2019-05-28 10:10:26 -07:00
|
|
|
/* SPDX-License-Identifier: GPL-2.0-only */
|
2008-07-02 15:53:13 +02:00
|
|
|
/*
|
|
|
|
|
* Copyright 2008 Vitaly Mayatskikh <vmayatsk@redhat.com>
|
|
|
|
|
* Copyright 2002 Andi Kleen, SuSE Labs.
|
|
|
|
|
*
|
|
|
|
|
* Functions to copy from and to user space.
|
|
|
|
|
*/
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-08-06 23:59:56 +09:00
|
|
|
#include <linux/export.h>
|
2006-09-26 10:52:32 +02:00
|
|
|
#include <linux/linkage.h>
|
2025-02-07 13:15:33 +01:00
|
|
|
#include <linux/cfi_types.h>
|
|
|
|
|
#include <linux/objtool.h>
|
x86: re-introduce support for ERMS copies for user space accesses
I tried to streamline our user memory copy code fairly aggressively in
commit adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory
copies"), in order to then be able to clean up the code and inline the
modern FSRM case in commit 577e6a7fd50d ("x86: inline the 'rep movs' in
user copies for the FSRM case").
We had reports [1] of that causing regressions earlier with blogbench,
but that turned out to be a horrible benchmark for that case, and not a
sufficient reason for re-instating "rep movsb" on older machines.
However, now Eric Dumazet reported [2] a regression in performance that
seems to be a rather more real benchmark, where due to the removal of
"rep movs" a TCP stream over a 100Gbps network no longer reaches line
speed.
And it turns out that with the simplified the calling convention for the
non-FSRM case in commit 427fda2c8a49 ("x86: improve on the non-rep
'copy_user' function"), re-introducing the ERMS case is actually fairly
simple.
Of course, that "fairly simple" is glossing over several missteps due to
having to fight our assembler alternative code. This code really wanted
to rewrite a conditional branch to have two different targets, but that
made objtool sufficiently unhappy that this instead just ended up doing
a choice between "jump to the unrolled loop, or use 'rep movsb'
directly".
Let's see if somebody finds a case where the kernel memory copies also
care (see commit 68674f94ffc9: "x86: don't use REP_GOOD or ERMS for
small memory copies"). But Eric does argue that the user copies are
special because networking tries to copy up to 32KB at a time, if
order-3 pages allocations are possible.
In-kernel memory copies are typically small, unless they are the special
"copy pages at a time" kind that still use "rep movs".
Link: https://lore.kernel.org/lkml/202305041446.71d46724-yujie.liu@intel.com/ [1]
Link: https://lore.kernel.org/lkml/CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com/ [2]
Reported-and-tested-by: Eric Dumazet <edumazet@google.com>
Fixes: adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory copies")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-05-26 12:34:20 -07:00
|
|
|
#include <asm/cpufeatures.h>
|
|
|
|
|
#include <asm/alternative.h>
|
2012-04-20 12:19:51 -07:00
|
|
|
#include <asm/asm.h>
|
2006-09-26 10:52:39 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
/*
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
* rep_movs_alternative - memory copy with exception handling.
|
|
|
|
|
* This version is for CPUs that don't have FSRM (Fast Short Rep Movs)
|
2008-07-02 15:53:13 +02:00
|
|
|
*
|
|
|
|
|
* Input:
|
2005-04-16 15:20:36 -07:00
|
|
|
* rdi destination
|
|
|
|
|
* rsi source
|
2023-04-15 19:31:34 -07:00
|
|
|
* rcx count
|
2005-04-16 15:20:36 -07:00
|
|
|
*
|
2008-07-02 15:53:13 +02:00
|
|
|
* Output:
|
2023-04-15 19:31:34 -07:00
|
|
|
* rcx uncopied bytes or 0 if successful.
|
|
|
|
|
*
|
|
|
|
|
* NOTE! The calling convention is very intentionally the same as
|
|
|
|
|
* for 'rep movs', so that we can rewrite the function call with
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
* just a plain 'rep movs' on machines that have FSRM. But to make
|
2023-08-30 16:03:15 +02:00
|
|
|
* it simpler for us, we can clobber rsi/rdi and rax freely.
