perf trace: Fix race condition at the end of started workloads

I get following crash on multiple systems and across several releases
(at least since v3.18).

	Core was generated by `/tmp/perf trace sleep 0.2 '.
	Program terminated with signal SIGSEGV, Segmentation fault.
	#0  perf_mmap__read_head (mm=0x3fff9bf30070) at util/evlist.h:195
	195		u64 head = ACCESS_ONCE(pc->data_head);
	(gdb) bt
	#0  perf_mmap__read_head (mm=0x3fff9bf30070) at util/evlist.h:195
	#1  perf_evlist__mmap_read (evlist=0x10027f11910, idx=<optimized out>)
	    at util/evlist.c:637
	#2  0x000000001003ce4c in trace__run (argv=<optimized out>,
	    argc=<optimized out>, trace=0x3fffd7b28288) at builtin-trace.c:2259
	#3  cmd_trace (argc=<optimized out>, argv=<optimized out>,
	    prefix=<optimized out>) at builtin-trace.c:2799
	#4  0x00000000100657b8 in run_builtin (p=0x10176798 <commands+480>, argc=3,
	    argv=0x3fffd7b2b550) at perf.c:370
	#5  0x00000000100063e8 in handle_internal_command (argv=0x3fffd7b2b550, argc=3)
	    at perf.c:429
	#6  run_argv (argv=0x3fffd7b2af70, argcp=0x3fffd7b2af7c) at perf.c:473
	#7  main (argc=3, argv=0x3fffd7b2b550) at perf.c:588

The problem seems to be a race condition, when the application has just
exited.  Some/all fds associated with the perf-events (tracepoints) go
into a POLLHUP/ POLLERR state and the mmap region associated with those
events are unmapped (in perf_evlist__filter_pollfd()).

But we go back and do a perf_evlist__mmap_read() which assumes that the
mmaps are still valid and we hit the crash.

If the mapping for an event is released, its refcnt is 0 (and ->base
is NULL), so ensure we have non-zero refcount before accessing the map.

Note that perf-record has a similar logic but unlike perf-trace, the
record__mmap_read_all() checks the evlist->mmap[i].base before accessing
the map.

Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Li Zhang <zhlcindy@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/20150612060003.GA19913@us.ibm.com
[ Fixed it up to use atomic_read() ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index dc1dc2c181ef22c2a8f33df66a3b0ed2bcbf9f2d..6b58a47a79ecf6789c41687607e73c0317e757d6 100644 (file)
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -634,11 +634,18 @@ static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 {
        struct perf_mmap *md = &evlist->mmap[idx];
-       u64 head = perf_mmap__read_head(md);
+       u64 head;
        u64 old = md->prev;
        unsigned char *data = md->base + page_size;
        union perf_event *event = NULL;
 
+       /*
+        * Check if event was unmapped due to a POLLHUP/POLLERR.
+        */
+       if (!atomic_read(&md->refcnt))
+               return NULL;
+
+       head = perf_mmap__read_head(md);
        if (evlist->overwrite) {
                /*
                 * If we're further behind than half the buffer, there's a chance