workqueue: skip the node_nr_active update for non-unbound workqueuesgrafted

apply_wqattrs_commit() updates node_nr_active->max unconditionally.
wq->node_nr_active[] is only allocated for unbound workqueues, so guard
the call before per-cpu workqueues start using this path.

No functional change: only unbound workqueues reach apply_wqattrs_*()
today.

Signed-off-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Tejun Heo <tj@kernel.org>

4 files changed, 934 insertions, 0 deletions

diff --git a/fs/smb/common/compress/compress.c b/fs/smb/common/compress/compress.c
new file mode 100644
index 000000000..b07a31759
--- /dev/null
+++ b/fs/smb/common/compress/compress.c
@@ -0,0 +1,399 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * SMB2 compression transform helpers.
+ *
+ * Copyright (C) 2026 Namjae Jeon <linkinjeon@kernel.org>
+ */
+#include <linux/module.h>
+#include <linux/overflow.h>
+#include <linux/string.h>
+#include <linux/unaligned.h>
+
+#include "compress.h"
+#include "lz77.h"
+
+#define SMB2_COMPRESSION_CHAINED_HDR_LEN \
+	offsetof(struct smb2_compression_hdr, CompressionAlgorithm)
+#define SMB2_COMPRESSION_PAYLOAD_BASE_LEN \
+	(sizeof(struct smb2_compression_payload_hdr) - sizeof(__le32))
+
+/*
+ * A NONE payload carries bytes verbatim. Keep both cursors and remaining
+ * lengths together so every chained payload handler applies identical bounds
+ * accounting.
+ */
+static int smb_decompress_none(const u8 **src, u32 *slen, u8 **dst, u32 *dlen,
+			       u32 len)
+{
+	if (len > *slen || len > *dlen)
+		return -EINVAL;
+
+	memcpy(*dst, *src, len);
+	*src += len;
+	*slen -= len;
+	*dst += len;
+	*dlen -= len;
+	return 0;
+}
+
+/*
+ * Pattern_V1 represents a run of one byte. Its wire payload is always the
+ * fixed-size smb2_compression_pattern_v1 structure.
+ */
+static int smb_decompress_pattern(const u8 **src, u32 *slen, u8 **dst,
+				  u32 *dlen, u32 len)
+{
+	const struct smb2_compression_pattern_v1 *pattern;
+	u32 repetitions;
+
+	if (len != sizeof(*pattern) || len > *slen)
+		return -EINVAL;
+
+	pattern = (const struct smb2_compression_pattern_v1 *)*src;
+	repetitions = le32_to_cpu(pattern->Repetitions);
+	if (repetitions > *dlen)
+		return -EINVAL;
+
+	memset(*dst, pattern->Pattern, repetitions);
+	*src += len;
+	*slen -= len;
+	*dst += repetitions;
+	*dlen -= repetitions;
+	return 0;
+}
+
+/*
+ * LZ77 payload Length includes the four-byte OriginalPayloadSize field.
+ * Consume that field before passing the compressed stream to the raw codec.
+ */
+static int smb_decompress_lz77_payload(const u8 **src, u32 *slen, u8 **dst,
+				       u32 *dlen, u32 len)
+{
+	u32 orig_size;
+	int rc;
+
+	if (len < sizeof(__le32) || len > *slen)
+		return -EINVAL;
+
+	orig_size = get_unaligned_le32(*src);
+	if (orig_size > *dlen)
+		return -EINVAL;
+
+	*src += sizeof(__le32);
+	*slen -= sizeof(__le32);
+	len -= sizeof(__le32);
+
+	rc = smb_lz77_decompress(*src, len, *dst, orig_size);
+	if (rc)
+		return rc;
+
+	*src += len;
+	*slen -= len;
+	*dst += orig_size;
+	*dlen -= orig_size;
+	return 0;
+}
+
+static int smb_decompress_chained(__le16 alg, bool allow_chained,
+				  const struct smb2_compression_hdr *hdr,
+				  u32 slen, void *dst, u32 dlen)
+{
+	const struct smb2_compression_payload_hdr *payload;
+	const u8 *src = (const u8 *)hdr + SMB2_COMPRESSION_CHAINED_HDR_LEN;
+	u32 orig_size = le32_to_cpu(hdr->OriginalCompressedSegmentSize);
+	u32 remaining = slen - SMB2_COMPRESSION_CHAINED_HDR_LEN;
+	u8 *out = dst;
+	u32 out_remaining = dlen;
+	bool first = true;
+	int rc;
+
+	if (!allow_chained || orig_size != dlen)
+		return -EINVAL;
+
+	/*
+	 * The chained transform has an eight-byte top-level header. The next
+	 * bytes are a sequence of payload headers whose Length fields account
+	 * for payload data, including OriginalPayloadSize where applicable.
