Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Fix PR19927: Avoid unwinder recursion if sniffer uses calls parse_and…
…_eval This fixes the problem exercised by Kevin's test at: https://sourceware.org/ml/gdb-patches/2016-08/msg00216.html This was originally exposed by the OpenJDK Python-based unwinder. If an unwinder attempts to call parse_and_eval from within its sniffing method, GDB's unwinding machinery enters infinite recursion. However, parse_and_eval is a pretty reasonable thing to call, because Python/Scheme-based unwinders will often need to read globals out of inferior memory. The recursion happens because: - get_current_frame() is called soon after the target stops. - current_frame is NULL, and so we unwind it from the sentinel frame (which is special and has level == -1). - We reach get_prev_frame_if_no_cycle, which does cycle detection based on frame id, and thus tries to compute the frame id of the new frame. - Frame id computation requires an unwinder, so we go through all unwinder sniffers trying to see if one accepts the new frame (the current frame). - the unwinder's sniffer calls parse_and_eval(). - parse_and_eval depends on the selected frame/block, and if not set yet, the selected frame is set to the current frame. - get_current_frame () is called again. current_frame is still NULL, so ... - recurse forever. In Kevin's test at: https://sourceware.org/ml/gdb-patches/2016-08/msg00216.html gdb doesn't recurse forever simply because the Python unwinder contains code to detect and stop the recursion itself. However, GDB goes downhill from here, e.g., by showing the sentinel frame as current frame (note the -1): Breakpoint 1, ccc (arg=<unavailable>) at py-recurse-unwind.c:23 23 } (gdb) bt #-1 ccc (arg=<unavailable>) at py-recurse-unwind.c:23 Backtrace stopped: previous frame identical to this frame (corrupt stack?) That "-1" frame level comes from this: if (catch_exceptions (current_uiout, unwind_to_current_frame, sentinel_frame, RETURN_MASK_ERROR) != 0) { /* Oops! Fake a current frame? Is this useful? It has a PC of zero, for instance. */ current_frame = sentinel_frame; } which is bogus. It's never correct to set the current frame to the sentinel frame. The only reason this has survived so long is that getting here normally indicates something wrong has already happened before and we fix that. And this case is no exception -- it doesn't really matter how precisely we managed to get to that bogus code (it has to do with the the stash), because anything after recursion happens is going to be invalid. So the fix is to avoid the recursion in the first place. Observations: #1 - The recursion happens because we try to do cycle detection from within get_prev_frame_if_no_cycle. That requires computing the frame id of the frame being unwound, and that itself requires calling into the unwinders. #2 - But, the first time we're unwinding from the sentinel frame, when we reach get_prev_frame_if_no_cycle, there's no frame chain at all yet: - current_frame is NULL. - the frame stash is empty. Thus, there's really no need to do cycle detection the first time we reach get_prev_frame_if_no_cycle, when building the current frame. So we can break the recursion by making get_current_frame call a simplified version of get_prev_frame_if_no_cycle that results in setting the current_frame global _before_ computing the current frame's id. But, we can go a little bit further. As there's really no reason anymore to compute the current frame's frame id immediately, we can defer computing it to when some caller of get_current_frame might need it. This was actually how the frame id was computed for all frames before the stash-based cycle detection was added. So in a way, this patch reintroduces the lazy frame id computation, but unlike before, only for the case of the current frame, which turns out to be special. This lazyness, however, requires adjusting gdb.python/py-unwind-maint.exp, because that assumes unwinders are immediately called as side effect of some commands. I didn't see a need to preserve the behavior expected by that test (all it would take is call get_frame_id inside get_current_frame), so I adjusted the test. gdb/ChangeLog: 2016-09-05 Pedro Alves <[email protected]> PR backtrace/19927 * frame.c (get_frame_id): Compute the frame id if not computed yet. (unwind_to_current_frame): Delete. (get_current_frame): Use get_prev_frame_always_1 to get the current frame and assert that that always succeeds. (get_prev_frame_if_no_cycle): Skip cycle detection if returning the current frame. gdb/testsuite/ChangeLog: 2016-09-05 Pedro Alves <[email protected]> PR backtrace/19927 * gdb.python/py-unwind-maint.exp: Adjust tests to not expect that unwinders are immediately called as side effect of "source" or "disable unwinder" commands. * gdb.python/py-recurse-unwind.exp: Remove setup_kfail calls.
- Loading branch information
