smb3: mark compression as CONFIG_EXPERIMENTAL and fix missing compression operation

Move SMB3.1.1 compression code into experimental config option,
and fix the compress mount option. Implement unchained LZ77
"plain" compression algorithm as per MS-XCA specification
section "2.3 Plain LZ77 Compression Algorithm Details".

Signed-off-by: Enzo Matsumiya <ematsumiya@suse.de>
Signed-off-by: Steve French <stfrench@microsoft.com>
This commit is contained in:
Steve French 2024-07-26 16:30:23 -05:00
parent 6795dab403
commit d14bbfff25
9 changed files with 685 additions and 4 deletions

View File

@ -204,4 +204,18 @@ config CIFS_ROOT
Most people say N here.
config CIFS_COMPRESSION
bool "SMB message compression (Experimental)"
depends on CIFS
default n
help
Enables over-the-wire message compression for SMB 3.1.1
mounts when negotiated with the server.
Only write requests with data size >= PAGE_SIZE will be
compressed to avoid wasting resources.
Say Y here if you want SMB traffic to be compressed.
If unsure, say N.
endif

View File

@ -33,3 +33,5 @@ cifs-$(CONFIG_CIFS_SMB_DIRECT) += smbdirect.o
cifs-$(CONFIG_CIFS_ROOT) += cifsroot.o
cifs-$(CONFIG_CIFS_ALLOW_INSECURE_LEGACY) += smb1ops.o cifssmb.o
cifs-$(CONFIG_CIFS_COMPRESSION) += compress.o compress/lz77.o

View File

@ -349,6 +349,9 @@ static int cifs_debug_data_proc_show(struct seq_file *m, void *v)
seq_printf(m, ",ACL");
#ifdef CONFIG_CIFS_SWN_UPCALL
seq_puts(m, ",WITNESS");
#endif
#ifdef CONFIG_CIFS_COMPRESSION
seq_puts(m, ",COMPRESSION");
#endif
seq_putc(m, '\n');
seq_printf(m, "CIFSMaxBufSize: %d\n", CIFSMaxBufSize);
@ -475,7 +478,9 @@ skip_rdma:
}
seq_puts(m, "\nCompression: ");
if (!server->compression.requested)
if (!IS_ENABLED(CONFIG_CIFS_COMPRESSION))
seq_puts(m, "no built-in support");
else if (!server->compression.requested)
seq_puts(m, "disabled on mount");
else if (server->compression.enabled)
seq_printf(m, "enabled (%s)", compression_alg_str(server->compression.alg));

View File

@ -555,7 +555,7 @@ struct smb_version_operations {
bool (*dir_needs_close)(struct cifsFileInfo *);
long (*fallocate)(struct file *, struct cifs_tcon *, int, loff_t,
loff_t);
/* init transform request - used for encryption for now */
/* init transform (compress/encrypt) request */
int (*init_transform_rq)(struct TCP_Server_Info *, int num_rqst,
struct smb_rqst *, struct smb_rqst *);
int (*is_transform_hdr)(void *buf);
@ -1876,6 +1876,7 @@ static inline bool is_replayable_error(int error)
#define CIFS_HAS_CREDITS 0x0400 /* already has credits */
#define CIFS_TRANSFORM_REQ 0x0800 /* transform request before sending */
#define CIFS_NO_SRV_RSP 0x1000 /* there is no server response */
#define CIFS_COMPRESS_REQ 0x4000 /* compress request before sending */
/* Security Flags: indicate type of session setup needed */
#define CIFSSEC_MAY_SIGN 0x00001