+	 */
+	while (remaining) {
+		__le16 payload_alg;
+		__le16 flags;
+		u32 len;
+
+		if (remaining < SMB2_COMPRESSION_PAYLOAD_BASE_LEN)
+			return -EINVAL;
+
+		payload = (const struct smb2_compression_payload_hdr *)src;
+		payload_alg = payload->CompressionAlgorithm;
+		flags = payload->Flags;
+		len = le32_to_cpu(payload->Length);
+
+		/*
+		 * CHAINED marks only the first payload. Requiring NONE on every
+		 * later payload rejects ambiguous or independently chained data.
+		 */
+		if ((first && flags != cpu_to_le16(SMB2_COMPRESSION_FLAG_CHAINED)) ||
+		    (!first && flags != cpu_to_le16(SMB2_COMPRESSION_FLAG_NONE)))
+			return -EINVAL;
+
+		src += SMB2_COMPRESSION_PAYLOAD_BASE_LEN;
+		remaining -= SMB2_COMPRESSION_PAYLOAD_BASE_LEN;
+
+		if (payload_alg == SMB3_COMPRESS_NONE) {
+			rc = smb_decompress_none(&src, &remaining, &out,
+						 &out_remaining, len);
+		} else if (payload_alg == SMB3_COMPRESS_PATTERN) {
+			rc = smb_decompress_pattern(&src, &remaining, &out,
+						    &out_remaining, len);
+		} else if (payload_alg == alg && alg == SMB3_COMPRESS_LZ77) {
+			rc = smb_decompress_lz77_payload(&src, &remaining, &out,
+							 &out_remaining, len);
+		} else {
+			return -EINVAL;
+		}
+		if (rc)
+			return rc;
+		first = false;
+	}
+
+	return out_remaining ? -EINVAL : 0;
+}
+
+static int smb_decompress_unchained(__le16 alg,
+				    const struct smb2_compression_hdr *hdr,
+				    u32 slen, void *dst, u32 dlen)
+{
+	u32 orig_size, offset, comp_size;
+
+	if (hdr->CompressionAlgorithm != alg ||
+	    !smb_compress_alg_valid(hdr->CompressionAlgorithm, false))
+		return -EINVAL;
+
+	orig_size = le32_to_cpu(hdr->OriginalCompressedSegmentSize);
+	offset = le32_to_cpu(hdr->Offset);
+	if (offset > slen - sizeof(*hdr) || offset > dlen ||
+	    orig_size > dlen - offset || orig_size + offset != dlen)
+		return -EINVAL;
+
+	memcpy(dst, (const u8 *)hdr + sizeof(*hdr), offset);
+	comp_size = slen - sizeof(*hdr) - offset;
+	return smb_lz77_decompress((const u8 *)hdr + sizeof(*hdr) + offset,
+				   comp_size, (u8 *)dst + offset, orig_size);
+}
+
+/**
+ * smb_compression_decompress() - decode an SMB2 compression transform
+ * @alg: negotiated general-purpose compression algorithm
+ * @allow_chained: whether chained transforms were negotiated
+ * @src: transform header followed by compressed payload data
+ * @slen: total number of bytes available at @src
+ * @dst: output buffer for the reconstructed SMB2 message
+ * @dlen: exact expected size of the reconstructed SMB2 message
+ *
+ * Validate the transform type and negotiated capabilities before dispatching
+ * to the chained or unchained decoder. The caller supplies the expected output
+ * size after applying its transport-specific message size limits.
+ *
+ * Return: 0 on success, otherwise a negative errno.
