In __copy_tofrom_user, if we encounter an exception on a store, we
stop copying and return the number of bytes not copied. However,
if the store is wider than one byte and is to an unaligned address,
it is possible that the store operand overlaps a page boundary
and the exception occurred on the latter part of the store operand,
meaning that it would be possible to copy a few more bytes. Since
copy_to_user is generally expected to copy as much as possible,
it would be better to copy those extra few bytes. This adds code
to do that. Since this edge case is not performance-critical,
the code has been written to be compact rather than as fast as
possible.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The hand-coded assembler 64-bit copy routines include feature sections
that select one code path or another depending on which CPU we are
executing on. The self-tests for these copy routines end up testing
just one path. This adds a mechanism for selecting any desired code
path at compile time, and makes 2 or 3 versions of each test, each
using a different code path, so as to cover all the possible paths.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
[mpe: Add -mcpu=power4 to CFLAGS for older compilers]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This aims to make the generation of exception table entries for the
loads and stores in __copy_tofrom_user_base clearer and easier to
verify. Instead of having a series of local labels on the loads and
stores, with a series of corresponding labels later for the exception
handlers, we now use macros to generate exception table entries at the
point of each load and store that could potentially trap. We do this
with the macros lex (load exception) and stex (store exception).
These macros are used right before the load or store to which they
apply.
Some complexity is introduced by the fact that we have some more work
to do after hitting an exception, because we need to calculate and
return the number of bytes not copied. The code uses r3 as the
current pointer into the destination buffer, that is, the address of
the first byte of the destination that has not been modified.
However, at various points in the copy loops, r3 can be 4, 8, 16 or 24
bytes behind that point.
To express this offset in an understandable way, we define a symbol
r3_offset which is updated at various points so that it equal to the
difference between the address of the first unmodified byte of the
destination and the value in r3. (In fact it only needs to be
accurate at the point of each lex or stex macro invocation.)
The rules for updating r3_offset are as follows:
* It starts out at 0
* An addi r3,r3,N instruction decreases r3_offset by N
* A store instruction (stb, sth, stw, std) to N(r3)
increases r3_offset by the width of the store (1, 2, 4, 8)
* A store with update instruction (stbu, sthu, stwu, stdu) to N(r3)
sets r3_offset to the width of the store.
There is some trickiness to the way that the lex and stex macros and
the associated exception handlers work. I would have liked to use
the current value of r3_offset in the name of the symbol used as
the exception handler, as in ".Lld_exc_$(r3_offset)" and then
have symbols .Lld_exc_0, .Lld_exc_8, .Lld_exc_16 etc. corresponding
to the offsets that needed to be added to r3. However, I couldn't
see a way to do that with gas.
Instead, the exception handler address is .Lld_exc - r3_offset or
.Lst_exc - r3_offset, that is, the distance ahead of .Lld_exc/.Lst_exc
that we start executing is equal to the amount that we need to add to
r3. This works because r3_offset is always a small multiple of 4,
and our instructions are 4 bytes long. This means that before
.Lld_exc and .Lst_exc, we have a sequence of instructions that
increments r3 by 4, 8, 16 or 24 depending on where we start. The
sequence increments r3 by 4 per instruction (on average).
We also replace the exception table for the 4k copy loop by a
macro per load or store. These loads and stores all use exactly
the same exception handler, which simply resets the argument registers
r3, r4 and r5 to there original values and re-does the whole copy
using the slower loop.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
files not using feature fixup don't need asm/feature-fixups.h
files using feature fixup need asm/feature-fixups.h
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch moves ASM_CONST() and stringify_in_c() into
dedicated asm-const.h, then cleans all related inclusions.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
[mpe: asm-compat.h should include asm-const.h]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Rather than override the machine type in .S code (which can hide wrong
or ambiguous code generation for the target), set the type to power4
for all assembly.
