This patch ensures that the frame id for the current frame is stashed
before that of the previous frame (to the current frame).
First, it should be noted that the frame id for the current frame is
not stashed by get_current_frame(). The current frame's frame id is
lazily computed and stashed via calls to get_frame_id(). However,
it's possible for get_prev_frame() to be called without first stashing
the current frame.
The frame stash is used not only to speed up frame lookups, but
also to detect cycles. When attempting to compute the frame id
for a "previous" frame (in get_prev_frame_if_no_cycle), a cycle
is detected if the computed frame id is already in the stash.
If it should happen that a previous frame id is stashed which should
represent a cycle for the current frame, then an assertion failure
will trigger should get_frame_id() be later called to determine
the frame id for the current frame.
As of late 2016, with the "Tweak meaning of VALUE_FRAME_ID" patch in
place, this actually occurs when running the
gdb.dwarf2/dw2-dup-frame.exp test. While attempting to generate a
backtrace, the python frame filter code is invoked, leading to
frame_info_to_frame_object() (in python/py-frame.c) being called.
That function will potentially call get_prev_frame() before
get_frame_id() is called. The call to get_prev_frame() can eventually
end up in get_prev_frame_if_no_cycle() which, in turn, calls
compute_frame_id(), after which the frame id is stashed for the
previous frame.
If the frame id for the current frame is stashed, the cycle detection
code (which relies on the frame stash) in get_prev_frame_if_no_cycle()
will be triggered for a cycle starting with the current frame. If the
current frame's id is not stashed, the cycle detecting code can't
operate as designed. Instead, when get_frame_id() is called on the
current frame at some later point, the current frame's id will found
to be already in the stash, triggering an assertion failure.
Below is an in depth examination of the failure which lead to this change.
I've shortened pathnames for brevity and readability.
Here's the portion of the log file showing the failure/internal error:
(gdb) break stop_frame
Breakpoint 1 at 0x40059a: file dw2-dup-frame.c, line 22.
(gdb) run
Starting program: testsuite/outputs/gdb.dwarf2/dw2-dup-frame/dw2-dup-frame
Breakpoint 1, stop_frame () at dw2-dup-frame.c:22
22 }
(gdb) bt
gdb/frame.c:544: internal-error: frame_id get_frame_id(frame_info*): Assertion `stashed' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
Quit this debugging session? (y or n)
FAIL: gdb.dwarf2/dw2-dup-frame.exp: backtrace from stop_frame (GDB internal error)
Here's a partial backtrace from the internal error, showing the frames
which I think are relevant, plus several extra to provide context:
#0 internal_error (
file=0x932b98 "gdb/frame.c", line=544,
fmt=0x932b20 "%s: Assertion `%s' failed.")
at gdb/common/errors.c:54
#1 0x000000000072207e in get_frame_id (fi=0xe5a760)
at gdb/frame.c:544
#2 0x00000000004eb50d in frame_info_to_frame_object (frame=0xe5a760)
at gdb/python/py-frame.c:390
#3 0x00000000004ef5be in bootstrap_python_frame_filters (frame=0xe5a760,
frame_low=0, frame_high=-1)
at gdb/python/py-framefilter.c:1453
#4 0x00000000004ef7a9 in gdbpy_apply_frame_filter (
extlang=0x8857e0 <extension_language_python>, frame=0xe5a760, flags=7,
args_type=CLI_SCALAR_VALUES, out=0xf6def0, frame_low=0, frame_high=-1)
at gdb/python/py-framefilter.c:1548
#5 0x00000000005f2c5a in apply_ext_lang_frame_filter (frame=0xe5a760,
flags=7, args_type=CLI_SCALAR_VALUES, out=0xf6def0, frame_low=0,
frame_high=-1)
at gdb/extension.c:572
#6 0x00000000005ea896 in backtrace_command_1 (count_exp=0x0, show_locals=0,
no_filters=0, from_tty=1)
at gdb/stack.c:1834
Examination of the code in frame_info_to_frame_object(), which is in
python/py-frame.c, is key to understanding this problem:
if (get_prev_frame (frame) == NULL
&& get_frame_unwind_stop_reason (frame) != UNWIND_NO_REASON
&& get_next_frame (frame) != NULL)
{
frame_obj->frame_id = get_frame_id (get_next_frame (frame));
frame_obj->frame_id_is_next = 1;
}
else
{
frame_obj->frame_id = get_frame_id (frame);
frame_obj->frame_id_is_next = 0;
}
I will first note that the frame id for frame has not been computed yet. (This
was verified by placing a breakpoint on compute_frame_id().)
The call to get_prev_frame() causes the the frame id to (eventually) be
computed for the previous frame. Here's a backtrace showing how we
get there:
#0 compute_frame_id (fi=0x10e2810)
at gdb/frame.c:496
#1 0x0000000000724a67 in get_prev_frame_if_no_cycle (this_frame=0xe5a760)
at gdb/frame.c:1871
#2 0x0000000000725136 in get_prev_frame_always_1 (this_frame=0xe5a760)
at gdb/frame.c:2045
#3 0x000000000072516b in get_prev_frame_always (this_frame=0xe5a760)
at gdb/frame.c:2061
#4 0x000000000072570f in get_prev_frame (this_frame=0xe5a760)
at gdb/frame.c:2303
#5 0x00000000004eb471 in frame_info_to_frame_object (frame=0xe5a760)
at gdb/python/py-frame.c:381
For this particular case, we end up in the else clause of the code above
which calls get_frame_id (frame). It's at this point that the frame id
for frame is computed. Again, here's a backtrace:
#0 compute_frame_id (fi=0xe5a760)
at gdb/frame.c:496
#1 0x000000000072203d in get_frame_id (fi=0xe5a760)
at gdb/frame.c:539
#2 0x00000000004eb50d in frame_info_to_frame_object (frame=0xe5a760)
at gdb/python/py-frame.c:390
The test in question, dw2-dup-frame.exp, deliberately creates a broken
(cyclic) stack. So, in this instance, the frame id for the prev
`frame' will be the same as that for `frame'. But that particular
frame id ended up in the stash during the previous frame operation.
When, just a few lines later, we compute the frame id for `frame', the
id in question is already in the stash, thus triggering the assertion
failure.
I considered two other solutions to solving this problem:
We could prevent get_prev_frame() from being called before
get_frame_id() in frame_info_to_frame_object(). (See above for the
snippet of code where this happens.) A call to get_frame_id (frame)
could be placed ahead of that code snippet above. I have tested this
approach and, while it does work, I can't be certain that
get_prev_frame() isn't called ahead of stashing the current frame
somewhere else in GDB, but in a less obvious way.
Another approach is to stash the current frame's id by calling
get_frame_id() in get_current_frame(). This approach is conceptually
simpler, but when importing a python unwinder, has the unwelcome side
effect of causing the unwinder to be called during import.
A cleaner looking fix would be to place this code after code
corresponding to the "Don't compute the frame id of the current frame
yet..." comment in get_prev_frame_if_no_cycle(). Sadly, this does not
work though; by the time we get to this point, the frame state for the
prev frame has been modified just enough to cause an internal error to
occur when attempting to compute the (current) frame id for inline
frames. (The unexpected failure count increases by roughly 130
failures.) Therefore, I decided to place it as early as possible
in get_prev_frame().
gdb/ChangeLog:
* frame.c (get_prev_frame): Stash frame id for current frame
prior to computing frame id for previous frame.
The VALUE_FRAME_ID macro provides access to a member in struct value
that's used to hold the frame id that's used when determining a
register's value or when assigning to a register. The underlying
member has a long and obscure name. I won't refer to it here, but
will simply refer to VALUE_FRAME_ID as if it's the struct value member
instead of being a convenient macro.
At the moment, without this patch in place, VALUE_FRAME_ID is set in
value_of_register_lazy() and several other locations to hold the frame
id of the frame passed to those functions.
VALUE_FRAME_ID is used in the lval_register case of
value_fetch_lazy(). To fetch the register's value, it calls
get_frame_register_value() which, in turn, calls
frame_unwind_register_value() with frame->next.
A python based unwinder may wish to determine the value of a register
or evaluate an expression containing a register. When it does this,
value_fetch_lazy() will be called under some circumstances. It will
attempt to determine the frame id associated with the frame passed to
it. In so doing, it will end up back in the frame sniffer of the very
same python unwinder that's attempting to learn the value of a
register as part of the sniffing operation. This recursion is not
desirable.
As noted above, when value_fetch_lazy() wants to fetch a register's
value, it does so (indirectly) by unwinding from frame->next.
With this in mind, a solution suggests itself: Change VALUE_FRAME_ID
to hold the frame id associated with the next frame. Then, when it
comes time to obtain the value associated with the register, we can
simply unwind from the frame corresponding to the frame id stored in
VALUE_FRAME_ID. This neatly avoids the python unwinder recursion
problem by changing when the "next" operation occurs. Instead of the
"next" operation occuring when the register value is fetched, it
occurs earlier on when assigning a frame id to VALUE_FRAME_ID.
(Thanks to Pedro for this suggestion.)
This patch implements this idea.
It builds on the patch "Distinguish sentinel frame from null frame".
Without that work in place, it's necessary to check for null_id at
several places and then obtain the sentinel frame.
It also renames most occurences of VALUE_FRAME_ID to
VALUE_NEXT_FRAME_ID to reflect the new meaning of this field.
