The generated mloop files can trigger compile time warnings. It can
be difficult to see/understand where the original code is coming from
as all the diagnostics point to the generated output. Using #line
pragmas, we can point people to the original source files.
Unfortunately, this code is written in POSIX shell, and that lacks
support for line number tracking. The $LINENO variable, even when
available, can just be plain wrong. For example, when using dash
and subshells, $LINENO can end up having negative values. Add a
wrapper script that will uses awk to rewrite the $LINENO variable
to the right value to avoid all that.
Basically lineno.sh takes an input script, rewrites all uses of
$LINENO into the actual line number (and $0 into the original file
name), and then executes the temporary script.
This commit doesn't actually add #line pragmas to any files. That
comes next.
The TUI status window is called the "locator" in the source, but
"status" in the documentation. Whenever I've needed to find the code,
I've had to search to "locate" it (ha, ha). This patch renames the
window to match the public name of the window.
The TUI status line is called the "status" window in the
documentation, but not in the source. There, the relevant files are
named "tui-stack", which to me makes it sound like they have something
to do with backtraces. This patch renames them to "tui-status".
GNU ld currently fails to link 32-bit executables on FreeBSD/amd64 when
the linked libraries have dependencies on shared objects themselves:
$ gcc -m32 -o ei ei.c -lexecinfo
/var/gcc/binutils/amd64/lib/gcc/amd64-pc-freebsd14.0/13.2.0/../../../../amd64-pc-freebsd14.0/bin/ld:
warning: libelf.so.2, needed by /usr/lib/../lib32/libexecinfo.so, not found
(try using -rpath or -rpath-link)
/var/gcc/binutils/amd64/lib/gcc/amd64-pc-freebsd14.0/13.2.0/../../../../amd64-pc-freebsd14.0/bin/ld:
/usr/lib/../lib32/libexecinfo.so: undefined reference to `elf_begin@R1.0'
[...]
Fixed by handling FreeBSD/amd64 like Linux/x86.
Tested on amd64-pc-freebsd14.0.
Commit d5cebea18e ("Make cached_reg_t own its data") added a
destructor to cached_reg_t.
That caused a build problem with Clang, which errors out like so:
> CXX python/py-unwind.o
> gdb/python/py-unwind.c:126:16: error: flexible array member 'reg' of type 'cached_reg_t[]' with non-trivial destruction
> 126 | cached_reg_t reg[];
> | ^
This is is not really a problem for our code, which allocates the
whole structure with xmalloc, and then initializes the array elements
with in-place new, and then takes care to call the destructor
manually. Like, commit d5cebea18e did:
@@ -928,7 +927,7 @@ pyuw_dealloc_cache (frame_info *this_frame, void *cache)
cached_frame_info *cached_frame = (cached_frame_info *) cache;
for (int i = 0; i < cached_frame->reg_count; i++)
- xfree (cached_frame->reg[i].data);
+ cached_frame->reg[i].~cached_reg_t ();
Maybe we should get rid of the flexible array member and use a bog
standard std::vector. I doubt this would cause any visible
performance issue.
Meanwhile, to unbreak the build, this commit switches from C99-style
flexible array member to 0-length array. It behaves the same, and
Clang doesn't complain. I got the idea from here:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70932#c11
GCC 9, our oldest support version, already supported this:
https://gcc.gnu.org/onlinedocs/gcc-9.1.0/gcc/Zero-Length.html
but the extension is actually much older than that. Note that
C99-style flexible array members are not standard C++ either.
Change-Id: I37dda18f367e238a41d610619935b2a0f2acacce
Replace some fall through comments with the attribute, and add some
default abort calls when the compiler can't figure out that the set
of values were already fully enumerated in the switch statement.
All funcs already call other funcs that don't return. The mips port is
the only exception because its generic exception handler can return in
the case of normal exceptions. So while the exceptions its signal handler
triggers doesn't return, we can't express that conditional logic. So add
some useless abort calls to make the compiler happy.
The compiler pointed out that we're testing LAST_TIMER_REG and
LAST_COUNTER which are the same value ... and that's because we
set LAST_TIMER_REG to the wrong register. Fix the typo.
The compiler pointed out we checked AZ twice. Sort by name to avoid
that in the future, and to make it clearer that we have coverage of
all the bits. And add the bits we were missing.
The order here doesn't matter as it's just turning a pointer into a
human readable string when store tracing is enabled.
The scache vars aren't used by ports in the pbb & fast codepaths,
nor are they documented as inputs to the callbacks, so delete them
to avoid unused variable compiler warnings.
