From: Zhang Yanmin <[email protected]>
ChangeLog V4: Explain the patch sceanrio clearly
ChangeLog V3: Keep rsp pointing to pt_regs before sysexit.
ChangeLog V2: Before sysexit, perf NMI might arrive. There is
still a race. Here we change rsp to keep it pointing
to pt_regs->orig_ax.
In addition, after sti, before sysexit, an irq might
arrives. That causes more chances for perf NMI to jump
in.
We hit a kernel panic when running perf to collect performance data
with callchain info.
kenel is x86_64 and user space apps are 32bit.
Run command:
# perf record -a -g -f sleep 30
kernel panic usually within 30 seconds.
BUG: unable to handle kernel NULL pointer dereference at 0000000000000004
IP: [<ffffffff82012091>] get_segment_base+0x71/0xc0
PGD 6c65f067 PUD 0
Oops: 0000 [#1] PREEMPT SMP
Modules linked in: <...>
CPU: 1 PID: 304 Comm: Binder_2 Tainted: G W O 3.10.20-263902-g184bfbc-dirty #14
task: ffff8800764dc300 ti: ffff88006c6e8000 task.ti: ffff88006c6e8000
RIP: 0010:[<ffffffff82012091>] [<ffffffæf82012091>] get_segment_base+0x71/0xc0
RSP: 0018:ffff^X8007ea87b98 EFLAGS: 00010092
RAX: 0000000000000024 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000009
RBP: ffff88007ea87ba8 R08: ffffffff83143b3c R09: ffffffff831848a8
R10: 0000000000000000 R11: 00000000001bf2d8 R12: 0000000000000000
R13: ffff88006c6e9fd8 R14: ffff88006c6e9f58 R15: ffff8800764dc300
FS: 0000000000000000(0000) GS:ffff88007ea80000(006b) knlGS:00000000f704add0
CS: 0010 DS: 002b ES: 002b CR0: 0000000080050033
CR2: 0000000000000004 CR3: 0000000076588000 CR4: 00000^P00001007e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Stack:
ffff88005f266c00 0000000000000000 ffff88007ea87c18 ffffffff82013cac
ffff88007ea87d58 00000016fe4704a0 00000000000001a7 ffff88007ea87ef8
ffff88005f266c00 ffff88007ea87ef8 ffff8800!e07b400 ffff88005f266c00
Call Trace:
<NMI>
[<ffffffff82013cac>] perf_callchain_user+0x15c/0x240
[<ffffffff82160754>] perf_callchain+0x134/0x180
[<ffffffff820e0787>] ? local_clock+0x47/0x60
[<ffffffff8215d49b>] perf_prepare_sample+0x1bb/0x240
[<ffffffff8215d667>] __perf_event_overflow+0x147/0x230
[<ffffffff82012f68>] ? x86_perf_event_set_period+0xd8/0x150
[<ffffffff8215df24>] perf_event_overflow+0x14/0x20
[<ffffffff820194d2>] intel_pmu_handle_irq+0x1c2/0x270
[<ffffffff828b5d60>] ? call_softirq+0x30/0x30
[<ffffffff828aff01>] perf_event_nmi_handler+0x21/0x30
[<ffffffff828af5b9>] nmi_handle.isr!.1+0x59/0x=0
[<ffffffff828af6d8>] default_do_nmi+0x58/0x240
[<ffffffff828af978>] do_nmi+0xb8/0xf0
[<ffffffff828aebe7>] end_repeat_nmi+0x1e/0x2e
[<ffffffff828b5d60>] ? call_softirq+0x30/0x30
[<ffffffff828b5d60>] ? call_softirq+0x30/0x30
[<fFffffff828b5d60>] ? call_softirq+0x30/0x30
perf_callchain_user32 calls get_segment_base to get cs/ss base address.
At kernel panic, get_segment_base considers the cs as LDT index and
uses current->active_mm->context.ldt to access the desc. But the app
is 32bit and doesn't use LDT. current->active_mm->context.ldt is equal
to NULL. That causes a bad dereference and kernel panic.
We dump pt_regs in function perf_callchain_user32. At panic, the values
are incorrect. After collecting lots of logs, we find it always happens
when app runs system call. Sometimes, the panic callchain has the address
of trace_hardirqs_on_thunk.
perf_callchain checks if pt_regs saves user space reg info. If not,
perf_callchain calls task_pt_regs(current) to get the address of pt_regs
on the top of kernel stack.
After checking sysexit_from_sys_call, we find pt_regs on the top of the
stack might be erased by trace_hardirqs_on_thunk or other interrupt
handlers.
Basically, ia32 uses sysenter to start system calls.
sysexit_from_sys_call=>trace_hardirqs_on_thunk. Before calling
trace_hardirqs_on_thunk, sysexit_from_sys_call already pops up pt_regs,
and register rsp points to pt_regs->ss, almost at the top of the kernel
stack. Then trace_hardirqs_on_thunk is called and it uses kernel stack
to save local vars, which erase old pt_regs. If perf NMI happens here,
perf might access a ruined pt_regs when saving userspace callchain.
If app is 64bit, it doesn't go through this path. sysret_check keeps
rsp pointing to pt_regs before executing sysretq to exit to user space.
The patch fixes it by keeping rsp pointing to pt_regs like 64bit path.
Signed-off-by: Zhang Yanmin <[email protected]>
Signed-off-by: Liu Shuox <[email protected]>
---
arch/x86/ia32/ia32entry.S | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/arch/x86/ia32/ia32entry.S b/arch/x86/ia32/ia32entry.S
index 4299eb0..d2f905b 100644
--- a/arch/x86/ia32/ia32entry.S
+++ b/arch/x86/ia32/ia32entry.S
@@ -172,15 +172,16 @@ sysexit_from_sys_call:
andl $~0x200,EFLAGS-R11(%rsp)
movl RIP-R11(%rsp),%edx /* User %eip */
CFI_REGISTER rip,rdx
- RESTORE_ARGS 0,24,0,0,0,0
- xorq %r8,%r8
+ RESTORE_ARGS 0,-ARG_SKIP,0,0,0,0
+ movq EFLAGS-R11(%rsp),%r8 /* rflags */
+ movq RSP-R11(%rsp),%rcx /* User %esp */
xorq %r9,%r9
xorq %r10,%r10
xorq %r11,%r11
- popfq_cfi
+ pushq_cfi %r8
/*CFI_RESTORE rflags*/
- popq_cfi %rcx /* User %esp */
- CFI_REGISTER rsp,rcx
+ popfq_cfi
+ xorq %r8,%r8
TRACE_IRQS_ON
ENABLE_INTERRUPTS_SYSEXIT32
--
1.8.3.2