Re: [GIT PULL] perf fixes

From: Steven Rostedt
Date: Fri Jun 22 2012 - 15:54:57 EST

On Fri, 2012-06-22 at 21:06 +0200, Hagen Paul Pfeifer wrote:

> >I may be more sensitive to this than most, because I look at profiles
> >and the function prologue just looks very ugly with the call mcount
> >thing. Ugh.
>
> Yes, ugh. Even Stevens -mfentry replacement do not change things here.

Why doesn't -mfentry help here? The link I showed still had frame
pointers enabled. With -mfentry, frame pointers do not need to be
enabled. And my latest patches do not automatically enable frame
pointers when enabling function tracing if -mfentry is supported.

I just ran a bunch of compiles against kernel/sched/core.c:

no pg, no mfentry, no fp:

0000000000000882 <schedule>:
882: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
889: 00 00
887: R_X86_64_32S current_task
88b: 57 push %rdi
88c: 48 8b 10 mov (%rax),%rdx
88f: 48 85 d2 test %rdx,%rdx
892: 74 45 je 8d9 <schedule+0x57>
894: 48 83 b8 60 06 00 00 cmpq $0x0,0x660(%rax)
89b: 00
89c: 75 3b jne 8d9 <schedule+0x57>
89e: 48 8b b8 60 0e 00 00 mov 0xe60(%rax),%rdi
8a5: 31 c0 xor %eax,%eax
8a7: 48 85 ff test %rdi,%rdi
8aa: 74 1a je 8c6 <schedule+0x44>
8ac: 48 8d 57 08 lea 0x8(%rdi),%rdx
8b0: 48 39 57 08 cmp %rdx,0x8(%rdi)
8b4: b0 01 mov $0x1,%al
8b6: 75 0e jne 8c6 <schedule+0x44>
8b8: 48 8d 47 18 lea 0x18(%rdi),%rax
8bc: 48 39 47 18 cmp %rax,0x18(%rdi)
8c0: 0f 95 c0 setne %al
8c3: 0f b6 c0 movzbl %al,%eax
8c6: 85 c0 test %eax,%eax
8c8: 74 0f je 8d9 <schedule+0x57>
8ca: 48 85 ff test %rdi,%rdi
8cd: 74 0a je 8d9 <schedule+0x57>
8cf: be 01 00 00 00 mov $0x1,%esi
8d4: e8 00 00 00 00 callq 8d9 <schedule+0x57>
8d5: R_X86_64_PC32 blk_flush_plug_list-0x4
8d9: e8 70 f9 ff ff callq 24e <__schedule>
8de: 5e pop %rsi
8df: c3 retq


no pg, no mfentry, with fp:

00000000000008cb <schedule>:
8cb: 55 push %rbp
8cc: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
8d3: 00 00
8d1: R_X86_64_32S current_task
8d5: 48 8b 10 mov (%rax),%rdx
8d8: 48 89 e5 mov %rsp,%rbp
8db: 48 85 d2 test %rdx,%rdx
8de: 74 45 je 925 <schedule+0x5a>
8e0: 48 83 b8 60 06 00 00 cmpq $0x0,0x660(%rax)
8e7: 00
8e8: 75 3b jne 925 <schedule+0x5a>
8ea: 48 8b b8 60 0e 00 00 mov 0xe60(%rax),%rdi
8f1: 31 c0 xor %eax,%eax
8f3: 48 85 ff test %rdi,%rdi
8f6: 74 1a je 912 <schedule+0x47>
8f8: 48 8d 57 08 lea 0x8(%rdi),%rdx
8fc: 48 39 57 08 cmp %rdx,0x8(%rdi)
900: b0 01 mov $0x1,%al
902: 75 0e jne 912 <schedule+0x47>
904: 48 8d 47 18 lea 0x18(%rdi),%rax
908: 48 39 47 18 cmp %rax,0x18(%rdi)
90c: 0f 95 c0 setne %al
90f: 0f b6 c0 movzbl %al,%eax
912: 85 c0 test %eax,%eax
914: 74 0f je 925 <schedule+0x5a>
916: 48 85 ff test %rdi,%rdi
919: 74 0a je 925 <schedule+0x5a>
91b: be 01 00 00 00 mov $0x1,%esi
920: e8 00 00 00 00 callq 925 <schedule+0x5a>
921: R_X86_64_PC32 blk_flush_plug_list-0x4
925: e8 41 f9 ff ff callq 26b <__schedule>
92a: 5d pop %rbp
92b: c3 retq

The above is our basis. Now lets look at the current -pg

with pg, no mfentry, with fp:

000000000000090c <schedule>:
90c: 55 push %rbp
90d: 48 89 e5 mov %rsp,%rbp
910: e8 00 00 00 00 callq 915 <schedule+0x9>
911: R_X86_64_PC32 mcount-0x4
915: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
91c: 00 00
91a: R_X86_64_32S current_task
91e: 48 8b 10 mov (%rax),%rdx
921: 48 85 d2 test %rdx,%rdx
924: 74 45 je 96b <schedule+0x5f>
926: 48 83 b8 60 06 00 00 cmpq $0x0,0x660(%rax)
92d: 00
92e: 75 3b jne 96b <schedule+0x5f>
930: 48 8b b8 60 0e 00 00 mov 0xe60(%rax),%rdi
937: 31 c0 xor %eax,%eax
939: 48 85 ff test %rdi,%rdi
93c: 74 1a je 958 <schedule+0x4c>
93e: 48 8d 57 08 lea 0x8(%rdi),%rdx
942: 48 39 57 08 cmp %rdx,0x8(%rdi)
946: b0 01 mov $0x1,%al
948: 75 0e jne 958 <schedule+0x4c>
94a: 48 8d 47 18 lea 0x18(%rdi),%rax
94e: 48 39 47 18 cmp %rax,0x18(%rdi)
952: 0f 95 c0 setne %al
955: 0f b6 c0 movzbl %al,%eax
958: 85 c0 test %eax,%eax
95a: 74 0f je 96b <schedule+0x5f>
95c: 48 85 ff test %rdi,%rdi
95f: 74 0a je 96b <schedule+0x5f>
961: be 01 00 00 00 mov $0x1,%esi
966: e8 00 00 00 00 callq 96b <schedule+0x5f>
967: R_X86_64_PC32 blk_flush_plug_list-0x4
96b: e8 37 f9 ff ff callq 2a7 <__schedule>
970: 5d pop %rbp
971: c3 retq

Looks like %rsp is saved in %rbp here as well as the call to mcount.

-pg must have frame pointers when -mfentry is not included, so there is
no 'with pg, no mfentry, no fp'. Now lets look at mfentry:

with pg, with mfentry, with fp:

000000000000090c <schedule>:
90c: e8 00 00 00 00 callq 911 <schedule+0x5>
90d: R_X86_64_PC32 __fentry__-0x4
911: 55 push %rbp
912: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
919: 00 00
917: R_X86_64_32S current_task
91b: 48 8b 10 mov (%rax),%rdx
91e: 48 89 e5 mov %rsp,%rbp
921: 48 85 d2 test %rdx,%rdx
924: 74 45 je 96b <schedule+0x5f>
926: 48 83 b8 60 06 00 00 cmpq $0x0,0x660(%rax)
92d: 00
92e: 75 3b jne 96b <schedule+0x5f>
930: 48 8b b8 60 0e 00 00 mov 0xe60(%rax),%rdi
937: 31 c0 xor %eax,%eax
939: 48 85 ff test %rdi,%rdi
93c: 74 1a je 958 <schedule+0x4c>
93e: 48 8d 57 08 lea 0x8(%rdi),%rdx
942: 48 39 57 08 cmp %rdx,0x8(%rdi)
946: b0 01 mov $0x1,%al
948: 75 0e jne 958 <schedule+0x4c>
94a: 48 8d 47 18 lea 0x18(%rdi),%rax
94e: 48 39 47 18 cmp %rax,0x18(%rdi)
952: 0f 95 c0 setne %al
955: 0f b6 c0 movzbl %al,%eax
958: 85 c0 test %eax,%eax
95a: 74 0f je 96b <schedule+0x5f>
95c: 48 85 ff test %rdi,%rdi
95f: 74 0a je 96b <schedule+0x5f>
961: be 01 00 00 00 mov $0x1,%esi
966: e8 00 00 00 00 callq 96b <schedule+0x5f>
967: R_X86_64_PC32 blk_flush_plug_list-0x4
96b: e8 37 f9 ff ff callq 2a7 <__schedule>
970: 5d pop %rbp
971: c3 retq


It is identical with non -pg and frame pointers, except that we added a
call to fentry in the start of the function.

with pg, with fentry, no fp:

00000000000008c3 <schedule>:
8c3: e8 00 00 00 00 callq 8c8 <schedule+0x5>
8c4: R_X86_64_PC32 __fentry__-0x4
8c8: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
8cf: 00 00
8cd: R_X86_64_32S current_task
8d1: 57 push %rdi
8d2: 48 8b 10 mov (%rax),%rdx
8d5: 48 85 d2 test %rdx,%rdx
8d8: 74 45 je 91f <schedule+0x5c>
8da: 48 83 b8 60 06 00 00 cmpq $0x0,0x660(%rax)
8e1: 00
8e2: 75 3b jne 91f <schedule+0x5c>
8e4: 48 8b b8 60 0e 00 00 mov 0xe60(%rax),%rdi
8eb: 31 c0 xor %eax,%eax
8ed: 48 85 ff test %rdi,%rdi
8f0: 74 1a je 90c <schedule+0x49>
8f2: 48 8d 57 08 lea 0x8(%rdi),%rdx
8f6: 48 39 57 08 cmp %rdx,0x8(%rdi)
8fa: b0 01 mov $0x1,%al
8fc: 75 0e jne 90c <schedule+0x49>
8fe: 48 8d 47 18 lea 0x18(%rdi),%rax
902: 48 39 47 18 cmp %rax,0x18(%rdi)
906: 0f 95 c0 setne %al
909: 0f b6 c0 movzbl %al,%eax
90c: 85 c0 test %eax,%eax
90e: 74 0f je 91f <schedule+0x5c>
910: 48 85 ff test %rdi,%rdi
913: 74 0a je 91f <schedule+0x5c>
915: be 01 00 00 00 mov $0x1,%esi
91a: e8 00 00 00 00 callq 91f <schedule+0x5c>
91b: R_X86_64_PC32 blk_flush_plug_list-0x4
91f: e8 66 f9 ff ff callq 28a <__schedule>
924: 5e pop %rsi
925: c3 retq

Now here's the big difference from -pg. This is identical to compiling
without frame pointers with the exception of the fentry call at the
start of the function.

Now what's the issue with function prologues with -mfentry?

-- Steve


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/