Re: [PATCH] tracing: Cleanup the convoluted softirq tracepoints

From: Jason Baron
Date: Tue Oct 19 2010 - 17:29:04 EST


On Tue, Oct 19, 2010 at 09:49:45PM +0200, Thomas Gleixner wrote:
> > > On Tue, 19 Oct 2010, Steven Rostedt wrote:
> > as an excuse for adding extra performance impact to kernel code, because when it
> > will be replaced by asm gotos, all that will be left is the performance impact
> > inappropriately justified as insignificant compared to the impact of the old
> > tracepoint scheme.
>
> Can you at one point just stop your tracing lectures and look at the
> facts ?
>
> The impact of a sensible tracepoint design on the code in question
> before kstat_incr_softirqs_this_cpu() was added would have been a mere
> _FIVE_ bytes of text. But the original tracepoint code itself is
> _TWENTY_ bytes of text larger.
>
> So we trade horrible code plus 20 bytes text against 5 bytes of text
> in the hotpath. And you tell me that these _FIVE_ bytes are impacting
> performance so much that it's significant.
>
> Now with kstat_incr_softirqs_this_cpu() the impact is zero, it even
> removes code.
>
> And talking about non impact of disabled trace points. The tracepoint
> in question which made me look at the code results in deinlining
> __raise_softirq_irqsoff() in net/dev/core.c. There goes your theory.
>
> So no, you _cannot_ tell what impact a tracepoint has in reality
> except by looking at the assembly output.
>
> And what scares me way more is the size of a single tracepoint in a
> code file.
>
> Just adding "trace_softirq_entry(nr);" adds 88 bytes of text. So
> that's optimized tracing code ?
>
> All it's supposed to do is:
>
> if (enabled)
> trace_foo(nr);
>
> Replace "if (enabled)" with your favourite code patching jump label
> whatever magic. The above stupid version takes about 28, but the
> "optimized" tracing code makes that 88. Brilliant. That's inlining
> utter shite for no good reason. WTF is it necessary to inline all that
> gunk ?
>
> Please spare me the "jump label will make this less intrusive"
> lecture. I'm not interested at all.
>
> Let's instead look at some more facts:
>
> #include <linux/interrupt.h>
> #include <linux/module.h>
>
> #include <trace/events/irq.h>
>
> static struct softirq_action softirq_vec[NR_SOFTIRQS];
>
> void test(struct softirq_action *h)
> {
> trace_softirq_entry(h - softirq_vec);
>
> h->action(h);
> }
>
> Compile this code with GCC 4.5 with and without jump labels (zap the
> select HAVE_ARCH_JUMP_LABEL line in arch/x86/Kconfig)
>
> So now the !jumplabel case gives us:
>
> ../build/kernel/soft.o: file format elf64-x86-64
>
> Disassembly of section .text:
>
> 0000000000000000 <test>:
> 0: 55 push %rbp
> 1: 48 89 e5 mov %rsp,%rbp
> 4: 41 55 push %r13
> 6: 49 89 fd mov %rdi,%r13
> 9: 49 81 ed 00 00 00 00 sub $0x0,%r13
> 10: 41 54 push %r12
> 12: 49 c1 ed 03 shr $0x3,%r13
> 16: 49 89 fc mov %rdi,%r12
> 19: 53 push %rbx
> 1a: 48 83 ec 08 sub $0x8,%rsp
> 1e: 83 3d 00 00 00 00 00 cmpl $0x0,0x0(%rip) # 25 <test+0x25>
> 25: 74 4d je 74 <test+0x74>
> 27: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
> 2e: 00 00
> 30: ff 80 44 e0 ff ff incl -0x1fbc(%rax)
> 36: 48 8b 1d 00 00 00 00 mov 0x0(%rip),%rbx # 3d <test+0x3d>
> 3d: 48 85 db test %rbx,%rbx
> 40: 74 13 je 55 <test+0x55>
> 42: 48 8b 7b 08 mov 0x8(%rbx),%rdi
> 46: 44 89 ee mov %r13d,%esi
> 49: ff 13 callq *(%rbx)
> 4b: 48 83 c3 10 add $0x10,%rbx
> 4f: 48 83 3b 00 cmpq $0x0,(%rbx)
> 53: eb eb jmp 40 <test+0x40>
> 55: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
> 5c: 00 00
> 5e: ff 88 44 e0 ff ff decl -0x1fbc(%rax)
> 64: 48 8b 80 38 e0 ff ff mov -0x1fc8(%rax),%rax
> 6b: a8 08 test $0x8,%al
> 6d: 74 05 je 74 <test+0x74>
> 6f: e8 00 00 00 00 callq 74 <test+0x74>
> 74: 4c 89 e7 mov %r12,%rdi
> 77: 41 ff 14 24 callq *(%r12)
> 7b: 58 pop %rax
> 7c: 5b pop %rbx
> 7d: 41 5c pop %r12
> 7f: 41 5d pop %r13
> 81: c9 leaveq
> 82: c3 retq
>
> The jumplabel=y case gives:
>
> ../build/kernel/soft.o: file format elf64-x86-64
>
> Disassembly of section .text:
>
> 0000000000000000 <test>:
> 0: 55 push %rbp
> 1: 48 89 e5 mov %rsp,%rbp
> 4: 41 55 push %r13
> 6: 49 89 fd mov %rdi,%r13
> 9: 49 81 ed 00 00 00 00 sub $0x0,%r13
> 10: 41 54 push %r12
> 12: 49 c1 ed 03 shr $0x3,%r13
> 16: 49 89 fc mov %rdi,%r12
> 19: 53 push %rbx
> 1a: 48 83 ec 08 sub $0x8,%rsp
> 1e: e9 00 00 00 00 jmpq 23 <test+0x23>
> 23: eb 4d jmp 72 <test+0x72>
> 25: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
> 2c: 00 00
> 2e: ff 80 44 e0 ff ff incl -0x1fbc(%rax)
> 34: 48 8b 1d 00 00 00 00 mov 0x0(%rip),%rbx # 3b <test+0x3b>
> 3b: 48 85 db test %rbx,%rbx
> 3e: 74 13 je 53 <test+0x53>
> 40: 48 8b 7b 08 mov 0x8(%rbx),%rdi
> 44: 44 89 ee mov %r13d,%esi
> 47: ff 13 callq *(%rbx)
> 49: 48 83 c3 10 add $0x10,%rbx
> 4d: 48 83 3b 00 cmpq $0x0,(%rbx)
> 51: eb eb jmp 3e <test+0x3e>
> 53: 65 48 8b 04 25 00 00 mov %gs:0x0,%rax
> 5a: 00 00
> 5c: ff 88 44 e0 ff ff decl -0x1fbc(%rax)
> 62: 48 8b 80 38 e0 ff ff mov -0x1fc8(%rax),%rax
> 69: a8 08 test $0x8,%al
> 6b: 74 05 je 72 <test+0x72>
> 6d: e8 00 00 00 00 callq 72 <test+0x72>
> 72: 4c 89 e7 mov %r12,%rdi
> 75: 41 ff 14 24 callq *(%r12)
> 79: 58 pop %rax
> 7a: 5b pop %rbx
> 7b: 41 5c pop %r12
> 7d: 41 5d pop %r13
> 7f: c9 leaveq
> 80: c3 retq
>
> So that saves _TWO_ bytes of text and replaces:
>
> - 1e: 83 3d 00 00 00 00 00 cmpl $0x0,0x0(%rip) # 25 <test+0x25>
> - 25: 74 4d je 74 <test+0x74>
> + 1e: e9 00 00 00 00 jmpq 23 <test+0x23>
> + 23: eb 4d jmp 72 <test+0x72>
>
> So it trades a conditional vs. two jumps ? WTF ??
>