|
2005-04-16 15:20:36 -07:00
|
|
|
*/
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
SYM_FUNC_START(rep_movs_alternative)
|
2025-02-07 13:15:33 +01:00
|
|
|
ANNOTATE_NOENDBR
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
cmpq $64,%rcx
|
x86: re-introduce support for ERMS copies for user space accesses
I tried to streamline our user memory copy code fairly aggressively in
commit adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory
copies"), in order to then be able to clean up the code and inline the
modern FSRM case in commit 577e6a7fd50d ("x86: inline the 'rep movs' in
user copies for the FSRM case").
We had reports [1] of that causing regressions earlier with blogbench,
but that turned out to be a horrible benchmark for that case, and not a
sufficient reason for re-instating "rep movsb" on older machines.
However, now Eric Dumazet reported [2] a regression in performance that
seems to be a rather more real benchmark, where due to the removal of
"rep movs" a TCP stream over a 100Gbps network no longer reaches line
speed.
And it turns out that with the simplified the calling convention for the
non-FSRM case in commit 427fda2c8a49 ("x86: improve on the non-rep
'copy_user' function"), re-introducing the ERMS case is actually fairly
simple.
Of course, that "fairly simple" is glossing over several missteps due to
having to fight our assembler alternative code. This code really wanted
to rewrite a conditional branch to have two different targets, but that
made objtool sufficiently unhappy that this instead just ended up doing
a choice between "jump to the unrolled loop, or use 'rep movsb'
directly".
Let's see if somebody finds a case where the kernel memory copies also
care (see commit 68674f94ffc9: "x86: don't use REP_GOOD or ERMS for
small memory copies"). But Eric does argue that the user copies are
special because networking tries to copy up to 32KB at a time, if
order-3 pages allocations are possible.
In-kernel memory copies are typically small, unless they are the special
"copy pages at a time" kind that still use "rep movs".
Link: https://lore.kernel.org/lkml/202305041446.71d46724-yujie.liu@intel.com/ [1]
Link: https://lore.kernel.org/lkml/CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com/ [2]
Reported-and-tested-by: Eric Dumazet <edumazet@google.com>
Fixes: adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory copies")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-05-26 12:34:20 -07:00
|
|
|
jae .Llarge
|
2006-02-03 21:51:02 +01:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
cmp $8,%ecx
|
|
|
|
|
jae .Lword
|
2006-02-03 21:51:02 +01:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
testl %ecx,%ecx
|
|
|
|
|
je .Lexit
|
2006-09-26 10:52:32 +02:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
.Lcopy_user_tail:
|
|
|
|
|
0: movb (%rsi),%al
|
|
|
|
|
1: movb %al,(%rdi)
|
|
|
|
|
inc %rdi
|
|
|
|
|
inc %rsi
|
|
|
|
|
dec %rcx
|
|
|
|
|
jne .Lcopy_user_tail
|
|
|
|
|
.Lexit:
|
2021-12-04 14:43:40 +01:00
|
|
|
RET
|
2019-03-01 15:24:33 +01:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
_ASM_EXTABLE_UA( 0b, .Lexit)
|
|
|
|
|
_ASM_EXTABLE_UA( 1b, .Lexit)
|
2021-12-23 12:07:01 -08:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
.p2align 4
|
|
|
|
|
.Lword:
|
|
|
|
|
2: movq (%rsi),%rax
|
|
|
|
|
3: movq %rax,(%rdi)
|
|
|
|
|
addq $8,%rsi
|
|
|
|
|
addq $8,%rdi
|
|
|
|
|
sub $8,%ecx
|
|
|
|
|
je .Lexit
|
|
|
|
|
cmp $8,%ecx
|
|
|
|
|
jae .Lword
|
|
|
|
|
jmp .Lcopy_user_tail
|
2021-11-10 11:01:05 +01:00
|
|
|
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
_ASM_EXTABLE_UA( 2b, .Lcopy_user_tail)
|
|
|
|
|
_ASM_EXTABLE_UA( 3b, .Lcopy_user_tail)
|
2021-11-10 11:01:05 +01:00
|
|
|
|
x86: re-introduce support for ERMS copies for user space accesses
I tried to streamline our user memory copy code fairly aggressively in
commit adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory
copies"), in order to then be able to clean up the code and inline the
modern FSRM case in commit 577e6a7fd50d ("x86: inline the 'rep movs' in
user copies for the FSRM case").