+ */
+int smb_compression_decompress(__le16 alg, bool allow_chained,
+			       const void *src, u32 slen, void *dst, u32 dlen)
+{
+	const struct smb2_compression_hdr *hdr = src;
+
+	if (!src || !dst || slen < sizeof(*hdr) ||
+	    hdr->ProtocolId != SMB2_COMPRESSION_TRANSFORM_ID ||
+	    alg == SMB3_COMPRESS_NONE)
+		return -EINVAL;
+
+	if (hdr->Flags == cpu_to_le16(SMB2_COMPRESSION_FLAG_CHAINED))
+		return smb_decompress_chained(alg, allow_chained, hdr, slen,
+					      dst, dlen);
+
+	if (hdr->Flags != cpu_to_le16(SMB2_COMPRESSION_FLAG_NONE))
+		return -EINVAL;
+
+	return smb_decompress_unchained(alg, hdr, slen, dst, dlen);
+}
+EXPORT_SYMBOL_GPL(smb_compression_decompress);
+
+struct smb_compression_builder {
+	u8 *pos;
+	u32 remaining;
+	bool first;
+};
+
+/*
+ * Reserve one chained payload header and initialize its common fields.
+ * OriginalPayloadSize is present only for LZNT1/LZ77/LZ77+Huffman payloads.
+ */
+static struct smb2_compression_payload_hdr *
+smb_compression_add_payload(struct smb_compression_builder *builder,
+			    __le16 alg, u32 payload_len, bool orig_size)
+{
+	struct smb2_compression_payload_hdr *payload;
+	u32 hdr_len = SMB2_COMPRESSION_PAYLOAD_BASE_LEN;
+	u32 total_len;
+
+	if (orig_size)
+		hdr_len += sizeof(payload->OriginalPayloadSize);
+	if (check_add_overflow(hdr_len, payload_len, &total_len) ||
+	    total_len > builder->remaining)
+		return NULL;
+
+	payload = (struct smb2_compression_payload_hdr *)builder->pos;
+	payload->CompressionAlgorithm = alg;
+	payload->Flags = cpu_to_le16(builder->first ?
+		SMB2_COMPRESSION_FLAG_CHAINED : SMB2_COMPRESSION_FLAG_NONE);
+	payload->Length = cpu_to_le32(payload_len +
+		(orig_size ? sizeof(payload->OriginalPayloadSize) : 0));
+
+	builder->pos += hdr_len;
+	builder->remaining -= hdr_len;
+	builder->first = false;
+	return payload;
+}
+
+static int smb_compression_add_pattern(struct smb_compression_builder *builder,
+				       u8 pattern, u32 repetitions)
+{
+	struct smb2_compression_pattern_v1 *payload;
+
+	if (!smb_compression_add_payload(builder, SMB3_COMPRESS_PATTERN,
+					 sizeof(*payload), false))
+		return -ENOSPC;
+
+	payload = (struct smb2_compression_pattern_v1 *)builder->pos;
+	payload->Pattern = pattern;
+	payload->Reserved1 = 0;
+	payload->Reserved2 = 0;
+	payload->Repetitions = cpu_to_le32(repetitions);
+	builder->pos += sizeof(*payload);
+	builder->remaining -= sizeof(*payload);
+	return 0;
+}
+
+static int smb_compression_add_none(struct smb_compression_builder *builder,
+				    const u8 *src, u32 len)
+{
+	if (!smb_compression_add_payload(builder, SMB3_COMPRESS_NONE, len, false))
+		return -ENOSPC;
+
+	memcpy(builder->pos, src, len);
+	builder->pos += len;
+	builder->remaining -= len;
+	return 0;
+}
+
+static int smb_compression_add_lz77(struct smb_compression_builder *builder,
+				    const u8 *src, u32 len)
+{
+	struct smb2_compression_payload_hdr *payload;
+	u32 comp_len;
+	int rc;
+
+	if (builder->remaining <= sizeof(*payload))
+		return -ENOSPC;
+
+	comp_len = builder->remaining - sizeof(*payload);
+	payload = smb_compression_add_payload(builder, SMB3_COMPRESS_LZ77,
+					      comp_len, true);
+	if (!payload)
+		return -ENOSPC;
+
+	rc = smb_lz77_compress(src, len, builder->pos, &comp_len);
+	if (rc)
+		return rc;
+
+	payload->Length = cpu_to_le32(comp_len +
+				      sizeof(payload->OriginalPayloadSize));
+	payload->OriginalPayloadSize = cpu_to_le32(len);
+	builder->pos += comp_len;
+	builder->remaining -= comp_len;
+	return 0;
+}
+
+/**
+ * smb_compression_compress_chained() - build a chained SMB2 transform
+ * @alg: negotiated general-purpose compression algorithm
+ * @allow_pattern: whether Pattern_V1 was negotiated
+ * @src: complete uncompressed SMB2 message
+ * @slen: size of @src
+ * @dst: output buffer for the transform
+ * @dlen: input capacity of @dst and output transform size
+ *
+ * Following the algorithm in [MS-SMB2] 3.1.4.4, encode sufficiently long
+ * repeated runs at the front and back as Pattern_V1 payloads. Compress a
+ * middle region larger than 1 KiB with LZ77; smaller middle regions are
+ * represented by a chained NONE payload.