This also means we need to be careful not to build power4-only code
when we're not building for Book3S, such as the "power7" versions of
copyuser/page/memcpy.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Fix Book3E build, don't build the "power7" variants for non-Book3S]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This macro is taken from s390, and allows more flexibility in
changing exception table format.
mpe: Put it in ppc_asm.h and only define one version using
stringinfy_in_c(). Add some empty definitions and headers to keep the
selftests happy.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Pull kbuild updates from Michal Marek:
- EXPORT_SYMBOL for asm source by Al Viro.
This does bring a regression, because genksyms no longer generates
checksums for these symbols (CONFIG_MODVERSIONS). Nick Piggin is
working on a patch to fix this.
Plus, we are talking about functions like strcpy(), which rarely
change prototypes.
- Fixes for PPC fallout of the above by Stephen Rothwell and Nick
Piggin
- fixdep speedup by Alexey Dobriyan.
- preparatory work by Nick Piggin to allow architectures to build with
-ffunction-sections, -fdata-sections and --gc-sections
- CONFIG_THIN_ARCHIVES support by Stephen Rothwell
- fix for filenames with colons in the initramfs source by me.
* 'kbuild' of git://git.kernel.org/pub/scm/linux/kernel/git/mmarek/kbuild: (22 commits)
initramfs: Escape colons in depfile
ppc: there is no clear_pages to export
powerpc/64: whitelist unresolved modversions CRCs
kbuild: -ffunction-sections fix for archs with conflicting sections
kbuild: add arch specific post-link Makefile
kbuild: allow archs to select link dead code/data elimination
kbuild: allow architectures to use thin archives instead of ld -r
kbuild: Regenerate genksyms lexer
kbuild: genksyms fix for typeof handling
fixdep: faster CONFIG_ search
ia64: move exports to definitions
sparc32: debride memcpy.S a bit
[sparc] unify 32bit and 64bit string.h
sparc: move exports to definitions
ppc: move exports to definitions
arm: move exports to definitions
s390: move exports to definitions
m68k: move exports to definitions
alpha: move exports to actual definitions
x86: move exports to actual definitions
...
Debugging a data corruption issue with virtio-net/vhost-net led to
the observation that __copy_tofrom_user was occasionally returning
a value 16 larger than it should. Since the return value from
__copy_tofrom_user is the number of bytes not copied, this means
that __copy_tofrom_user can occasionally return a value larger
than the number of bytes it was asked to copy. In turn this can
cause higher-level copy functions such as copy_page_to_iter_iovec
to corrupt memory by copying data into the wrong memory locations.
It turns out that the failing case involves a fault on the store
at label 79, and at that point the first unmodified byte of the
destination is at R3 + 16. Consequently the exception handler
for that store needs to add 16 to R3 before using it to work out
how many bytes were not copied, but in this one case it was not
adding the offset to R3. To fix it, this moves the label 179 to
the point where we add 16 to R3. I have checked manually all the
exception handlers for the loads and stores in this code and the
rest of them are correct (it would be excellent to have an
automated test of all the exception cases).
This bug has been present since this code was initially
committed in May 2002 to Linux version 2.5.20.
Cc: stable@vger.kernel.org
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
There are still a few occurences where it remains, because it helps to
explain something that persists.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
If an assembly function that calls back into c code is exported to
modules, we need to ensure r2 is setup correctly. There are only
two places crazy enough to do it (two of which are my fault).
Signed-off-by: Anton Blanchard <anton@samba.org>
The powerpc 64-bit __copy_tofrom_user() function uses shifts to handle
unaligned invocations. However, these shifts were designed for
big-endian systems: On little-endian systems, they must shift in the
opposite direction.
This commit relies on the C preprocessor to insert the correct shifts
into the assembly code.
[ This is a rare but nasty LE issue. Most of the time we use the POWER7
optimised __copy_tofrom_user_power7 loop, but when it hits an exception
we fall back to the base __copy_tofrom_user loop. - Anton ]
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Remove CONFIG_POWER4_ONLY, the option is badly named and only does two
things:
- It wraps the MMU segment table code. With feature fixups there is
little downside to compiling this in.
- It uses the newer mtocrf instruction in various assembly functions.
