Em Mon, Jul 24, 2023 at 10:28:15PM +0530, Athira Rajeev escreveu:
> perf record with callchain recording fails as below
> in powerpc:
>
> ./perf record -a -gR sleep 10
> ./perf report
> perf: Segmentation fault
>
> gdb trace points to thread__find_map
>
> 0 0x00000000101df314 in atomic_cmpxchg (newval=1818846826,
> oldval=1818846827, v=0x1001a8f3) at
> /home/athira/linux/tools/include/asm-generic/atomic-gcc.h:70
> 1 refcount_sub_and_test (i=1, r=0x1001a8f3) at
> /home/athira/linux/tools/include/linux/refcount.h:135
> 2 refcount_dec_and_test (r=0x1001a8f3) at
> /home/athira/linux/tools/include/linux/refcount.h:148
> 3 map__put (map=0x1001a8b3) at util/map.c:311
> 4 0x000000001016842c in __map__zput (map=0x7fffffffa368) at
> util/map.h:190
> 5 thread__find_map (thread=0x105b92f0, cpumode=<optimized out>,
> addr=13835058055283572736, al=al@entry=0x7fffffffa358) at util/event.c:582
> 6 0x000000001016882c in thread__find_symbol (thread=<optimized out>,
> cpumode=<optimized out>, addr=<optimized out>, al=0x7fffffffa358) at
> util/event.c:656
> 7 0x00000000102e12b4 in arch_skip_callchain_idx (thread=<optimized out>,
> chain=<optimized out>) at arch/powerpc/util/skip-callchain-idx.c:255
> 8 0x00000000101d3bf4 in thread__resolve_callchain_sample
> (thread=0x105b92f0, cursor=0x1053d160, evsel=<optimized out>,
> sample=0x7fffffffa908, parent=0x7fffffffa778, root_al=0x7fffffffa710,
> max_stack=<optimized out>) at util/machine.c:2940
> 9 0x00000000101cd210 in sample__resolve_callchain (sample=<optimized
> out>, cursor=<optimized out>, parent=<optimized out>, evsel=<optimized out>,
> al=<optimized out>, max_stack=<optimized out>)
> at util/callchain.c:1112
> 10 0x000000001022a9d8 in hist_entry_iter__add (iter=0x7fffffffa750,
> al=0x7fffffffa710, max_stack_depth=<optimized out>, arg=0x7fffffffbbd0) at
> util/hist.c:1232
> 11 0x0000000010056d98 in process_sample_event (tool=0x7fffffffbbd0,
> event=0x7ffff6223c38, sample=0x7fffffffa908, evsel=<optimized out>,
> machine=0x10524ef8) at builtin-report.c:332
>
> Here arch_skip_callchain_idx calls thread__find_symbol and which
> invokes thread__find_map with uninitialised "addr_location".
> Snippet:
>
> thread__find_symbol(thread, PERF_RECORD_MISC_USER, ip, &al);
>
> Recent change with commit 0dd5041c9a0ea ("perf addr_location:
> Add init/exit/copy functions"), introduced "maps__zput" in the
> function thread__find_map. This could result in segfault while
> accessing uninitialised map from "struct addr_location". Fix this
> by adding addr_location__init and addr_location__exit in
> arch_skip_callchain_idx.
Thanks, applied.
> Fixes: 0dd5041c9a0ea ("perf addr_location: Add init/exit/copy functions")
> Reported-by: Aneesh Kumar K.V <aneesh.ku...@linux.ibm.com>
> Signed-off-by: Athira Rajeev <atraj...@linux.vnet.ibm.com>
I'll also do a audit of all calls to thread__find_map() and its callers
to check for other such cases :-\
For instance, this one seem buggy as well, Adrian?
diff --git a/tools/perf/util/dlfilter.c b/tools/perf/util/dlfilter.c
index 46f74b2344dbb34c..798a53d7e6c9dfc5 100644
--- a/tools/perf/util/dlfilter.c
+++ b/tools/perf/util/dlfilter.c
@@ -166,6 +166,7 @@ static __s32 dlfilter__resolve_address(void *ctx, __u64
address, struct perf_dlf
if (!thread)
return -1;
+ addr_location__init(&al);
thread__find_symbol_fb(thread, d->sample->cpumode, address, &al);
al_to_d_al(&al, &d_al);