right, so the 'jmpq' on boot on x86 gets patched with 5 byte no-op
sequence. So in the disabled case we have no-op followed by a jump
around the disabled code.

> I thought that jumplabel magic was supposed to get rid of the jump
> over the tracing code ? In fact it adds another jump. Whatfor ?
>

yes, that is the plan. gcc does not yet support hot/cold labels...once
it does the second jump will go away and the entire tracepoint code will
be moved to a 'cold' section. It's not quite completely optimal yet, but
we are getting there.

> Now even worse, when you NOP out the jmpq then your tracepoint is
> still not enabled. Brilliant !
>

The 'jmpq' in the enabled case is patched with a jmpq to the body of the
tracepoint itself.

> Did you guys ever look at the assembly output of that insane shite you
> are advertising with lengthy explanations ?
>
> Obviously _NOT_
>
> Come back when you can show me a clean imlementation of all this crap
> which reproduces with my jumplabel enabled stock compiler. And please
> just send me a patch w/o the blurb.
>
> And sane looks like:
>
> jmpq 2f <---- This gets noped out
> 1:
> mov %r12,%rdi
> callq *(%r12)
> [whatever cleanup it takes ]
> leaveq
> retq
>
> 2f:
> [tracing gunk]
> jmp 1b
>

yes, this is what the code should look like when we get support for
hot/cold labels. I've discussed this support with gcc folk, and its the
next step here. So yes, this is exacatly where we are headed.

thanks,

-Jason
--
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/