This is a faster copy_to/from_user function for athlon/duron.
(300% faster in file read/write?)
L2 cache size for Athlon(256)/Duron(64) should be adjusted on run-time.
I appreciate trying this patch who owns Athlon/Duron CPU.
1)
I have been using Taka’s file or socket benchmark program for
this purpose, if any better testing tools?
On the benchmark, athlon copy function is extremely fast.
read/write file test
http://dns.suna-asobi.com/~akira-t/linux/taka-bench/fileio2.c
org-copy: file read/write
buf(0x886e000) copied 24.0 Mbytes in 0.080 seconds at 299.4 Mbytes/sec
buf(0x886e001) copied 24.0 Mbytes in 0.129 seconds at 186.7 Mbytes/sec
buf(0x886e002) copied 24.0 Mbytes in 0.129 seconds at 186.4 Mbytes/sec
buf(0x886e003) copied 24.0 Mbytes in 0.129 seconds at 185.7 Mbytes/sec
(Entire log is here, http://dns.suna-asobi.com/~akira-t/linux/taka-bench/org-copy-file.log)
athlon-fast_copy: file read/write
buf(0x886e000) copied 24.0 Mbytes in 0.025 seconds at 959.2 Mbytes/sec
buf(0x886e001) copied 24.0 Mbytes in 0.032 seconds at 745.8 Mbytes/sec
buf(0x886e002) copied 24.0 Mbytes in 0.033 seconds at 731.4 Mbytes/sec
buf(0x886e003) copied 24.0 Mbytes in 0.032 seconds at 742.7 Mbytes/sec
(Entire log is here, http://dns.suna-asobi.com/~akira-t/linux/taka-bench/aki-copy-file.log)
network test
http://dns.suna-asobi.com/~akira-t/linux/taka-bench/netio2.c
org-copy: socket
0x846e000+0 -> 0x804e000+0 send/recv: 0.034387 seconds at 116.3 Mbytes/sec
0x846e000+1 -> 0x804e000+1 send/recv: 0.043644 seconds at 91.7 Mbytes/sec
0x846e000+2 -> 0x804e000+2 send/recv: 0.044038 seconds at 90.8 Mbytes/sec
0x846e000+3 -> 0x804e000+3 send/recv: 0.043457 seconds at 92.0 Mbytes/sec
(Entire log is here, http://dns.suna-asobi.com/~akira-t/linux/taka-bench/org-copy-net.log)
athlon-fast_copy: socket
0x846e000+0 -> 0x804e000+0 send/recv: 0.019374 seconds at 206.5 Mbytes/sec
0x846e000+1 -> 0x804e000+1 send/recv: 0.036772 seconds at 108.8 Mbytes/sec
0x846e000+2 -> 0x804e000+2 send/recv: 0.037353 seconds at 107.1 Mbytes/sec
0x846e000+3 -> 0x804e000+3 send/recv: 0.040598 seconds at 98.5 Mbytes/sec
(Entire log is here, http://dns.suna-asobi.com/~akira-t/linux/taka-bench/aki-copy-net.log)
2) Last Friday was my first day to touch gcc assembler, and I really
appreciate if there any suggestion for the aki_copy().
Should I save all the mmx registers? I didn't but seems to work ok.
3) From the comment from Andi and Andrew last time, statically compiling
for each CPU is not really optimal for the Linux distributors.
I saw Denis Vlasenko’s csum_copy routines with boot-time selection, and
it looked good. I think it is a good idea to do the same for the copy_user
functions.
4) I think this is just a mistake in the current kernel.
config X86_INTEL_USERCOPY
bool
- depends on MPENTIUM4 || MPENTIUMIII || M586MMX
+ depends on MPENTIUM4 || MPENTIUMIII || M686
default y
The intel_faster_copy is slower for Pentium MMX and faster for PenII.
