lz4: Fix kernel decompression speed
PATCH 1/1 lz4: Fix kernel decompression speed
From: Nick Terrell <[email protected]>
This patch replaces all memcpy() calls with LZ4_memcpy() which calls __builtin_memcpy() so the compiler can inline it.
LZ4 relies heavily on memcpy() with a constant size being inlined. In x86 and i386 pre-boot environments memcpy() cannot be inlined because memcpy() doesn't get defined as __builtin_memcpy().
An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the x86_64 and i386 architectures. The speed-up is about 10x as shown below.
Code Arch Kernel Size Time Speed
v5.8 x86_64 11504832 B 148 ms 79 MB/s
patch x86_64 11503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/s
I also measured the time to decompress the initramfs on x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.
[1] https://github.com/lz4/lz4/pull/890
--- lib/lz4/lz4_compress.c | 4 ++-- lib/lz4/lz4_decompress.c | 18 +++++++++--------- lib/lz4/lz4defs.h | 10 ++++++++++ lib/lz4/lz4hc_compress.c | 2 +- 4 files changed, 22 insertions(+), 12 deletions(-)
lib/lz4/lz4_compress.c +2 -2
@@ -446,7 +446,7 @@ static FORCE_INLINE int LZ4_compress_generic( *op++ = (BYTE )(lastRun <<ML_BITS ); } -memcpy (op, anchor, lastRun); +LZ4_memcpy (op, anchor, lastRun); op += lastRun; }
@@ -708,7 +708,7 @@ static int LZ4_compress_destSize_generic( }else { *op++ = (BYTE )(lastRunSize<<ML_BITS ); } -memcpy (op, anchor, lastRunSize); +LZ4_memcpy (op, anchor, lastRunSize); op += lastRunSize; }
lib/lz4/lz4_decompress.c +9 -9
@@ -153,7 +153,7 @@ static FORCE_INLINE int LZ4_decompress_generic( && likely((endOnInput ? ip < shortiend : 1) & (op <= shortoend))) {/* Copy the literals */ -memcpy (op, ip, endOnInput ?16 :8 ); +LZ4_memcpy (op, ip, endOnInput ?16 :8 ); op += length; ip += length;/*
@@ -172,9 +172,9 @@ static FORCE_INLINE int LZ4_decompress_generic( (offset >= 8) && (dict == withPrefix64k || match >= lowPrefix)) {/* Copy the match. */ -memcpy (op +0 , match +0 ,8 ); -memcpy (op +8 , match +8 ,8 ); -memcpy (op +16 , match +16 ,2 ); +LZ4_memcpy (op +0 , match +0 ,8 ); +LZ4_memcpy (op +8 , match +8 ,8 ); +LZ4_memcpy (op +16 , match +16 ,2 ); op += length +MINMATCH ;/* Both stages worked, load the next token. */ continue ;
@@ -263,7 +263,7 @@ static FORCE_INLINE int LZ4_decompress_generic( } } -memcpy (op, ip, length); +LZ4_memcpy (op, ip, length); ip += length; op += length;
@@ -350,7 +350,7 @@ static FORCE_INLINE int LZ4_decompress_generic(size_t const copySize = (size_t )(lowPrefix - match);size_t const restSize = length - copySize; -memcpy (op, dictEnd - copySize, copySize); +LZ4_memcpy (op, dictEnd - copySize, copySize); op += copySize;if (restSize > (size_t )(op - lowPrefix)) {/* overlap copy */
@@ -360,7 +360,7 @@ static FORCE_INLINE int LZ4_decompress_generic(while (op < endOfMatch) *op++ = *copyFrom++; }else { -memcpy (op, lowPrefix, restSize); +LZ4_memcpy (op, lowPrefix, restSize); op += restSize; } }
@@ -386,7 +386,7 @@ static FORCE_INLINE int LZ4_decompress_generic(while (op < copyEnd) *op++ = *match++; }else { -memcpy (op, match, mlen); +LZ4_memcpy (op, match, mlen); } op = copyEnd;if (op == oend)
@@ -400,7 +400,7 @@ static FORCE_INLINE int LZ4_decompress_generic( op[2 ] = match[2 ]; op[3 ] = match[3 ]; match += inc32table[offset]; -memcpy (op +4 , match,4 ); +LZ4_memcpy (op +4 , match,4 ); match -= dec64table[offset]; }else {LZ4_copy8 (op, match);
lib/lz4/lz4defs.h +10 -0
@@ -137,6 +137,16 @@ static FORCE_INLINE void LZ4_writeLE16(void *memPtr, U16 value)return put_unaligned_le16 (value, memPtr); } +/* +* LZ4 relies on memcpy with a constant size being inlined. In freestanding +* environments, the compiler can't assume the implementation of memcpy() is +* standard compliant, so apply its specialized memcpy() inlining logic. When +* possible, use __builtin_memcpy() to tell the compiler to analyze memcpy() +* as-if it were standard compliant, so it can inline it in freestanding +* environments. This is needed when decompressing the Linux Kernel, for example. +*/ +#define LZ4_memcpy (dst, src, size) __builtin_memcpy(dst, src, size) + static FORCE_INLINE void LZ4_copy8(void *dst, const void *src) {#if LZ4_ARCH64