We had reports [1] of that causing regressions earlier with blogbench,
but that turned out to be a horrible benchmark for that case, and not a
sufficient reason for re-instating "rep movsb" on older machines.
However, now Eric Dumazet reported [2] a regression in performance that
seems to be a rather more real benchmark, where due to the removal of
"rep movs" a TCP stream over a 100Gbps network no longer reaches line
speed.
And it turns out that with the simplified the calling convention for the
non-FSRM case in commit 427fda2c8a49 ("x86: improve on the non-rep
'copy_user' function"), re-introducing the ERMS case is actually fairly
simple.
Of course, that "fairly simple" is glossing over several missteps due to
having to fight our assembler alternative code. This code really wanted
to rewrite a conditional branch to have two different targets, but that
made objtool sufficiently unhappy that this instead just ended up doing
a choice between "jump to the unrolled loop, or use 'rep movsb'
directly".
Let's see if somebody finds a case where the kernel memory copies also
care (see commit 68674f94ffc9: "x86: don't use REP_GOOD or ERMS for
small memory copies"). But Eric does argue that the user copies are
special because networking tries to copy up to 32KB at a time, if
order-3 pages allocations are possible.
In-kernel memory copies are typically small, unless they are the special
"copy pages at a time" kind that still use "rep movs".
Link: https://lore.kernel.org/lkml/202305041446.71d46724-yujie.liu@intel.com/ [1]
Link: https://lore.kernel.org/lkml/CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com/ [2]
Reported-and-tested-by: Eric Dumazet <edumazet@google.com>
Fixes: adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory copies")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-05-26 12:34:20 -07:00
|
|
|
.Llarge:
|
2023-08-30 16:03:15 +02:00
|
|
|
0: ALTERNATIVE "jmp .Llarge_movsq", "rep movsb", X86_FEATURE_ERMS
|
x86: re-introduce support for ERMS copies for user space accesses
I tried to streamline our user memory copy code fairly aggressively in
commit adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory
copies"), in order to then be able to clean up the code and inline the
modern FSRM case in commit 577e6a7fd50d ("x86: inline the 'rep movs' in
user copies for the FSRM case").
We had reports [1] of that causing regressions earlier with blogbench,
but that turned out to be a horrible benchmark for that case, and not a
sufficient reason for re-instating "rep movsb" on older machines.
However, now Eric Dumazet reported [2] a regression in performance that
seems to be a rather more real benchmark, where due to the removal of
"rep movs" a TCP stream over a 100Gbps network no longer reaches line
speed.
And it turns out that with the simplified the calling convention for the
non-FSRM case in commit 427fda2c8a49 ("x86: improve on the non-rep
'copy_user' function"), re-introducing the ERMS case is actually fairly
simple.
Of course, that "fairly simple" is glossing over several missteps due to
having to fight our assembler alternative code. This code really wanted
to rewrite a conditional branch to have two different targets, but that
made objtool sufficiently unhappy that this instead just ended up doing
a choice between "jump to the unrolled loop, or use 'rep movsb'
directly".
Let's see if somebody finds a case where the kernel memory copies also
care (see commit 68674f94ffc9: "x86: don't use REP_GOOD or ERMS for
small memory copies"). But Eric does argue that the user copies are
special because networking tries to copy up to 32KB at a time, if
order-3 pages allocations are possible.
In-kernel memory copies are typically small, unless they are the special
"copy pages at a time" kind that still use "rep movs".