There are several uses of VALUE_FRAME_ID which were not changed. In
each case, the original meaning of VALUE_FRAME_ID is required to get
correct results. In all but one of these uses, either
put_frame_register_bytes() or get_frame_register_bytes() is being
called with the frame value obtained from VALUE_FRAME_ID. Both of
these functions perform some unwinding by performing a "->next"
operation on the frame passed to it. If we were to use the new
VALUE_NEXT_FRAME_ID macro, this would effectively do two "->next"
operations, which is not what we want.
The VALUE_FRAME_ID macro has been redefined in terms of
VALUE_NEXT_FRAME_ID. It simply fetches the previous frame's id,
providing this id as the value of the macro.
gdb/ChangeLog:
* value.h (VALUE_FRAME_ID): Rename to VALUE_NEXT_FRAME_ID. Update
comment. Create new VALUE_FRAME_ID which is defined in terms of
VALUE_NEXT_FRAME_ID.
(deprecated_value_frame_id_hack): Rename to
deprecated_value_next_frame_id_hack.
* dwarf2loc.c, findvar.c, frame-unwind.c, sentinel-frame.c,
valarith.c, valops.c, value.c: Adjust nearly all occurences of
VALUE_FRAME_ID to VALUE_NEXT_FRAME_ID. Add comments for those
which did not change.
* value.c (struct value): Rename frame_id field to next_frame_id.
Update comment.
(deprecated_value_frame_id_hack): Rename to
deprecated_value_next_frame_id_hack.
(value_fetch_lazy): Call frame_unwind_register_value()
instead of get_frame_register_value().
* frame.c (get_prev_frame_id_by_id): New function.
* frame.h (get_prev_frame_id_by_id): Declare.
* dwarf2loc.c (dwarf2_evaluate_loc_desc_full): Make
VALUE_NEXT_FRAME_ID refer to the next frame.
* findvar.c (value_of_register_lazy): Likewise.
(default_value_from_register): Likewise.
(value_from_register): Likewise.
* frame_unwind.c (frame_unwind_got_optimized): Likewise.
* sentinel-frame.c (sentinel_frame_prev_register): Likewise.
* value.h (VALUE_FRAME_ID): Update comment describing this macro.
This patch replaces the `current_frame' static global in frame.c with
`sentinel_frame'. It also makes the sentinel frame id unique and
different from the null frame.
By itself, there is not much point to this patch, but it makes
the code cleaner for the VALUE_FRAME_ID changes in another patch.
Since we now allow "navigation" to the sentinel frame, it removes
the necessity of adding special cases to other parts of GDB.
Note that a new function, get_next_frame_sentinel_okay, is introduced
in this patch. It will be used by the VALUE_FRAME_ID changes that
I've made.
Thanks to Pedro Alves for this suggestion.
gdb/ChangeLog:
* frame.h (enum frame_id_stack_status): Add FID_STACK_SENTINEL.
(struct frame_id): Increase number of bits required for storing
stack status to 3 from 2.
(sentinel_frame_id): New declaration.
(get_next_frame_sentinel_okay): Declare.
(frame_find_by_id_sentinel_okay): Declare.
* frame.c (current_frame): Rename this static global to...
(sentinel_frame): ...this static global, which has also been
moved an earlier location in the file.
(fprint_frame_id): Add case for sentinel frame id.
(get_frame_id): Return early for sentinel frame.
(sentinel_frame_id): Define.
(frame_find_by_id): Add case for sentinel_frame_id.
(create_sentinel_frame): Use sentinel_frame_id for this_id.value
instead of null_frame_id.
(get_current_frame): Add local declaration for `current_frame'.
Remove local declaration for `sentinel_frame.'
(get_next_frame_sentinel_okay): New function.
(reinit_frame_cache): Use `sentinel_frame' in place of
`current_frame'.
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 <palves@redhat.com>
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 <palves@redhat.com>
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.
Nowadays, GDB can't insert breakpoint on the return address of the
exception handler on ARM M-profile, because the address is a magic
one 0xfffffff9,
(gdb) bt
#0 CT32B1_IRQHandler () at ../src/timer.c:67
#1 <signal handler called>
#2 main () at ../src/timer.c:127
(gdb) info frame
Stack level 0, frame at 0x200ffa8:
pc = 0x4ec in CT32B1_IRQHandler (../src/timer.c:67); saved pc = 0xfffffff9
called by frame at 0x200ffc8
source language c.
Arglist at 0x200ffa0, args:
Locals at 0x200ffa0, Previous frame's sp is 0x200ffa8
Saved registers:
r7 at 0x200ffa0, lr at 0x200ffa4
(gdb) x/x 0xfffffff9
0xfffffff9: Cannot access memory at address 0xfffffff9
(gdb) finish
Run till exit from #0 CT32B1_IRQHandler () at ../src/timer.c:67
Ed:15: Target error from Set break/watch: Et:96: Pseudo-address (0xFFFFFFxx) for EXC_RETURN is invalid (GDB error?)
Warning:
Cannot insert hardware breakpoint 0.
Could not insert hardware breakpoints:
You may have requested too many hardware breakpoints/watchpoints.
Command aborted.
even some debug probe can't set hardware breakpoint on the magic
address too,
(gdb) hbreak *0xfffffff9
Hardware assisted breakpoint 2 at 0xfffffff9
(gdb) c
Continuing.
Ed:15: Target error from Set break/watch: Et:96: Pseudo-address (0xFFFFFFxx) for EXC_RETURN is invalid (GDB error?)
Warning:
Cannot insert hardware breakpoint 2.
Could not insert hardware breakpoints:
You may have requested too many hardware breakpoints/watchpoints.
Command aborted.
The problem described above is quite similar to PR 8841, in which GDB
can't set breakpoint on signal trampoline, which is mapped to a read-only
page by kernel. The rationale of this patch is to skip "unwritable"
frames when looking for caller frames in command "finish", and a new
gdbarch method code_of_frame_writable is added. This patch fixes
the problem on ARM cortex-m target, but it can be used to fix
PR 8841 too.
gdb:
2016-05-10 Yao Qi <yao.qi@arm.com>
* arch-utils.c (default_code_of_frame_writable): New function.
* arch-utils.h (default_code_of_frame_writable): Declare.
* arm-tdep.c (arm_code_of_frame_writable): New function.
(arm_gdbarch_init): Install gdbarch method
code_of_frame_writable if the target is M-profile.
* frame.c (skip_unwritable_frames): New function.
* frame.h (skip_unwritable_frames): Declare.
* gdbarch.sh (code_of_frame_writable): New.
* gdbarch.c, gdbarch.h: Re-generated.
* infcmd.c (finish_command): Call skip_unwritable_frames.
In skip_artificial_frames we repeatedly call get_prev_frame_always until we get
a non-inline and non-tailcall frame assuming that there must be such a frame
eventually.
For record targets, however, we may have a frame chain that consists only of
artificial frames. This leads to a crash in get_frame_type when dereferencing a
NULL frame pointer.
Change skip_artificial_frames and skip_tailcall_frames to return NULL in such a
case and modify each caller to cope with a NULL return.
In frame_unwind_caller_pc and frame_unwind_caller_arch, we simply assert that
the returned value is not NULL. Their caller was supposed to check
frame_unwind_caller_id before calling those functions.
In other cases, we thrown an error.
In infcmd further move the skip_tailcall_frames call to the forward-stepping
case since we don't need a frame for reverse execution and we don't want to fail
because of that. Reverse-finish does make sense for a tailcall frame.
gdb/
* frame.h (skip_tailcall_frames): Update comment.
* frame.c (skip_artificial_frames, skip_tailcall_frames): Return NULL
if only artificial frames are found. Update comment.
(frame_unwind_caller_id): Handle NULL return.
(frame_unwind_caller_pc, frame_unwind_caller_arch): Assert that
skip_artificial_frames does not return NULL.
(frame_pop): Add an error if only tailcall frames are found.
* infcmd.c (finish_command): Move skip_tailcall_frames call into forward-
execution case. Add an error if only tailcall frames are found.
testsuite/
* gdb.btrace/tailcall-only.exp: New.
* gdb.btrace/tailcall-only.c: New.
* gdb.btrace/x86_64-tailcall-only.S: New.
* gdb.btrace/i686-tailcall-only.S: New.
This commit merges both the registers and $_siginfo "thread
running/executing" checks into a single function.
Accessing $_siginfo from a "catch signal" breakpoint condition doesn't
work. The condition always fails with "Selected thread is running":
(gdb) catch signal
Catchpoint 3 (standard signals)
(gdb)
condition $bpnum $_siginfo.si_signo == 5
(gdb) continue
Continuing.
Error in testing breakpoint condition:
Selected thread is running.
Catchpoint 3 (signal SIGUSR1), 0x0000003615e35877 in __GI_raise (sig=10) at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
56 return INLINE_SYSCALL (tgkill, 3, pid, selftid, sig);
(gdb)
When accessing the $_siginfo object, we check whether the thread is
marked running (external/public) state and refuse the access if so.
This is so "print $_siginfo" at the prompt fails nicelly when the
current thread is running. While evaluating breakpoint conditionals,
we haven't decided yet whether the thread is going to stop, so
is_running still returns true, and we thus always error out.
Evaluating an expression that requires registers access is really
conceptually the same -- we could think of $_siginfo as a pseudo
register. However, in that case we check whether the thread is marked
executing (internal/private state), not running (external/public
state). Changing the $_siginfo validation to check is_executing as
well fixes the bug in question.
Note that checking is_executing is not fully correct, not even for
registers. See PR 19389. However, I think this is the lesser of two
evils and ends up as an improvement. We at least now have a single
place to fix.