Pull out the common parts of the genmloop invocation into the common
code. This will make it easier to add more, and make the per-port
differences a little more obvious.
gprofng incorrectly reads the form of the DW_FORM_ref_addr attribute for DWARF
Version 3 or later.
From DWARF specification:
References that use the attribute form DW_FORM_ref_addr are specified to
be four bytes in the DWARF 32-bit format and eight bytes in the DWARF
64-bit format, while DWARF Version 2 specifies that such references have
the same size as an address on the target system.
2023-12-18 Vladimir Mezentsev <vladimir.mezentsev@oracle.com>
PR gprofng/31169
* src/DwarfLib.cc: Fix the reader for DW_FORM_ref_addr.
Downstream, AMD is carrying a testcase
(gdb.rocm/continue-over-kernel-exit.exp) that exposes a couple issues
with the amd-dbgapi target's handling of exited threads. The test
can't be added upstream yet, unfortunately, due to dependency on DWARF
extensions that can't be upstreamed yet. However, it can be found on
the mailing list on the same series as this patch.
The test spawns a kernel with a number of waves. The waves do nothing
but exit. There is a breakpoint on the s_endpgm instruction. Once
that breakpoint is hit, the test issues a "continue" command. We
should see one breakpoint hit per wave, and then the whole program
exiting. We do see that, however we also see this:
[New AMDGPU Wave ?:?:?:1 (?,?,?)/?]
[AMDGPU Wave ?:?:?:1 (?,?,?)/? exited]
*repeat for other waves*
...
[Thread 0x7ffff626f640 (LWP 3048491) exited]
[Thread 0x7fffeb7ff640 (LWP 3048488) exited]
[Inferior 1 (process 3048475) exited normally]
That "New AMDGPU Wave" output comes from infrun.c itself adding the
thread to the GDB thread list, because it got an event for a thread
not on the thread list yet. The output shows "?"s instead of proper
coordinates, because the event was a TARGET_WAITKIND_THREAD_EXITED,
i.e., the wave was already gone when infrun.c added the thread to the
thread list.
That shouldn't ever happen for the amd-dbgapi target, threads should
only ever be added by the backend.
Note "New AMDGPU Wave ?:?:?:1" is for wave 1. What happened was that
wave 1 terminated previously, and a previous call to
amd_dbgapi_target::update_thread_list() noticed the wave had vanished
and removed it from the GDB thread list. However, because the wave
was stepping when it terminated (due to the displaced step over the
s_endpgm) instruction, it is guaranteed that the amd-dbgapi library
queues a WAVE_COMMAND_TERMINATED event for the exit.
When we process that WAVE_COMMAND_TERMINATED event, in
amd-dbgapi-target.c:process_one_event, we return it to the core as a
TARGET_WAITKIND_THREAD_EXITED event:
static void
process_one_event (amd_dbgapi_event_id_t event_id,
amd_dbgapi_event_kind_t event_kind)
{
...
if (status == AMD_DBGAPI_STATUS_ERROR_INVALID_WAVE_ID
&& event_kind == AMD_DBGAPI_EVENT_KIND_WAVE_COMMAND_TERMINATED)
ws.set_thread_exited (0);
...
}
Recall the wave is already gone from the GDB thread list. So when GDB
sees that TARGET_WAITKIND_THREAD_EXITED event for a thread it doesn't
know about, it adds the thread to the thread list, resulting in that:
[New AMDGPU Wave ?:?:?:1 (?,?,?)/?]
and then, because it was a TARGET_WAITKIND_THREAD_EXITED event, GDB
marks the thread exited right afterwards:
[AMDGPU Wave ?:?:?:1 (?,?,?)/? exited]
The fix is to make amd_dbgapi_target::update_thread_list() _not_
delete vanishing waves iff they were stepping or in progress of being
stopped. These two cases are the ones dbgapi guarantees will result
in a WAVE_COMMAND_TERMINATED event if the wave terminates:
/**
* A command for a wave was not able to complete because the wave has
* terminated.
*
* Commands that can result in this event are ::amd_dbgapi_wave_stop and
* ::amd_dbgapi_wave_resume in single step mode. Since the wave terminated
* before stopping, this event will be reported instead of
* ::AMD_DBGAPI_EVENT_KIND_WAVE_STOP.
*
* The wave that terminated is available by the ::AMD_DBGAPI_EVENT_INFO_WAVE
* query. However, the wave will be invalid since it has already terminated.
* It is the client's responsibility to know what command was being performed
* and was unable to complete due to the wave terminating.