Akira
--- linux-2.5.47/arch/i386/lib/usercopy.c Thu Oct 31 22:40:01 2002
+++ linux-2.5.47-aki/arch/i386/lib/usercopy.c Sat Nov 16 00:42:46 2002
@@ -337,6 +337,354 @@
__copy_user_intel(void *to, const void *from,unsigned long size);
#endif /* CONFIG_X86_INTEL_USERCOPY */
+#ifdef CONFIG_MK7
+/* Ahtlon version */
+/* akira version, specific to Athlon/Duron CPU
+ * 'size' must be larger than 256 */
+static unsigned long
+aki_copy(void *to, const void *from, unsigned long size) {
+ __asm__ __volatile(
+ /* These are just saving it for later use */
+ " movl %4, %%edx\n"
+ " push %%edi\n"
+ " movl %%ecx, %%ebx\n"
+ /* Athlon speeds up a lot when the address to read is
+ * aligned at 32bit(8bytes) boundary */
+ " movl %%esi, %%ecx\n"
+ " negl %%ecx\n"
+ " andl $7, %%ecx\n"
+ " subl %%ecx, %%ebx\n"
+ "80: rep; movsb\n"
+ /* Here is one trick to speed up, it is prefetching
+ * entire 'from' or up to L2 cache size, which ever is
+ * smaller.
+ * I used movl instead of special instruction, such as
+ * prefetch or prefetchnta because movl was faster
+ * when combining with movq, used in later. */
+ " movl %%ebx, %%eax\n"
+ " shrl $8, %%eax\n"
+ " cmpl %%edx, %%eax\n"
+ " jbe 10f\n"
+ " movl %%edx, %%eax\n"
+ " .align 2\n"
+ "10: movl 0(%%esi, %%ecx), %%edx\n"
+ "11: movl 64(%%esi, %%ecx), %%edx\n"
+ "12: movl 128(%%esi, %%ecx), %%edx\n"
+ "13: movl 192(%%esi, %%ecx), %%edx\n"
+ " addl $256, %%ecx\n"
+ " decl %%eax\n"
+ " jnz 10b\n"
+ // " femms\n"
+ /* Bulk transfer by using movq and movntq, every
+ * iteration 64 bytes. There is movl in the first line
+ * but it is not a mistake. I put movl there because it
+ * is a second prefetch for the next reading
+ * iteration, speeds up about 4 or 5%. */
+ "21: movl %%ebx, %%ecx\n"
+ " shrl $6, %%ecx\n"
+ " .align 2\n"
+ "20:\n"
+ " movl 64(%%esi), %%eax\n"
+ "30: movq (%%esi), %%mm0\n"
+ "31: movq 8(%%esi), %%mm1\n"
+ "32: movq 16(%%esi), %%mm2\n"
+ "33: movq 24(%%esi), %%mm3\n"
+ "34: movq 32(%%esi), %%mm4\n"
+ "35: movq 40(%%esi), %%mm5\n"
+ "36: movq 48(%%esi), %%mm6\n"
+ "37: movq 56(%%esi), %%mm7\n"
+ "40: movntq %%mm0, (%%edi)\n"
+ "41: movntq %%mm1, 8(%%edi)\n"
+ "42: movntq %%mm2, 16(%%edi)\n"
+ "43: movntq %%mm3, 24(%%edi)\n"
+ "44: movntq %%mm4, 32(%%edi)\n"
+ "45: movntq %%mm5, 40(%%edi)\n"
+ "46: movntq %%mm6, 48(%%edi)\n"
+ "47: movntq %%mm7, 56(%%edi)\n"
+ " addl $64, %%esi\n"
+ " addl $64, %%edi\n"
+ " decl %%ecx\n"
+ " jnz 20b\n"
+ " sfence\n"
+ " femms\n"
+ /* Finish the remaining data, which did not fit
+ * into 64 bytes.*/
+ " movl %%ebx, %%ecx\n"
+ " andl $0x3f, %%ecx\n"
+ "90: rep; movsb\n"
+ /* This is a postfetch for the written data,
+ * because the moved data is likely to be used right
