lkml.org 
[lkml]   [2017]   [Sep]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: Does perf-annotate work correctly?
    On Tue, Sep 12, 2017 at 11:33:50AM -0300, Arnaldo Carvalho de Melo wrote:
    > Em Tue, Sep 12, 2017 at 06:10:35PM +0800, Du, Changbin escreveu:
    > > When a annotate a symbol, I find the annotated C source code doesn't match assembly code.
    > > So I cannot determine which line of C code has much overhead withou gdb's help.
    > >
    > > Here is a example result of function apic_has_interrupt_for_ppr() in kvm module.
    >
    > Ok, was this using the module .ko file or /proc/kcore? You forgot to
    > cut'n'paste the first line on the screen.
    >
    It is arch/x86/kvm/kvm.ko.

    > Also, how did you use gdb?
    >
    $ gdb arch/x86/kvm/kvm.ko
    $ (gdb) disassemble /s apic_has_interrupt_for_ppr

    > perf uses objdump to do the disassembly, and depending on how it is used
    > (live system, post processing on a different machine, permissions) it
    > may use different files to do the disassembly.
    >
    But objdump has same out as gdb. (Always on same machine, and no binary changed.)

    $ objdump -d -S arch/x86/kvm/kvm.o
    ...
    static int apic_has_interrupt_for_ppr(struct kvm_lapic *apic, u32 ppr)
    {
    3b4a0: e8 00 00 00 00 callq 3b4a5 <apic_has_interrupt_for_ppr+0x5>
    3b4a5: 55 push %rbp
    3b4a6: 48 89 e5 mov %rsp,%rbp
    3b4a9: 48 83 ec 08 sub $0x8,%rsp
    int highest_irr;
    if (kvm_x86_ops->sync_pir_to_irr && apic->vcpu->arch.apicv_active)
    3b4ad: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # 3b4b4 <apic_has_interrupt_for_ppr+0x14>
    3b4b4: 48 8b 80 38 02 00 00 mov 0x238(%rax),%rax
    3b4bb: 48 85 c0 test %rax,%rax
    3b4be: 74 10 je 3b4d0 <apic_has_interrupt_for_ppr+0x30>
    3b4c0: 48 8b 97 88 00 00 00 mov 0x88(%rdi),%rdx
    3b4c7: 80 ba 28 03 00 00 00 cmpb $0x0,0x328(%rdx)
    3b4ce: 75 3a jne 3b50a <apic_has_interrupt_for_ppr+0x6a>

    /*
    * Note that irr_pending is just a hint. It will be always
    * true with virtual interrupt delivery enabled.
    */
    if (!apic->irr_pending)
    3b4d0: 80 bf 91 00 00 00 00 cmpb $0x0,0x91(%rdi)
    3b4d7: 74 2a je 3b503 <apic_has_interrupt_for_ppr+0x63>
    3b4d9: 48 8b 8f a0 00 00 00 mov 0xa0(%rdi),%rcx
    static int find_highest_vector(void *bitmap)
    {
    int vec;
    u32 *reg;

    for (vec = MAX_APIC_VECTOR - APIC_VECTORS_PER_REG;
    3b4e0: b8 e0 00 00 00 mov $0xe0,%eax
    vec >= 0; vec -= APIC_VECTORS_PER_REG) {
    reg = bitmap + REG_POS(vec);
    if (*reg)
    3b4e5: 89 c2 mov %eax,%edx
    3b4e7: c1 fa 05 sar $0x5,%edx
    3b4ea: c1 e2 04 shl $0x4,%edx
    3b4ed: 48 63 d2 movslq %edx,%rdx
    3b4f0: 8b 94 11 00 02 00 00 mov 0x200(%rcx,%rdx,1),%edx
    3b4f7: 85 d2 test %edx,%edx
    3b4f9: 75 2d jne 3b528 <apic_has_interrupt_for_ppr+0x88>



    > Please provide more detailed information on the exact command line
    > arguments and usage scenario.
    >
    > - Arnaldo