Instead of making this a compile option just do it at runtime via
a feature fixup.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Implement a POWER7 optimised copy_to_user/copy_from_user using VMX.
For large aligned copies this new loop is over 10% faster, and for
large unaligned copies it is over 200% faster.
If we take a fault we fall back to the old version, this keeps
things relatively simple and easy to verify.
On POWER7 unaligned stores rarely slow down - they only flush when
a store crosses a 4KB page boundary. Furthermore this flush is
handled completely in hardware and should be 20-30 cycles.
Unaligned loads on the other hand flush much more often - whenever
crossing a 128 byte cache line, or a 32 byte sector if either sector
is an L1 miss.
Considering this information we really want to get the loads aligned
and not worry about the alignment of the stores. Microbenchmarks
confirm that this approach is much faster than the current unaligned
copy loop that uses shifts and rotates to ensure both loads and
stores are aligned.
We also want to try and do the stores in cacheline aligned, cacheline
sized chunks. If the store queue is unable to merge an entire
cacheline of stores then the L2 cache will have to do a
read/modify/write. Even worse, we will serialise this with the stores
in the next iteration of the copy loop since both iterations hit
the same cacheline.
Based on this, the new loop does the following things:
1 - 127 bytes
Get the source 8 byte aligned and use 8 byte loads and stores. Pretty
boring and similar to how the current loop works.
128 - 4095 bytes
Get the source 8 byte aligned and use 8 byte loads and stores,
1 cacheline at a time. We aren't doing the stores in cacheline
aligned chunks so we will potentially serialise once per cacheline.
Even so it is much better than the loop we have today.
4096 - bytes
If both source and destination have the same alignment get them both
16 byte aligned, then get the destination cacheline aligned. Do
cacheline sized loads and stores using VMX.
If source and destination do not have the same alignment, we get the
destination cacheline aligned, and use permute to do aligned loads.
In both cases the VMX loop should be optimal - we always do aligned
loads and stores and are always doing stores in cacheline aligned,
cacheline sized chunks.
To be able to use VMX we must be careful about interrupts and
sleeping. We don't use the VMX loop when in an interrupt (which should
be rare anyway) and we wrap the VMX loop in disable/enable_pagefault
and fall back to the existing copy_tofrom_user loop if we do need to
sleep.
The VMX breakpoint of 4096 bytes was chosen using this microbenchmark:
http://ozlabs.org/~anton/junkcode/copy_to_user.c
Since we are using VMX and there is a cost to saving and restoring
the user VMX state there are two broad cases we need to benchmark:
- Best case - userspace never uses VMX
- Worst case - userspace always uses VMX
In reality a userspace process will sit somewhere between these two
extremes. Since we need to test both aligned and unaligned copies we
end up with 4 combinations. The point at which the VMX loop begins to
win is:
0% VMX
aligned 2048 bytes
unaligned 2048 bytes
100% VMX
aligned 16384 bytes
unaligned 8192 bytes
Considering this is a microbenchmark, the data is hot in cache and
the VMX loop has better store queue merging properties we set the
breakpoint to 4096 bytes, a little below the unaligned breakpoints.
Some future optimisations we can look at:
- Looking at the perf data, a significant part of the cost when a
task is always using VMX is the extra exception we take to restore
the VMX state. As such we should do something similar to the x86
optimisation that restores FPU state for heavy users. ie:
/*
* If the task has used fpu the last 5 timeslices, just do a full
* restore of the math state immediately to avoid the trap; the
* chances of needing FPU soon are obviously high now
*/
preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5;
and
/*
* fpu_counter contains the number of consecutive context switches
* that the FPU is used. If this is over a threshold, the lazy fpu
* saving becomes unlazy to save the trap. This is an unsigned char
* so that after 256 times the counter wraps and the behavior turns
* lazy again; this to deal with bursty apps that only use FPU for
* a short time
*/
- We could create a paca bit to mirror the VMX enabled MSR bit and check
that first, avoiding multiple calls to calling enable_kernel_altivec.
That should help with iovec based system calls like readv.