Tested on x86_64 GNU/Linux.
gdb/ChangeLog:
2016-01-13 Pedro Alves <palves@redhat.com>
PR breakpoints/19388
* frame.c (get_current_frame): Use validate_registers_access.
* gdbthread.h (validate_registers_access): Declare.
* infrun.c (validate_siginfo_access): Delete.
(siginfo_value_read, siginfo_value_write): Use
validate_registers_access.
* thread.c (validate_registers_access): New function.
gdb/testsuite/ChangeLog:
2016-01-13 Pedro Alves <palves@redhat.com>
PR breakpoints/19388
* gdb.base/catch-signal-siginfo-cond.c: New file.
* gdb.base/catch-signal-siginfo-cond.exp: New file.
The get_frame_language feels like it would be more at home in frame.c
rather than in stack.c, while the declaration, that is currently in
language.h can be moved into frame.h to match.
A couple of new includes are added, but otherwise no substantial change
here.
gdb/ChangeLog:
* stack.c (get_frame_language): Moved ...
* frame.c (get_frame_language): ... to here.
* language.h (get_frame_language): Declaration moved to frame.h.
* frame.h: Add language.h include, for language enum.
(get_frame_language): Declaration moved from language.h.
* language.c: Add frame.h include.
* top.c: Add frame.h include.
* symtab.h (struct obj_section): Declare.
(struct cmd_list_element): Declare.
The select_frame hook is used by TUI to update TUI's frame and register
information following changes to the selected frame. The problem with
this hook is that it gets called after every single frame change, even
if the frame change is only temporary or internal. This is the primary
cause of flickering and slowdown when running the inferior under TUI
with conditional breakpoints set. Internal GDB events are the source of
many calls to select_frame and these internal events are triggered
frequently, especially when a few conditional breakpoints are set.
This patch removes the select_frame hook altogether and instead makes
the frame and register information get updated in two key places (using
observers): after an inferior stops, and right before displaying a
prompt. The latter hook covers the case when frame information must be
updated following a call to "up", "down" or "frame", and the former
covers the case when frame and register information must be updated
after a call to "continue", "step", etc. or after the inferior stops in
async execution mode. Together these hooks should cover all the cases
when frame information ought to be refreshed (and when the relevant
windows ought to be subsequently updated).
The print_frame_info_listing hook is also effectively obsolete now, but
it still must be set while the TUI is active because its caller
print_frame_info will otherwise assume that the CLI is active, and will
print the frame informaion accordingly. So this patch also sets the
print_frame_info_listing hook to a dummy callback, in lieu of outright
removing it yet.
Effectively, with this patch, frame/PC changes that do not immediately
precede an inferior-stop event or a prompt display event no longer cause
TUI's frame and register information to be updated.
And as a result of this change and of the previous change to
tui_show_frame_info, the TUI is much more disciplined about updating the
screen, and so the flicker as described in the PR is totally gone.
gdb/ChangeLog:
PR tui/13378
* frame.c (select_frame): Remove reference to
deprecated_selected_frame_level_changed_hook.
* frame.h (deprecated_selected_frame_level_changed_hook): Remove
declaration.
* stack.c (deprecated_selected_frame_level_changed_hook):
Likewise.
* tui/tui-hooks.c (tui_selected_frame_level_changed_hook):
Rename to ...
(tui_refresh_frame_and_register_information): ... this. Bail
out if there is no stack. Don't update register information
unless registers_too_p is true.
(tui_print_frame_info_listing_hook): Rename to ...
(tui_dummy_print_frame_info_listing_hook): ... this.
(tui_before_prompt): New function.
(tui_normal_stop): New function.
(tui_before_prompt_observer): New observer.
(tui_normal_stop_observer): New observer.
(tui_install_hooks): Set
deprecated_print_frame_info_listing_hook to
tui_dummy_print_frame_info_listing_hook. Register
tui_before_prompt_observer to call tui_before_prompt and
tui_normal_stop_observer to call tui_normal_stop. Remove
reference to deprecated_selected_frame_level_changed_hook.
(tui_remove_hooks): Detach and unset tui_before_prompt_observer
and tui_normal_stop_observer. Remove reference to
deprecated_selected_frame_level_changed_hook.
This patch splits the TRY_CATCH macro into three, so that we go from
this:
~~~
volatile gdb_exception ex;
TRY_CATCH (ex, RETURN_MASK_ERROR)
{
}
if (ex.reason < 0)
{
}
~~~
to this:
~~~
TRY
{
}
CATCH (ex, RETURN_MASK_ERROR)
{
}
END_CATCH
~~~
Thus, we'll be getting rid of the local volatile exception object, and
declaring the caught exception in the catch block.
This allows reimplementing TRY/CATCH in terms of C++ exceptions when
building in C++ mode, while still allowing to build GDB in C mode
(using setjmp/longjmp), as a transition step.
TBC, after this patch, is it _not_ valid to have code between the TRY
and the CATCH blocks, like:
TRY
{
}
// some code here.
CATCH (ex, RETURN_MASK_ERROR)
{
}
END_CATCH
Just like it isn't valid to do that with C++'s native try/catch.
By switching to creating the exception object inside the CATCH block
scope, we can get rid of all the explicitly allocated volatile
exception objects all over the tree, and map the CATCH block more
directly to C++'s catch blocks.
The majority of the TRY_CATCH -> TRY+CATCH+END_CATCH conversion was
done with a script, rerun from scratch at every rebase, no manual
editing involved. After the mechanical conversion, a few places
needed manual intervention, to fix preexisting cases where we were
using the exception object outside of the TRY_CATCH block, and cases
where we were using "else" after a 'if (ex.reason) < 0)' [a CATCH
after this patch]. The result was folded into this patch so that GDB
still builds at each incremental step.
END_CATCH is necessary for two reasons:
First, because we name the exception object in the CATCH block, which
requires creating a scope, which in turn must be closed somewhere.
Declaring the exception variable in the initializer field of a for
block, like:
#define CATCH(EXCEPTION, mask) \
for (struct gdb_exception EXCEPTION; \
exceptions_state_mc_catch (&EXCEPTION, MASK); \
EXCEPTION = exception_none)
would avoid needing END_CATCH, but alas, in C mode, we build with C90,
which doesn't allow mixed declarations and code.
Second, because when TRY/CATCH are wired to real C++ try/catch, as
long as we need to handle cleanup chains, even if there's no CATCH
block that wants to catch the exception, we need for stop at every
frame in the unwind chain and run cleanups, then rethrow. That will
be done in END_CATCH.
After we require C++, we'll still need TRY/CATCH/END_CATCH until
cleanups are completely phased out -- TRY/CATCH in C++ mode will
save/restore the current cleanup chain, like in C mode, and END_CATCH
catches otherwise uncaugh exceptions, runs cleanups and rethrows, so
that C++ cleanups and exceptions can coexist.
IMO, this still makes the TRY/CATCH code look a bit more like a
newcomer would expect, so IMO worth it even if we weren't considering
C++.
gdb/ChangeLog.
2015-03-07 Pedro Alves <palves@redhat.com>
* common/common-exceptions.c (struct catcher) <exception>: No
longer a pointer to volatile exception. Now an exception value.
<mask>: Delete field.
(exceptions_state_mc_init): Remove all parameters. Adjust.
(exceptions_state_mc): No longer pop the catcher here.
(exceptions_state_mc_catch): New function.
(throw_exception): Adjust.
* common/common-exceptions.h (exceptions_state_mc_init): Remove
all parameters.
(exceptions_state_mc_catch): Declare.
(TRY_CATCH): Rename to ...
(TRY): ... this. Remove EXCEPTION and MASK parameters.
(CATCH, END_CATCH): New.
All callers adjusted.
gdb/gdbserver/ChangeLog:
2015-03-07 Pedro Alves <palves@redhat.com>
Adjust all callers of TRY_CATCH to use TRY/CATCH/END_CATCH
instead.
This normalizes some exception catch blocks that check for ex.reason
to look like this:
~~~
volatile gdb_exception ex;
TRY_CATCH (ex, RETURN_MASK_ALL)
{
...
}
if (ex.reason < 0)
{
...
}
~~~
This is a preparation step for running a script that converts all
TRY_CATCH uses to look like this instead:
~~~
TRY
{
...
}
CATCH (ex, RETURN_MASK_ALL)
{
...
}
END_CATCH
~~~
The motivation for that change is being able to reimplent TRY/CATCH in
terms of C++ try/catch.
This commit makes it so that:
- no condition other than ex.reason < 0 is checked in the if
predicate
- there's no "else" block to check whether no exception was caught
- there's no code between the TRY_CATCH (TRY) block and the
'if (ex.reason < 0)' block (CATCH).
- the exception object is no longer referred to outside the if/catch
block. Note the local volatile exception objects that are
currently defined inside functions that use TRY_CATCH will
disappear. In cases it's more convenient to still refer to the
exception outside the catch block, a new non-volatile local is
added and copy to that object is made within the catch block.
The following patches should make this all clearer.
gdb/ChangeLog:
2015-03-07 Pedro Alves <palves@redhat.com>
* amd64-tdep.c (amd64_frame_cache, amd64_sigtramp_frame_cache)
(amd64_epilogue_frame_cache): Normal exception handling code.
* break-catch-throw.c (check_status_exception_catchpoint)
(re_set_exception_catchpoint): Ditto.
* cli/cli-interp.c (safe_execute_command):
* cli/cli-script.c (script_from_file): Ditto.
* compile/compile-c-symbols.c (generate_c_for_for_one_variable):
Ditto.
* compile/compile-object-run.c (compile_object_run): Ditto.
