Messages in this thread | | | Date | Fri, 17 Apr 2009 00:52:29 -0400 | From | Mathieu Desnoyers <> | Subject | Re: [PATCH 2/2] tracing/events/lockdep: move tracepoints within recursive protection |
| |
* Steven Rostedt (rostedt@goodmis.org) wrote: > > Here's the dump that I get that triggers the lockdep warning: > > WARNING: at kernel/lockdep.c:2893 check_flags+0x1a7/0x1d0() > Hardware name: Precision WorkStation 470 > Modules linked in: radeon drm autofs4 hidp rfcomm l2cap bluetooth sunrpc > > [snip] > > Pid: 3768, comm: sshd Not tainted 2.6.30-rc1 #1036 > Call Trace: > [<ffffffff8025c321>] warn_slowpath+0xe1/0x110 > [<ffffffff80287f7c>] ? __lock_acquire+0x3ac/0xb20 > [<ffffffff80286dca>] ? validate_chain+0x4ca/0x12d0 > [<ffffffff80286dca>] ? validate_chain+0x4ca/0x12d0 > [<ffffffff80286dca>] ? validate_chain+0x4ca/0x12d0 > [<ffffffff80286dca>] ? validate_chain+0x4ca/0x12d0 > [<ffffffff80287f7c>] ? __lock_acquire+0x3ac/0xb20 > [<ffffffff8051a5a9>] ? __alloc_skb+0x49/0x160 > [<ffffffff80282407>] check_flags+0x1a7/0x1d0 > [<ffffffff80284d63>] lockdep_trace_alloc+0x33/0xe0 > [<ffffffff802f3682>] kmem_cache_alloc+0x32/0x160 > [<ffffffff8051a5a9>] __alloc_skb+0x49/0x160 > [<ffffffff8059192d>] tcp_send_ack+0x2d/0xe0 > [<ffffffff8058e941>] __tcp_ack_snd_check+0x61/0xb0 > [<ffffffff80590408>] tcp_rcv_established+0x398/0x600 > [<ffffffff80596a58>] tcp_v4_do_rcv+0x228/0x380 > [<ffffffff80228ed6>] ? ftrace_call+0x5/0x2b > [<ffffffff8058454e>] ? tcp_prequeue_process+0x2e/0xa0 > [<ffffffff8058458d>] tcp_prequeue_process+0x6d/0xa0 > [<ffffffff8058734a>] tcp_recvmsg+0x49a/0x880 > [<ffffffff80514267>] sock_common_recvmsg+0x37/0x50 > [<ffffffff805116b9>] sock_aio_read+0x109/0x110 > [<ffffffff802f75f1>] do_sync_read+0xf1/0x130 > [<ffffffff8022ec33>] ? sched_clock+0x13/0x20 > [<ffffffff8022ec5d>] ? native_sched_clock+0x1d/0x50 > [<ffffffff802737d0>] ? autoremove_wake_function+0x0/0x40 > [<ffffffff80228ed6>] ? ftrace_call+0x5/0x2b > [<ffffffff80381469>] ? cap_file_permission+0x9/0x10 > [<ffffffff80380116>] ? security_file_permission+0x16/0x20 > [<ffffffff802f7fc9>] vfs_read+0x159/0x170 > [<ffffffff802f8285>] sys_read+0x55/0x90 > [<ffffffff802291af>] system_call_fastpath+0x16/0x1b > ---[ end trace 03d889e04bc7a9a7 ]--- > possible reason: unannotated irqs-on. > irq event stamp: 12569 > hardirqs last enabled at (12567): [<ffffffff8026206a>] local_bh_enable+0xaa/0x110 > hardirqs last disabled at (12569): [<ffffffff80610c76>] int3+0x16/0x40 > softirqs last enabled at (12566): [<ffffffff80514d2b>] lock_sock_nested+0xfb/0x110 > softirqs last disabled at (12568): [<ffffffff8058454e>] tcp_prequeue_process+0x2e/0xa0 > > > Note, for some reason we hit int3 ?? > > Tracepoints do not use int3 does it? >
Not in your ftrace tree. My LTTng tree includes the immediate values, which brienfly uses the int3 handler when enabling/disabling tracepoints. But this seems unrelated to your problem.
> I have kprobes defined but not any kprobe self tests on. >
Could this be a userspace breakpoint then ?
> Anyway, let me describe what the above is and what I found in my > investigation. > > The lockdep took a check_flags error when it noticed that interrupts were > enabled, but the current->hardirqs_enabled was 0. Lockdep thought > interrupts were disabled but they were in fact enabled. > > > The last 4 lines of the warning have the numbers in the parenthesis > annotate the order of events: (Here they are in order) > > softirqs last enabled at (12566): [<ffffffff80514d2b>] lock_sock_nested+0xfb/0x110 > hardirqs last enabled at (12567): [<ffffffff8026206a>] local_bh_enable+0xaa/0x110 > softirqs last disabled at (12568): [<ffffffff8058454e>] tcp_prequeue_process+0x2e/0xa0 > hardirqs last disabled at (12569): [<ffffffff80610c76>] int3+0x16/0x40 > > The last change that lockdep saw was interrupts being disabled by int3. I > still don't understand why int3 was enabled. I have startup tests for > ftrace and the event tracer, but this blob happened when I first ssh'd > into the box.
Also note that maybe some entry.S annotation could be missing, making kallsyms think it was running within int3 when in fact it was running in a different function. I would double-check with objdump givin the ffffffff80610c76 address to make sure.
Just giving my 2 cents before going to bed. I don't have any more brain left for tonight.
Cheers,
Mathieu
> > In arch/x86/kernel/entry_64.S we have: > > paranoidzeroentry_ist int3 do_int3 DEBUG_STACK > > .macro paranoidzeroentry sym do_sym > ENTRY(\sym) > INTR_FRAME > PARAVIRT_ADJUST_EXCEPTION_FRAME > pushq $-1 /* ORIG_RAX: no syscall to restart */ > CFI_ADJUST_CFA_OFFSET 8 > subq $15*8, %rsp > call save_paranoid > TRACE_IRQS_OFF > movq %rsp,%rdi /* pt_regs pointer */ > xorl %esi,%esi /* no error code */ > call \do_sym > jmp paranoid_exit /* %ebx: no swapgs flag */ > CFI_ENDPROC > END(\sym) > .endm > > > ENTRY(paranoid_exit) > INTR_FRAME > DISABLE_INTERRUPTS(CLBR_NONE) > TRACE_IRQS_OFF > testl %ebx,%ebx /* swapgs needed? */ > jnz paranoid_restore > testl $3,CS(%rsp) > jnz paranoid_userspace > paranoid_swapgs: > TRACE_IRQS_IRETQ 0 > SWAPGS_UNSAFE_STACK > paranoid_restore: > RESTORE_ALL 8 > jmp irq_return > > > irq_return: > INTERRUPT_RETURN > > > INTERRUPT_RETURN is simply defined as iretq > > > I see that we call TRACE_IRQS_OFF when entering paranoid_exit, but if we > do not need to swapgs (we don't because int3 looks like it happened in > kernel space) we just call irq_return and jump back. We miss the fact that > the irq_return enables interrupts for us. > > I'd try to come up with a fix, but this gets a bit complex, and I figured > I let the lockdep irq-tracing guru's play with this magic. I'm just > reporting the problem ;-) > > -- Steve >
-- Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
| |