+ * after this copy function but the movntq does not
+ * leave them in the L2 cache, so doing it here. */
+#if 0
+ " pop %%edi\n"
+ " prefetchnta (%%edi, %%ecx)\n"
+ " prefetchnta 64(%%edi, %%ecx)\n"
+ " prefetchnta 128(%%edi, %%ecx)\n"
+ " prefetchnta 192(%%edi, %%ecx)\n"
+#endif
+#if 1 /* 0 turn off postfetch */
+ " movl %%edi, %%eax\n"
+ " popl %%edi\n"
+ " subl %%edi, %%eax\n"
+ // " movl %%edx, %%edi\n"
+ " shrl $8, %%eax\n"
+ " jz 100f\n"
+ " cmpl $256, %%eax\n"
+ " jbe 50f\n"
+ " movl $256, %%eax\n"
+ " .align 2\n"
+ "50:\n"
+#if 0 /* 1 use 'prefetch' or 0 use movl for postfetch */
+ " prefetcht0 (%%edi, %%ecx)\n"
+ " prefetcht0 64(%%edi, %%ecx)\n"
+ " prefetcht0 128(%%edi, %%ecx)\n"
+ " prefetcht0 192(%%edi, %%ecx)\n"
+#else
+ // " prefetchnta 256(%%edi, %%ecx)\n"
+ " movl 0(%%edi, %%ecx), %%edx\n"
+ " movl 64(%%edi, %%ecx), %%edx\n"
+ " movl 128(%%edi, %%ecx), %%edx\n"
+ " movl 192(%%edi, %%ecx), %%edx\n"
+#endif
+ " addl $256, %%ecx\n"
+ " decl %%eax\n"
+ " jnz 50b\n"
+ " xorl %%ecx, %%ecx\n"
+#endif
+ "100:\n"
+
+ ".section .fixup,\"ax\"\n"
+ /* Page fault occured during prefetch, go back and
+ * start data transfer. */
+ "1:\n"
+ " jmp 21b\n"
+ "2:\n"
+ "3:\n"
+ " sfence\n"
+ " femms\n"
+ " shll $6, %%ecx\n"
+ " andl $0x3f, %%ebx\n"
+ " addl %%ebx, %%ecx\n"
+ " jmp 90b\n"
+ "8:\n"
+ " addl %%ebx, %%ecx\n"
+ "99:\n"
+ " popl %%edi\n"
+ " jmp 100b\n"
+
+ ".previous\n"
+ ".section __ex_table,\"a\"\n"
+ " .align 4\n"
+ " .long 10b, 1b\n"
+ " .long 11b, 1b\n"
+ " .long 12b, 1b\n"
+ " .long 13b, 1b\n"
+#if 0
+ " .long 20b, 2b\n"
+ " .long 30b, 3b\n"
+ " .long 31b, 3b\n"
+ " .long 32b, 3b\n"
+ " .long 33b, 3b\n"
+ " .long 34b, 3b\n"
+ " .long 35b, 3b\n"
+ " .long 36b, 3b\n"
+ " .long 37b, 3b\n"
+#endif
+ " .long 40b, 3b\n"
+ " .long 41b, 3b\n"
+ " .long 42b, 3b\n"
+ " .long 43b, 3b\n"
+ " .long 44b, 3b\n"
+ " .long 45b, 3b\n"
+ " .long 46b, 3b\n"
+ " .long 47b, 3b\n"
+ " .long 80b, 8b\n"
+ " .long 90b, 99b\n"
+ ".previous"
+
+ : "=&c"(size)
+ : "0"(size), "D"(to), "S"(from),
+ "r"(current_cpu_data.x86_cache_size * 1024 / 256)
+ : "eax", "ebx", "edx", "memory");
+ return size;
+}
+
+static unsigned long
+aki_copy_zeroing(void *to, const void *from, unsigned long size) {
+ __asm__ __volatile(
+ /* These are just saving it for later use */
+ " movl %4, %%edx\n"
+ " push %%edi\n"
+ " movl %%ecx, %%ebx\n"
+ /* Athlon speeds up a lot when the address to read is
+ * aligned at 32bit(8bytes) boundary */
+ " movl %%esi, %%ecx\n"
+ " negl %%ecx\n"
+ " andl $7, %%ecx\n"
+ " subl %%ecx, %%ebx\n"
+ "80: rep; movsb\n"
+ /* Here is one trick to speed up, it is prefetching
+ * entire 'from' or up to L2 cache size, which ever is
+ * smaller.