Link: https://lore.kernel.org/lkml/202305041446.71d46724-yujie.liu@intel.com/ [1]
Link: https://lore.kernel.org/lkml/CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com/ [2]
Reported-and-tested-by: Eric Dumazet <edumazet@google.com>
Fixes: adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory copies")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-05-26 12:34:20 -07:00
|
|
|
1: RET
|
|
|
|
|
|
2023-08-30 16:03:15 +02:00
|
|
|
_ASM_EXTABLE_UA( 0b, 1b)
|
x86: re-introduce support for ERMS copies for user space accesses
I tried to streamline our user memory copy code fairly aggressively in
commit adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory
copies"), in order to then be able to clean up the code and inline the
modern FSRM case in commit 577e6a7fd50d ("x86: inline the 'rep movs' in
user copies for the FSRM case").
We had reports [1] of that causing regressions earlier with blogbench,
but that turned out to be a horrible benchmark for that case, and not a
sufficient reason for re-instating "rep movsb" on older machines.
However, now Eric Dumazet reported [2] a regression in performance that
seems to be a rather more real benchmark, where due to the removal of
"rep movs" a TCP stream over a 100Gbps network no longer reaches line
speed.
And it turns out that with the simplified the calling convention for the
non-FSRM case in commit 427fda2c8a49 ("x86: improve on the non-rep
'copy_user' function"), re-introducing the ERMS case is actually fairly
simple.
Of course, that "fairly simple" is glossing over several missteps due to
having to fight our assembler alternative code. This code really wanted
to rewrite a conditional branch to have two different targets, but that
made objtool sufficiently unhappy that this instead just ended up doing
a choice between "jump to the unrolled loop, or use 'rep movsb'
directly".
Let's see if somebody finds a case where the kernel memory copies also
care (see commit 68674f94ffc9: "x86: don't use REP_GOOD or ERMS for
small memory copies"). But Eric does argue that the user copies are
special because networking tries to copy up to 32KB at a time, if
order-3 pages allocations are possible.
In-kernel memory copies are typically small, unless they are the special
"copy pages at a time" kind that still use "rep movs".
Link: https://lore.kernel.org/lkml/202305041446.71d46724-yujie.liu@intel.com/ [1]
Link: https://lore.kernel.org/lkml/CANn89iKUbyrJ=r2+_kK+sb2ZSSHifFZ7QkPLDpAtkJ8v4WUumA@mail.gmail.com/ [2]
Reported-and-tested-by: Eric Dumazet <edumazet@google.com>
Fixes: adfcf4231b8c ("x86: don't use REP_GOOD or ERMS for user memory copies")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-05-26 12:34:20 -07:00
|
|
|
|
2023-08-30 16:03:15 +02:00
|
|
|
.Llarge_movsq:
|
x86/uaccess: Improve performance by aligning writes to 8 bytes in copy_user_generic(), on non-FSRM/ERMS CPUs
History of the performance regression:
======================================
Since the following series of user copy updates were merged upstream
~2 years ago via:
a5624566431d ("Merge branch 'x86-rep-insns': x86 user copy clarifications")
.. copy_user_generic() on x86_64 stopped doing alignment of the
writes to the destination to a 8 byte boundary for the non FSRM case.
Previously, this was done through the ALIGN_DESTINATION macro that
was used in the now removed copy_user_generic_unrolled function.
Turns out this change causes some loss of performance/throughput on
some use cases and specific CPU/platforms without FSRM and ERMS.
Lately I got two reports of performance/throughput issues after a
RHEL 9 kernel pulled the same upstream series with updates to user
copy functions. Both reports consisted of running specific
networking/TCP related testing using iperf3.
Partial upstream fix
====================
The first report was related to a Linux Bridge testing using VMs on a
specific machine with an AMD CPU (EPYC 7402), and after a brief
investigation it turned out that the later change via:
ca96b162bfd2 ("x86: bring back rep movsq for user access on CPUs without ERMS")
... helped/fixed the performance issue.
However, after the later commit/fix was applied, then I got another
regression reported in a multistream TCP test on a 100Gbit mlx5 nic, also
running on an AMD based platform (AMD EPYC 7302 CPU), again that was using
iperf3 to run the test. That regression was after applying the later
fix/commit, but only this didn't help in telling the whole history.
Testing performed to pinpoint residual regression
=================================================
So I narrowed down the second regression use case, but running it
without traffic through a NIC, on localhost, in trying to narrow down
CPU usage and not being limited by other factor like network bandwidth.
