Re: perf: fuzzer causes crash in new XMM code
From: Liang, Kan
Date: Thu May 23 2019 - 08:51:23 EST
On 5/22/2019 6:50 PM, Jiri Olsa wrote:
On Wed, May 22, 2019 at 05:54:58PM -0400, Liang, Kan wrote:
SNIP
[ 9680.197779] RSP: 002b:00007fff595603a8 EFLAGS: 00000206 ORIG_RAX: ffffffffffffff13
[ 9680.205489] RAX: 0000000000004985 RBX: 000000000000000c RCX: 00000000000365ca
[ 9680.212748] RDX: 00001e15d36cec84 RSI: 0000000000000000 RDI: 0000000000000001
[ 9680.220059] RBP: 00007fff595603c0 R08: 0000000000000000 R09: 00007fccbb62e540
[ 9680.227362] R10: fffffffffffffd4e R11: 0000000000000246 R12: 000055dad779a4c0
[ 9680.234630] R13: 00007fff595627b0 R14: 0000000000000000 R15: 0000000000000000
[ 9680.310017] ---[ end trace 511b9368cf14c65a ]---
Hi Vince,
Thanks for the test.
XMM registers can only collected by hardware PEBS events. We should disable
it for all software/probe events.
I think you should also include HW non-PEBS events
in those checks below
The HW non-PEBS events have been checked in x86_pmu_hw_config().
Only the software/probe events are missed.
Thanks,
Kan
jirka
Could you please try the patch as below?
Thanks,
Kan
From 0136d8374c2db65b125c8d92b661c96e8d21adb0 Mon Sep 17 00:00:00 2001
From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
Date: Wed, 22 May 2019 12:14:16 -0700
Subject: [PATCH] perf/x86: Disable non generic regs for software/probe
events
The perf fuzzer caused skylake machine to crash.
[ 9680.085831] Call Trace:
[ 9680.088301] <IRQ>
[ 9680.090363] perf_output_sample_regs+0x43/0xa0
[ 9680.094928] perf_output_sample+0x3aa/0x7a0
[ 9680.099181] perf_event_output_forward+0x53/0x80
[ 9680.103917] __perf_event_overflow+0x52/0xf0
[ 9680.108266] ? perf_trace_run_bpf_submit+0xc0/0xc0
[ 9680.113108] perf_swevent_hrtimer+0xe2/0x150
[ 9680.117475] ? check_preempt_wakeup+0x181/0x230
[ 9680.122091] ? check_preempt_curr+0x62/0x90
[ 9680.126361] ? ttwu_do_wakeup+0x19/0x140
[ 9680.130355] ? try_to_wake_up+0x54/0x460
[ 9680.134366] ? reweight_entity+0x15b/0x1a0
[ 9680.138559] ? __queue_work+0x103/0x3f0
[ 9680.142472] ? update_dl_rq_load_avg+0x1cd/0x270
[ 9680.147194] ? timerqueue_del+0x1e/0x40
[ 9680.151092] ? __remove_hrtimer+0x35/0x70
[ 9680.155191] __hrtimer_run_queues+0x100/0x280
[ 9680.159658] hrtimer_interrupt+0x100/0x220
[ 9680.163835] smp_apic_timer_interrupt+0x6a/0x140
[ 9680.168555] apic_timer_interrupt+0xf/0x20
[ 9680.172756] </IRQ>
The XMM registers can only be collected by hardware PEBS events, not
software/probe events.
Add has_non_generic_regs() to check if non-generic regs, e.g. XMM on
X86, are applied for software/probe events. If yes, return -EOPNOTSUPP.
Add __weak function non_generic_regs_mask() to return the mask of
non-generic regs. For X86, the mask of non-generic regs equals to the
mask of XMM registers.
Fixes: 878068ea270e ("perf/x86: Support outputting XMM registers")
Reported-by: Vince Weaver <vincent.weaver@xxxxxxxxx>
Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
---
arch/x86/kernel/perf_regs.c | 5 +++++
include/linux/perf_regs.h | 2 ++
kernel/events/core.c | 37 +++++++++++++++++++++++++++++++++++++
3 files changed, 44 insertions(+)
diff --git a/arch/x86/kernel/perf_regs.c b/arch/x86/kernel/perf_regs.c
index 07c30ee..86ffe5a 100644
--- a/arch/x86/kernel/perf_regs.c
+++ b/arch/x86/kernel/perf_regs.c
@@ -57,6 +57,11 @@ static unsigned int pt_regs_offset[PERF_REG_X86_MAX] = {
#endif
};
+u64 non_generic_regs_mask(void)
+{
+ return (~((1ULL << PERF_REG_X86_XMM0) - 1));
+}
+
u64 perf_reg_value(struct pt_regs *regs, int idx)
{
struct x86_perf_regs *perf_regs;
diff --git a/include/linux/perf_regs.h b/include/linux/perf_regs.h
index 4767474..c1c3454 100644
--- a/include/linux/perf_regs.h
+++ b/include/linux/perf_regs.h
@@ -9,6 +9,8 @@ struct perf_regs {
struct pt_regs *regs;
};
+u64 non_generic_regs_mask(void);
+
#ifdef CONFIG_HAVE_PERF_REGS
#include <asm/perf_regs.h>
u64 perf_reg_value(struct pt_regs *regs, int idx);
diff --git a/kernel/events/core.c b/kernel/events/core.c
index abbd4b3..14da1d9 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -8457,6 +8457,19 @@ static void sw_perf_event_destroy(struct perf_event
*event)
swevent_hlist_put();
}
+u64 __weak non_generic_regs_mask(void)
+{
+ return 0;
+}
+
+static inline bool has_non_generic_regs(struct perf_event *event)
+{
+ u64 mask = non_generic_regs_mask();
+
+ return ((event->attr.sample_regs_user & mask) ||
+ (event->attr.sample_regs_intr & mask));
+}
+
static int perf_swevent_init(struct perf_event *event)
{
u64 event_id = event->attr.config;
@@ -8470,6 +8483,10 @@ static int perf_swevent_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
switch (event_id) {
case PERF_COUNT_SW_CPU_CLOCK:
case PERF_COUNT_SW_TASK_CLOCK:
@@ -8633,6 +8650,10 @@ static int perf_tp_event_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
err = perf_trace_init(event);
if (err)
return err;
@@ -8722,6 +8743,10 @@ static int perf_kprobe_event_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE;
err = perf_kprobe_init(event, is_retprobe);
if (err)
@@ -8782,6 +8807,10 @@ static int perf_uprobe_event_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
is_retprobe = event->attr.config & PERF_PROBE_CONFIG_IS_RETPROBE;
ref_ctr_offset = event->attr.config >> PERF_UPROBE_REF_CTR_OFFSET_SHIFT;
err = perf_uprobe_init(event, ref_ctr_offset, is_retprobe);
@@ -9562,6 +9591,10 @@ static int cpu_clock_event_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
perf_swevent_init_hrtimer(event);
return 0;
@@ -9643,6 +9676,10 @@ static int task_clock_event_init(struct perf_event
*event)
if (has_branch_stack(event))
return -EOPNOTSUPP;
+ /* only support generic regs */
+ if (has_non_generic_regs(event))
+ return -EOPNOTSUPP;
+
perf_swevent_init_hrtimer(event);
return 0;
--
2.7.4