (for lz77_match_len()) */
+	rlim = end - COMPRESS_MSTEP_SIZE; /* read limit (for mem_match_len()) */
 	dstp = dst;
 	flag_pos = dstp;
 	dstp += sizeof(u32);
 	nib = NULL;
 
-	htable = kvcalloc(LZ77_HASH_SIZE, sizeof(*htable), GFP_KERNEL);
+	htable = kvcalloc(COMPRESS_HASH_SIZE, sizeof(*htable), GFP_KERNEL);
 	if (!htable)
 		return -ENOMEM;
 
-	LZ77_PREFETCH(srcp + LZ77_RSTEP_SIZE);
+	MEM_PREFETCH(srcp + COMPRESS_RSTEP_SIZE);
 
 	/*
 	 * Adjust @srcp so we don't get a false positive match on first iteration.
 	 * Then prepare hash for first loop iteration (don't advance @srcp again).
 	 */
-	hash = lz77_hash(lz77_read32(srcp++));
+	hash = compress_hash_ptr(srcp++);
 	htable[hash] = 0;
-	hash = lz77_hash(lz77_read32(srcp));
+	hash = compress_hash_ptr(srcp);
 
 	/*
 	 * Main loop.
@@ -278,11 +202,11 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 			if (unlikely(next > rlim))
 				goto out;
 
-			hash = lz77_hash(lz77_read32(next));
+			hash = compress_hash_ptr(next);
 			match = src + htable[cur_hash];
 			htable[cur_hash] = srcp - src;
 		} while (likely(match + LZ77_MATCH_MAX_DIST < srcp) ||
-			 lz77_read32(match) != lz77_read32(srcp));
+			 mem_read32(match) != mem_read32(srcp));
 
 		/*
 		 * Match found.  Warm/cold path; begin parsing @srcp and writing to @dstp:
@@ -295,17 +219,17 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 		 * redundantly compute it again in lz77_match_len() than to adjust pointers/len.
 		 */
 		dstp = lz77_encode_literals(anchor, srcp, dstp, &flag, &flag_count, &flag_pos);
-		len = lz77_match_len(match, srcp, end);
+		len = mem_match_len(match, srcp, end);
 		dstp = lz77_encode_match(dstp, &nib, srcp - match, len);
 		srcp += len;
 		anchor = srcp;
 
-		LZ77_PREFETCH(srcp);
+		MEM_PREFETCH(srcp);
 
 		flag = (flag << 1) | 1;
 		flag_count++;
 		if (flag_count == LZ77_FLAG_MAX) {
-			lz77_write32(flag_pos, flag);
+			mem_write32(flag_pos, flag);
 			flag_count = 0;
 			flag_pos = dstp;
 			dstp += sizeof(u32);
@@ -315,7 +239,7 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 			break;
 
 		/* Prepare for next loop. */
-		hash = lz77_hash(lz77_read32(srcp));
+		hash = compress_hash_ptr(srcp);
 	} while (srcp < end);
 out:
 	dstp = lz77_encode_literals(anchor, end, dstp, &flag, &flag_count, &flag_pos);
@@ -323,7 +247,7 @@ noinline int lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen
 	flag_count = LZ77_FLAG_MAX - flag_count;
 	flag <<= flag_count;
 	flag |= (1UL << flag_count) - 1;
-	lz77_write32(flag_pos, flag);
+	mem_write32(flag_pos, flag);
 
 	*dlen = dstp - dst;
 	kvfree(htable);
-- 
2.53.0

Re: [PATCH 7/8] smb: client: compress: add compress/common.h

[Nathan Chancellor]

Hi Enzo,

On Mon, Apr 13, 2026 at 04:07:12PM -0300, Enzo Matsumiya wrote:
> Add compress/common.h to aggregate helpers and definitions that will be
> shared with other compression algorithms.
> 
> Also add a few build time checks for proper support.
> 
> Changes:
> - update affected call paths in compress/lz77.c
> 
> Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
> ---
...
> diff --git a/fs/smb/client/compress/common.h b/fs/smb/client/compress/common.h
> new file mode 100644
> index 000000000000..b5ccf5debd22
> --- /dev/null
> +++ b/fs/smb/client/compress/common.h
...
> +/*
> + * Build time checks/asserts.
> + * These are assumptions/expectations for all algorithms implemented.
> + */
> +#ifndef __LITTLE_ENDIAN /* TODO */
> +# error "SMB3 compression is only supported on little endian architectures"
> +#endif /* !__LITTLE_ENDIAN */
> +
> +#if BITS_PER_LONG < 64 /* TODO */
> +# error "SMB3 compression is only supported on 64-bit architectures"
> +#endif /* BITS_PER_LONG < 64 */
> +
> +/* Build time double checks (probably unnecessary) */
> +static_assert(sizeof(u16) == 2);
> +static_assert(sizeof(u32) == 4);
> +static_assert(sizeof(u64) == 8);
> +static_assert(sizeof(size_t) == 8);

Steve pushed this change into -next as commit 20a248d5e77a ("smb:
client: compress: add compress/common.h"), where it breaks the build for
many 32-bit architectures for obvious reasons:

