leandrof/linux
1
0
Fork 0
forked from Minki/linux
Code Issues Pull Requests Packages Projects Releases Wiki Activity
4e72dbe135
linux/arch/powerpc/lib/copyuser_64.S
Anton Blanchard a66086b819 powerpc: POWER7 optimised copy_to_user/copy_from_user using VMX 
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>
2011-12-19 14:40:40 +11:00

652 lines
10 KiB
ArmAsm
Raw Blame History

/*
* Copyright (C) 2002 Paul Mackerras, IBM Corp.
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*/
#include <asm/processor.h>
#include <asm/ppc_asm.h>
.align 7
_GLOBAL(__copy_tofrom_user)
BEGIN_FTR_SECTION
nop
FTR_SECTION_ELSE
b __copy_tofrom_user_power7
ALT_FTR_SECTION_END_IFCLR(CPU_FTR_VMX_COPY)
_GLOBAL(__copy_tofrom_user_base)
/* first check for a whole page copy on a page boundary */
cmpldi cr1,r5,16
cmpdi cr6,r5,4096
or r0,r3,r4
neg r6,r3 /* LS 3 bits = # bytes to 8-byte dest bdry */
andi. r0,r0,4095
std r3,-24(r1)
crand cr0*4+2,cr0*4+2,cr6*4+2
std r4,-16(r1)
std r5,-8(r1)
dcbt 0,r4
beq .Lcopy_page_4K
andi. r6,r6,7
PPC_MTOCRF 0x01,r5
blt cr1,.Lshort_copy
/* Below we want to nop out the bne if we're on a CPU that has the
* CPU_FTR_UNALIGNED_LD_STD bit set and the CPU_FTR_CP_USE_DCBTZ bit
* cleared.
* At the time of writing the only CPU that has this combination of bits
* set is Power6.
*/
BEGIN_FTR_SECTION
nop
FTR_SECTION_ELSE
bne .Ldst_unaligned
ALT_FTR_SECTION_END(CPU_FTR_UNALIGNED_LD_STD | CPU_FTR_CP_USE_DCBTZ, \
CPU_FTR_UNALIGNED_LD_STD)
.Ldst_aligned:
addi r3,r3,-16
BEGIN_FTR_SECTION
andi. r0,r4,7
bne .Lsrc_unaligned
END_FTR_SECTION_IFCLR(CPU_FTR_UNALIGNED_LD_STD)
blt cr1,.Ldo_tail /* if < 16 bytes to copy */
srdi r0,r5,5
cmpdi cr1,r0,0
20: ld r7,0(r4)
220: ld r6,8(r4)
addi r4,r4,16
mtctr r0
andi. r0,r5,0x10
beq 22f
addi r3,r3,16
addi r4,r4,-16
mr r9,r7
mr r8,r6
beq cr1,72f
21: ld r7,16(r4)
221: ld r6,24(r4)
addi r4,r4,32
70: std r9,0(r3)
270: std r8,8(r3)
22: ld r9,0(r4)
222: ld r8,8(r4)
71: std r7,16(r3)
271: std r6,24(r3)
addi r3,r3,32
bdnz 21b
72: std r9,0(r3)
272: std r8,8(r3)
andi. r5,r5,0xf
beq+ 3f
addi r4,r4,16
.Ldo_tail:
addi r3,r3,16
bf cr7*4+0,246f
244: ld r9,0(r4)
addi r4,r4,8
245: std r9,0(r3)
addi r3,r3,8
246: bf cr7*4+1,1f
23: lwz r9,0(r4)
addi r4,r4,4
73: stw r9,0(r3)
addi r3,r3,4
1: bf cr7*4+2,2f
44: lhz r9,0(r4)
addi r4,r4,2
74: sth r9,0(r3)
addi r3,r3,2
2: bf cr7*4+3,3f
45: lbz r9,0(r4)
75: stb r9,0(r3)