50
fs/smb/client/compress.c Normal file
View File

@ -0,0 +1,50 @@
// SPDX-License-Identifier: GPL-2.0-only
/*
* Copyright (C) 2024, SUSE LLC
*
* Authors: Enzo Matsumiya <ematsumiya@suse.de>
*
* This file implements I/O compression support for SMB2 messages (SMB 3.1.1 only).
* See compress/ for implementation details of each algorithm.
*
* References:
* MS-SMB2 "3.1.4.4 Compressing the Message"
* MS-SMB2 "3.1.5.3 Decompressing the Chained Message"
* MS-XCA - for details of the supported algorithms
*/
#include <linux/slab.h>
#include <linux/kernel.h>
#include <linux/uio.h>
#include "cifsglob.h"
#include "../common/smb2pdu.h"
#include "cifsproto.h"
#include "smb2proto.h"
#include "compress/lz77.h"
#include "compress.h"
int smb_compress(void *buf, const void *data, size_t *len)
{
struct smb2_compression_hdr *hdr;
size_t buf_len, data_len;
int ret;
buf_len = sizeof(struct smb2_write_req);
data_len = *len;
*len = 0;
hdr = buf;
hdr->ProtocolId = SMB2_COMPRESSION_TRANSFORM_ID;
hdr->OriginalCompressedSegmentSize = cpu_to_le32(data_len);
hdr->Offset = cpu_to_le32(buf_len);
hdr->Flags = SMB2_COMPRESSION_FLAG_NONE;
hdr->CompressionAlgorithm = SMB3_COMPRESS_LZ77;
/* XXX: add other algs here as they're implemented */
ret = lz77_compress(data, data_len, buf + SMB_COMPRESS_HDR_LEN + buf_len, &data_len);
if (!ret)
*len = SMB_COMPRESS_HDR_LEN + buf_len + data_len;
return ret;
}

109
fs/smb/client/compress.h Normal file
View File

@ -0,0 +1,109 @@
/* SPDX-License-Identifier: GPL-2.0-only */
/*
* Copyright (C) 2024, SUSE LLC
*
* Authors: Enzo Matsumiya <ematsumiya@suse.de>
*
* This file implements I/O compression support for SMB2 messages (SMB 3.1.1 only).
* See compress/ for implementation details of each algorithm.
*
* References:
* MS-SMB2 "3.1.4.4 Compressing the Message" - for compression details
* MS-SMB2 "3.1.5.3 Decompressing the Chained Message" - for decompression details
* MS-XCA - for details of the supported algorithms
*/
#ifndef _SMB_COMPRESS_H
#define _SMB_COMPRESS_H
#include <linux/uio.h>
#include <linux/kernel.h>
#include "../common/smb2pdu.h"
#include "cifsglob.h"
/* sizeof(smb2_compression_hdr) - sizeof(OriginalPayloadSize) */
#define SMB_COMPRESS_HDR_LEN 16
/* sizeof(smb2_compression_payload_hdr) - sizeof(OriginalPayloadSize) */
#define SMB_COMPRESS_PAYLOAD_HDR_LEN 8