I used another system also with an AMD CPU (AMD EPYC 7742). Basically,
I run iperf3 in server and client mode in the same system, for example:
- Start the server binding it to CPU core/thread 19:
$ taskset -c 19 iperf3 -D -s -B 127.0.0.1 -p 12000
- Start the client always binding/running on CPU core/thread 17, using
perf to get statistics:
$ perf stat -o stat.txt taskset -c 17 iperf3 -c 127.0.0.1 -b 0/1000 -V \
-n 50G --repeating-payload -l 16384 -p 12000 --cport 12001 2>&1 \
> stat-19.txt
For the client, always running/pinned to CPU 17. But for the iperf3 in
server mode, I did test runs using CPUs 19, 21, 23 or not pinned to any
specific CPU. So it basically consisted with four runs of the same
commands, just changing the CPU which the server is pinned, or without
pinning by removing the taskset call before the server command. The CPUs
were chosen based on NUMA node they were on, this is the relevant output
of lscpu on the system:
$ lscpu
...
Model name: AMD EPYC 7742 64-Core Processor
...
Caches (sum of all):
L1d: 2 MiB (64 instances)
L1i: 2 MiB (64 instances)
L2: 32 MiB (64 instances)
L3: 256 MiB (16 instances)
NUMA:
NUMA node(s): 4
NUMA node0 CPU(s): 0,1,8,9,16,17,24,25,32,33,40,41,48,49,56,57,64,65,72,73,80,81,88,89,96,97,104,105,112,113,120,121
NUMA node1 CPU(s): 2,3,10,11,18,19,26,27,34,35,42,43,50,51,58,59,66,67,74,75,82,83,90,91,98,99,106,107,114,115,122,123
NUMA node2 CPU(s): 4,5,12,13,20,21,28,29,36,37,44,45,52,53,60,61,68,69,76,77,84,85,92,93,100,101,108,109,116,117,124,125
NUMA node3 CPU(s): 6,7,14,15,22,23,30,31,38,39,46,47,54,55,62,63,70,71,78,79,86,87,94,95,102,103,110,111,118,119,126,127
...
So for the server run, when picking a CPU, I chose CPUs to be not on the same
node. The reason is with that I was able to get/measure relevant
performance differences when changing the alignment of the writes to the
destination in copy_user_generic.
Testing shows up to +81% performance improvement under iperf3
=============================================================
Here's a summary of the iperf3 runs:
# Vanilla upstream alignment:
CPU RATE SYS TIME sender-receiver
Server bind 19: 13.0Gbits/sec 28.371851000 33.233499566 86.9%-70.8%
Server bind 21: 12.9Gbits/sec 28.283381000 33.586486621 85.8%-69.9%
Server bind 23: 11.1Gbits/sec 33.660190000 39.012243176 87.7%-64.5%
Server bind none: 18.9Gbits/sec 19.215339000 22.875117865 86.0%-80.5%
# With the attached patch (aligning writes in non ERMS/FSRM case):
CPU RATE SYS TIME sender-receiver
Server bind 19: 20.8Gbits/sec 14.897284000 20.811101382 75.7%-89.0%
Server bind 21: 20.4Gbits/sec 15.205055000 21.263165909 75.4%-89.7%
Server bind 23: 20.2Gbits/sec 15.433801000 21.456175000 75.5%-89.8%
Server bind none: 26.1Gbits/sec 12.534022000 16.632447315 79.8%-89.6%
So I consistently got better results when aligning the write. The
results above were run on 6.14.0-rc6/rc7 based kernels. The sys is sys
time and then the total time to run/transfer 50G of data. The last
field is the CPU usage of sender/receiver iperf3 process. It's also
worth to note that each pair of iperf3 runs may get slightly different
results on each run, but I always got consistent higher results with
the write alignment for this specific test of running the processes
on CPUs in different NUMA nodes.