lib/lz4/lz4hc_compress.c +1 -1
@@ -570,7 +570,7 @@ static int LZ4HC_compress_generic( *op++ = (BYTE ) lastRun; }else *op++ = (BYTE )(lastRun<<ML_BITS ); -memcpy (op, anchor, iend - anchor); +LZ4_memcpy (op, anchor, iend - anchor); op += iend - anchor; }
Signature · 1 line
2.28.0
Quoted 22 lines · Nick Terrell
From: Nick Terrell <[email protected]>
This patch replaces all
memcpy()calls withLZ4_memcpy()which calls__builtin_memcpy()so the compiler can inline it.LZ4 relies heavily on
memcpy()with a constant size being inlined. In x86 and i386 pre-boot environmentsmemcpy()cannot be inlined becausememcpy()doesn't get defined as__builtin_memcpy().An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/s
Quoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.
Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in misc.c which causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.
Thanks.
From 10f8d939fc367e3127e2d72ba099678debcae422 Mon Sep 17 00:00:00 2001
From: Arvind Sankar <[email protected]>
Date: Mon, 3 Aug 2020 17:07:37 -0400
Subject: [PATCH] x86/boot/compressed: Use builtin mem functions for decompressor
Since commits
c041b5ad8640 ("x86, boot: Create a separate string.h file to provide standard string functions") fb4cac573ef6 ("x86, boot: Move memcmp() into string.h and string.c")
the decompressor stub has been using the compiler's builtin memcpy, memset and memcmp functions, _except_ where it would likely have the largest impact, in the decompression code itself.
Remove the #undef's of memcpy and memset in misc.c so that the decompressor code also uses the compiler builtins.
The rationale given in the comment doesn't really apply: just because some functions use the out-of-line version is no reason to not use the builtin version in the rest.
Replace the comment with an explanation of why memzero and memmove are being #define'd.
Drop the suggestion to #undef in boot/string.h as well: the out-of-line versions are not really optimized versions, they're generic code that's good enough for the preboot environment. The compiler will likely generate better code for constant-size memcpy/memset/memcmp if it is allowed to.
--- arch/x86/boot/compressed/misc.c | 7 ++----- arch/x86/boot/string.h | 5 +---- 2 files changed, 3 insertions(+), 9 deletions(-)
arch/x86/boot/compressed/misc.c +2 -5
@@ -30,12 +30,9 @@#define STATIC static/* -* Use normal definitions of mem*() from string.c. There are already -* included header files which expect a definition of memset() and by -* the time we define memset macro, it is too late. +* Provide definitions of memzero and memmove as some of the decompressors will +* try to define their own functions if these are not defined as macros. */ -#undef memcpy -#undef memset#define memzero (s, n) memset((s), 0, (n))#define memmove memmove
arch/x86/boot/string.h +1 -4
@@ -11,10 +11,7 @@ void *memcpy(void *dst, const void *src, size_t len);void *memset (void *dst,int c,size_t len);int memcmp (const void *s1,const void *s2,size_t len); -/* -* Access builtin version by default. If one needs to use optimized version, -* do "undef memcpy" in .c file and link against right string.c -*/ +/* Access builtin version by default. */ #define memcpy (d,s,l) __builtin_memcpy(d,s,l)#define memset (d,c,l) __builtin_memset(d,c,l)#define memcmp __builtin_memcmp
Signature · 1 line
2.26.2
Quoted 34 lines · Arvind Sankar
Quoted 22 lines · Nick Terrell
From: Nick Terrell <[email protected]>
This patch replaces all
memcpy()calls withLZ4_memcpy()which calls__builtin_memcpy()so the compiler can inline it.LZ4 relies heavily on
memcpy()with a constant size being inlined. In x86 and i386 pre-boot environmentsmemcpy()cannot be inlined becausememcpy()doesn't get defined as__builtin_memcpy().An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.