#define SMB_COMPRESS_MIN_LEN PAGE_SIZE
struct smb_compress_ctx {
struct TCP_Server_Info *server;
struct work_struct work;
struct mid_q_entry *mid;
void *buf; /* compressed data */
void *data; /* uncompressed data */
size_t len;
};
#ifdef CONFIG_CIFS_COMPRESSION
int smb_compress(void *buf, const void *data, size_t *len);
/**
* smb_compress_alg_valid() - Validate a compression algorithm.
* @alg: Compression algorithm to check.
* @valid_none: Conditional check whether NONE algorithm should be
* considered valid or not.
*
* If @alg is SMB3_COMPRESS_NONE, this function returns @valid_none.
*
* Note that 'NONE' (0) compressor type is considered invalid in protocol
* negotiation, as it's never requested to/returned from the server.
*
* Return: true if @alg is valid/supported, false otherwise.
*/
static __always_inline int smb_compress_alg_valid(__le16 alg, bool valid_none)
{
if (alg == SMB3_COMPRESS_NONE)
return valid_none;
if (alg == SMB3_COMPRESS_LZ77 || alg == SMB3_COMPRESS_PATTERN)
return true;
return false;
}
/**
* should_compress() - Determines if a request (write) or the response to a
* request (read) should be compressed.
* @tcon: tcon of the request is being sent to
* @buf: buffer with an SMB2 READ/WRITE request
*
* Return: true iff:
* - compression was successfully negotiated with server
* - server has enabled compression for the share
* - it's a read or write request
* - if write, request length is >= SMB_COMPRESS_MIN_LEN
*
* Return false otherwise.
*/
static __always_inline bool should_compress(const struct cifs_tcon *tcon, const void *buf)
{
const struct smb2_hdr *shdr = buf;
if (!tcon || !tcon->ses || !tcon->ses->server)
return false;
if (!tcon->ses->server->compression.enabled)
return false;
if (!(tcon->share_flags & SMB2_SHAREFLAG_COMPRESS_DATA))
return false;
if (shdr->Command == SMB2_WRITE) {
const struct smb2_write_req *req = buf;
return (req->Length >= SMB_COMPRESS_MIN_LEN);
}
return (shdr->Command == SMB2_READ);
}
/*
* #else !CONFIG_CIFS_COMPRESSION ...
* These routines should not be called when CONFIG_CIFS_COMPRESSION disabled
* #define smb_compress(arg1, arg2, arg3) (-EOPNOTSUPP)
* #define smb_compress_alg_valid(arg1, arg2) (-EOPNOTSUPP)
* #define should_compress(arg1, arg2) (false)
*/
#endif /* !CONFIG_CIFS_COMPRESSION */
#endif /* _SMB_COMPRESS_H */