* cp-abi.c (baseclass_offset): Ditto.
* cp-valprint.c (cp_print_value): Ditto.
* exceptions.c (catch_exceptions_with_msg):
* frame-unwind.c (frame_unwind_try_unwinder): Ditto.
* frame.c (get_frame_address_in_block_if_available): Ditto.
* i386-tdep.c (i386_frame_cache, i386_epilogue_frame_cache)
(i386_sigtramp_frame_cache): Ditto.
* infcmd.c (post_create_inferior): Ditto.
* linespec.c (parse_linespec, find_linespec_symbols):
* p-valprint.c (pascal_object_print_value): Ditto.
* parse.c (parse_expression_for_completion): Ditto.
* python/py-finishbreakpoint.c (bpfinishpy_init): Ditto.
* remote.c (remote_get_noisy_reply): Ditto.
* s390-linux-tdep.c (s390_frame_unwind_cache): Ditto.
* solib-svr4.c (solib_svr4_r_map): Ditto.
This patch renames symbols that happen to have names which are
reserved keywords in C++.
Most of this was generated with Tromey's cxx-conversion.el script.
Some places where later hand massaged a bit, to fix formatting, etc.
And this was rebased several times meanwhile, along with re-running
the script, so re-running the script from scratch probably does not
result in the exact same output. I don't think that matters anyway.
gdb/
2015-02-27 Tom Tromey <tromey@redhat.com>
Pedro Alves <palves@redhat.com>
Rename symbols whose names are reserved C++ keywords throughout.
gdb/gdbserver/
2015-02-27 Tom Tromey <tromey@redhat.com>
Pedro Alves <palves@redhat.com>
Rename symbols whose names are reserved C++ keywords throughout.
I noticed in frame_id_eq() we were checking for the "l" frame_id being
invalid twice instead of checking both "l" and "r", so this patch
corrects it.
gdb/ChangeLog:
* frame.c (frame_id_eq): Fix the check for FID_STACK_INVALID.
Currently "symtabs" in gdb are stored as a single linked list of
struct symtab that contains both symbol symtabs (the blockvectors)
and file symtabs (the linetables).
This has led to confusion, bugs, and performance issues.
This patch is conceptually very simple: split struct symtab into
two pieces: one part containing things common across the entire
compilation unit, and one part containing things specific to each
source file.
Example.
For the case of a program built out of these files:
foo.c
foo1.h
foo2.h
bar.c
foo1.h
bar.h
Today we have a single list of struct symtabs:
objfile -> foo.c -> foo1.h -> foo2.h -> bar.c -> foo1.h -> bar.h -> NULL
where "->" means the "next" pointer in struct symtab.
With this patch, that turns into:
objfile -> foo.c(cu) -> bar.c(cu) -> NULL
| |
v v
foo.c bar.c
| |
v v
foo1.h foo1.h
| |
v v
foo2.h bar.h
| |
v v
NULL NULL
where "foo.c(cu)" and "bar.c(cu)" are struct compunit_symtab objects,
and the files foo.c, etc. are struct symtab objects.
So now, for example, when we want to iterate over all blockvectors
we can now just iterate over the compunit_symtab list.
Plus a lot of the data that was either unused or replicated for each
symtab in a compilation unit now lives in struct compunit_symtab.
E.g., the objfile pointer, the producer string, etc.
I thought of moving "language" out of struct symtab but there is
logic to try to compute the language based on previously seen files,
and I think that's best left as is for now.
With my standard monster benchmark with -readnow (which I can't actually
do, but based on my calculations), whereas today the list requires
77MB to store all the struct symtabs, it now only requires 37MB.
A modest space savings given the gigabytes needed for all the debug info,
etc. Still, it's nice. Plus, whereas today we create a copy of dirname
for each source file symtab in a compilation unit, we now only create one
for the compunit.
So this patch is basically just a data structure reorg,
I don't expect significant performance improvements from it.
Notes:
1) A followup patch can do a similar split for struct partial_symtab.
I have left that until after I get the changes I want in to
better utilize .gdb_index (it may affect how we do partial syms).
2) Another followup patch *could* rename struct symtab.
The term "symtab" is ambiguous and has been a source of confusion.
In this patch I'm leaving it alone, calling it the "historical" name
of "filetabs", which is what they are now: just the file-name + line-table.
gdb/ChangeLog:
Split struct symtab into two: struct symtab and compunit_symtab.
* amd64-tdep.c (amd64_skip_xmm_prologue): Fetch producer from compunit.
* block.c (blockvector_for_pc_sect): Change "struct symtab *" argument
to "struct compunit_symtab *". All callers updated.
(set_block_compunit_symtab): Renamed from set_block_symtab. Change
"struct symtab *" argument to "struct compunit_symtab *".
All callers updated.
(get_block_compunit_symtab): Renamed from get_block_symtab. Change
result to "struct compunit_symtab *". All callers updated.
(find_iterator_compunit_symtab): Renamed from find_iterator_symtab.
Change result to "struct compunit_symtab *". All callers updated.
* block.h (struct global_block) <compunit_symtab>: Renamed from symtab.
hange type to "struct compunit_symtab *". All uses updated.
(struct block_iterator) <d.compunit_symtab>: Renamed from "d.symtab".
Change type to "struct compunit_symtab *". All uses updated.
* buildsym.c (struct buildsym_compunit): New struct.
(subfiles, buildsym_compdir, buildsym_objfile, main_subfile): Delete.
(buildsym_compunit): New static global.
(finish_block_internal): Update to fetch objfile from
buildsym_compunit.
(make_blockvector): Delete objfile argument.
(start_subfile): Rewrite to use buildsym_compunit. Don't initialize
debugformat, producer.
(start_buildsym_compunit): New function.
(free_buildsym_compunit): Renamed from free_subfiles_list.
All callers updated.
(patch_subfile_names): Rewrite to use buildsym_compunit.
(get_compunit_symtab): New function.
(get_macro_table): Delete argument comp_dir. All callers updated.
(start_symtab): Change result to "struct compunit_symtab *".
All callers updated. Create the subfile of the main source file.
(watch_main_source_file_lossage): Rewrite to use buildsym_compunit.
(reset_symtab_globals): Update.
(end_symtab_get_static_block): Update to use buildsym_compunit.
(end_symtab_without_blockvector): Rewrite.
(end_symtab_with_blockvector): Change result to
"struct compunit_symtab *". All callers updated.
Update to use buildsym_compunit. Don't set symtab->dirname,
instead set it in the compunit.
Explicitly make sure main symtab is first in its list.
Set debugformat, producer, blockvector, block_line_section, and
macrotable in the compunit.
(end_symtab_from_static_block): Change result to
"struct compunit_symtab *". All callers updated.
(end_symtab, end_expandable_symtab): Ditto.
(set_missing_symtab): Change symtab argument to
"struct compunit_symtab *". All callers updated.
(augment_type_symtab): Ditto.
(record_debugformat): Update to use buildsym_compunit.
(record_producer): Update to use buildsym_compunit.
* buildsym.h (struct subfile) <dirname>: Delete.
<producer, debugformat>: Delete.
<buildsym_compunit>: New member.
(get_compunit_symtab): Declare.
* dwarf2read.c (struct type_unit_group) <compunit_symtab>: Renamed
from primary_symtab. Change type to "struct compunit_symtab *".
All uses updated.
(dwarf2_start_symtab): Change result to "struct compunit_symtab *".
All callers updated.
(dwarf_decode_macros): Delete comp_dir argument. All callers updated.
(struct dwarf2_per_cu_quick_data) <compunit_symtab>: Renamed from
symtab. Change type to "struct compunit_symtab *". All uses updated.
(dw2_instantiate_symtab): Change result to "struct compunit_symtab *".
All callers updated.
(dw2_find_last_source_symtab): Ditto.
(dw2_lookup_symbol): Ditto.
(recursively_find_pc_sect_compunit_symtab): Renamed from
recursively_find_pc_sect_symtab. Change result to
"struct compunit_symtab *". All callers updated.
(dw2_find_pc_sect_compunit_symtab): Renamed from
dw2_find_pc_sect_symtab. Change result to
"struct compunit_symtab *". All callers updated.
(get_compunit_symtab): Renamed from get_symtab. Change result to
"struct compunit_symtab *". All callers updated.
(recursively_compute_inclusions): Change type of immediate_parent
argument to "struct compunit_symtab *". All callers updated.
(compute_compunit_symtab_includes): Renamed from
compute_symtab_includes. All callers updated. Rewrite to compute
includes of compunit_symtabs and not symtabs.
(process_full_comp_unit): Update to work with struct compunit_symtab.
(process_full_type_unit): Ditto.
(dwarf_decode_lines_1): Delete argument comp_dir. All callers updated.
(dwarf_decode_lines): Remove special case handling of main subfile.
(macro_start_file): Delete argument comp_dir. All callers updated.
(dwarf_decode_macro_bytes): Ditto.
* guile/scm-block.c (bkscm_print_block_syms_progress_smob): Update to
use struct compunit_symtab.
* i386-tdep.c (i386_skip_prologue): Fetch producer from compunit.
* jit.c (finalize_symtab): Build compunit_symtab.
* jv-lang.c (get_java_class_symtab): Change result to
"struct compunit_symtab *". All callers updated.
* macroscope.c (sal_macro_scope): Fetch macro table from compunit.
* macrotab.c (struct macro_table) <compunit_symtab>: Renamed from
comp_dir. Change type to "struct compunit_symtab *".
All uses updated.
(new_macro_table): Change comp_dir argument to cust,
"struct compunit_symtab *". All callers updated.