Quoted 6 lines
Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
But, I believe that this is also a problem for ARM, though I have a hard time measuring because I can’t get pre-boot print statements in QEMU. I will attempt to take a look at the assembly, because I’m fairly certain that memcpy() isn’t inlined in master.
Even if we fix all the architectures, I would still like to merge the LZ4 patch. It seems like it is pretty easy to merge a patch that is a boot speed regression, because people aren’t actively measuring it. So I prefer a layered defense.
Additionally, this is following upstream. At some point LZ4 will be imported from upstream and get an equivalent of this patch https://github.com/lz4/lz4/pull/890.
Best,
Nick
Quoted 39 lines · Nick Terrell
Quoted 39 lines · Arvind Sankar
Quoted 22 lines · Nick Terrell
From: Nick Terrell <[email protected]>
This patch replaces all
memcpy()calls withLZ4_memcpy()which calls__builtin_memcpy()so the compiler can inline it.LZ4 relies heavily on
memcpy()with a constant size being inlined. In x86 and i386 pre-boot environmentsmemcpy()cannot be inlined becausememcpy()doesn't get defined as__builtin_memcpy().An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.
Quoted 4 lines
Quoted 1 line
Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
Thanks. I tried using RDTSC to get some timings under QEMU, and I get similar speedup as you have for LZ4, and around 15-20% or so for ZSTD (on 64-bit) -- I see that ZSTD_copy8 is already using __builtin_memcpy, but there must be more that can be optimized? There's a couple 1/2-byte sized copies in huf_decompress.c.
For the rest of the algos there seems to be no difference, likely because some casual grepping doesn't show any constant-size copies in the other algorithms.
Quoted 44 lines · Arvind Sankar
Quoted 39 lines · Nick Terrell
Quoted 39 lines · Arvind Sankar
Quoted 22 lines · Nick Terrell
From: Nick Terrell <[email protected]>
This patch replaces all
memcpy()calls withLZ4_memcpy()which calls__builtin_memcpy()so the compiler can inline it.LZ4 relies heavily on
memcpy()with a constant size being inlined. In x86 and i386 pre-boot environmentsmemcpy()cannot be inlined becausememcpy()doesn't get defined as__builtin_memcpy().An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
Thanks. I tried using RDTSC to get some timings under QEMU, and I get similar speedup as you have for LZ4, and around 15-20% or so for ZSTD (on 64-bit)
By the way, I was using this script for performance testing [0].
Quoted 3 lines
-- I see that
ZSTD_copy8is already using__builtin_memcpy, but there must be more that can be optimized? There's a couple 1/2-byte sized copies inhuf_decompress.c.
Oh wow, I totally missed that, I guess I stopped looking once performance was about what I expected, nice find!
I suspect it is mostly the memcpy inside of HUF_decodeSymbolX4(), since that should be the only hot one [1].
Do you want to put up the patch to fix the memcpy’s in zstd Huffman, or should I?
I will be submitting a patch upstream to migrate all of zstd’s memcpy() calls to use __builtin_memcpy(), since I plan on updating the version in the kernel to upstream zstd in the next few months. I was waiting until the compressed kernel patch set landed, so I didn't distract from it.
[0] https://gist.github.com/terrelln/9bd53321a669f62683c608af8944fbc2
[1] https://github.com/torvalds/linux/blob/master/lib/zstd/huf_decompress.c#L598
Best,
Nick
Quoted 22 lines · Nick Terrell
Quoted 8 lines · Arvind Sankar
-- I see that
ZSTD_copy8is already using__builtin_memcpy, but there must be more that can be optimized? There's a couple 1/2-byte sized copies inhuf_decompress.c.Oh wow, I totally missed that, I guess I stopped looking once performance was about what I expected, nice find!