3: li r3,0
blr
.Lsrc_unaligned:
srdi r6,r5,3
addi r5,r5,-16
subf r4,r0,r4
srdi r7,r5,4
sldi r10,r0,3
cmpldi cr6,r6,3
andi. r5,r5,7
mtctr r7
subfic r11,r10,64
add r5,r5,r0
bt cr7*4+0,28f
24: ld r9,0(r4) /* 3+2n loads, 2+2n stores */
25: ld r0,8(r4)
sld r6,r9,r10
26: ldu r9,16(r4)
srd r7,r0,r11
sld r8,r0,r10
or r7,r7,r6
blt cr6,79f
27: ld r0,8(r4)
b 2f
28: ld r0,0(r4) /* 4+2n loads, 3+2n stores */
29: ldu r9,8(r4)
sld r8,r0,r10
addi r3,r3,-8
blt cr6,5f
30: ld r0,8(r4)
srd r12,r9,r11
sld r6,r9,r10
31: ldu r9,16(r4)
or r12,r8,r12
srd r7,r0,r11
sld r8,r0,r10
addi r3,r3,16
beq cr6,78f
1: or r7,r7,r6
32: ld r0,8(r4)
76: std r12,8(r3)
2: srd r12,r9,r11
sld r6,r9,r10
33: ldu r9,16(r4)
or r12,r8,r12
77: stdu r7,16(r3)
srd r7,r0,r11
sld r8,r0,r10
bdnz 1b
78: std r12,8(r3)
or r7,r7,r6
79: std r7,16(r3)
5: srd r12,r9,r11
or r12,r8,r12
80: std r12,24(r3)
bne 6f
li r3,0
blr
6: cmpwi cr1,r5,8
addi r3,r3,32
sld r9,r9,r10
ble cr1,7f
34: ld r0,8(r4)
srd r7,r0,r11
or r9,r7,r9
7:
bf cr7*4+1,1f
rotldi r9,r9,32
94: stw r9,0(r3)
addi r3,r3,4
1: bf cr7*4+2,2f
rotldi r9,r9,16
95: sth r9,0(r3)
addi r3,r3,2
2: bf cr7*4+3,3f
rotldi r9,r9,8
96: stb r9,0(r3)
3: li r3,0
blr
.Ldst_unaligned:
PPC_MTOCRF 0x01,r6 /* put #bytes to 8B bdry into cr7 */
subf r5,r6,r5
li r7,0
cmpldi cr1,r5,16
bf cr7*4+3,1f
35: lbz r0,0(r4)
81: stb r0,0(r3)
addi r7,r7,1
1: bf cr7*4+2,2f
36: lhzx r0,r7,r4
82: sthx r0,r7,r3
addi r7,r7,2
2: bf cr7*4+1,3f
37: lwzx r0,r7,r4
83: stwx r0,r7,r3
3: PPC_MTOCRF 0x01,r5
add r4,r6,r4
add r3,r6,r3
b .Ldst_aligned
.Lshort_copy:
bf cr7*4+0,1f
38: lwz r0,0(r4)
39: lwz r9,4(r4)
addi r4,r4,8
84: stw r0,0(r3)
85: stw r9,4(r3)
addi r3,r3,8
1: bf cr7*4+1,2f
40: lwz r0,0(r4)
addi r4,r4,4
86: stw r0,0(r3)
addi r3,r3,4
2: bf cr7*4+2,3f
41: lhz r0,0(r4)
addi r4,r4,2
87: sth r0,0(r3)
addi r3,r3,2
3: bf cr7*4+3,4f
42: lbz r0,0(r4)
88: stb r0,0(r3)
4: li r3,0
blr
/*
* exception handlers follow
* we have to return the number of bytes not copied
* for an exception on a load, we set the rest of the destination to 0
*/
136:
137:
add r3,r3,r7
b 1f
130:
131:
addi r3,r3,8
120:
320:
122:
322:
124:
125:
126:
127:
128:
129:
133:
addi r3,r3,8
132:
addi r3,r3,8
121:
321:
344:
134:
135:
138:
139:
140:
141:
142:
123:
144:
145:
/*
* here we have had a fault on a load and r3 points to the first
* unmodified byte of the destination
*/
1: ld r6,-24(r1)
ld r4,-16(r1)