* maint.c (struct cmd_stats) <nr_compunit_symtabs>: Renamed from
nr_primary_symtabs. All uses updated.
(count_symtabs_and_blocks): Update to handle compunits.
(report_command_stats): Update output, "primary symtabs" renamed to
"compunits".
* mdebugread.c (new_symtab): Change result to
"struct compunit_symtab *". All callers updated.
(parse_procedure): Change type of search_symtab argument to
"struct compunit_symtab *". All callers updated.
* objfiles.c (objfile_relocate1): Loop over blockvectors in a
separate loop.
* objfiles.h (struct objfile) <compunit_symtabs>: Renamed from
symtabs. Change type to "struct compunit_symtab *". All uses updated.
(ALL_OBJFILE_FILETABS): Renamed from ALL_OBJFILE_SYMTABS.
All uses updated.
(ALL_OBJFILE_COMPUNITS): Renamed from ALL_OBJFILE_PRIMARY_SYMTABS.
All uses updated.
(ALL_FILETABS): Renamed from ALL_SYMTABS. All uses updated.
(ALL_COMPUNITS): Renamed from ALL_PRIMARY_SYMTABS. All uses updated.
* psympriv.h (struct partial_symtab) <compunit_symtab>: Renamed from
symtab. Change type to "struct compunit_symtab *". All uses updated.
* psymtab.c (psymtab_to_symtab): Change result type to
"struct compunit_symtab *". All callers updated.
(find_pc_sect_compunit_symtab_from_partial): Renamed from
find_pc_sect_symtab_from_partial. Change result type to
"struct compunit_symtab *". All callers updated.
(lookup_symbol_aux_psymtabs): Change result type to
"struct compunit_symtab *". All callers updated.
(find_last_source_symtab_from_partial): Ditto.
* python/py-symtab.c (stpy_get_producer): Fetch producer from compunit.
* source.c (forget_cached_source_info_for_objfile): Fetch debugformat
and macro_table from compunit.
* symfile-debug.c (debug_qf_find_last_source_symtab): Change result
type to "struct compunit_symtab *". All callers updated.
(debug_qf_lookup_symbol): Ditto.
(debug_qf_find_pc_sect_compunit_symtab): Renamed from
debug_qf_find_pc_sect_symtab, change result type to
"struct compunit_symtab *". All callers updated.
* symfile.c (allocate_symtab): Delete objfile argument.
New argument cust.
(allocate_compunit_symtab): New function.
(add_compunit_symtab_to_objfile): New function.
* symfile.h (struct quick_symbol_functions) <lookup_symbol>:
Change result type to "struct compunit_symtab *". All uses updated.
<find_pc_sect_compunit_symtab>: Renamed from find_pc_sect_symtab.
Change result type to "struct compunit_symtab *". All uses updated.
* symmisc.c (print_objfile_statistics): Compute blockvector count in
separate loop.
(dump_symtab_1): Update test for primary source symtab.
(maintenance_info_symtabs): Update to handle compunit symtabs.
(maintenance_check_symtabs): Ditto.
* symtab.c (set_primary_symtab): Delete.
(compunit_primary_filetab): New function.
(compunit_language): New function.
(iterate_over_some_symtabs): Change type of arguments "first",
"after_last" to "struct compunit_symtab *". All callers updated.
Update to loop over symtabs in each compunit.
(error_in_psymtab_expansion): Rename symtab argument to cust,
and change type to "struct compunit_symtab *". All callers updated.
(find_pc_sect_compunit_symtab): Renamed from find_pc_sect_symtab.
Change result type to "struct compunit_symtab *". All callers updated.
(find_pc_compunit_symtab): Renamed from find_pc_symtab.
Change result type to "struct compunit_symtab *". All callers updated.
(find_pc_sect_line): Only loop over symtabs within selected compunit
instead of all symtabs in the objfile.
* symtab.h (struct symtab) <blockvector>: Moved to compunit_symtab.
<compunit_symtab> New member.
<block_line_section>: Moved to compunit_symtab.
<locations_valid>: Ditto.
<epilogue_unwind_valid>: Ditto.
<macro_table>: Ditto.
<dirname>: Ditto.
<debugformat>: Ditto.
<producer>: Ditto.
<objfile>: Ditto.
<call_site_htab>: Ditto.
<includes>: Ditto.
<user>: Ditto.
<primary>: Delete
(SYMTAB_COMPUNIT): New macro.
(SYMTAB_BLOCKVECTOR): Update definition.
(SYMTAB_OBJFILE): Update definition.
(SYMTAB_DIRNAME): Update definition.
(struct compunit_symtab): New type. Common members among all source
symtabs within a compilation unit moved here. All uses updated.
(COMPUNIT_OBJFILE): New macro.
(COMPUNIT_FILETABS): New macro.
(COMPUNIT_DEBUGFORMAT): New macro.
(COMPUNIT_PRODUCER): New macro.
(COMPUNIT_DIRNAME): New macro.
(COMPUNIT_BLOCKVECTOR): New macro.
(COMPUNIT_BLOCK_LINE_SECTION): New macro.
(COMPUNIT_LOCATIONS_VALID): New macro.
(COMPUNIT_EPILOGUE_UNWIND_VALID): New macro.
(COMPUNIT_CALL_SITE_HTAB): New macro.
(COMPUNIT_MACRO_TABLE): New macro.
(ALL_COMPUNIT_FILETABS): New macro.
(compunit_symtab_ptr): New typedef.
(DEF_VEC_P (compunit_symtab_ptr)): New vector type.
gdb/testsuite/ChangeLog:
* gdb.base/maint.exp: Update expected output.
This patch is to add ptid into dummy_frame and extend frame_id to
dummy_frame_id (which has a ptid field). With this change, GDB uses
dummy_frame_id (thread ptid and frame_id) to find the dummy frames.
Currently, dummy frames are looked up by frame_id, which isn't
accurate in non-stop or multi-process mode. The test case
gdb.multi/dummy-frame-restore.exp shows the problem and this patch can
fix it.
Test dummy-frame-restore.exp makes two inferiors stop at
different functions, say, inferior 1 stops at f1 while inferior 2
stops at f2. Set a breakpoint to a function, do the inferior call
in two inferiors, and GDB has two dummy frames of the same frame_id.
When the inferior call is finished, GDB will look up a dummy frame
from its stack/list and restore the inferior's regcache. Two
inferiors are finished in different orders, the inferiors' states are
restored differently, which is wrong. Running dummy-frame-restore.exp
under un-patched GDB, we'll get two fails:
FAIL: gdb.multi/dummy-frame-restore.exp: inf 2 first: after infcall: bt in inferior 2
FAIL: gdb.multi/dummy-frame-restore.exp: inf 2 first: after infcall: bt in inferior 1
With this patch applied, GDB will choose the correct dummy_frame to
restore for a given inferior, because ptid is considered when looking up
dummy frames. Two fails above are fixed.
Regression tested on x86_64-linux, both native and gdbserver.
gdb:
2014-06-27 Yao Qi <yao@codesourcery.com>
* breakpoint.c (check_longjmp_breakpoint_for_call_dummy):
Change parameter type to 'struct thread_info *'. Caller
updated.
* breakpoint.h (check_longjmp_breakpoint_for_call_dummy):
Update declaration.
* dummy-frame.c (struct dummy_frame_id): New.
(dummy_frame_id_eq): New function.
(struct dummy_frame) <id>: Change its type to 'struct
dummy_frame_id'.
(dummy_frame_push): Add parameter ptid and save it in
dummy_frame_id.
(pop_dummy_frame_bpt): Use ptid of dummy_frame instead of
inferior_ptid.
(pop_dummy_frame): Assert that the ptid of dummy_frame equals
to inferior_ptid.
(lookup_dummy_frame): Change parameter type to 'struct
dummy_frame_id *'. Callers updated. Call dummy_frame_id_eq
instead of frame_id_eq.
(dummy_frame_pop): Add parameter ptid. Callers updated.
Update comments. Compose dummy_frame_id and pass it to
lookup_dummy_frame.
(dummy_frame_discard): Add parameter ptid.
(dummy_frame_sniffer): Compose dummy_frame_id and call
dummy_frame_id_eq instead of frame_id_eq.
(fprint_dummy_frames): Print ptid.
* dummy-frame.h: Remove comments.
(dummy_frame_push): Add ptid in declaration.
(dummy_frame_pop, dummy_frame_discard): Likewise.
gdb/testsuite:
2014-06-27 Yao Qi <yao@codesourcery.com>
* gdb.multi/dummy-frame-restore.exp: New.
* gdb.multi/dummy-frame-restore.c: New.
gdb/doc:
2014-06-27 Yao Qi <yao@codesourcery.com>
* gdb.texinfo (Maintenance Commands): Update the output of
'maint print dummy-frames' command.
Currently there are many calls to help_list that pass the constant -1
as the "class" value. However, the parameter is declared as being of
type enum command_class, and uses of the constant violate this
abstraction.
This patch fixes the error everywhere it occurs in the gdb sources.
Tested by rebuilding.
2014-06-13 Tom Tromey <tromey@redhat.com>
* cp-support.c (maint_cplus_command): Pass all_commands, not -1,
to help_list.
* guile/guile.c (info_guile_command): Pass all_commands, not -1,
to help_list.
* tui/tui-win.c (tui_command): Pass all_commands, not -1, to
help_list.
* tui/tui-regs.c (tui_reg_command): Pass all_commands, not -1, to
help_list.Pass all_commands, not -1, to help_list.