I suspect it is mostly the memcpy inside of
HUF_decodeSymbolX4(), since that should be the only hot one [1].Do you want to put up the patch to fix the memcpy’s in zstd Huffman, or should I?
I will be submitting a patch upstream to migrate all of zstd’s
memcpy()calls to use__builtin_memcpy(), since I plan on updating the version in the kernel to upstream zstd in the next few months. I was waiting until the compressed kernel patch set landed, so I didn't distract from it.
Quoted 6 lines
[0] https://gist.github.com/terrelln/9bd53321a669f62683c608af8944fbc2
[1] https://github.com/torvalds/linux/blob/master/lib/zstd/huf_decompress.c#L598Best,
Nick
It's better if you do the zstd changes I think, as I'm not familiar with the code at all.
Thanks.
Quoted 46 lines · Arvind Sankar
Quoted 41 lines · Nick Terrell
Quoted 40 lines · Arvind Sankar
Quoted 22 lines · Nick Terrell
From: Nick Terrell <[email protected]>
This patch replaces all
memcpy()calls withLZ4_memcpy()which calls__builtin_memcpy()so the compiler can inline it.LZ4 relies heavily on
memcpy()with a constant size being inlined. In x86 and i386 pre-boot environmentsmemcpy()cannot be inlined becausememcpy()doesn't get defined as__builtin_memcpy().An equivalent patch has been applied upstream so that the next import won't lose this change [1].
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.Thanks.
Quoted 8 lines
Quoted 2 lines
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
Thanks. I tried using RDTSC to get some timings under QEMU, and I get similar speedup as you have for LZ4, and around 15-20% or so for ZSTD (on 64-bit) -- I see that
ZSTD_copy8is already using__builtin_memcpy, but there must be more that can be optimized? There's a couple 1/2-byte sized copies inhuf_decompress.c.
I tested your patch. I saw the same performance as my patch for LZ4, as expected. I saw that zstd decompression time went from 84 ms -> 63 ms, a 25% improvement.
I will get started on a zstd patch. I look forward to your x86 patch getting merged!
Thanks,
Nick
cover x86/boot/compressed: Use builtin mem functions for decompressor
This patch drops a couple of lines that prevent the decompressor code from using __builtin_memcpy (and __builtin_memset).
Note that Nick Terrell's patch [0] will apply similar changes directly in the decompressor code (in-kernel and upstream) for LZ4, and the ZSTD code already is partially optimized to use __builtin_memcpy for hot memcpy's, but still improves a little bit more with this patch.
[0] https://lore.kernel.org/lkml/[email protected]/
Arvind Sankar (1):
x86/boot/compressed: Use builtin mem functions for decompressor
arch/x86/boot/compressed/misc.c | 7 ++----- arch/x86/boot/string.h | 5 +---- 2 files changed, 3 insertions(+), 9 deletions(-)
Signature · 1 line
2.26.2
PATCH 1/1 x86/boot/compressed: Use builtin mem functions for decompressor
Since commits
c041b5ad8640 ("x86, boot: Create a separate string.h file to provide standard string functions") fb4cac573ef6 ("x86, boot: Move memcmp() into string.h and string.c")
the decompressor stub has been using the compiler's builtin memcpy, memset and memcmp functions, _except_ where it would likely have the largest impact, in the decompression code itself.
Remove the #undef's of memcpy and memset in misc.c so that the decompressor code also uses the compiler builtins.
The rationale given in the comment doesn't really apply: just because some functions use the out-of-line version is no reason to not use the builtin version in the rest.
Replace the comment with an explanation of why memzero and memmove are being #define'd.
Drop the suggestion to #undef in boot/string.h as well: the out-of-line versions are not really optimized versions, they're generic code that's good enough for the preboot environment. The compiler will likely generate better code for constant-size memcpy/memset/memcmp if it is allowed to.
Most decompressors' performance is unchanged, with the exception of LZ4 and 64-bit ZSTD.
Before After ARCH
LZ4 73ms 10ms 32
LZ4 120ms 10ms 64
ZSTD 90ms 74ms 64
Measurements on QEMU on 2.2GHz Broadwell Xeon, using defconfig kernels.
Decompressor code size has small differences, with the largest being that 64-bit ZSTD decreases just over 2k. The largest code size increase was on 64-bit XZ, of about 400 bytes.