+ * I used movl instead of special instruction, such as
+ * prefetch or prefetchnta because movl was faster
+ * when combining with movq, used in later. */
+ " movl %%ebx, %%eax\n"
+ " shrl $8, %%eax\n"
+ " jz 11f\n"
+ " cmpl %%edx, %%eax\n"
+ " jbe 10f\n"
+ " movl %%edx, %%eax\n"
+ " .align 2\n"
+ "10: movl 192(%%esi, %%ecx), %%edx\n"
+ "11: movl 0(%%esi, %%ecx), %%edx\n"
+ "12: movl 64(%%esi, %%ecx), %%edx\n"
+ "13: movl 128(%%esi, %%ecx), %%edx\n"
+ " addl $256, %%ecx\n"
+ " decl %%eax\n"
+ " jnz 10b\n"
+ // " femms\n"
+ /* Bulk transfer by using movq and movntq, every
+ * iteration 64 bytes. There is movl in the first line
+ * but it is not a mistake. I put movl there because it
+ * is a second prefetch for the next reading
+ * iteration, speeds up about 4 or 5%. */
+ "21: movl %%ebx, %%ecx\n"
+ " shrl $6, %%ecx\n"
+ " .align 2\n"
+ "20:\n"
+ " movl 64(%%esi), %%eax\n"
+ "30: movq (%%esi), %%mm0\n"
+ "31: movq 8(%%esi), %%mm1\n"
+ "32: movq 16(%%esi), %%mm2\n"
+ "33: movq 24(%%esi), %%mm3\n"
+ "34: movq 32(%%esi), %%mm4\n"
+ "35: movq 40(%%esi), %%mm5\n"
+ "36: movq 48(%%esi), %%mm6\n"
+ "37: movq 56(%%esi), %%mm7\n"
+ "40: movntq %%mm0, (%%edi)\n"
+ "41: movntq %%mm1, 8(%%edi)\n"
+ "42: movntq %%mm2, 16(%%edi)\n"
+ "43: movntq %%mm3, 24(%%edi)\n"
+ "44: movntq %%mm4, 32(%%edi)\n"
+ "45: movntq %%mm5, 40(%%edi)\n"
+ "46: movntq %%mm6, 48(%%edi)\n"
+ "47: movntq %%mm7, 56(%%edi)\n"
+ " addl $64, %%esi\n"
+ " addl $64, %%edi\n"
+ " decl %%ecx\n"
+ " jnz 20b\n"
+ " sfence\n"
+ " femms\n"
+ /* Finish the remaining data, which did not fit
+ * into 64 bytes.*/
+ " movl %%ebx, %%ecx\n"
+ " andl $0x3f, %%ecx\n"
+ "90: rep; movsb\n"
+ /* This is a postfetch for the written data,
+ * because the moved data is likely to be used right
+ * after this copy function but the movntq does not
+ * leave them in the L2 cache, so doing it here. */
+#if 0
+ " popl %%edi\n"
+ " prefetchnta (%%edi, %%ecx)\n"
+ " prefetchnta 64(%%edi, %%ecx)\n"
+ " prefetchnta 128(%%edi, %%ecx)\n"
+ " prefetchnta 192(%%edi, %%ecx)\n"
+#endif
+#if 1 /* 0 turn off postfetch */
+ " movl %%edi, %%eax\n"
+ " popl %%edi\n"
+ " subl %%edi, %%eax\n"
+ // " movl %%edx, %%edi\n"
+ " shrl $8, %%eax\n"
+ " jz 100f\n"
+ " cmpl $256, %%eax\n"
+ " jbe 50f\n"
+ " movl $256, %%eax\n"
+ " .align 2\n"
+ "50:\n"
+#if 1 /* 1 use 'prefetch' or 0 use movl for postfetch */
+ " prefetcht0 (%%edi, %%ecx)\n"
+ " prefetcht0 64(%%edi, %%ecx)\n"
+ " prefetcht0 128(%%edi, %%ecx)\n"
