lkml.org 
[lkml]   [2008]   [Nov]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [Bug #11308] tbench regression on each kernel release from 2.6.22 -> 2.6.28

* Ingo Molnar <mingo@elte.hu> wrote:

> 100.000000 total
> ................
> 3.038025 skb_release_data

hits (303802 total)
.........
ffffffff80488c7e: 780 <skb_release_data>:
ffffffff80488c7e: 780 55 push %rbp
ffffffff80488c7f: 267141 53 push %rbx
ffffffff80488c80: 0 48 89 fb mov %rdi,%rbx
ffffffff80488c83: 3552 48 83 ec 08 sub $0x8,%rsp
ffffffff80488c87: 604 8a 47 7c mov 0x7c(%rdi),%al
ffffffff80488c8a: 2644 a8 02 test $0x2,%al
ffffffff80488c8c: 49 74 2a je ffffffff80488cb8 <skb_release_data+0x3a>
ffffffff80488c8e: 0 83 e0 10 and $0x10,%eax
ffffffff80488c91: 2079 8b 97 c8 00 00 00 mov 0xc8(%rdi),%edx
ffffffff80488c97: 53 3c 01 cmp $0x1,%al
ffffffff80488c99: 0 19 c0 sbb %eax,%eax
ffffffff80488c9b: 870 48 03 97 d0 00 00 00 add 0xd0(%rdi),%rdx
ffffffff80488ca2: 65 66 31 c0 xor %ax,%ax
ffffffff80488ca5: 0 05 01 00 01 00 add $0x10001,%eax
ffffffff80488caa: 888 f7 d8 neg %eax
ffffffff80488cac: 49 89 c1 mov %eax,%ecx
ffffffff80488cae: 0 f0 0f c1 0a lock xadd %ecx,(%rdx)
ffffffff80488cb2: 1909 01 c8 add %ecx,%eax
ffffffff80488cb4: 1040 85 c0 test %eax,%eax
ffffffff80488cb6: 0 75 6d jne ffffffff80488d25 <skb_release_data+0xa7>
ffffffff80488cb8: 0 8b 93 c8 00 00 00 mov 0xc8(%rbx),%edx
ffffffff80488cbe: 4199 48 8b 83 d0 00 00 00 mov 0xd0(%rbx),%rax
ffffffff80488cc5: 4995 31 ed xor %ebp,%ebp
ffffffff80488cc7: 0 66 83 7c 10 04 00 cmpw $0x0,0x4(%rax,%rdx,1)
ffffffff80488ccd: 983 75 15 jne ffffffff80488ce4 <skb_release_data+0x66>
ffffffff80488ccf: 15 eb 28 jmp ffffffff80488cf9 <skb_release_data+0x7b>
ffffffff80488cd1: 665 48 63 c5 movslq %ebp,%rax
ffffffff80488cd4: 546 ff c5 inc %ebp
ffffffff80488cd6: 328 48 c1 e0 04 shl $0x4,%rax
ffffffff80488cda: 356 48 8b 7c 02 20 mov 0x20(%rdx,%rax,1),%rdi
ffffffff80488cdf: 95 e8 be 87 de ff callq ffffffff802714a2 <put_page>
ffffffff80488ce4: 66 8b 93 c8 00 00 00 mov 0xc8(%rbx),%edx
ffffffff80488cea: 1321 48 03 93 d0 00 00 00 add 0xd0(%rbx),%rdx
ffffffff80488cf1: 439 0f b7 42 04 movzwl 0x4(%rdx),%eax
ffffffff80488cf5: 0 39 c5 cmp %eax,%ebp
ffffffff80488cf7: 1887 7c d8 jl ffffffff80488cd1 <skb_release_data+0x53>
ffffffff80488cf9: 2187 8b 93 c8 00 00 00 mov 0xc8(%rbx),%edx
ffffffff80488cff: 1784 48 8b 83 d0 00 00 00 mov 0xd0(%rbx),%rax
ffffffff80488d06: 422 48 83 7c 10 18 00 cmpq $0x0,0x18(%rax,%rdx,1)
ffffffff80488d0c: 110 74 08 je ffffffff80488d16 <skb_release_data+0x98>
ffffffff80488d0e: 0 48 89 df mov %rbx,%rdi
ffffffff80488d11: 0 e8 52 ff ff ff callq ffffffff80488c68 <skb_drop_fraglist>
ffffffff80488d16: 14 48 8b bb d0 00 00 00 mov 0xd0(%rbx),%rdi
ffffffff80488d1d: 715 5e pop %rsi
ffffffff80488d1e: 109 5b pop %rbx
ffffffff80488d1f: 20 5d pop %rbp
ffffffff80488d20: 980 e9 b7 66 e0 ff jmpq ffffffff8028f3dc <kfree>
ffffffff80488d25: 0 59 pop %rcx
ffffffff80488d26: 1948 5b pop %rbx
ffffffff80488d27: 0 5d pop %rbp
ffffffff80488d28: 0 c3 retq