    >
    > > │580 __clear_bit(KVM_APIC_PV_EOI_PENDING, &vcpu->arch.apic_attention); ▒
    > > │581 } ▒
    > > │ ▒
    > > │583 static int apic_has_interrupt_for_ppr(struct kvm_lapic *apic, u32 ppr) ▒
    > > │584 { ▒
    > > 0.88 │30: cmpb $0x0,0x91(%rdi) ▒
    > > 2.54 │ ↓ je 63 ▒
    > > 0.20 │ mov 0xa0(%rdi),%rcx ▒
    > > │581 int highest_irr; ▒
    > > │582 if (kvm_x86_ops->sync_pir_to_irr && apic->vcpu->arch.apicv_active) ▒
    > > 4.91 │ mov $0xe0,%eax x ▒
    > > 1.46 │45: mov %eax,%edx x ▒
    > > 0.02 │ sar $0x5,%edx x ▒
    > > 3.57 │ shl $0x4,%edx x ▒
    > > 3.34 │ movslq %edx,%rdx x ▒
    > > 1.25 │ mov 0x200(%rcx,%rdx,1),%edx x ▒
    > > 42.44 │ test %edx,%edx x ▒
    > > 0.01 │ ┌──jne 88 x ▒
    > > 3.48 │ │ sub $0x20,%eax x ▒
    > > 2.24 │ │ cmp $0xffffffe0,%eax x ▒
    > > │586│apic_find_highest_irr(): ▒
    > > │ │ ▒
    > > │407│ /* ▒
    > > │408│ * Note that irr_pending is just a hint. It will be always ▒
    > > │409│ * true with virtual interrupt delivery enabled. ▒
    > > │410│ */ ▒
    > > │411│ if (!apic->irr_pending) ▒
    > > │ │↑ jne 45 ▒
    > > 0.62 │63:│ mov $0xffffffff,%eax ◆
    > > 0.83 │ │ leaveq ▒
    > > 13.52 │ │← retq ▒
    > > │6a:│ mov %esi,-0x4(%rbp) ▒
    > > │ │ mov %rdx,%rdi ▒
    > > │418│find_highest_vector(): ▒
    > > │340│static int find_highest_vector(void *bitmap) ▒
    > > │341│{ ▒
    > > │342│ int vec; ▒
    > > │343│ u32 *reg; ▒
    > > │ │ ▒
    > > │345│ for (vec = MAX_APIC_VECTOR - APIC_VECTORS_PER_REG; ▒
    > > │ │→ callq *%rax ▒
    > > │ │ mov -0x4(%rbp),%esi ▒
    > > │343│ vec >= 0; vec -= APIC_VECTORS_PER_REG) { ▒
    > > │344│ reg = bitmap + REG_POS(vec); ▒
    > > │345│ if (*reg) ▒
    > > 0.05 │75:│ cmp $0xffffffff,%eax ▒
    > > │ │↑ je 63 ▒
    > > 1.95 │ │ mov %eax,%edx ▒
    > > 1.45 │ │ and $0xf0,%edx
    > >
    > >
    > > Look at the assembly code block where I have put a 'x' on the right. Apparently the
    > > assembly code doesn't match the C source code arrounded. Let's look the correct disassemble
    > > result from gdb:
    > >
    > > 340 for (vec = MAX_APIC_VECTOR - APIC_VECTORS_PER_REG;
    > > 0x000000000003b4e0 <+64>: mov $0xe0,%eax
    > >
    > > 342 reg = bitmap + REG_POS(vec);
    > > 343 if (*reg)
    > > 0x000000000003b4e5 <+69>: mov %eax,%edx
    > > 0x000000000003b4e7 <+71>: sar $0x5,%edx
    > > 0x000000000003b4ea <+74>: shl $0x4,%edx
    > > 0x000000000003b4ed <+77>: movslq %edx,%rdx
    > > 0x000000000003b4f0 <+80>: mov 0x200(%rcx,%rdx,1),%edx
    > > 0x000000000003b4f7 <+87>: test %edx,%edx
    > > 0x000000000003b4f9 <+89>: jne 0x3b528 <apic_has_interrupt_for_ppr+136>
    > >
    > > 341 vec >= 0; vec -= APIC_VECTORS_PER_REG) {
    > > 0x000000000003b4fb <+91>: sub $0x20,%eax
    > >
    > > 340 for (vec = MAX_APIC_VECTOR - APIC_VECTORS_PER_REG;
    > > 0x000000000003b4fe <+94>: cmp $0xffffffe0,%eax
    > > 0x000000000003b501 <+97>: jne 0x3b4e5 <apic_has_interrupt_for_ppr+69>
    > >
    > >
    > > Compared to gdb, perf-annoate has messed up. is it a bug or just perf is not as perfect as gdb?
    > >
    > > --
    > > Thanks,
    > > Changbin Du
    >
    >

    --
    Thanks,
    Changbin Du
    [unhandled content-type:application/pgp-signature]
    \
     
     \ /
      Last update: 2017-09-13 04:01    [W:3.238 / U:0.124 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site