+ " prefetcht0 192(%%edi, %%ecx)\n"
+#else
+ // " prefetchnta 256(%%edi, %%ecx)\n"
+ " movl 0(%%edi, %%ecx), %%edx\n"
+ " movl 64(%%edi, %%ecx), %%edx\n"
+ " movl 128(%%edi, %%ecx), %%edx\n"
+ " movl 192(%%edi, %%ecx), %%edx\n"
+#endif
+ " addl $256, %%ecx\n"
+ " decl %%eax\n"
+ " jnz 50b\n"
+ " xorl %%ecx, %%ecx\n"
+#endif
+ "100:\n"
+
+ ".section .fixup,\"ax\"\n"
+ /* Page fault occured during prefetch, go back and
+ * start data transfer. */
+ "1:\n"
+ " jmp 21b\n"
+ "2:\n"
+ "3:\n"
+ " sfence\n"
+ " femms\n"
+ " shll $6, %%ecx\n"
+ " andl $0x3f, %%ebx\n"
+ " addl %%ebx, %%ecx\n"
+ " jmp 90b\n"
+ "8:\n"
+ " addl %%ebx, %%ecx\n"
+ "99:\n"
+ " movl %%ecx, %%ebx\n"
+ " xorl %%eax,%%eax\n"
+ " rep; stosb\n"
+ " movl %%ebx, %%ecx\n"
+ " popl %%edi\n"
+ " jmp 100b\n"
+
+ ".previous\n"
+ ".section __ex_table,\"a\"\n"
+ " .align 4\n"
+ " .long 10b, 1b\n"
+ " .long 11b, 1b\n"
+ " .long 12b, 1b\n"
+ " .long 13b, 1b\n"
+ " .long 20b, 2b\n"
+ " .long 30b, 3b\n"
+ " .long 31b, 3b\n"
+ " .long 32b, 3b\n"
+ " .long 33b, 3b\n"
+ " .long 34b, 3b\n"
+ " .long 35b, 3b\n"
+ " .long 36b, 3b\n"
+ " .long 37b, 3b\n"
+#if 0
+ " .long 40b, 3b\n"
+ " .long 41b, 3b\n"
+ " .long 42b, 3b\n"
+ " .long 43b, 3b\n"
+ " .long 44b, 3b\n"
+ " .long 45b, 3b\n"
+ " .long 46b, 3b\n"
+ " .long 47b, 3b\n"
+#endif
+ " .long 80b, 8b\n"
+ " .long 90b, 99b\n"
+ ".previous"
+
+ : "=&c"(size)
+ : "0"(size), "D"(to), "S"(from),
+ "r"(current_cpu_data.x86_cache_size * 1024 / 256)
+ : "eax", "ebx", "edx", "memory");
+ return size;
+}
+#endif /* CONFIG_MK7 */
+
/* Generic arbitrary sized copy. */
#define __copy_user(to,from,size) \
do { \
@@ -416,7 +764,25 @@
: "memory"); \
} while (0)
+#ifdef CONFIG_MK7
+unsigned long __copy_to_user(void *to, const void *from, unsigned long n)
+{
+ if (n < 256)
+ __copy_user(to, from, n);
+ else
+ n = aki_copy(to, from, n);
+ return n;
+}
+unsigned long __copy_from_user(void *to, const void *from, unsigned long n)
+{
+ if (n < 256)
+ __copy_user_zeroing(to, from, n);
+ else
+ n = aki_copy_zeroing(to, from, n);
+ return n;
+}
+#else
unsigned long __copy_to_user(void *to, const void *from, unsigned long n)
{
if (movsl_is_ok(to, from, n))
@@ -434,6 +800,7 @@
n = __copy_user_zeroing_intel(to, from, n);
return n;
}
+#endif /* CONFIG_MK7 */
unsigned long copy_to_user(void *to, const void *from, unsigned long n)
{
You don't seem to save/restore the FPU state, so it will be likely
corrupted after your copy runs.