* cli/cli-dump.c (dump_command, append_command)
(srec_dump_command, ihex_dump_command, tekhex_dump_command)
(binary_dump_command, binary_append_command): Pass all_commands,
not -1, to help_list.
* cli/cli-cmds.c (info_command, set_debug): Pass all_commands, not
-1, to help_list.
* valprint.c (set_print, set_print_raw): Pass all_commands, not
-1, to help_list.
* typeprint.c (set_print_type): Pass all_commands, not -1, to
help_list.
* top.c (set_history): Pass all_commands, not -1, to help_list.
* target-descriptions.c (set_tdesc_cmd, unset_tdesc_cmd): Pass
all_commands, not -1, to help_list.
* symfile.c (overlay_command): Pass all_commands, not -1, to
help_list.
* spu-tdep.c (info_spu_command): Pass all_commands, not -1, to
help_list.
* serial.c (serial_set_cmd): Pass all_commands, not -1, to
help_list.
* ser-tcp.c (set_tcp_cmd, show_tcp_cmd): Pass all_commands, not
-1, to help_list.
* remote.c (remote_command, set_remote_cmd): Pass all_commands,
not -1, to help_list.
* ravenscar-thread.c (set_ravenscar_command): Pass all_commands,
not -1, to help_list.
* maint.c (maintenance_command, maintenance_info_command)
(maintenance_print_command, maintenance_set_cmd): Pass
all_commands, not -1, to help_list.
* macrocmd.c (macro_command): Pass all_commands, not -1, to
help_list.
* language.c (set_check): Pass all_commands, not -1, to help_list.
* infcmd.c (unset_command): Pass all_commands, not -1, to
help_list.
* frame.c (set_backtrace_cmd): Pass all_commands, not -1, to
help_list.
* dwarf2read.c (set_dwarf2_cmd): Pass all_commands, not -1, to
help_list.
* dcache.c (set_dcache_command): Pass all_commands, not -1, to
help_list.
* breakpoint.c (save_command): Pass all_commands, not -1, to
help_list.
* ada-lang.c (maint_set_ada_cmd, set_ada_command): Pass
all_commands, not -1, to help_list.
https://sourceware.org/ml/gdb-patches/2014-05/msg00737.html
Currently a MEMORY_ERROR raised during unwinding a frame will cause the
unwind to stop with an error message, for example:
(gdb) bt
#0 breakpt () at amd64-invalid-stack-middle.c:27
#1 0x00000000004008f0 in func5 () at amd64-invalid-stack-middle.c:32
#2 0x0000000000400900 in func4 () at amd64-invalid-stack-middle.c:38
#3 0x0000000000400910 in func3 () at amd64-invalid-stack-middle.c:44
#4 0x0000000000400928 in func2 () at amd64-invalid-stack-middle.c:50
Cannot access memory at address 0x2aaaaaab0000
However, frame #4 is marked as being the end of the stack unwind, so a
subsequent request for the backtrace looses the error message, such as:
(gdb) bt
#0 breakpt () at amd64-invalid-stack-middle.c:27
#1 0x00000000004008f0 in func5 () at amd64-invalid-stack-middle.c:32
#2 0x0000000000400900 in func4 () at amd64-invalid-stack-middle.c:38
#3 0x0000000000400910 in func3 () at amd64-invalid-stack-middle.c:44
#4 0x0000000000400928 in func2 () at amd64-invalid-stack-middle.c:50
When fetching the backtrace, or requesting the stack depth using the MI
interface the situation is even worse, the first time a request is made
we encounter the memory error and so the MI returns an error instead of
the correct result, for example:
(gdb) -stack-info-depth
^error,msg="Cannot access memory at address 0x2aaaaaab0000"
Or,
(gdb) -stack-list-frames
^error,msg="Cannot access memory at address 0x2aaaaaab0000"
However, once one of these commands has been used gdb has, internally,
walked the stack and figured that out that frame #4 is the bottom of the
stack, so the second time an MI command is tried you'll get the "expected"
result:
(gdb) -stack-info-depth
^done,depth="5"
Or,
(gdb) -stack-list-frames
^done,stack=[frame={level="0", .. snip lots .. }]
After this patch the MEMORY_ERROR encountered during the frame unwind is
attached to frame #4 as the stop reason, and is displayed in the CLI each
time the backtrace is requested. In the MI, catching the error means that
the "expected" result is returned the first time the MI command is issued.
So, from the CLI the results of the backtrace will be:
(gdb) bt
#0 breakpt () at amd64-invalid-stack-middle.c:27
#1 0x00000000004008f0 in func5 () at amd64-invalid-stack-middle.c:32
#2 0x0000000000400900 in func4 () at amd64-invalid-stack-middle.c:38
#3 0x0000000000400910 in func3 () at amd64-invalid-stack-middle.c:44
#4 0x0000000000400928 in func2 () at amd64-invalid-stack-middle.c:50
Backtrace stopped: Cannot access memory at address 0x2aaaaaab0000
Each and every time that the backtrace is requested, while the MI output
will similarly be consistently:
(gdb) -stack-info-depth
^done,depth="5"
Or,
(gdb) -stack-list-frames
^done,stack=[frame={level="0", .. snip lots .. }]
gdb/ChangeLog:
* frame.c (struct frame_info): Add stop_string field.
(get_prev_frame_always_1): Renamed from get_prev_frame_always.
(get_prev_frame_always): Old content moved into
get_prev_frame_always_1. Call get_prev_frame_always_1 inside
TRY_CATCH, handle MEMORY_ERROR exceptions.
(frame_stop_reason_string): New function definition.
* frame.h (unwind_stop_reason_to_string): Extend comment to
mention frame_stop_reason_string.
(frame_stop_reason_string): New function declaration.
* stack.c (frame_info): Switch to frame_stop_reason_string.
(backtrace_command_1): Switch to frame_stop_reason_string.
* unwind_stop_reason.def: Add UNWIND_MEMORY_ERROR.
(LAST_ENTRY): Changed to UNWIND_MEMORY_ERROR.
* guile/lib/gdb.scm: Add FRAME_UNWIND_MEMORY_ERROR to export list.
gdb/doc/ChangeLog:
* guile.texi (Frames In Guile): Mention FRAME_UNWIND_MEMORY_ERROR.
* python.texi (Frames In Python): Mention
gdb.FRAME_UNWIND_MEMORY_ERROR.
gdb/testsuite/ChangeLog:
* gdb.arch/amd64-invalid-stack-middle.exp: Update expected results.
* gdb.arch/amd64-invalid-stack-top.exp: Likewise.
https://sourceware.org/ml/gdb-patches/2014-05/msg00721.html
This function is confusingly named, the "frame_" in the name implies it
somehow is frame dependent, when in reality the function just converts an
'enum unwind_stop_reason' value to a string.
gdb/ChangeLog:
* frame.c (frame_stop_reason_string): Rename to ...
(unwind_stop_reason_to_string): this.
* frame.h (frame_stop_reason_string): Rename to ...
(unwind_stop_reason_to_string): this.
* stack.c (frame_info): Update call to frame_stop_reason_string.
(backtrace_command_1): Likewise.
* guile/scm-frame.c (gdbscm_unwind_stop_reason_string): Likewise.
* python/py-frame.c (gdbpy_frame_stop_reason_string): Likewise.
https://sourceware.org/ml/gdb-patches/2014-05/msg00712.html
If an error is thrown during computing a frame id then the frame is left
in existence but without a valid frame id, this will trigger internal
errors if/when the frame is later visited (for example in a backtrace).
This patch catches errors raised while computing the frame id, and
arranges for the new frame, the one without a frame id, to be removed
from the linked list of frames.
gdb/ChangeLog:
* frame.c (remove_prev_frame): New function.
(get_prev_frame_if_no_cycle): Create / discard cleanup using
remove_prev_frame.
gdb/testsuite/ChangeLog:
* gdb.arch/amd64-invalid-stack-middle.S: New file.
* gdb.arch/amd64-invalid-stack-middle.c: New file.
* gdb.arch/amd64-invalid-stack-middle.exp: New file.
* gdb.arch/amd64-invalid-stack-top.c: New file.
* gdb.arch/amd64-invalid-stack-top.exp: New file.
This PR is about an assertion failure in GDB that can be triggered by
setting "backtrace limit" to a value that causes GDB to stop unwinding
after an inline frame. In this case, an assertion in
inline_frame_this_id will trigger:
/* We need a valid frame ID, so we need to be based on a valid
frame. (...). */
gdb_assert (frame_id_p (*this_id));
Looking at the function:
static void
inline_frame_this_id (struct frame_info *this_frame,
void **this_cache,
struct frame_id *this_id)
{
struct symbol *func;
/* In order to have a stable frame ID for a given inline function,
we must get the stack / special addresses from the underlying
real frame's this_id method. So we must call get_prev_frame.
Because we are inlined into some function, there must be previous
frames, so this is safe - as long as we're careful not to
create any cycles. */
*this_id = get_frame_id (get_prev_frame (this_frame));
we see we're computing the frame id for the inline frame. If this is
an inline frame, which is a virtual frame constructed based on debug
info, on top of a real stack frame, we should _always_ be able to find
where the frame was inlined into, as that ultimately just means
peeling off the virtual frames on top of the real stack frame. If
there ultimately was no prev (real) stack frame, then we wouldn't have
been able to construct the inline frame either, by design. That's
what the assertion catches.
So we have an inline frame, we should _always_ be able to compute its
ID, even if that means bypassing the user backtrace limits to get at
the real stack frame's info. The problem is that inline_frame_id
calls get_prev_frame, and that takes user backtrace limits into
account. Code that wants to bypass the limits calls get_prev_frame_1
instead.