this is a short function, and 90% of the overhead is false leaked-in
overhead from callsites:

ffffffff80488c7f: 267141 53 push %rbx

unfortunately i have a hard time mapping its callsites.
pskb_expand_head() is the only static callsite, but it's not active in
the profile.

The _usual_ callsite is normally skb_release_all(), which does have
overhead:

ffffffff80489449: 925 <skb_release_all>:
ffffffff80489449: 925 53 push %rbx
ffffffff8048944a: 5249 48 89 fb mov %rdi,%rbx
ffffffff8048944d: 4 e8 3c ff ff ff callq ffffffff8048938e <skb_release_head_state>
ffffffff80489452: 1149 48 89 df mov %rbx,%rdi
ffffffff80489455: 13163 5b pop %rbx
ffffffff80489456: 0 e9 23 f8 ff ff jmpq ffffffff80488c7e <skb_release_data>

it is also tail-optimized, which explains why i found little
callsites. The main callsite of skb_release_all() is:

ffffffff80488b86: 26 e8 be 08 00 00 callq ffffffff80489449 <skb_release_all>

which is __kfree_skb(). That is a frequently referenced function, and
in my profile there's a single callsite active:

ffffffff804c1027: 432 e8 56 7b fc ff callq ffffffff80488b82 <__kfree_skb>

which is tcp_ack() - subject of a later email. The wider context is:

ffffffff804c0ffc: 433 41 2b 85 e0 00 00 00 sub 0xe0(%r13),%eax
ffffffff804c1003: 4843 89 85 f0 00 00 00 mov %eax,0xf0(%rbp)
ffffffff804c1009: 1730 48 8b 45 30 mov 0x30(%rbp),%rax
ffffffff804c100d: 311 41 8b 95 e0 00 00 00 mov 0xe0(%r13),%edx
ffffffff804c1014: 0 48 83 b8 b0 00 00 00 cmpq $0x0,0xb0(%rax)
ffffffff804c101b: 0 00
ffffffff804c101c: 418 74 06 je ffffffff804c1024 <tcp_ack+0x50d>
ffffffff804c101e: 37 01 95 f4 00 00 00 add %edx,0xf4(%rbp)
ffffffff804c1024: 2 4c 89 ef mov %r13,%rdi
ffffffff804c1027: 432 e8 56 7b fc ff callq ffffffff80488b82 <__kfree_skb>

this is a good, top-of-the-line x86 CPU with a really good BTB
implementation that seems to be able to fall through calls and tail
optimizations as if they werent there.

some guesses are:

(gdb) list *0xffffffff804c1003
0xffffffff804c1003 is in tcp_ack (include/net/sock.h:789).
784
785 static inline void sk_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
786 {
787 skb_truesize_check(skb);
788 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
789 sk->sk_wmem_queued -= skb->truesize;
790 sk_mem_uncharge(sk, skb->truesize);
791 __kfree_skb(skb);
792 }
793

both sk and skb should be cache-hot here so this seems unlikely.

(gdb) list *0xffffffff804c10090xffffffff804c1009 is in tcp_ack (include/net/sock.h:736).
731 }
732
733 static inline int sk_has_account(struct sock *sk)
734 {
735 /* return true if protocol supports memory accounting */
736 return !!sk->sk_prot->memory_allocated;
737 }
738
739 static inline int sk_wmem_schedule(struct sock *sk, int size)
740 {

this cannot be it - unless sk_prot somehow ends up being dirtied or
false-shared?

Still, my guess would be on ffffffff804c1009 and a
sk_prot->memory_allocated cachemiss: look at how this instruction uses
%ebp, and the one that shows the many hits in skb_release_data()
pushes %ebp to the stack - that's where the CPU's OOO trick ends: it
has to compute the result and serialize on the cachemiss.

Ingo



\
 
 \ /
  Last update: 2008-11-17 21:51    [W:0.362 / U:3.040 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site