*/
AMD_DBGAPI_EVENT_KIND_WAVE_COMMAND_TERMINATED = 2,
As the comment says, it's GDB's responsability to know whether the
wave was stepping or being stopped. Since we now have a wave_info map
with one entry for each wave, that seems like the place to store that
information. However, I still decided to put all the coordinate
information in its own structure. I.e., basically renamed the
existing wave_info to wave_coordinates, and then added a new wave_info
structure that holds the new state, plus a wave_coordinates object.
This seemed cleaner as there are places where we only need to
instantiate a wave_coordinates object.
There's an extra twist. The testcase also exercises stopping at a new
kernel right after the first kernel fully exits. In that scenario, we
were hitting this assertion after the first kernel fully exits and the
hit of the breakpoint at the second kernel is handled:
[amd-dbgapi] process_event_queue: Pulled event from dbgapi: event_id.handle = 26, event_kind = WAVE_STOP
[amd-dbgapi-lib] suspending queue_3, queue_2, queue_1 (refresh wave list)
../../src/gdb/amd-dbgapi-target.c:1625: internal-error: amd_dbgapi_thread_deleted: Assertion `it != info->wave_info_map.end ()' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
This is the exact same problem as above, just a different
manifestation. In this scenario, we end up in update_thread_list
successfully deleting the exited thread (because it was no longer the
current thread) that was incorrectly added by infrun.c. Because it
was added by infrun.c and not by amd-dbgapi-target.c:add_gpu_thread,
it doesn't have an entry in the wave_info map, so
amd_dbgapi_thread_deleted trips on this assertion:
gdb_assert (it != info->wave_info_map.end ());
here:
...
-> stop_all_threads
-> update_thread_list
-> target_update_thread_list
-> amd_dbgapi_target::update_thread_list
-> thread_db_target::update_thread_list
-> linux_nat_target::update_thread_list
-> delete_exited_threads
-> delete_thread
-> delete_thread_1
-> gdb::observers::observable<thread_info*>::notify
-> amd_dbgapi_thread_deleted
-> internal_error_loc
The testcase thus tries both running to exit after the first kernel
exits, and running to a breakpoint in a second kernel after the first
kernel exits.
Approved-By: Lancelot Six <lancelot.six@amd.com> (amdgpu)
Change-Id: I43a66f060c35aad1fe0d9ff022ce2afd0537f028
Currently, if you step over kernel exit, you see:
stepi
[AMDGPU Wave ?:?:?:1 (?,?,?)/? exited]
Command aborted, thread exited.
(gdb)
Those '?' are because the thread/wave is already gone by the time GDB
prints the "exited" notification, we can't ask dbgapi for any info
about the wave anymore.
This commit fixes it by caching the wave's coordinates as soon as GDB
sees the wave for the first time, and making
amd_dbgapi_target::pid_to_str use the cached info.
At first I thought of clearing the wave_info object from a
thread_exited observer. However, that is too soon, resulting in this:
(gdb) si
[AMDGPU Wave 1:4:1:1 (0,0,0)/0 exited]
Command aborted, thread exited.
(gdb) thread
[Current thread is 6 (AMDGPU Wave ?:?:?:0 (?,?,?)/?) (exited)]
We need instead to clear the wave info when the thread is ultimately
deleted, so we get:
(gdb) si
[AMDGPU Wave 1:4:1:1 (0,0,0)/0 exited]
Command aborted, thread exited.
(gdb) thread
[Current thread is 6 (AMDGPU Wave 1:4:1:1 (0,0,0)/0) (exited)]
And for that, we need a new thread_deleted observable.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Approved-By: Lancelot Six <lancelot.six@amd.com> (amdgpu)
Change-Id: I6c3e22541f051e1205f75eb657b04dc15e547580
With AMD GPU debugging, I noticed that when stepping over a breakpoint
placed on top of the s_endpgm instruction inline (displaced=off), GDB
would behave differently -- it wouldn't print the wave exit. E.g:
With displaced stepping, or no breakpoint at all:
stepi
[AMDGPU Wave 1:4:1:1 (0,0,0)/0 exited]
Command aborted, thread exited.
(gdb)
With inline stepping:
stepi
Command aborted, thread exited.
(gdb)
In the cases we see the "exited" notification, handle_thread_exit is
what first called delete_thread on the exiting thread, which is
non-silent.
With inline stepping, however, handle_thread_exit ends up in
update_thread_list (via restart_threads) before any delete_thread
call. Thus, amd_dbgapi_target::update_thread_list notices that the
wave is gone and deletes it with delete_thread_silent.