View File

@ -0,0 +1,211 @@
// SPDX-License-Identifier: GPL-2.0-only
/*
* Copyright (C) 2024, SUSE LLC
*
* Authors: Enzo Matsumiya <ematsumiya@suse.de>
*
* Implementation of the LZ77 "plain" compression algorithm, as per MS-XCA spec.
*/
#include <linux/slab.h>
#include "lz77.h"
static __always_inline u32 hash3(const u8 *ptr)
{
return lz77_hash32(lz77_read32(ptr) & 0xffffff, LZ77_HASH_LOG);
}
static u8 *write_match(u8 *dst, u8 **nib, u32 dist, u32 len)
{
len -= 3;
dist--;
dist <<= 3;
if (len < 7) {
lz77_write16(dst, dist + len);
return dst + 2;
}
dist |= 7;
lz77_write16(dst, dist);
dst += 2;
len -= 7;
if (!*nib) {
*nib = dst;
lz77_write8(dst, min_t(unsigned int, len, 15));
dst++;
} else {
**nib |= min_t(unsigned int, 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 + 2;
}
lz77_write16(dst, 0);
dst += 2;
lz77_write32(dst, len);
return dst + 4;
}
static u8 *write_literals(u8 *dst, const u8 *dst_end, const u8 *src, size_t count,
struct lz77_flags *flags)
{
const u8 *end = src + count;
while (src < end) {
size_t c = lz77_min(count, 32 - flags->count);
if (dst + c >= dst_end)
return ERR_PTR(-EFAULT);
if (lz77_copy(dst, src, c))
return ERR_PTR(-EFAULT);
dst += c;
src += c;
count -= c;
flags->val <<= c;
flags->count += c;
if (flags->count == 32) {
lz77_write32(flags->pos, flags->val);
flags->count = 0;
flags->pos = dst;
dst += 4;
}
}
return dst;
}
static __always_inline bool is_valid_match(const u32 dist, const u32 len)
{
return (dist >= LZ77_MATCH_MIN_DIST && dist < LZ77_MATCH_MAX_DIST) &&
(len >= LZ77_MATCH_MIN_LEN && len < LZ77_MATCH_MAX_LEN);
}
static __always_inline const u8 *find_match(u32 *htable, const u8 *base, const u8 *cur,
const u8 *end, u32 *best_len)
{
const u8 *match;
u32 hash;
size_t offset;
hash = hash3(cur);
offset = cur - base;
if (htable[hash] >= offset)
return cur;
match = base + htable[hash];
*best_len = lz77_match(match, cur, end);
if (is_valid_match(cur - match, *best_len))
return match;
return cur;
}
int lz77_compress(const u8 *src, size_t src_len, u8 *dst, size_t *dst_len)
{
const u8 *srcp, *src_end, *anchor;
struct lz77_flags flags = { 0 };
u8 *dstp, *dst_end, *nib;
u32 *htable;
int ret;
srcp = src;
anchor = srcp;
src_end = src + src_len;
dstp = dst;
dst_end = dst + *dst_len;
flags.pos = dstp;
nib = NULL;
memset(dstp, 0, *dst_len);
dstp += 4;
htable = kvcalloc(LZ77_HASH_SIZE, sizeof(u32), GFP_KERNEL);
if (!htable)
return -ENOMEM;
/* fill hashtable with invalid offsets */
memset(htable, 0xff, LZ77_HASH_SIZE * sizeof(u32));
/* from here on, any error is because @dst_len reached >= @src_len */
ret = -EMSGSIZE;
/* main loop */
while (srcp < src_end) {
u32 hash, dist, len;
const u8 *match;
while (srcp + 3 < src_end) {
len = LZ77_MATCH_MIN_LEN - 1;
match = find_match(htable, src, srcp, src_end, &len);
hash = hash3(srcp);
htable[hash] = srcp - src;
if (likely(match < srcp)) {
dist = srcp - match;
break;
}
srcp++;
}
dstp = write_literals(dstp, dst_end, anchor, srcp - anchor, &flags);
if (IS_ERR(dstp))
goto err_free;
if (srcp + 3 >= src_end)
goto leftovers;
dstp = write_match(dstp, &nib, dist, len);
srcp += len;
anchor = srcp;
flags.val = (flags.val << 1) | 1;
flags.count++;
if (flags.count == 32) {
lz77_write32(flags.pos, flags.val);
flags.count = 0;
flags.pos = dstp;
dstp += 4;
}
}
leftovers:
if (srcp < src_end) {
dstp = write_literals(dstp, dst_end, srcp, src_end - srcp, &flags);
if (IS_ERR(dstp))
goto err_free;
}
flags.val <<= (32 - flags.count);
flags.val |= (1 << (32 - flags.count)) - 1;
lz77_write32(flags.pos, flags.val);
*dst_len = dstp - dst;
ret = 0;
err_free:
kvfree(htable);
return ret;
}