Note how get_prev_frame_1 already skips all checks for inline frames:
/* If we are unwinding from an inline frame, all of the below tests
were already performed when we unwound from the next non-inline
frame. We must skip them, since we can not get THIS_FRAME's ID
until we have unwound all the way down to the previous non-inline
frame. */
if (get_frame_type (this_frame) == INLINE_FRAME)
return get_prev_frame_if_no_cycle (this_frame);
And note how the related frame_unwind_caller_id function also uses
get_prev_frame_1:
struct frame_id
frame_unwind_caller_id (struct frame_info *next_frame)
{
struct frame_info *this_frame;
/* Use get_prev_frame_1, and not get_prev_frame. The latter will truncate
the frame chain, leading to this function unintentionally
returning a null_frame_id (e.g., when a caller requests the frame
ID of "main()"s caller. */
next_frame = skip_artificial_frames (next_frame);
this_frame = get_prev_frame_1 (next_frame);
if (this_frame)
return get_frame_id (skip_artificial_frames (this_frame));
else
return null_frame_id;
}
get_prev_frame_1 is currently static in frame.c. As a _1 suffix is
not a good name for an extern function, I've renamed it.
Tested on x86-64 Fedora 17.
gdb/
2014-04-18 Pedro alves <palves@redhat.com>
Tom Tromey <tromey@redhat.com>
PR backtrace/15558
* frame.c (get_prev_frame_1): Rename to ...
(get_prev_frame_always): ... this, and make extern. Adjust.
(skip_artificial_frames): Use get_prev_frame_always.
(frame_unwind_caller_id, frame_pop, get_prev_frame)
(get_frame_unwind_stop_reason): Adjust to rename.
* frame.h (get_prev_frame_always): Declare.
* inline-frame.c: Include frame.h.
(inline_frame_this_id): Use get_prev_frame_always.
gdb/testsuite/
2014-04-18 Tom Tromey <palves@redhat.com>
Pedro alves <tromey@redhat.com>
PR backtrace/15558
* gdb.opt/inline-bt.exp: Test backtracing from an inline function
with a backtrace limit.
* gdb.python/py-frame-inline.exp: Test running to an inline
function with a backtrace limit, and printing the newest frame.
* gdb.python/py-frame-inline.c (main): Call f.
In get_frame_unwind_stop_reason, remove the assumption that further frame
unwinding will succeed.
gdb/
2013-02-11 Jan Kratochvil <jan.kratochvil@redhat.com>
* frame.c (get_frame_unwind_stop_reason): Unconditionally call
get_prev_frame_1.
Add a function to build a frame_id for a frame with unavailable stack
and with a special identifier address.
2014-01-16 Markus Metzger <markus.t.metzger@intel.com>
* frame.h (frame_id_build_unavailable_stack_special): New.
* frame.c (frame_id_build_unavailable_stack_special): New.
Like when stepping, the current stack frame location is expected to be
printed as result of tfind command, if that results in moving to a
different function. In tfind_1 we see:
if (from_tty
&& (has_stack_frames () || traceframe_number >= 0))
{
enum print_what print_what;
/* NOTE: in imitation of the step command, try to determine
whether we have made a transition from one function to
another. If so, we'll print the "stack frame" (ie. the new
function and it's arguments) -- otherwise we'll just show the
new source line. */
if (frame_id_eq (old_frame_id,
get_frame_id (get_current_frame ())))
print_what = SRC_LINE;
else
print_what = SRC_AND_LOC;
print_stack_frame (get_selected_frame (NULL), 1, print_what, 1);
do_displays ();
}
However, when we haven't collected any registers in the tracepoint
(collect $regs), that doesn't actually work:
(gdb) tstart
(gdb) info tracepoints
Num Type Disp Enb Address What
1 tracepoint keep y 0x080483b7 in func0
at ../.././../git/gdb/testsuite/gdb.trace/circ.c:28
collect testload
installed on target
2 tracepoint keep y 0x080483bc in func1
at ../.././../git/gdb/testsuite/gdb.trace/circ.c:32
collect testload
installed on target
(gdb) c
Continuing.
Breakpoint 3, end () at ../.././../git/gdb/testsuite/gdb.trace/circ.c:72
72 }
(gdb) tstop
(gdb) tfind start
Found trace frame 0, tracepoint 1
#0 func0 () at ../.././../git/gdb/testsuite/gdb.trace/circ.c:28
28 }
(gdb) tfind
Found trace frame 1, tracepoint 2
32 }
(gdb)
When we don't have info about the stack available
(UNWIND_UNAVAILABLE), frames end up with outer_frame_id as frame ID.
And in the scenario above, the issue is that both frames before and
after the second tfind (the frames for func0 an func1) have the same
id (outer_frame_id), so the frame_id_eq check returns false, even
though the frames were of different functions. GDB knows that,
because the PC is inferred from the tracepoint's address, even if no
registers were collected.
To fix this, this patch adds support for frame ids with a valid code
address, but <unavailable> stack address, and then makes the unwinders
use that instead of the catch-all outer_frame_id for such frames. The
frame_id_eq check in tfind_1 then automatically does the right thing
as expected.
I tested with --directory=gdb.trace/ , before/after the patch, and
compared the resulting gdb.logs, then adjusted the tests to expect the
extra output that came out. Turns out that was only circ.exp, the
original test that actually brought this issue to light.
Tested on x86_64 Fedora 17, native and gdbserver.
gdb/
2013-12-17 Pedro Alves <palves@redhat.com>
* frame.h (enum frame_id_stack_status): New enum.
(struct frame_id) <stack_addr>: Adjust comment.
<stack_addr_p>: Delete field, replaced with ...
<stack_status>: ... this new field.
(frame_id_build_unavailable_stack): Declare.
* frame.c (frame_addr_hash, fprint_field, outer_frame_id)
(frame_id_build_special): Adjust.
(frame_id_build_unavailable_stack): New function.
(frame_id_build, frame_id_build_wild): Adjust.
(frame_id_p, frame_id_eq, frame_id_inner): Adjust to take into
account frames with unavailable stack.
* amd64-tdep.c (amd64_frame_this_id)
(amd64_sigtramp_frame_this_id, amd64_epilogue_frame_this_id): Use
frame_id_build_unavailable_stack.
* dwarf2-frame.c (dwarf2_frame_this_id): Likewise.
* i386-tdep.c (i386_frame_this_id, i386_epilogue_frame_this_id)
(i386_sigtramp_frame_this_id): Likewise.
gdb/testsuite/
2013-12-17 Pedro Alves <palves@redhat.com>
* gdb.trace/circ.exp: Expect frame info to be printed when
switching between frames with unavailable stack, but different
functions.
Doing "info frame" in the outermost frame, when that was indicated by
the next frame saying the unwound PC is undefined/not saved, results
in error and incomplete output:
(gdb) bt
#0 thread_function0 (arg=0x0) at threads.c:63
#1 0x00000034cf407d14 in start_thread (arg=0x7ffff7fcb700) at pthread_create.c:309
#2 0x000000323d4f168d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:115
(gdb) frame 2
#2 0x000000323d4f168d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:115
115 call *%rax
(gdb) info frame
Stack level 2, frame at 0x0:
rip = 0x323d4f168d in clone (../sysdeps/unix/sysv/linux/x86_64/clone.S:115); saved rip Register 16 was not saved
(gdb)
Not saved register values are treated as optimized out values
internally throughout. stack.c:frame_info is handing unvailable
values, but not optimized out ones. The patch deletes the
frame_unwind_caller_pc_if_available wrapper function and instead lets
errors propagate to frame_info (it's only user).
As frame_unwind_pc now needs to be able to handle and cache two
different error scenarios, the prev_pc.p variable is replaced with an
enumeration.
(FWIW, I looked into making gdbarch_unwind_pc or a variant return
struct value's instead, but it results in lots of boxing and unboxing
for no real gain -- e.g., the mips and arm implementations need to do
computation on the unboxed PC value. Might as well throw an error on
first attempt to get at invalid contents.)
After the patch, we get:
(gdb) info frame
Stack level 2, frame at 0x0:
rip = 0x323d4f168d in clone (../sysdeps/unix/sysv/linux/x86_64/clone.S:115); saved rip = <not saved>
Outermost frame: outermost
caller of frame at 0x7ffff7fcafc0
source language asm.
Arglist at 0x7ffff7fcafb8, args:
Locals at 0x7ffff7fcafb8, Previous frame's sp is 0x7ffff7fcafc8
(gdb)
A new test is added. It's based off dw2-reg-undefined.exp, and tweaked to
mark the return address (rip) of "stop_frame" as undefined.
Tested on x86_64 Fedora 17.
gdb/
2013-12-06 Pedro Alves <palves@redhat.com>
* frame.c (enum cached_copy_status): New enum.
(struct frame_info) <prev_pc.p>: Change type to enum
cached_copy_status.
(fprint_frame): Handle not saved and unavailable prev_pc values.
(frame_unwind_pc_if_available): Delete and merge contents into ...
(frame_unwind_pc): ... here. Handle OPTIMIZED_OUT_ERROR. Adjust
to use enum cached_copy_status.
(frame_unwind_caller_pc_if_available): Delete.
(create_new_frame): Adjust.
* frame.h (frame_unwind_caller_pc_if_available): Delete
declaration.
* stack.c (frame_info): Use frame_unwind_caller_pc instead of
frame_unwind_caller_pc_if_available, and handle
NOT_AVAILABLE_ERROR and OPTIMIZED_OUT_ERROR errors.
