Hello,
On Thu, Nov 30, 2017 at 04:37:12PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Thu, Nov 30, 2017 at 09:20:26AM +0100, Peter Zijlstra escreveu:
> > On Thu, Nov 30, 2017 at 10:32:19AM +0800, Fengguang Wu wrote:
> > > Hello,
> > >
> > > FYI this happens in mainline kernel 4.15.0-rc1.
> > > It looks like a new regression and hard to bisect.
> > >
> > > It occurs in 1 out of 57 boots.
> > >
> > > [ 10.009610] chown (367) used greatest stack depth: 26944 bytes left
> > > Kernel tests: Boot OK!
> > > [ 30.357729] trinity-main uses obsolete (PF_INET,SOCK_PACKET)
> > > [ 31.301433] sock: process `trinity-main' is using obsolete setsockopt
> > > SO_BSDCOMPAT
> > > [ 31.310289]
> > > ==================================================================
> > > [ 31.311490] BUG: KASAN: slab-out-of-bounds in
> > > perf_callchain_user+0x494/0x530:
> > > perf_callchain_store at
> > > include/linux/perf_event.h:1128
> > > (inlined by)
> > > perf_callchain_user at arch/x86/events/core.c:2485
> >
> > I don't think we recently changed anything here...
> >
> > But I do have vague memories of something being off here; I never quite
> > could penetrate the max_stack / contexts_maxed stuff, and istr acme was
> > going to have a peek.
>
> Sure, but I saw some backward ring buffer stuff in there as well, no?
> IIRC that came after the max-stack code, Adding Wang to the CC list.
I think it's because of per-event max-stack not being checked for the
first event. Please see the patch below..
Also I'm not sure that the allocation failure check would work
correctly since it decrements nr_callchain_events when it fails.
Thanks,
Namhyung
>From c12126c4ff9835f0899619db3ee7b4a3151ff2bb Mon Sep 17 00:00:00 2001
From: Namhyung Kim <namhy...@kernel.org>
Date: Tue, 5 Dec 2017 16:54:50 +0900
Subject: [PATCH] perf/core: Fix overflow on perf_callchain_entry
The commit 97c79a38cd45 add a check whether per-event max stack is
greater than the global max. But it missed to do it for the first
event. So if the event had a stack depth greater than the global max,
it could overflow the callchain entry list.
Reported-by: Fengguang Wu <fengguang...@intel.com>
Fixes: 97c79a38cd45 ("perf core: Per event callchain limit")
Signed-off-by: Namhyung Kim <namhy...@kernel.org>
---
kernel/events/callchain.c | 21 ++++++++++++---------
1 file changed, 12 insertions(+), 9 deletions(-)
diff --git a/kernel/events/callchain.c b/kernel/events/callchain.c
index 1b2be63c8528..e449e23802eb 100644
--- a/kernel/events/callchain.c
+++ b/kernel/events/callchain.c
@@ -119,19 +119,22 @@ int get_callchain_buffers(int event_max_stack)
goto exit;
}
+ /*
+ * If requesting per event more than the global cap,
+ * return a different error to help userspace figure this out.
+ *
+ * And also do it here so that we have &callchain_mutex held.
+ */
+ if (event_max_stack > sysctl_perf_event_max_stack) {
+ err = -EOVERFLOW;
+ goto exit;
+ }
+
if (count > 1) {
/* If the allocation failed, give up */
if (!callchain_cpus_entries)
err = -ENOMEM;
- /*
- * If requesting per event more than the global cap,
- * return a different error to help userspace figure
- * this out.
- *
- * And also do it here so that we have &callchain_mutex held.
- */
- if (event_max_stack > sysctl_perf_event_max_stack)
- err = -EOVERFLOW;
+
goto exit;
}
--
2.15.0
