perf: fix kernel panic when parsing user space CS

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 some performance data.
kenel is x86_64 and user space apps are 32bit.

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
 [<fffff&ff820e0787>] ? 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
 [|ffffffgf828aebe7>] end_repeat_nmi+0x1e/0x2e
 [<ffffffff828b5d60>] ? call_softirq+0x30/0x30
 [<ffffffff828b5d60>] ? call_softirq+0x30/0x30
 [<fFffffff828b5d60>] ? call_softirq+0x30/0x30

Basically, ia32 uses sysenter to start system calls.

sysexit_from_sys_call=>trace_hardirqs_on_thunk. Before calling,
sysexit_from_sys_call already pops up pt_regs, then trace_hardirqs_on_thunk
would reuse pt_regs space. If perf NMI happens here, perf might use a bad pt_regs.

The patch fixes it by keeping rsp pointing to pt_regs.

Change-Id: I7574260c8e2786601dabd7a7010d5c58fcc2201d

---
 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 474dc1b..1704282 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
 
--
1.7.9.5