+ *
+ * This helper does not decide whether the final transform is smaller than the
+ * original message. The transport caller owns that policy decision.
+ *
+ * Return: 0 on success, otherwise a negative errno.
+ */
+int smb_compression_compress_chained(__le16 alg, bool allow_pattern,
+				     const void *src, u32 slen,
+				     void *dst, u32 *dlen)
+{
+	struct smb2_compression_hdr *hdr = dst;
+	struct smb_compression_builder builder;
+	const u8 *input = src;
+	u32 forward = 0, backward = 0, middle_len;
+	int rc;
+
+	if (!src || !dst || !dlen || alg != SMB3_COMPRESS_LZ77 ||
+	    *dlen <= SMB2_COMPRESSION_CHAINED_HDR_LEN || !slen)
+		return -EINVAL;
+
+	hdr->ProtocolId = SMB2_COMPRESSION_TRANSFORM_ID;
+	hdr->OriginalCompressedSegmentSize = cpu_to_le32(slen);
+	builder.pos = (u8 *)dst + SMB2_COMPRESSION_CHAINED_HDR_LEN;
+	builder.remaining = *dlen - SMB2_COMPRESSION_CHAINED_HDR_LEN;
+	builder.first = true;
+
+	if (allow_pattern && slen > 32) {
+		for (forward = 1; forward < slen; forward++) {
+			if (input[forward] != input[0])
+				break;
+		}
+		if (forward <= 32)
+			forward = 0;
+
+		for (backward = 1; backward < slen - forward; backward++) {
+			if (input[slen - backward - 1] != input[slen - 1])
+				break;
+		}
+		if (backward <= 32)
+			backward = 0;
+	}
+
+	if (forward) {
+		rc = smb_compression_add_pattern(&builder, input[0], forward);
+		if (rc)
+			return rc;
+	}
+
+	middle_len = slen - forward - backward;
+	if (middle_len > 1024)
+		rc = smb_compression_add_lz77(&builder, input + forward,
+					      middle_len);
+	else if (middle_len)
+		rc = smb_compression_add_none(&builder,
+					      input + forward, middle_len);
+	else
+		rc = 0;
+	if (rc)
+		return rc;
+
+	if (backward) {
+		rc = smb_compression_add_pattern(&builder, input[slen - 1],
+						 backward);
+		if (rc)
+			return rc;
+	}
+
+	*dlen = builder.pos - (u8 *)dst;
+	return 0;
+}
+EXPORT_SYMBOL_GPL(smb_compression_compress_chained);
diff --git a/fs/smb/common/compress/compress.h b/fs/smb/common/compress/compress.h
new file mode 100644
index 000000000..7ace3bf4b
--- /dev/null
+++ b/fs/smb/common/compress/compress.h
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2026 Namjae Jeon <linkinjeon@kernel.org>
+ */
+#ifndef _COMMON_SMB_COMPRESS_H
+#define _COMMON_SMB_COMPRESS_H
+
+#include "../smb2pdu.h"
+
+/*
+ * SMB3_COMPRESS_NONE is valid only in chained payload headers. It is never
+ * negotiated as a compression algorithm.