This commit fixes it, by making handle_thread_exited call
set_thread_exited (with the default silent=false) early, which emits
the user-visible notification.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Change-Id: I22ab3145e18d07c99dace45576307b9f9d5d966f
displaced_step_finish can be called with event_status.kind ==
TARGET_WAITKIND_THREAD_EXITED, and in that case it is not possible to
get at the already-exited thread's registers.
This patch moves the get_thread_regcache calls to branches that
actually need it, where we know the thread is still alive.
It also adds an assertion to get_thread_regcache, to help catching
these broken cases sooner.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Change-Id: I63b5eacb3e02a538fc5087c270d8025adfda88c3
While making step over thread exit work properly on AMDGPU, I noticed
that if there's a breakpoint on top of the exit syscall, and,
displaced stepping is off, then when GDB reports "Command aborted,
thread exited.", GDB also switches focus to a random thread, instead
of leaving the exited thread as selected:
(gdb) thread
[Current thread is 6, lane 0 (AMDGPU Lane 1:4:1:1/0 (0,0,0)[0,0,0])]
(gdb) si
Command aborted, thread exited.
(gdb) thread
[Current thread is 5 (Thread 0x7ffff626f640 (LWP 3248392))]
(gdb)
The previous patch extended gdb.threads/step-over-thread-exit.exp to
exercise this on GNU/Linux (on the CPU side), and there, after that
"si", we always end up with the exiting thread as selected even
without this fix, but that's just a concidence, there's a code path
that happens to select the exiting thread for an unrelated reason.
This commit add the explict switch, fixing the latent problem for
GNU/Linux, and the actual problem on AMDGPU. I wrote a gdb.rocm/
testcase for this, but it can't be upstreamed yet, until more pieces
of the DWARF machinery are upstream as well.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Change-Id: I6ff57a79514ac0142bba35c749fe83d53d9e4e51
This commit makes the following improvements to
gdb.threads/step-over-thread-exit.exp:
- Add a third axis to stepping over the breakpoint with displaced vs
inline stepping -- also test with no breakpoint at all.
- Check that when GDB reports "Command aborted, thread exited.", the
selected thread is the thread that exited. This is always true
currently on GNU/Linux by coincidence, but a similar testcase on AMD
GPU exposed a problem here. Better make the testcase catch any
potential regression.
- Fixes a race that Simon ran into with GDBserver testing.
(gdb) next
[New Thread 2143071.2143438]
Thread 3 "step-over-threa" hit Breakpoint 2, 0x000055555555524e in my_exit_syscall () at .../testsuite/lib/my-syscalls.S:74
74 SYSCALL (my_exit, __NR_exit)
(gdb) FAIL: gdb.threads/step-over-thread-exit.exp: displaced-stepping=auto: non-stop=on: target-non-stop=on: schedlock=off: cmd=next: ns_stop_all=0: command aborts when thread exits
I was not able to reproduce it, but I believe that what happens is
the following:
Once we continue, the thread 2 exits, and the main thread thus
unblocks from its pthread_join, and spawns a new thread. That new
thread may hit the breakpoint at my_exit_syscall very quickly. GDB
could then see/process that breakpoint event before the thread exit
event for the thread we care about, which would result in the
failure seen above.
The fix here is to not loop and start a new thread at all in the
scenario where the race can happen. We only need to loop and spawn
new threads when testing with "cmd=continue" and schedlock off, in
which case GDB doesn't abort the command when the thread exits.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Change-Id: I90c95c32f00630a3f682b1541c23aff52451f9b6
By inspection, I noticed that the previous patch went too far, here:
@@ -7705,7 +7713,8 @@ remote_target::discard_pending_stop_replies (struct inferior *inf)
if (rs->remote_desc == NULL)
return;
- reply = (struct stop_reply *) rns->pending_event[notif_client_stop.id];
+ stop_reply_up reply
+ = as_stop_reply_up (std::move (rns->pending_event[notif_client_stop.id]));
/* Discard the in-flight notification. */
if (reply != NULL && reply->ptid.pid () == inf->pid)
That is always moving the stop reply from pending_event, when we only
really want to peek into it. The code further below that even says:
/* Discard the in-flight notification. */
if (reply != NULL && reply->ptid.pid () == inf->pid)
{
/* Leave the notification pending, since the server expects that
we acknowledge it with vStopped. But clear its contents, so
that later on when we acknowledge it, we also discard it. */
This commit reverts that hunk back, adjusted to use unique_ptr::get().
Change-Id: Ifc809d1a8225150a4656889f056d51267100ee24