View File

@ -0,0 +1,286 @@
/* SPDX-License-Identifier: GPL-2.0-only */
/*
* Copyright (C) 2024, SUSE LLC
*
* Authors: Enzo Matsumiya <ematsumiya@suse.de>
*
* Definitions and optmized helpers for LZ77 compression.
*/
#ifndef _SMB_COMPRESS_LZ77_H
#define _SMB_COMPRESS_LZ77_H
#include <linux/uaccess.h>
#ifdef CONFIG_CIFS_COMPRESSION
#include <asm/ptrace.h>
#include <linux/kernel.h>
#include <linux/string.h>
#ifndef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
#include <asm-generic/unaligned.h>
#endif
#define LZ77_HASH_LOG 13
#define LZ77_HASH_SIZE (1 << LZ77_HASH_LOG)
#define LZ77_HASH_MASK lz77_hash_mask(LZ77_HASH_LOG)
/* We can increase this for better compression (but worse performance). */
#define LZ77_MATCH_MIN_LEN 3
/* From MS-XCA, but it's arbitrarily chosen. */
#define LZ77_MATCH_MAX_LEN S32_MAX
/*
* Check this to ensure we don't match the current position, which would
* end up doing a verbatim copy of the input, and actually overflowing
* the output buffer because of the encoded metadata.
*/
#define LZ77_MATCH_MIN_DIST 1
/* How far back in the buffer can we try to find a match (i.e. window size) */
#define LZ77_MATCH_MAX_DIST 8192
#define LZ77_STEPSIZE_16 sizeof(u16)
#define LZ77_STEPSIZE_32 sizeof(u32)
#define LZ77_STEPSIZE_64 sizeof(u64)
struct lz77_flags {
u8 *pos;
size_t count;
long val;
};
static __always_inline u32 lz77_hash_mask(const unsigned int log2)
{
return ((1 << log2) - 1);
}
static __always_inline u32 lz77_hash64(const u64 v, const unsigned int log2)
{
const u64 prime5bytes = 889523592379ULL;
return (u32)(((v << 24) * prime5bytes) >> (64 - log2));
}
static __always_inline u32 lz77_hash32(const u32 v, const unsigned int log2)
{
return ((v * 2654435769LL) >> (32 - log2)) & lz77_hash_mask(log2);
}
static __always_inline u32 lz77_log2(unsigned int x)
{
return x ? ((u32)(31 - __builtin_clz(x))) : 0;
}
#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
static __always_inline u8 lz77_read8(const void *ptr)
{
return *(u8 *)ptr;
}
static __always_inline u16 lz77_read16(const void *ptr)
{
return *(u16 *)ptr;
}
static __always_inline u32 lz77_read32(const void *ptr)
{
return *(u32 *)ptr;
}
static __always_inline u64 lz77_read64(const void *ptr)
{
return *(u64 *)ptr;
}
static __always_inline void lz77_write8(void *ptr, const u8 v)
{
*(u8 *)ptr = v;
}
static __always_inline void lz77_write16(void *ptr, const u16 v)
{
*(u16 *)ptr = v;
}
static __always_inline void lz77_write32(void *ptr, const u32 v)
{
*(u32 *)ptr = v;
}
static __always_inline void lz77_write64(void *ptr, const u64 v)
{
*(u64 *)ptr = v;
}
static __always_inline void lz77_write_ptr16(void *ptr, const void *vp)
{
*(u16 *)ptr = *(const u16 *)vp;
}
static __always_inline void lz77_write_ptr32(void *ptr, const void *vp)
{
*(u32 *)ptr = *(const u32 *)vp;
}
static __always_inline void lz77_write_ptr64(void *ptr, const void *vp)
{
*(u64 *)ptr = *(const u64 *)vp;
}
static __always_inline long lz77_copy(u8 *dst, const u8 *src, size_t count)
{
return copy_from_kernel_nofault(dst, src, count);
}
#else /* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
static __always_inline u8 lz77_read8(const void *ptr)
{
return get_unaligned((u8 *)ptr);
}
static __always_inline u16 lz77_read16(const void *ptr)
{
return lz77_read8(ptr) | (lz77_read8(ptr + 1) << 8);
}
static __always_inline u32 lz77_read32(const void *ptr)
{
return lz77_read16(ptr) | (lz77_read16(ptr + 2) << 16);
}