+ */
+static __always_inline bool smb_compress_alg_valid(__le16 alg, bool valid_none)
+{
+	if (alg == SMB3_COMPRESS_NONE)
+		return valid_none;
+
+	return alg == SMB3_COMPRESS_LZ77 || alg == SMB3_COMPRESS_PATTERN;
+}
+
+int smb_compression_decompress(__le16 alg, bool allow_chained,
+			       const void *src, u32 slen, void *dst, u32 dlen);
+int smb_compression_compress_chained(__le16 alg, bool allow_pattern,
+				     const void *src, u32 slen,
+				     void *dst, u32 *dlen);
+
+#endif /* _COMMON_SMB_COMPRESS_H */
diff --git a/fs/smb/common/compress/lz77.c b/fs/smb/common/compress/lz77.c
new file mode 100644
index 000000000..9216d973d
--- /dev/null
+++ b/fs/smb/common/compress/lz77.c
@@ -0,0 +1,461 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2024-2026, SUSE LLC
+ * Copyright (C) 2026 Namjae Jeon <linkinjeon@kernel.org>
+ *
+ * Authors: Enzo Matsumiya <ematsumiya@suse.de>
+ *          Namjae Jeon <linkinjeon@kernel.org>
+ *
+ * Implementation of the LZ77 "plain" compression algorithm, as per MS-XCA spec.
+ */
+#include <linux/slab.h>
+#include <linux/sizes.h>
+#include <linux/count_zeros.h>
+#include <linux/unaligned.h>
+#include <linux/module.h>
+#include <linux/overflow.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
+ *				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
+#define LZ77_HASH_SIZE		BIT(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
+ * @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;
+	dist--;
+	dist <<= 3;
+
+	if (len < 7) {
+		lz77_write16(dst, dist + len);
+
+		return dst + sizeof(u16);
+	}
+
+	dist |= 7;
+	lz77_write16(dst, dist);
+	dst += sizeof(u16);
+	len -= 7;
+
+	if (!*nib) {
+		lz77_write8(dst, umin(len, 15));
+		*nib = dst;
+		dst++;
+	} else {
+		u8 *b = *nib;
+
+		lz77_write8(b, *b | umin(len, 15) << 4);
+		*nib = NULL;
+	}
+
+	if (len < 15)
+		return dst;
+
+	len -= 15;
+	if (len < 255) {
+		lz77_write8(dst, len);
+
+		return dst + 1;
+	}
+
+	lz77_write8(dst, 0xff);
+	dst++;
+	len += 7 + 15;
+	if (len <= 0xffff) {
+		lz77_write16(dst, len);
+
+		return dst + sizeof(u16);
+	}
+
+	lz77_write16(dst, 0);
+	dst += sizeof(u16);
+	lz77_write32(dst, len);
+
+	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)
+{
+	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 += sizeof(u32);
+		}
+	} while (start < end);
+
+	return dst;
+}
+
+static __always_inline u32 lz77_hash(const u32 v)
+{
+	return ((v ^ 0x9E3779B9) * 0x85EBCA6B) >> (32 - LZ77_HASH_LOG);
+}
+
+noinline int smb_lz77_compress(const void *src, const u32 slen,
+			       void *dst, u32 *dlen)
+{
+	const void *srcp, *rlim, *end, *anchor;
+	u32 *htable, hash, flag_count = 0;
+	void *dstp, *nib, *flag_pos;
+	long flag = 0;
+
+	/* This is probably a bug, so throw a warning. */
+	if (WARN_ON_ONCE(*dlen < smb_lz77_compressed_alloc_size(slen)))
+		return -EINVAL;
+
+	srcp = src;
+	anchor = src;
+	end = srcp + slen; /* absolute end */
+	rlim = end - LZ77_MSTEP_SIZE; /* read limit (for lz77_match_len()) */
+	dstp = dst;
+	flag_pos = dstp;
+	dstp += sizeof(u32);
+	nib = NULL;
+
+	htable = kvcalloc(LZ77_HASH_SIZE, sizeof(*htable), GFP_KERNEL);
+	if (!htable)
+		return -ENOMEM;
+
+	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));
+
+	/*
+	 * Main loop.