- We could have two VMX breakpoints, one for when we know the user VMX
state is loaded into the registers and one when it isn't. This could
be a second bit in the paca so we can calculate the break points quickly.
- One suggestion from Ben was to save and restore the VSX registers
we use inline instead of using enable_kernel_altivec.
[BenH: Fixed a problem with preempt and fixed build without CONFIG_ALTIVEC]
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Here is a patch from Paul Mackerras that improves the ppc64 copy_tofrom_user.
The loop now does 32 bytes at a time and as well as pairing loads and stores.
A quick test case that reads 8kB over and over shows the improvement:
POWER6: 53% faster
POWER7: 51% faster
#define _XOPEN_SOURCE 500
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/stat.h>
#define BUFSIZE (8 * 1024)
#define ITERATIONS 10000000
int main()
{
char tmpfile[] = "/tmp/copy_to_user_testXXXXXX";
int fd;
char *buf[BUFSIZE];
unsigned long i;
fd = mkstemp(tmpfile);
if (fd < 0) {
perror("open");
exit(1);
}
if (write(fd, buf, BUFSIZE) != BUFSIZE) {
perror("open");
exit(1);
}
for (i = 0; i < 10000000; i++) {
if (pread(fd, buf, BUFSIZE, 0) != BUFSIZE) {
perror("pread");
exit(1);
}
}
unlink(tmpfile);
return 0;
}
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This fixes a regression introduced by commit
a4e22f02f5 ("powerpc: Update 64bit
__copy_tofrom_user() using CPU_FTR_UNALIGNED_LD_STD").
The same bug that existed in the 64bit memcpy() also exists here so fix
it here too. The fix is the same as that applied to memcpy() with the
addition of fixes for the exception handling code required for
__copy_tofrom_user().
This stops us reading beyond the end of the source region we were told
to copy.
Signed-off-by: Mark Nelson <markn@au1.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
In exactly the same way that we updated memcpy() with new feature
sections in commit 25d6e2d7c5 ("powerpc:
Update 64bit memcpy() using CPU_FTR_UNALIGNED_LD_STD"), we do the same
thing here for __copy_tofrom_user(). Once again this is purely a
performance tweak for Cell and Power6 - this has no effect on all the
other 64bit powerpc chips.
We can make these same changes to __copy_tofrom_user() because the
basic copy algorithm is the same as in memcpy() - this version just
has all the exception handling logic needed when copying to or from
userspace as well as a special case for copying whole 4K pages that
are page aligned.
CPU_FTR_UNALIGNED_LD_STD CPU was added in commit
4ec577a289 ("powerpc: Add new CPU
feature: CPU_FTR_UNALIGNED_LD_STD").
We also make the same simple one line change from cmpldi r1,... to
cmpldi cr1,... for consistency.
Signed-off-by: Mark Nelson <markn@au1.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
mtocrf is a faster single-field mtcrf (move to condition register
fields) instruction available in POWER4 and later processors. It can
make quite a difference in performance on some implementations, so use
it for CONFIG_POWER4_ONLY builds.
Signed-off-by: Olof Johansson <olof@lixom.net>
Signed-off-by: Paul Mackerras <paulus@samba.org>
This patch removes all self references and fixes references to files
in the now defunct arch/ppc64 tree. I think this accomplises
everything wanted, though there might be a few references I missed.
Signed-off-by: Jon Mason <jdmason@us.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Adds a new CONFIG_PPC_64K_PAGES which, when enabled, changes the kernel
base page size to 64K. The resulting kernel still boots on any
hardware. On current machines with 4K pages support only, the kernel
will maintain 16 "subpages" for each 64K page transparently.
Note that while real 64K capable HW has been tested, the current patch
will not enable it yet as such hardware is not released yet, and I'm
still verifying with the firmware architects the proper to get the
information from the newer hypervisors.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This doesn't change any code, just renames things so we consistently
have foo_32.c and foo_64.c where we have separate 32- and 64-bit
versions.
Signed-off-by: Paul Mackerras <paulus@samba.org>