static __always_inline u64 lz77_read64(const void *ptr)
{
return lz77_read32(ptr) | ((u64)lz77_read32(ptr + 4) << 32);
}
static __always_inline void lz77_write8(void *ptr, const u8 v)
{
put_unaligned(v, (u8 *)ptr);
}
static __always_inline void lz77_write16(void *ptr, const u16 v)
{
lz77_write8(ptr, v & 0xff);
lz77_write8(ptr + 1, (v >> 8) & 0xff);
}
static __always_inline void lz77_write32(void *ptr, const u32 v)
{
lz77_write16(ptr, v & 0xffff);
lz77_write16(ptr + 2, (v >> 16) & 0xffff);
}
static __always_inline void lz77_write64(void *ptr, const u64 v)
{
lz77_write32(ptr, v & 0xffffffff);
lz77_write32(ptr + 4, (v >> 32) & 0xffffffff);
}
static __always_inline void lz77_write_ptr16(void *ptr, const void *vp)
{
const u16 v = lz77_read16(vp);
lz77_write16(ptr, v);
}
static __always_inline void lz77_write_ptr32(void *ptr, const void *vp)
{
const u32 v = lz77_read32(vp);
lz77_write32(ptr, v);
}
static __always_inline void lz77_write_ptr64(void *ptr, const void *vp)
{
const u64 v = lz77_read64(vp);
lz77_write64(ptr, v);
}
static __always_inline long lz77_copy(u8 *dst, const u8 *src, size_t count)
{
memcpy(dst, src, count);
return 0;
}
#endif /* !CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS */
static __always_inline unsigned int __count_common_bytes(const unsigned long diff)
{
#ifdef __has_builtin
# if __has_builtin(__builtin_ctzll)
return (unsigned int)__builtin_ctzll(diff) >> 3;
# endif
#else
/* count trailing zeroes */
unsigned long bits = 0, i, z = 0;
bits |= diff;
for (i = 0; i < 64; i++) {
if (bits[i])
break;
z++;
}
return (unsigned int)z >> 3;
#endif
}
static __always_inline size_t lz77_match(const u8 *match, const u8 *cur, const u8 *end)
{
const u8 *start = cur;
if (cur == match)
return 0;
if (likely(cur < end - (LZ77_STEPSIZE_64 - 1))) {
u64 const diff = lz77_read64(cur) ^ lz77_read64(match);
if (!diff) {
cur += LZ77_STEPSIZE_64;
match += LZ77_STEPSIZE_64;
} else {
return __count_common_bytes(diff);
}
}
while (likely(cur < end - (LZ77_STEPSIZE_64 - 1))) {
u64 const diff = lz77_read64(cur) ^ lz77_read64(match);
if (!diff) {
cur += LZ77_STEPSIZE_64;
match += LZ77_STEPSIZE_64;
continue;
}
cur += __count_common_bytes(diff);
return (size_t)(cur - start);
}
if (cur < end - 3 && !(lz77_read32(cur) ^ lz77_read32(match))) {
cur += LZ77_STEPSIZE_32;
match += LZ77_STEPSIZE_32;
}
if (cur < end - 1 && lz77_read16(cur) == lz77_read16(match)) {
cur += LZ77_STEPSIZE_16;
match += LZ77_STEPSIZE_16;
}
if (cur < end && *cur == *match)
cur++;
return (size_t)(cur - start);
}
static __always_inline unsigned long lz77_max(unsigned long a, unsigned long b)
{
int m = (a < b) - 1;
return (a & m) | (b & ~m);
}
static __always_inline unsigned long lz77_min(unsigned long a, unsigned long b)
{
int m = (a > b) - 1;
return (a & m) | (b & ~m);
}
int lz77_compress(const u8 *src, size_t src_len, u8 *dst, size_t *dst_len);
/* when CONFIG_CIFS_COMPRESSION not set lz77_compress() is not called */
#endif /* !CONFIG_CIFS_COMPRESSION */
#endif /* _SMB_COMPRESS_LZ77_H */

View File

@ -978,9 +978,12 @@ static int smb3_fs_context_parse_param(struct fs_context *fc,
switch (opt) {
case Opt_compress:
if (!IS_ENABLED(CONFIG_CIFS_COMPRESSION)) {
cifs_errorf(fc, "CONFIG_CIFS_COMPRESSION kernel config option is unset\n");
goto cifs_parse_mount_err;
}
ctx->compress = true;
cifs_dbg(VFS,
"SMB3 compression support is experimental\n");
cifs_dbg(VFS, "SMB3 compression support is experimental\n");
break;
case Opt_nodfs:
ctx->nodfs = 1;