+	 *
+	 * @dlen is >= smb_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;
+
+			/*
+			 * 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;
+
+			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));
+
+		/*
+		 * 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);
+		srcp += len;
+		anchor = srcp;
+
+		LZ77_PREFETCH(srcp);
+
+		flag = (flag << 1) | 1;
+		flag_count++;
+		if (flag_count == LZ77_FLAG_MAX) {
+			lz77_write32(flag_pos, flag);
+			flag_count = 0;
+			flag_pos = dstp;
+			dstp += sizeof(u32);
+		}
+
+		if (unlikely(srcp > rlim))
+			break;
+
+		/* 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_count = LZ77_FLAG_MAX - flag_count;
+	flag <<= flag_count;
+	flag |= (1UL << flag_count) - 1;
+	lz77_write32(flag_pos, flag);
+
+	*dlen = dstp - dst;
+	kvfree(htable);
+
+	if (*dlen < slen)
+		return 0;
+
+	return -EMSGSIZE;
+}
+EXPORT_SYMBOL_GPL(smb_lz77_compress);
+
+static int lz77_decode_match_len(const u8 **src, const u8 *end, u16 token,
+				 u8 *nibble, bool *have_nibble, u32 *len)
+{
+	u8 extra;
+
+	*len = (token & 0x7) + 3;
+	if ((token & 0x7) != 0x7)
+		return 0;
+
+	if (!*have_nibble) {
+		if (*src >= end)
+			return -EINVAL;
+		*nibble = *(*src)++;
+		extra = *nibble & 0xf;
+		*have_nibble = true;
+	} else {
+		extra = *nibble >> 4;
+		*have_nibble = false;
+	}
+
+	*len += extra;
+	if (extra == 0xf) {
+		u8 b;
+
+		if (*src >= end)
+			return -EINVAL;
+		b = *(*src)++;
+		if (b != 0xff) {
+			*len += b;
+		} else {
+			u16 w;
+
+			if (end - *src < 2)
+				return -EINVAL;
+			w = get_unaligned_le16(*src);
+			*src += 2;
+			if (w) {
+				*len = w + 3;
+			} else {
+				u32 long_len;
+
+				if (end - *src < 4)
+					return -EINVAL;
+				long_len = get_unaligned_le32(*src);
+				*src += 4;
+				if (check_add_overflow(long_len, 3, len))
+					return -EINVAL;
+			}
+		}
+	}
+
+	return 0;
+}
+
+int smb_lz77_decompress(const void *src, const u32 slen, void *dst,
+			const u32 dlen)
+{
+	const u8 *sp = src, *send = sp + slen;
+	u8 *dp = dst, *dend = dp + dlen;
+	u32 flags = 0;
+	int flag_count = 0;
+	u8 nibble = 0;
+	bool have_nibble = false;
+
+	while (dp < dend) {
+		u32 len, dist;
+		u16 token;
+
+		if (!flag_count) {
+			if (send - sp < 4)
+				return -EINVAL;
+			flags = get_unaligned_le32(sp);
+			sp += 4;
+			flag_count = 32;
+		}
+
+		if (!(flags & 0x80000000)) {
+			if (sp >= send)
+				return -EINVAL;
+			*dp++ = *sp++;
+			flags <<= 1;
+			flag_count--;
+			continue;
+		}
+
+		flags <<= 1;
+		flag_count--;
+
+		if (send - sp < 2)
+			return -EINVAL;
+
+		token = get_unaligned_le16(sp);
+		sp += 2;
+
+		dist = (token >> 3) + 1;
+		if (dist > dp - (u8 *)dst)
+			return -EINVAL;
+
+		if (lz77_decode_match_len(&sp, send, token, &nibble,
+					  &have_nibble, &len))
+			return -EINVAL;
+
+		if (len > dend - dp)
+			return -EINVAL;
+
+		while (len--) {
+			*dp = *(dp - dist);
+			dp++;
+		}
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(smb_lz77_decompress);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("SMB plain LZ77 compression");
diff --git a/fs/smb/common/compress/lz77.h b/fs/smb/common/compress/lz77.h
new file mode 100644
index 000000000..e032c0f1b
--- /dev/null
+++ b/fs/smb/common/compress/lz77.h
@@ -0,0 +1,46 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2024-2026, SUSE LLC
+ *
+ * Authors: Enzo Matsumiya <ematsumiya@suse.de>
+ *
+ * Implementation of the LZ77 "plain" compression algorithm, as per MS-XCA spec.
+ */
+#ifndef _SMB_COMPRESS_LZ77_H
+#define _SMB_COMPRESS_LZ77_H
+
+#include <linux/kernel.h>
+
+/**
+ * smb_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 smb_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 smb_lz77_compress()
+ */
+static __always_inline u32 smb_lz77_compressed_alloc_size(const u32 size)
+{
+	return size + (size >> 3) + 8;
+}
+
+int smb_lz77_compress(const void *src, const u32 slen, void *dst, u32 *dlen);
+int smb_lz77_decompress(const void *src, const u32 slen, void *dst,
+			const u32 dlen);
+#endif /* _SMB_COMPRESS_LZ77_H */