ld r5,-8(r1)
subf r6,r6,r3
add r4,r4,r6
subf r5,r6,r5 /* #bytes left to go */
/*
* first see if we can copy any more bytes before hitting another exception
*/
mtctr r5
43: lbz r0,0(r4)
addi r4,r4,1
89: stb r0,0(r3)
addi r3,r3,1
bdnz 43b
li r3,0 /* huh? all copied successfully this time? */
blr
/*
* here we have trapped again, need to clear ctr bytes starting at r3
*/
143: mfctr r5
li r0,0
mr r4,r3
mr r3,r5 /* return the number of bytes not copied */
1: andi. r9,r4,7
beq 3f
90: stb r0,0(r4)
addic. r5,r5,-1
addi r4,r4,1
bne 1b
blr
3: cmpldi cr1,r5,8
srdi r9,r5,3
andi. r5,r5,7
blt cr1,93f
mtctr r9
91: std r0,0(r4)
addi r4,r4,8
bdnz 91b
93: beqlr
mtctr r5
92: stb r0,0(r4)
addi r4,r4,1
bdnz 92b
blr
/*
* exception handlers for stores: we just need to work
* out how many bytes weren't copied
*/
182:
183:
add r3,r3,r7
b 1f
371:
180:
addi r3,r3,8
171:
177:
addi r3,r3,8
370:
372:
176:
178:
addi r3,r3,4
185:
addi r3,r3,4
170:
172:
345:
173:
174:
175:
179:
181:
184:
186:
187:
188:
189:
194:
195:
196:
1:
ld r6,-24(r1)
ld r5,-8(r1)
add r6,r6,r5
subf r3,r3,r6 /* #bytes not copied */
190:
191:
192:
blr /* #bytes not copied in r3 */
.section __ex_table,"a"
.align 3
.llong 20b,120b
.llong 220b,320b
.llong 21b,121b
.llong 221b,321b
.llong 70b,170b
.llong 270b,370b
.llong 22b,122b
.llong 222b,322b
.llong 71b,171b
.llong 271b,371b
.llong 72b,172b
.llong 272b,372b
.llong 244b,344b
.llong 245b,345b
.llong 23b,123b
.llong 73b,173b
.llong 44b,144b
.llong 74b,174b
.llong 45b,145b
.llong 75b,175b
.llong 24b,124b
.llong 25b,125b
.llong 26b,126b
.llong 27b,127b
.llong 28b,128b
.llong 29b,129b
.llong 30b,130b
.llong 31b,131b
.llong 32b,132b
.llong 76b,176b
.llong 33b,133b
.llong 77b,177b
.llong 78b,178b
.llong 79b,179b
.llong 80b,180b
.llong 34b,134b
.llong 94b,194b
.llong 95b,195b
.llong 96b,196b
.llong 35b,135b
.llong 81b,181b
.llong 36b,136b
.llong 82b,182b
.llong 37b,137b
.llong 83b,183b
.llong 38b,138b
.llong 39b,139b
.llong 84b,184b
.llong 85b,185b
.llong 40b,140b
.llong 86b,186b
.llong 41b,141b
.llong 87b,187b
.llong 42b,142b
.llong 88b,188b
.llong 43b,143b
.llong 89b,189b
.llong 90b,190b
.llong 91b,191b
.llong 92b,192b
.text
/*
* Routine to copy a whole page of data, optimized for POWER4.
* On POWER4 it is more than 50% faster than the simple loop
* above (following the .Ldst_aligned label) but it runs slightly
* slower on POWER3.