--- arch/x86/boot/compressed/misc.c | 7 ++----- arch/x86/boot/string.h | 5 +---- 2 files changed, 3 insertions(+), 9 deletions(-)
arch/x86/boot/compressed/misc.c +2 -5
@@ -30,12 +30,9 @@#define STATIC static/* -* Use normal definitions of mem*() from string.c. There are already -* included header files which expect a definition of memset() and by -* the time we define memset macro, it is too late. +* Provide definitions of memzero and memmove as some of the decompressors will +* try to define their own functions if these are not defined as macros. */ -#undef memcpy -#undef memset#define memzero (s, n) memset((s), 0, (n))#define memmove memmove
arch/x86/boot/string.h +1 -4
@@ -11,10 +11,7 @@ void *memcpy(void *dst, const void *src, size_t len);void *memset (void *dst,int c,size_t len);int memcmp (const void *s1,const void *s2,size_t len); -/* -* Access builtin version by default. If one needs to use optimized version, -* do "undef memcpy" in .c file and link against right string.c -*/ +/* Access builtin version by default. */ #define memcpy (d,s,l) __builtin_memcpy(d,s,l)#define memset (d,c,l) __builtin_memset(d,c,l)#define memcmp __builtin_memcmp
Signature · 1 line
2.26.2
Re: [PATCH 1/1] x86/boot/compressed: Use builtin mem functions for decompressor
Quoted 35 lines · Arvind Sankar
Since commits
c041b5ad8640("x86, boot: Create a separatestring.hfile to provide standard string functions")fb4cac573ef6("x86, boot: Movememcmp()intostring.handstring.c")the decompressor stub has been using the compiler's builtin memcpy, memset and memcmp functions, _except_ where it would likely have the largest impact, in the decompression code itself.
Remove the #undef's of memcpy and memset in
misc.cso that the decompressor code also uses the compiler builtins.The rationale given in the comment doesn't really apply: just because some functions use the out-of-line version is no reason to not use the builtin version in the rest.
Replace the comment with an explanation of why memzero and memmove are being #define'd.
Drop the suggestion to #undef in
boot/string.has well: the out-of-line versions are not really optimized versions, they're generic code that's good enough for the preboot environment. The compiler will likely generate better code for constant-size memcpy/memset/memcmp if it is allowed to.Most decompressors' performance is unchanged, with the exception of LZ4 and 64-bit ZSTD.
Before After ARCH
LZ4 73ms 10ms 32
LZ4 120ms 10ms 64
ZSTD 90ms 74ms 64Measurements on QEMU on 2.2GHz Broadwell Xeon, using defconfig kernels.
Quoted 7 lines
Decompressor code size has small differences, with the largest being that 64-bit ZSTD decreases just over 2k. The largest code size increase was on 64-bit XZ, of about 400 bytes.
Signed-off-by Arvind Sankar <[email protected]>Suggested-by Nick Terrell <[email protected]>Tested-by Nick Terrell <[email protected]>
Did anyone pick this up? (Ingo can you snag it, or maybe akpm who took the LZ4-specific patch already?) This looks sane to me and provides some surprising performance benefits. :)
-Kees
Quoted 43 lines
--- arch/x86/boot/compressed/misc.c | 7 ++----- arch/x86/boot/string.h | 5 +---- 2 files changed, 3 insertions(+), 9 deletions(-)arch/x86/boot/compressed/misc.c +2
-5@@ -30,12 +30,9 @@#define STATIC static/* -* Use normal definitions of mem*() from string.c. There are already -* included header files which expect a definition of memset() and by -* the time we define memset macro, it is too late. +* Provide definitions of memzero and memmove as some of the decompressors will +* try to define their own functions if these are not defined as macros. */ -#undef memcpy -#undef memset#define memzero (s, n) memset((s), 0, (n))#define memmove memmovearch/x86/boot/string.h +1
-4@@ -11,10 +11,7 @@ void *memcpy(void *dst, const void *src, size_t len);void *memset (void *dst,int c,size_t len);int memcmp (const void *s1,const void *s2,size_t len); -/* -* Access builtin version by default. If one needs to use optimized version, -* do "undef memcpy" in .c file and link against right string.c -*/ +/* Access builtin version by default. */ #define memcpy (d,s,l) __builtin_memcpy(d,s,l)#define memset (d,c,l) __builtin_memset(d,c,l)#define memcmp __builtin_memcmpSignature · 1 line
2.26.2
Signature · 1 line
Kees Cook
Re: [PATCH 1/1] x86/boot/compressed: Use builtin mem functions for decompressor
Quoted 4 lines · Kees Cook
Did anyone pick this up? (Ingo can you snag it, or maybe akpm who took the LZ4-specific patch already?) This looks sane to me and provides some surprising performance benefits. :)
I'll take this and Arvind's freestanding patch too.
