libperf mmap: In user mmap rdpmc avoid undefined behavior
Ian Rogers
reviewerPATCH 1/1 libperf mmap: In user mmap rdpmc avoid undefined behavior
A shift left of a signed 64-bit s64 may overflow and result in undefined behavior caught by ubsan. Switch to a u64 instead.
--- tools/lib/perf/mmap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/lib/perf/mmap.c +1 -1
@@ -508,7 +508,7 @@ int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count idx =READ_ONCE (pc->index); cnt =READ_ONCE (pc->offset);if (pc->cap_user_rdpmc && idx) { -s64 evcnt =read_perf_counter (idx -1 ); +u64 evcnt =read_perf_counter (idx -1 );u16 width =READ_ONCE (pc->pmc_width); evcnt <<=64 - width;
Signature · 1 line
2.51.0.384.g4c02a37b29-goog
Quoted 2 lines · Ian Rogers
A shift left of a signed 64-bit s64 may overflow and result in undefined behavior caught by ubsan. Switch to a u64 instead.
Thanks, applied to perf-tools-next,
- Arnaldo
Quoted 21 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/mmap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)tools/lib/perf/mmap.c +1
-1@@ -508,7 +508,7 @@ int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count idx =READ_ONCE (pc->index); cnt =READ_ONCE (pc->offset);if (pc->cap_user_rdpmc && idx) { -s64 evcnt =read_perf_counter (idx -1 ); +u64 evcnt =read_perf_counter (idx -1 );u16 width =READ_ONCE (pc->pmc_width); evcnt <<=64 - width;Signature · 1 line
2.51.0.384.g4c02a37b29-goog
Quoted 2 lines · Ian Rogers
A shift left of a signed 64-bit s64 may overflow and result in undefined behavior caught by ubsan. Switch to a u64 instead.
This change was just brought to my attention...
Quoted 11 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/lib/perf/mmap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)tools/lib/perf/mmap.c +0
-0
tools/lib/perf/mmap.c int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count · line 511
Quoted 5 lines
tools/lib/perf/mmap.c +1
-1@@ -508,7 +508,7 @@ int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count idx =READ_ONCE (pc->index); cnt =READ_ONCE (pc->offset);if (pc->cap_user_rdpmc && idx) { -s64 evcnt =read_perf_counter (idx -1 ); +u64 evcnt =read_perf_counter (idx -1 );
According to uapi/linux/perf_event.h, this needs to be signed to sign extend the counter value:
/*
* If cap_user_rdpmc this field provides the bit-width of the value
* read using the rdpmc() or equivalent instruction. This can be used
* to sign extend the result like:
*
* pmc <<= 64 - width;
* pmc >>= 64 - width; // signed shift right
* count += pmc;
*/
It looks like a left/right shift has gone from implementation defined to undefined behavior. According to the internet, the right way would be something like this:
u64 sign_bit = 1ULL << (width - 1);
s64 val64 = (s64)(((raw & ((1ULL << width) - 1)) ^ sign_bit) - sign_bit);
tools/lib/perf/mmap.c int perf_mmap__read_self(struct perf_mmap *map, struct perf_counts_values *count · line 514
Quoted 3 lines
tools/lib/perf/mmap.c +0
-0u16 width =READ_ONCE (pc->pmc_width); evcnt <<=64 - width;
The left and then right shift here is probably redundant. I suppose there could be garbage in the upper bits to zero out, but I think read_perf_counter() should ensure there isn't.
Did the need for signed values go away or was wrong to begin with?
At a minimum, the header comments which documents the proper sequence no longer matches the implementation.
Rob