*/
.Lcopy_page_4K:
std r31,-32(1)
std r30,-40(1)
std r29,-48(1)
std r28,-56(1)
std r27,-64(1)
std r26,-72(1)
std r25,-80(1)
std r24,-88(1)
std r23,-96(1)
std r22,-104(1)
std r21,-112(1)
std r20,-120(1)
li r5,4096/32 - 1
addi r3,r3,-8
li r0,5
0: addi r5,r5,-24
mtctr r0
20: ld r22,640(4)
21: ld r21,512(4)
22: ld r20,384(4)
23: ld r11,256(4)
24: ld r9,128(4)
25: ld r7,0(4)
26: ld r25,648(4)
27: ld r24,520(4)
28: ld r23,392(4)
29: ld r10,264(4)
30: ld r8,136(4)
31: ldu r6,8(4)
cmpwi r5,24
1:
32: std r22,648(3)
33: std r21,520(3)
34: std r20,392(3)
35: std r11,264(3)
36: std r9,136(3)
37: std r7,8(3)
38: ld r28,648(4)
39: ld r27,520(4)
40: ld r26,392(4)
41: ld r31,264(4)
42: ld r30,136(4)
43: ld r29,8(4)
44: std r25,656(3)
45: std r24,528(3)
46: std r23,400(3)
47: std r10,272(3)
48: std r8,144(3)
49: std r6,16(3)
50: ld r22,656(4)
51: ld r21,528(4)
52: ld r20,400(4)
53: ld r11,272(4)
54: ld r9,144(4)
55: ld r7,16(4)
56: std r28,664(3)
57: std r27,536(3)
58: std r26,408(3)
59: std r31,280(3)
60: std r30,152(3)
61: stdu r29,24(3)
62: ld r25,664(4)
63: ld r24,536(4)
64: ld r23,408(4)
65: ld r10,280(4)
66: ld r8,152(4)
67: ldu r6,24(4)
bdnz 1b
68: std r22,648(3)
69: std r21,520(3)
70: std r20,392(3)
71: std r11,264(3)
72: std r9,136(3)
73: std r7,8(3)
74: addi r4,r4,640
75: addi r3,r3,648
bge 0b
mtctr r5
76: ld r7,0(4)
77: ld r8,8(4)
78: ldu r9,16(4)
3:
79: ld r10,8(4)
80: std r7,8(3)
81: ld r7,16(4)
82: std r8,16(3)
83: ld r8,24(4)
84: std r9,24(3)
85: ldu r9,32(4)
86: stdu r10,32(3)
bdnz 3b
4:
87: ld r10,8(4)
88: std r7,8(3)
89: std r8,16(3)
90: std r9,24(3)
91: std r10,32(3)
9: ld r20,-120(1)
ld r21,-112(1)
ld r22,-104(1)
ld r23,-96(1)
ld r24,-88(1)
ld r25,-80(1)
ld r26,-72(1)
ld r27,-64(1)
ld r28,-56(1)
ld r29,-48(1)
ld r30,-40(1)
ld r31,-32(1)
li r3,0
blr
/*
* on an exception, reset to the beginning and jump back into the
* standard __copy_tofrom_user
*/
100: ld r20,-120(1)
ld r21,-112(1)
ld r22,-104(1)
ld r23,-96(1)
ld r24,-88(1)
ld r25,-80(1)
ld r26,-72(1)
ld r27,-64(1)
ld r28,-56(1)
ld r29,-48(1)
ld r30,-40(1)
ld r31,-32(1)
ld r3,-24(r1)
ld r4,-16(r1)
li r5,4096
b .Ldst_aligned
.section __ex_table,"a"
.align 3
.llong 20b,100b
.llong 21b,100b
.llong 22b,100b
.llong 23b,100b
.llong 24b,100b
.llong 25b,100b
.llong 26b,100b
.llong 27b,100b
.llong 28b,100b
.llong 29b,100b
.llong 30b,100b
.llong 31b,100b
.llong 32b,100b
.llong 33b,100b
.llong 34b,100b
.llong 35b,100b
.llong 36b,100b
.llong 37b,100b
.llong 38b,100b
.llong 39b,100b
.llong 40b,100b
.llong 41b,100b
.llong 42b,100b
.llong 43b,100b
.llong 44b,100b
.llong 45b,100b
.llong 46b,100b
.llong 47b,100b
.llong 48b,100b
.llong 49b,100b
.llong 50b,100b
.llong 51b,100b
.llong 52b,100b
.llong 53b,100b
.llong 54b,100b
.llong 55b,100b
.llong 56b,100b
.llong 57b,100b
.llong 58b,100b
.llong 59b,100b
.llong 60b,100b
.llong 61b,100b
.llong 62b,100b
.llong 63b,100b
.llong 64b,100b
.llong 65b,100b
.llong 66b,100b
.llong 67b,100b
.llong 68b,100b
.llong 69b,100b
.llong 70b,100b
.llong 71b,100b
.llong 72b,100b
.llong 73b,100b
.llong 74b,100b
.llong 75b,100b
.llong 76b,100b
.llong 77b,100b
.llong 78b,100b
.llong 79b,100b
.llong 80b,100b
.llong 81b,100b
.llong 82b,100b
.llong 83b,100b
.llong 84b,100b
.llong 85b,100b
.llong 86b,100b
.llong 87b,100b
.llong 88b,100b
.llong 89b,100b
.llong 90b,100b
.llong 91b,100b
Reference in New Issue View Git Blame Copy Permalink