Also I'm pretty sure that using movntq (= forcing destination out of
cache) is not a good strategy for generic copy_from_user(). It may
be a win for the copies in write ( user space -> page cache ), but
will hurt for all the ioctls and other things that actually need the
data in cache afterwards. I am afraid it is not enough to do micro benchmarks
here.
-Andi
On Sat, 16 Nov 2002 11:56:52 +0100
Andi Kleen <[email protected]> mentioned:
>
> You don't seem to save/restore the FPU state, so it will be likely
> corrupted after your copy runs.
This is the main question for me that I was wondering for all week.
My first version was using fsave and frstore, so
just changing three lines will accomplish this.
Is it all I need? Any thing elase needed to consider using fpu register?
>
> Also I'm pretty sure that using movntq (= forcing destination out of
> cache) is not a good strategy for generic copy_from_user(). It may
> be a win for the copies in write ( user space -> page cache ),
Yes, that why I included postfetch in the code because movntq does not leave
them in the L2 cache.
Anygood idea to
> but
> will hurt for all the ioctls and other things that actually need the
> data in cache afterwards. I am afraid it is not enough to do micro benchmarks
> here.
check above?
>
>
> -Andi
On Sat, Nov 16, 2002 at 01:22:51PM -0500, Akira Tsukamoto wrote:
> On Sat, 16 Nov 2002 11:56:52 +0100
> Andi Kleen <[email protected]> mentioned:
> >
> > You don't seem to save/restore the FPU state, so it will be likely
> > corrupted after your copy runs.
>
> This is the main question for me that I was wondering for all week.
> My first version was using fsave and frstore, so
> just changing three lines will accomplish this.
> Is it all I need? Any thing elase needed to consider using fpu register?
You are currently corrupting the user's FPU state.
The proper way to save it is to use kernel_fpu_begin()
> >
> > Also I'm pretty sure that using movntq (= forcing destination out of
> > cache) is not a good strategy for generic copy_from_user(). It may
> > be a win for the copies in write ( user space -> page cache ),
>
> Yes, that why I included postfetch in the code because movntq does not leave
> them in the L2 cache.
That looks rather wasteful - first force it out and then trying to get it in
again. I have my doubts on it being a good strategy for speed.
> > but
> > will hurt for all the ioctls and other things that actually need the
> > data in cache afterwards. I am afraid it is not enough to do micro benchmarks
> > here.
>
> check above?
Use special function calls for them, don't put it into generic
copy_*_user
Also you should really check for small copy and not use the FPU based
copy for them. Best is probably to use a relatively simply copy_*_user
(no FPU tricks, just an unrolled integer core) and change the VFS
and the file systems to call a special function from write(), but only
when the write is big.
-Andi
Hi,
On Sat, 16 Nov 2002 19:30:03 +0100
Andi Kleen <[email protected]> mentioned:
>
> The proper way to save it is to use kernel_fpu_begin()
Thanks! I will look into it. This is what I was looking for.
I been running this kernel with my copy for three days, and never had
oops, but I was really worried.
> > > Also I'm pretty sure that using movntq (= forcing destination out of
> > > cache) is not a good strategy for generic copy_from_user(). It may
> > > be a win for the copies in write ( user space -> page cache ),
> >
> > Yes, that why I included postfetch in the code because movntq does not leave
> > them in the L2 cache.
>
> That looks rather wasteful - first force it out and then trying to get it in
> again. I have my doubts on it being a good strategy for speed.
It tried both, use just normal mov or movq <-> use movntq + postfetch, and the later
was much much faster, because postfetch needs to read only every 64 bytes.
I will ckeck kernel_fpu_begin() fisrt and if using fpu register is too much
overhead than I will remove them.
Akira
On Sat, 16 Nov 2002 19:30:03 +0100
Andi Kleen <[email protected]> mentioned:
> On Sat, Nov 16, 2002 at 01:22:51PM -0500, Akira Tsukamoto wrote:
> > This is the main question for me that I was wondering for all week.
