On Thu, Jan 4, 2018 at 8:38 AM, Pavel Tatashin soleen@gmail.com wrote:
I am getting the following panic when trying to boot 4.4.110rc1 on Intel(R) Xeon(R) CPU E5-2630:
[ 5.923489] BUG: unable to handle kernel NULL pointer dereference at 000000000000000d [ 5.932259] IP: [<ffffffff810e70d2>] dyntick_save_progress_counter+0x12/0x50
Hmm. You don't have the "Code:" line in this oops anywhere, do you?
[ 5.977905] RIP: dyntick_save_progress_counter+0x12/0x50 [ 5.988505] RSP: 0000:ffff881ff2f27dc0 EFLAGS: 00010046 [ 5.994434] RAX: 0000000000000001 RBX: ffffffff81b02140 RCX: ffff883fec768000 [ 6.002403] RDX: 0000000000000000 RSI: ffff881ff2f27e5f RDI: ffff88407e958140 [ 6.010368] RBP: ffff881ff2f27dc0 R08: ffff881ff2f27e78 R09: 000000016110f359 [ 6.018333] R10: 0000000000000b10 R11: 0000000000000000 R12: ffffffff81b02140 [ 6.026297] R13: 00000000ffffffdf R14: 0000000000000021 R15: 0000000200000000 [ 6.034262] FS: 0000000000000000(0000) GS:ffff881fff940000(0000) knlGS:0000000000000000 [ 6.043293] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 6.049707] CR2: 000000000000000d CR3: 0000000001aa6000 CR4: 0000000000360670 [ 6.057672] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 6.065638] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 6.073603] Stack: [ 6.075847] ffff881ff2f27e18 ffffffff810e8fac 0000000000000202 ffff881ff2f27e60 [ 6.084158] ffff881ff2f27e5f ffffffff810e70c0 ffffffff81b02140 ffffffff81b127a0 [ 6.092465] 0000000000000001 0000000000000000 0000000000000003 ffff881ff2f27eb8 [ 6.100768] Call Trace: [ 6.103501] [<ffffffff810e8fac>] force_qs_rnp+0xdc/0x150
The oops looks like it *might* be this:
lock xadd %edx,0xc(%rax)
which is from the
int snap = atomic_add_return(0, &rdtp->dynticks);
in rcu_dynticks_snap() because %rax is 1 and that would give you the invalid page fault and the right faulting address.
But that would be complete rcu data structure corruption (that rdtp pointer comes from
per_cpu_ptr(rsp->rda, cpu)
in force_qs_rnp(), afaik.
The PTI patches obviously change percpu stuff, but this looks like an odd place for that to manifest.
Linus