  $ make -skj"$(nproc)" ARCH=arm CROSS_COMPILE=arm-linux-gnueabi- mrproper allmodconfig fs/smb/
  In file included from fs/smb/client/compress/lz77.c:14:
  fs/smb/client/compress/common.h:34:3: error: #error "SMB3 compression is only supported on 64-bit architectures"
     34 | # error "SMB3 compression is only supported on 64-bit architectures"
        |   ^~~~~
  In file included from include/linux/init.h:5,
                   from include/linux/printk.h:6,
                   from include/asm-generic/bug.h:31,
                   from arch/arm/include/asm/bug.h:60,
                   from include/linux/bug.h:5,
                   from include/linux/slab.h:15,
                   from fs/smb/client/compress/lz77.c:9:
  include/linux/build_bug.h:80:41: error: static assertion failed: "sizeof(size_t) == 8"
     80 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
        |                                         ^~~~~~~~~~~~~~
  include/linux/build_bug.h:79:34: note: in expansion of macro '__static_assert'
     79 | #define static_assert(expr, ...) __static_assert(expr, ##__VA_ARGS__, #expr)
        |                                  ^~~~~~~~~~~~~~~
  fs/smb/client/compress/common.h:41:1: note: in expansion of macro 'static_assert'
     41 | static_assert(sizeof(size_t) == 8);
        | ^~~~~~~~~~~~~

It is unacceptable for the build to break in this manner in my opinion.
You should really use Kconfig to gate this code in a proper manner and
not emit any errors or warnings when it is unsupported.

Cheers,
Nathan

Re: [PATCH 7/8] smb: client: compress: add compress/common.h

[Steve French]

On Mon, Apr 20, 2026 at 4:24 PM Nathan Chancellor <nathan@kernel.org> wrote:
>
> Hi Enzo,
>
> On Mon, Apr 13, 2026 at 04:07:12PM -0300, Enzo Matsumiya wrote:
> > Add compress/common.h to aggregate helpers and definitions that will be
> > shared with other compression algorithms.
> >
> > Also add a few build time checks for proper support.
> >
> > Changes:
> > - update affected call paths in compress/lz77.c
> >
> > Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
> > ---
> ...
> > diff --git a/fs/smb/client/compress/common.h b/fs/smb/client/compress/common.h
> > new file mode 100644
> > index 000000000000..b5ccf5debd22
> > --- /dev/null
> > +++ b/fs/smb/client/compress/common.h
> ...
> > +/*
> > + * Build time checks/asserts.
> > + * These are assumptions/expectations for all algorithms implemented.
> > + */
> > +#ifndef __LITTLE_ENDIAN /* TODO */
> > +# error "SMB3 compression is only supported on little endian architectures"
> > +#endif /* !__LITTLE_ENDIAN */
> > +
> > +#if BITS_PER_LONG < 64 /* TODO */
> > +# error "SMB3 compression is only supported on 64-bit architectures"
> > +#endif /* BITS_PER_LONG < 64 */
> > +
> > +/* Build time double checks (probably unnecessary) */
> > +static_assert(sizeof(u16) == 2);
> > +static_assert(sizeof(u32) == 4);
> > +static_assert(sizeof(u64) == 8);
> > +static_assert(sizeof(size_t) == 8);
>
> Steve pushed this change into -next as commit 20a248d5e77a ("smb:
> client: compress: add compress/common.h"), where it breaks the build for
> many 32-bit architectures for obvious reasons:
>
>   $ make -skj"$(nproc)" ARCH=arm CROSS_COMPILE=arm-linux-gnueabi- mrproper allmodconfig fs/smb/
>   In file included from fs/smb/client/compress/lz77.c:14:
>   fs/smb/client/compress/common.h:34:3: error: #error "SMB3 compression is only supported on 64-bit architectures"
>      34 | # error "SMB3 compression is only supported on 64-bit architectures"
>         |   ^~~~~
>   In file included from include/linux/init.h:5,
>                    from include/linux/printk.h:6,
>                    from include/asm-generic/bug.h:31,
>                    from arch/arm/include/asm/bug.h:60,
>                    from include/linux/bug.h:5,
>                    from include/linux/slab.h:15,
>                    from fs/smb/client/compress/lz77.c:9:
>   include/linux/build_bug.h:80:41: error: static assertion failed: "sizeof(size_t) == 8"
>      80 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
>         |                                         ^~~~~~~~~~~~~~
>   include/linux/build_bug.h:79:34: note: in expansion of macro '__static_assert'
>      79 | #define static_assert(expr, ...) __static_assert(expr, ##__VA_ARGS__, #expr)
>         |                                  ^~~~~~~~~~~~~~~
>   fs/smb/client/compress/common.h:41:1: note: in expansion of macro 'static_assert'
>      41 | static_assert(sizeof(size_t) == 8);
>         | ^~~~~~~~~~~~~
>
> It is unacceptable for the build to break in this manner in my opinion.
> You should really use Kconfig to gate this code in a proper manner and
> not emit any errors or warnings when it is unsupported.
>

Thx for pointing this out.  I have removed the patch from cifs-2.6.git for-next

Let me know if you spot any additional problems


-- 
Thanks,

Steve

[PATCH 8/8] smb: common: add SMB3_COMPRESS_MAX_ALGS

[Enzo Matsumiya]

Set it to number of currently defined algorithms (6 as of now).

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
---
 fs/smb/common/smb2pdu.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/smb/common/smb2pdu.h b/fs/smb/common/smb2pdu.h
index e482c86ceb00..cda71964ed01 100644
--- a/fs/smb/common/smb2pdu.h
+++ b/fs/smb/common/smb2pdu.h
@@ -510,6 +510,8 @@ struct smb2_encryption_neg_context {
 /* Pattern scanning algorithm See MS-SMB2 3.1.4.4.1 */
 #define SMB3_COMPRESS_PATTERN	cpu_to_le16(0x0004) /* Pattern_V1 */
 #define SMB3_COMPRESS_LZ4	cpu_to_le16(0x0005)
+/* Account for NONE for easier array indexing */
+#define SMB3_COMPRESS_MAX_ALGS	6
 
 /* Compression Flags */
 #define SMB2_COMPRESSION_CAPABILITIES_FLAG_NONE		cpu_to_le32(0x00000000)
-- 
2.53.0