> > My first version was using fsave and frstore, so
> > just changing three lines will accomplish this.
> > Is it all I need? Any thing elase needed to consider using fpu register?
>
> You are currently corrupting the user's FPU state.
fsave and frstor should solve this problem, doesn't it?
> The proper way to save it is to use kernel_fpu_begin()
I looked into it. kernel_fpu_begin/end are basically doing:
1)preempt enable/disable
2)fsave and frstor
It does not look a lot of overhead.
So what is missing in my patch is:
1)Surround with kernel_fpu_begin/end.
2)Change the threshold of the size from 256 to somewhere around 512.
I removed the fsave/frstor, which was in my first version, to lower the
threshold because they had some overhead and if the copying size
was smaller than 512, the org_copy became faster.
I just need to reverse it.
Please let me know if anything esle is missing.
> > > > Also I'm pretty sure that using movntq (= forcing destination out of
> > > > cache) is not a good strategy for generic copy_from_user(). It may
> > > > be a win for the copies in write ( user space -> page cache ),
> > >
> > > Yes, that why I included postfetch in the code because movntq does not leave
> > > them in the L2 cache.
> > That looks rather wasteful - first force it out and then trying to get it in
> > again. I have my doubts on it being a good strategy for speed.
>
> It tried both, use just normal mov or movq <-> use movntq + postfetch, and the later
> was much much faster, because postfetch needs to read only every 64 bytes.
This is bench for read with my patch on 2.5.47
read: buf(0x804e000) copied 24.0 Mbytes in 0.040 seconds at 604.7 Mbytes/sec
read: buf(0x804e001) copied 24.0 Mbytes in 0.047 seconds at 509.5 Mbytes/sec
read: buf(0x804e002) copied 24.0 Mbytes in 0.046 seconds at 516.8 Mbytes/sec
read: buf(0x804e003) copied 24.0 Mbytes in 0.046 seconds at 516.4 Mbytes/sec
This is stock 2.5.47
read: buf(0x804e000) copied 24.0 Mbytes in 0.086 seconds at 279.8 Mbytes/sec
read: buf(0x804e001) copied 24.0 Mbytes in 0.105 seconds at 229.2 Mbytes/sec
read: buf(0x804e002) copied 24.0 Mbytes in 0.104 seconds at 230.8 Mbytes/sec
read: buf(0x804e003) copied 24.0 Mbytes in 0.105 seconds at 229.2 Mbytes/sec
About 200% faster.
Akira
Hello,
> > The proper way to save it is to use kernel_fpu_begin()
>
> Thanks! I will look into it. This is what I was looking for.
>
> I been running this kernel with my copy for three days, and never had
> oops, but I was really worried.
kernel_fpu_begin() is not enough for this purpose as user_*_copy may
cause pagefault and sleep in it. You should know lazy FPU swtich mechanism.
The mechanism will be confued by your your copy routine and may save
broken FPU context or another process may use broken FPU registers.
You should also care about do_page_fault like this:
do_page_fault()
{
if (fault happens in user_*_copy) {
save FPU context on stack;
recover FPU context to the previous state;
}
.........
.........
.........
if (fault happens in user_*_copy) {
recover FPU context from stack;
}
}
> > > > Also I'm pretty sure that using movntq (= forcing destination out of
> > > > cache) is not a good strategy for generic copy_from_user(). It may
> > > > be a win for the copies in write ( user space -> page cache ),
> > >
> > > Yes, that why I included postfetch in the code because movntq does not leave
> > > them in the L2 cache.
> >
> > That looks rather wasteful - first force it out and then trying to get it in
> > again. I have my doubts on it being a good strategy for speed.
>
> It tried both, use just normal mov or movq <-> use movntq + postfetch, and the later
> was much much faster, because postfetch needs to read only every 64 bytes.
>
> I will ckeck kernel_fpu_begin() fisrt and if using fpu register is too much
> overhead than I will remove them.
I guess kernel_fpu_begin() might not be heavy as many processes don't use
fpu registers so much.
Thank you,
Hirokazu Takahashi.