* valprint.c (val_print_optimized_out): Use val_print_not_saved.
(val_print_not_saved): New function.
* valprint.h (val_print_not_saved): Declare.
gdb/testsuite/
2013-12-06 Pedro Alves <palves@redhat.com>
* gdb.dwarf2/dw2-undefined-ret-addr.S: New file.
* gdb.dwarf2/dw2-undefined-ret-addr.c: New file.
* gdb.dwarf2/dw2-undefined-ret-addr.exp: New file.
In order to catch <optimized out> errors like we catch <unavailable>
errors, this adds a new OPTIMIZED_OUT_ERROR error code, and throws it
in various places.
gdb/ChangeLog
2013-12-06 Andrew Burgess <aburgess@broadcom.com>
Pedro Alves <palves@redhat.com>
* exceptions.h (errors): Add OPTIMIZED_OUT_ERROR.
* dwarf2loc.c (write_pieced_value): Throw OPTIMIZED_OUT_ERROR.
* frame.c (frame_unwind_register): Throw OPTIMIZED_OUT_ERROR.
* spu-tdep.c (spu_software_single_step): Throw
OPTIMIZED_OUT_ERROR.
* valops.c (value_assign): Throw OPTIMIZED_OUT_ERROR.
With a simple Ada program where I have 3 functions, one just calling
the next, the backtrace is currently broken when GDB is compiled
at -O2:
#0 hello.first () at hello.adb:5
#1 0x0000000100001475 in hello.second () at hello.adb:10
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
It turns out that a recent patch deleted the assignment of variable
this_id, making it an unitialized variable:
* frame-unwind.c (default_frame_unwind_stop_reason): Return
UNWIND_OUTERMOST if the frame's ID is outer_frame_id.
* frame.c (get_prev_frame_1): Remove outer_frame_id check.
The hunk in question starts with:
- /* Check that this frame is not the outermost. If it is, don't try
- to unwind to the prev frame. */
- this_id = get_frame_id (this_frame);
- if (frame_id_eq (this_id, outer_frame_id))
(the code was removed as redundant - but removing the assignment
was in fact not intentional).
There is no other code in this function that sets the variable.
Instead of re-adding the statement in the lone section where it is
actually used, I inlined it, and then got rid of the variable
altogether. This way, and until we start needing this frame ID
in another location within that function, we dont' have to worry
about the variable's validity/lifetime.
gdb/ChangeLog:
* frame.c (get_prev_frame_1): Delete variable "this_id".
Replace its use by a call to get_frame_id.
The stop_reason != UNWIND_NO_REASON doesn't currently have "set debug
frame" output. This patch makes it print the stop_reason enum value
as a string.
gdb/
2013-11-28 Pedro Alves <palves@redhat.com>
* frame.c (get_prev_frame_1) <stop_reason != UNWIND_NO_REASON>:
Add "set debug frame" output.
(frame_stop_reason_symbol_string): New function.
After the previous patch, it should be clear that the
this_frame->unwind->stop_reason check is redundant with the
outer_frame_id check just below. We can now move the frame_id_eq
comparison to the default this_frame->unwind->stop_reason callback.
Tested on x86_64 Fedora 17.
gdb/
2013-11-28 Pedro Alves <palves@redhat.com>
* frame-unwind.c (default_frame_unwind_stop_reason): Return
UNWIND_OUTERMOST if the frame's ID is outer_frame_id.
* frame.c (get_prev_frame_1): Remove outer_frame_id check.
- The UNWIND_NULL_ID check in get_prev_frame_1 used to really be
against null_frame_id, back before we had outer_frame_id. We didn't
have UNWIND_OUTERMOST when outer_frame_id was added, but we do now,
and it's more accurate.
- It used to be necessary to check for the sentinel frame explicitly
because that uses null_frame_id for frame id. Since no other frame
can have that id nowadays (it's asserted by compute_frame_id), we
don't need that explicit check.
Tested on x86_64 Fedora 17.
gdb/
2013-11-28 Pedro Alves <palves@redhat.com>
* frame.c (get_prev_frame_1): If the frame id is outer_frame_id,
set the unwind stop reason to UNWIND_OUTERMOST, not
UNWIND_NULL_ID. Remove explicit check for sentinel frame.
Another spot that missed the previous related text adjustments.
Tested on x86_64 Fedora 17.
gdb/
2013-11-28 Pedro Alves <palves@redhat.com>
* frame.c (frame_unwind_register): Say the register was "not
saved" instead of "optimized out".
Given we already have the frame id stash, which holds the ids of all
frames in the chain, detecting corrupted stacks with wide stack cycles
with non-consecutive dup frame ids is just as cheap as just detecting
cycles in consecutive frames:
#0 frame_id1
#1 frame_id2
#2 frame_id3
#3 frame_id1
#4 frame_id2
#5 frame_id3
#6 frame_id1
... forever ...
We just need to check whether the stash already knows about a given
frame id instead of comparing the ids of the previous/this frames.
Tested on x86_64 Fedora 17.
gdb/
2013-11-22 Pedro Alves <palves@redhat.com>
Tom Tromey <tromey@redhat.com>
* frame.c (frame_stash_add): Now returns whether a frame with the
same ID was already known.
(compute_frame_id): New function, factored out from get_frame_id.
(get_frame_id): No longer lazilly compute the frame id here.
(get_prev_frame_if_no_cycle): New function. Detects wider stack
cycles.
(get_prev_frame_1): Use it instead of get_prev_frame_raw directly,
and checking for stack cycles here.
The UNWIND_SAME_ID check is done between THIS_FRAME and the next frame
when we go try to unwind the previous frame. But at this point, it's
already too late -- we ended up with two frames with the same ID in
the frame chain. Each frame having its own ID is an invariant assumed
throughout GDB. This patch applies the UNWIND_SAME_ID detection
earlier, right after the previous frame is unwound, discarding the dup
frame if a cycle is detected.
The patch includes a new test that fails before the change. Before
the patch, the test causes an infinite loop in GDB, after the patch,
the UNWIND_SAME_ID logic kicks in and makes the backtrace stop with:
Backtrace stopped: previous frame identical to this frame (corrupt stack?)
The test uses dwarf CFI to emulate a corrupted stack with a cycle. It
has a function with registers marked DW_CFA_same_value (most
importantly RSP/RIP), so that GDB computes the same ID for that frame
and its caller. IOW, something like this:
#0 - frame_id_1
#1 - frame_id_2
#2 - frame_id_3
#3 - frame_id_4
#4 - frame_id_4 <<<< outermost (UNWIND_SAME_ID).
(The test's code is just a copy of dw2-reg-undefined.S /
dw2-reg-undefined.c, adjusted to use DW_CFA_same_value instead of
DW_CFA_undefined, and to mark a different set of registers.)
The infinite loop is here, in value_fetch_lazy:
while (VALUE_LVAL (new_val) == lval_register && value_lazy (new_val))
{
frame = frame_find_by_id (VALUE_FRAME_ID (new_val));
...
new_val = get_frame_register_value (frame, regnum);
}
get_frame_register_value can return a lazy register value pointing to
the next frame. This means that the register wasn't clobbered by
FRAME; the debugger should therefore retrieve its value from the next
frame.
To be clear, get_frame_register_value unwinds the value in question
from the next frame:
struct value *
get_frame_register_value (struct frame_info *frame, int regnum)
{
return frame_unwind_register_value (frame->next, regnum);
^^^^^^^^^^^
}
In other words, if we get a lazy lval_register, it should have the
frame ID of the _next_ frame, never of FRAME.
At this point in value_fetch_lazy, the whole relevant chunk of the
stack up to frame #4 has already been unwound. The loop always
"unlazies" lval_registers in the "next/innermost" direction, not in
the "prev/unwind further/outermost" direction.
So say we're looking at frame #4. get_frame_register_value in frame
#4 can return a lazy register value of frame #3. So the next
iteration, frame_find_by_id tries to read the register from frame #3.
But, since frame #4 happens to have same id as frame #3,
frame_find_by_id returns frame #4 instead. Rinse, repeat, and we have
an infinite loop.
This is an old latent problem, exposed by the recent addition of the
frame stash. Before we had a stash, frame_find_by_id(frame_id_4)
would walk over all frames starting at the current frame, and would
always find #3 first. The stash happens to return #4 instead:
struct frame_info *
frame_find_by_id (struct frame_id id)
{
struct frame_info *frame, *prev_frame;
...
/* Try using the frame stash first. Finding it there removes the need
to perform the search by looping over all frames, which can be very
CPU-intensive if the number of frames is very high (the loop is O(n)
and get_prev_frame performs a series of checks that are relatively
expensive). This optimization is particularly useful when this function
is called from another function (such as value_fetch_lazy, case
VALUE_LVAL (val) == lval_register) which already loops over all frames,
making the overall behavior O(n^2). */
frame = frame_stash_find (id);
if (frame)
return frame;
for (frame = get_current_frame (); ; frame = prev_frame)
{
gdb/
2013-11-22 Pedro Alves <palves@redhat.com>
PR 16155
* frame.c (get_prev_frame_1): Do the UNWIND_SAME_ID check between
this frame and the new previous frame, not between this frame and
the next frame.
gdb/testsuite/
2013-11-22 Pedro Alves <palves@redhat.com>
PR 16155
* gdb.dwarf2/dw2-dup-frame.S: New file.
* gdb.dwarf2/dw2-dup-frame.c: New file.
* gdb.dwarf2/dw2-dup-frame.exp: New file.
Kevin Buettner