Linus
Hi!
Quoted 27 lines
Quoted 17 lines
Quoted 10 lines
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Quoted 6 lines
Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
Quoted 7 lines
But, I believe that this is also a problem for ARM, though I have a hard time measuring because I can’t get pre-boot print statements in QEMU. I will attempt to take a look at the assembly, because I’m fairly certain that
memcpy()isn’t inlined in master.Even if we fix all the architectures, I would still like to merge the LZ4 patch. It seems like it is pretty easy to merge a patch that is a boot speed regression, because people aren’t actively measuring it. So I prefer a layered defense.
Layered defense against performance-only problem, happening on emulation-only?
IMO that's a bit of overkill.
Best regards,
Pavel
Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Quoted 38 lines · Pavel Machek
Hi!
Quoted 27 lines
Quoted 17 lines
Quoted 10 lines
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Quoted 6 lines
Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
Quoted 7 lines
But, I believe that this is also a problem for ARM, though I have a hard time measuring because I can’t get pre-boot print statements in QEMU. I will attempt to take a look at the assembly, because I’m fairly certain that
memcpy()isn’t inlined in master.Even if we fix all the architectures, I would still like to merge the LZ4 patch. It seems like it is pretty easy to merge a patch that is a boot speed regression, because people aren’t actively measuring it. So I prefer a layered defense.
Quoted 10 lines
Layered defense against performance-only problem, happening on emulation-only?
IMO that's a bit of overkill.
Best regards,
PavelSignature · 2 lines
(english) http://www.livejournal.com/~pavelmachek (cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
Why would it be emulation-only? QEMU is just being used for ease of testing, but the performance impact should be similar on bare metal.
Thanks.
Quoted 43 lines · Arvind Sankar
Quoted 38 lines · Pavel Machek
Hi!
Quoted 34 lines
Quoted 17 lines
Quoted 10 lines
I've measured the kernel decompression speed using QEMU before and after this patch for the
x86_64and i386 architectures. The speed-up is about 10x as shown below.Code Arch Kernel Size Time Speed
v5.8x86_6411504832 B 148 ms 79 MB/s
patchx86_6411503872 B 13 ms 885 MB/s
v5.8 i386 9621216 B 91 ms 106 MB/s
patch i386 9620224 B 10 ms 962 MB/sQuoted 6 lines
I also measured the time to decompress the initramfs on
x86_64, i386, and arm. All three show the same decompression speed before and after, as expected.Quoted 6 lines
Hi Nick, would you be able to test the below patch's performance to verify it gives the same speedup? It removes the #undef in
misc.cwhich causes the decompressors to not use the builtin version. It should be equivalent to yours except for applying it to all the decompressors.Thanks.
I will measure it. I would expect it to provide the same speed up. It would be great to fix the problem for x86/i386 in general.
But, I believe that this is also a problem for ARM, though I have a hard time measuring because I can’t get pre-boot print statements in QEMU. I will attempt to take a look at the assembly, because I’m fairly certain that
memcpy()isn’t inlined in master.Even if we fix all the architectures, I would still like to merge the LZ4 patch. It seems like it is pretty easy to merge a patch that is a boot speed regression, because people aren’t actively measuring it. So I prefer a layered defense.
Quoted 7 lines
Quoted 4 lines
Layered defense against performance-only problem, happening on emulation-only?
IMO that's a bit of overkill.
Why would it be emulation-only? QEMU is just being used for ease of testing, but the performance impact should be similar on bare metal.
In addition, I want the decompressors to be fast in the pre-boot for all architectures. Not everyone is going to know that zstd and lz4 require memcpy to be inlined or they are 10x slower, that is an implementation detail of the library. It is a performance gotcha that I’d rather not have.