From patchwork Tue Aug 23 18:45:49 2016 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Pedro Alves X-Patchwork-Id: 14888 Received: (qmail 25374 invoked by alias); 23 Aug 2016 18:46:04 -0000 Mailing-List: contact gdb-patches-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: gdb-patches-owner@sourceware.org Delivered-To: mailing list gdb-patches@sourceware.org Received: (qmail 25364 invoked by uid 89); 23 Aug 2016 18:46:02 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-2.1 required=5.0 tests=BAYES_00, RP_MATCHES_RCVD, SPF_HELO_PASS autolearn=ham version=3.3.2 spammy=ccc, onto, sentinel, recurse X-HELO: mx1.redhat.com Received: from mx1.redhat.com (HELO mx1.redhat.com) (209.132.183.28) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Tue, 23 Aug 2016 18:45:52 +0000 Received: from int-mx11.intmail.prod.int.phx2.redhat.com (int-mx11.intmail.prod.int.phx2.redhat.com [10.5.11.24]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mx1.redhat.com (Postfix) with ESMTPS id 6EE2D3F724 for ; Tue, 23 Aug 2016 18:45:51 +0000 (UTC) Received: from cascais.lan (ovpn01.gateway.prod.ext.ams2.redhat.com [10.39.146.11]) by int-mx11.intmail.prod.int.phx2.redhat.com (8.14.4/8.14.4) with ESMTP id u7NIjoEh005151 for ; Tue, 23 Aug 2016 14:45:50 -0400 From: Pedro Alves To: gdb-patches@sourceware.org Subject: [PATCH] Fix PR19927: Avoid unwinder recursion if sniffer uses calls parse_and_eval Date: Tue, 23 Aug 2016 19:45:49 +0100 Message-Id: <1471977949-10539-1-git-send-email-palves@redhat.com> 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. Observations: - The first time we're reaching get_prev_frame_if_no_cycle, there's no frame chain at all yet: #1 - current_frame is NULL. #2 - the frame stash is empty. 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=) at py-recurse-unwind.c:23 23 } (gdb) bt #-1 ccc (arg=) 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. But 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 the current frame's id 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: 2016-08-23 Pedro Alves 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: 2016-08-23 Pedro Alves 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/frame.c | 79 ++++++++++++++++++---------- gdb/testsuite/gdb.python/py-unwind-maint.exp | 8 +-- 2 files changed, 57 insertions(+), 30 deletions(-) diff --git a/gdb/frame.c b/gdb/frame.c index c25ce4c..d3f3056 100644 --- a/gdb/frame.c +++ b/gdb/frame.c @@ -511,7 +511,26 @@ get_frame_id (struct frame_info *fi) if (fi == NULL) return null_frame_id; - gdb_assert (fi->this_id.p); + if (!fi->this_id.p) + { + int stashed; + + /* If we haven't computed the frame id yet, then it must be that + this is the current frame. Compute it now, and stash the + result. The IDs of other frames are computed as soon as + they're created, in order to detect cycles. See + get_prev_frame_if_no_cycle. */ + gdb_assert (fi->level == 0); + + /* Compute. */ + compute_frame_id (fi); + + /* Since this is the first frame in the chain, this should + always succeed. */ + stashed = frame_stash_add (fi); + gdb_assert (stashed); + } + return fi->this_id.value; } @@ -1487,23 +1506,7 @@ frame_obstack_zalloc (unsigned long size) return data; } -/* Return the innermost (currently executing) stack frame. This is - split into two functions. The function unwind_to_current_frame() - is wrapped in catch exceptions so that, even when the unwind of the - sentinel frame fails, the function still returns a stack frame. */ - -static int -unwind_to_current_frame (struct ui_out *ui_out, void *args) -{ - struct frame_info *frame = get_prev_frame ((struct frame_info *) args); - - /* A sentinel frame can fail to unwind, e.g., because its PC value - lands in somewhere like start. */ - if (frame == NULL) - return 1; - current_frame = frame; - return 0; -} +static struct frame_info *get_prev_frame_always_1 (struct frame_info *this_frame); struct frame_info * get_current_frame (void) @@ -1525,16 +1528,28 @@ get_current_frame (void) if (current_frame == NULL) { + int stashed; struct frame_info *sentinel_frame = create_sentinel_frame (current_program_space, get_current_regcache ()); - 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; - } + + /* Set the current frame before computing the frame id, to avoid + recursion inside compute_frame_id, in case the frame's + unwinder decides to do a symbol lookup (which depends on the + selected frame's block). + + This call must always succeed. In particular, nothing inside + get_prev_frame_always_1 should try to unwind from the + sentinel frame, because that could fail/throw, and we always + want to leave with the current frame created and linked in -- + we should never end up with the sentinel frame as outermost + frame. */ + current_frame = get_prev_frame_always_1 (sentinel_frame); + gdb_assert (current_frame != NULL); + + /* No need to compute the frame id yet. That'll be done lazily + from inside get_frame_id instead. */ } + return current_frame; } @@ -1812,8 +1827,18 @@ get_prev_frame_if_no_cycle (struct frame_info *this_frame) struct cleanup *prev_frame_cleanup; prev_frame = get_prev_frame_raw (this_frame); - if (prev_frame == NULL) - return NULL; + + /* Don't compute the frame id of the current frame yet. Unwinding + the sentinel frame can fail (e.g., if the thread is gone and we + can't thus read its registers). If we let the cycle detection + code below try to compute a frame ID, then an error thrown from + within the frame ID computation would result in the sentinel + frame as outermost frame, which is bogus. Instead, we'll compute + the current frame's ID lazily in get_frame_id. Note that there's + no point in doing cycle detection when there's only one frame, so + nothing is lost here. */ + if (prev_frame->level == 0) + return prev_frame; /* The cleanup will remove the previous frame that get_prev_frame_raw linked onto THIS_FRAME. */ diff --git a/gdb/testsuite/gdb.python/py-unwind-maint.exp b/gdb/testsuite/gdb.python/py-unwind-maint.exp index 3a98cb1..1253057 100644 --- a/gdb/testsuite/gdb.python/py-unwind-maint.exp +++ b/gdb/testsuite/gdb.python/py-unwind-maint.exp @@ -34,10 +34,13 @@ if ![runto_main ] then { return -1 } -gdb_test_sequence "source ${pyfile}" "import python scripts" { - "Python script imported" +gdb_test "source ${pyfile}" "Python script imported" \ + "import python scripts" + +gdb_test_sequence "frame" "All unwinders enabled" { "py_unwind_maint_ps_unwinder called" "global_unwinder called" + "#0 main" } gdb_test_sequence "info unwinder" "Show all unwinders" { @@ -56,7 +59,6 @@ gdb_test_sequence "continue" "Unwinders called" { gdb_test_sequence "disable unwinder global .*" "Unwinder disabled" { "1 unwinder disabled" - "py_unwind_maint_ps_unwinder called" } gdb_test_sequence "info unwinder" "Show with global unwinder disabled" {