Linus Torvalds helped/provided this version of the patch. Initially I
proposed a version which aligned writes for all cases in
rep_movs_alternative, however it used two extra registers and thus
Linus provided an enhanced version that only aligns the write on the
large_movsq case, which is sufficient since the problem happens only
on those AMD CPUs like ones mentioned above without ERMS/FSRM, and
also doesn't require using extra registers. Also, I validated that
aligning only on large_movsq case is really enough for getting the
performance back.
I also tested this patch on an old Intel based non-ERMS/FRMS system
(with Xeon E5-2667 - Sandy Bridge based) and didn't get any problems:
no performance enhancement but also no regression either, using the
same iperf3 based benchmark. Also newer Intel processors after
Sandy Bridge usually have ERMS and should not be affected by this change.
[ mingo: Updated the changelog. ]
Fixes: ca96b162bfd2 ("x86: bring back rep movsq for user access on CPUs without ERMS")
Fixes: 034ff37d3407 ("x86: rewrite '__copy_user_nocache' function")
Reported-by: Ondrej Lichtner <olichtne@redhat.com>
Co-developed-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Herton R. Krzesinski <herton@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20250320142213.2623518-1-herton@redhat.com
2025-03-20 11:22:13 -03:00
|
|
|
/* Do the first possibly unaligned word */
|
|
|
|
|
0: movq (%rsi),%rax
|
|
|
|
|
1: movq %rax,(%rdi)
|
|
|
|
|
|
|
|
|
|
_ASM_EXTABLE_UA( 0b, .Lcopy_user_tail)
|
|
|
|
|
_ASM_EXTABLE_UA( 1b, .Lcopy_user_tail)
|
|
|
|
|
|
|
|
|
|
/* What would be the offset to the aligned destination? */
|
|
|
|
|
leaq 8(%rdi),%rax
|
|
|
|
|
andq $-8,%rax
|
|
|
|
|
subq %rdi,%rax
|
|
|
|
|
|
|
|
|
|
/* .. and update pointers and count to match */
|
|
|
|
|
addq %rax,%rdi
|
|
|
|
|
addq %rax,%rsi
|
|
|
|
|
subq %rax,%rcx
|
|
|
|
|
|
|
|
|
|
/* make %rcx contain the number of words, %rax the remainder */
|
2023-08-30 16:03:15 +02:00
|
|
|
movq %rcx,%rax
|
|
|
|
|
shrq $3,%rcx
|
|
|
|
|
andl $7,%eax
|
|
|
|
|
0: rep movsq
|
|
|
|
|
movl %eax,%ecx
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
testl %ecx,%ecx
|
|
|
|
|
jne .Lcopy_user_tail
|
|
|
|
|
RET
|
|
|
|
|
|
2023-08-30 16:03:15 +02:00
|
|
|
1: leaq (%rax,%rcx,8),%rcx
|
|
|
|
|
jmp .Lcopy_user_tail
|
|
|
|
|
|
|
|
|
|
_ASM_EXTABLE_UA( 0b, 1b)
|
x86: improve on the non-rep 'copy_user' function
The old 'copy_user_generic_unrolled' function was oddly implemented for
largely historical reasons: it had been largely based on the uncached
copy case, which has some other concerns.
For example, the __copy_user_nocache() function uses 'movnti' for the
destination stores, and those want the destination to be aligned. In
contrast, the regular copy function doesn't really care, and trying to
align things only complicates matters.
Also, like the clear_user function, the copy function had some odd
handling of the repeat counts, complicating the exception handling for
no really good reason. So as with clear_user, just write it to keep all
the byte counts in the %rcx register, exactly like the 'rep movs'
functionality that this replaces.
Unlike a real 'rep movs', we do allow for this to trash a few temporary
registers to not have to unnecessarily save/restore registers on the
stack.
And like the clearing case, rename this to what it now clearly is:
'rep_movs_alternative', and make it one coherent function, so that it
shows up as such in profiles (instead of the odd split between
"copy_user_generic_unrolled" and "copy_user_short_string", the latter of
which was not about strings at all, and which was shared with the
uncached case).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-16 18:23:06 -07:00
|
|
|
SYM_FUNC_END(rep_movs_alternative)
|
|
|
|
|
EXPORT_SYMBOL(rep_movs_alternative)
|