On 03/23/2015 06:18 PM, Takashi Iwai wrote: > At Mon, 23 Mar 2015 17:07:15 +0100, Denys Vlasenko wrote: >>>> I pulled tip tree on top of 4.0-rc5, built with your patch and now >>>> succeeded to get a better message: >>>> >>>> kvm: zapping shadow pages for mmio generation wraparound >>>> kvm [5126]: vcpu0 disabled perfctr wrmsr: 0xc1 data 0xffff >>>> Exception on user stack 00007ffd22c23ef0: RSP: 0018:00007ffd22c23f28 >>>> EFLAGS: 00010006 >>>> RIP: 0010:[<ffffffff8162681d>] [<ffffffff8162681d>] >>>> netlink_attachskb+0x1d/0x1d0 >>>> PANIC: double fault, error_code: 0x0 >>>> CPU: 1 PID: 10819 Comm: cc1 Tainted: G W 4.0.0-rc5-debug1+ #2 >>>> Hardware name: Dell Inc. OptiPlex 9010/0M9KCM, BIOS A12 01/10/2013 >>>> task: ffff8800d1b34b10 ti: ffff8800d1b30000 task.ti: ffff8800d1b30000 >>>> RIP: 0010:[<ffffffff8162681d>] [<ffffffff8162681d>] >>>> netlink_attachskb+0x1d/0x1d0 >>>> RSP: 0018:00007ffd22c23f28 EFLAGS: 00010006 >>>> RAX: 0000000000000000 RBX: 0000000000000005 RCX: 00000000c0000101 >>>> RDX: 0000000000000000 RSI: 0000000000000001 RDI: 00007ffd22c23ef0
>> FYI: the disassembly of netlink_attachskb (from "Code:" line) is: >> >> 0: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1) >> 5: 55 push %rbp >> 6: 48 89 e5 mov %rsp,%rbp >> 9: 41 56 push %r14 >> b: 41 55 push %r13 >> d: 49 89 d5 mov %rdx,%r13 >> 10: 41 54 push %r12 >> 12: 49 89 f4 mov %rsi,%r12 >> 15: 53 push %rbx >> 16: 48 89 fb mov %rdi,%rbx >> 19: 48 83 ec 30 sub $0x30,%rsp >> 1d: 8b 87 68 01 00 00 mov 0x168(%rdi),%eax >> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >> 23: 39 87 9c 01 00 00 cmp %eax,0x19c(%rdi) >> 29: 7c 25 jl 50 <_start+0x50> >> 2b: 48 8b 87 88 04 00 00 mov 0x488(%rdi),%rax >> >> The ^^^^^ instruction is the one which faults. Since you said it >> consistently happens here, this should be a page fault, not an external >> hardware interrupt. >> >> The code corresponds to the comparison in if(): >> >> int netlink_attachskb(struct sock *sk, struct sk_buff *skb, >> long *timeo, struct sock *ssk) >> { >> struct netlink_sock *nlk; >> >> nlk = nlk_sk(sk); >> >> if ((atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || >>> - Another piece is that the bug happens only when a KVM is running. >>> The kernel ran without problem over days with similar tasks >>> (compiling kernel, etc) when no KVM was used. >> >> Conceivably virtualization support in CPUs can have nasty erratas. >> However, you and other reporter have different CPUs - yours >> is Ivy Bridge, his CPU is a Penryn. >> >> I don't see the path how KVM helps to trigger this. >> >>> - And now I get the trace as above, pointing netlink_attachskb(). >>> >>> I have a difficulty to imagine how all these pieces fit into a single >>> picture. Is something already screwed up before that? >> >> Well, a tiny bit more info will be seen if you'd change %rdi >> to, say, %r15 in these two lines in my patch: >> >> /* Save bogus RSP value */ >> movq %rsp,%rdi >> ... >> push %rdi /* pt_regs->sp */ >> >> Then original %rdi will be visible in the crash message. > > OK, here we go. > > kvm: zapping shadow pages for mmio generation wraparound > kvm [5490]: vcpu0 disabled perfctr wrmsr: 0xc1 data 0xffff > Exception on user stack 00007fff1d7e5ec0: RSP: 0018:00007fff1d7e5ef8 > EFLAGS: 00010002 > RIP: 0010:[<ffffffff8162681d>] [<ffffffff8162681d>] > netlink_attachskb+0x1d/0x1d0 > PANIC: double fault, error_code: 0x0 > CPU: 5 PID: 14285 Comm: fixdep Tainted: G W 4.0.0-rc5-debug1+ #3 > Hardware name: Dell Inc. OptiPlex 9010/0M9KCM, BIOS A12 01/10/2013 > task: ffff88020ba1c690 ti: ffff880206ba4000 task.ti: ffff880206ba4000 > RIP: 0010:[<ffffffff8162681d>] [<ffffffff8162681d>] > netlink_attachskb+0x1d/0x1d0 > RSP: 0018:00007fff1d7e5ef8 EFLAGS: 00010002 > RAX: 0000000000000000 RBX: 0000000000000000 RCX: 00000000c0000101 > RDX: 0000000000000000 RSI: 0000000000001ebb RDI: 0000000000000000 Thanks for your testing. So the %rdi was NULL... not very informative. Notice that your every crash is preceded by kvm: zapping shadow pages for mmio generation wraparound kvm [5490]: vcpu0 disabled perfctr wrmsr: 0xc1 data 0xffff This hints that kvm _is_ somehow responsible. I'm no expert on kvm, I need to take a look around that code... -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/