diff options
author | Simon Marchi <simon.marchi@efficios.com> | 2021-01-24 23:57:29 -0500 |
---|---|---|
committer | Pedro Alves <pedro@palves.net> | 2021-03-26 15:58:47 +0000 |
commit | 1192f124a308601f5fef7a35715ccd6f904e7b17 (patch) | |
tree | 0aec8738ccfc8f3948a521f30257dae75e342e65 /gdb/target.h | |
parent | e5b9b39f8872fa01efb9c7f8ce7283fb9cd5122d (diff) | |
download | fsf-binutils-gdb-1192f124a308601f5fef7a35715ccd6f904e7b17.zip fsf-binutils-gdb-1192f124a308601f5fef7a35715ccd6f904e7b17.tar.gz fsf-binutils-gdb-1192f124a308601f5fef7a35715ccd6f904e7b17.tar.bz2 |
gdb: generalize commit_resume, avoid commit-resuming when threads have pending statuses
The rationale for this patch comes from the ROCm port [1], the goal
being to reduce the number of back and forths between GDB and the
target when doing successive operations. I'll start with explaining
the rationale and then go over the implementation. In the ROCm / GPU
world, the term "wave" is somewhat equivalent to a "thread" in GDB.
So if you read if from a GPU stand point, just s/thread/wave/.
ROCdbgapi, the library used by GDB [2] to communicate with the GPU
target, gives the illusion that it's possible for the debugger to
control (start and stop) individual threads. But in reality, this is
not how it works. Under the hood, all threads of a queue are
controlled as a group. To stop one thread in a group of running ones,
the state of all threads is retrieved from the GPU, all threads are
destroyed, and all threads but the one we want to stop are re-created
from the saved state. The net result, from the point of view of GDB,
is that the library stopped one thread. The same thing goes if we
want to resume one thread while others are running: the state of all
running threads is retrieved from the GPU, they are all destroyed, and
they are all re-created, including the thread we want to resume.
This leads to some inefficiencies when combined with how GDB works,
here are two examples:
- Stopping all threads: because the target operates in non-stop mode,
when the user interface mode is all-stop, GDB must stop all threads
individually when presenting a stop. Let's suppose we have 1000
threads and the user does ^C. GDB asks the target to stop one
thread. Behind the scenes, the library retrieves 1000 thread
states and restores the 999 others still running ones. GDB asks
the target to stop another one. The target retrieves 999 thread
states and restores the 998 remaining ones. That means that to
stop 1000 threads, we did 1000 back and forths with the GPU. It
would have been much better to just retrieve the states once and
stop there.
- Resuming with pending events: suppose the 1000 threads hit a
breakpoint at the same time. The breakpoint is conditional and
evaluates to true for the first thread, to false for all others.
GDB pulls one event (for the first thread) from the target, decides
that it should present a stop, so stops all threads using
stop_all_threads. All these other threads have a breakpoint event
to report, which is saved in `thread_info::suspend::waitstatus` for
later. When the user does "continue", GDB resumes that one thread
that did hit the breakpoint. It then processes the pending events
one by one as if they just arrived. It picks one, evaluates the
condition to false, and resumes the thread. It picks another one,
evaluates the condition to false, and resumes the thread. And so
on. In between each resumption, there is a full state retrieval
and re-creation. It would be much nicer if we could wait a little
bit before sending those threads on the GPU, until it processed all
those pending events.
To address this kind of performance issue, ROCdbgapi has a concept
called "forward progress required", which is a boolean state that
allows its user (i.e. GDB) to say "I'm doing a bunch of operations,
you can hold off putting the threads on the GPU until I'm done" (the
"forward progress not required" state). Turning forward progress back
on indicates to the library that all threads that are supposed to be
running should now be really running on the GPU.
It turns out that GDB has a similar concept, though not as general,
commit_resume. One difference is that commit_resume is not stateful:
the target can't look up "does the core need me to schedule resumed
threads for execution right now". It is also specifically linked to
the resume method, it is not used in other contexts. The target
accumulates resumption requests through target_ops::resume calls, and
then commits those resumptions when target_ops::commit_resume is
called. The target has no way to check if it's ok to leave resumed
threads stopped in other target methods.
To bridge the gap, this patch generalizes the commit_resume concept in
GDB to match the forward progress concept of ROCdbgapi. The current
name (commit_resume) can be interpreted as "commit the previous resume
calls". I renamed the concept to "commit_resumed", as in "commit the
threads that are resumed".
In the new version, we have two things:
- the commit_resumed_state field in process_stratum_target: indicates
whether GDB requires target stacks using this target to have
resumed threads committed to the execution target/device. If
false, an execution target is allowed to leave resumed threads
un-committed at the end of whatever method it is executing.
- the commit_resumed target method: called when commit_resumed_state
transitions from false to true. While commit_resumed_state was
false, the target may have left some resumed threads un-committed.
This method being called tells it that it should commit them back
to the execution device.
Let's take the "Stopping all threads" scenario from above and see how
it would work with the ROCm target with this change. Before stopping
all threads, GDB would set the target's commit_resumed_state field to
false. It would then ask the target to stop the first thread. The
target would retrieve all threads' state from the GPU and mark that
one as stopped. Since commit_resumed_state is false, it leaves all
the other threads (still resumed) stopped. GDB would then proceed to
call target_stop for all the other threads. Since resumed threads are
not committed, this doesn't do any back and forth with the GPU.
To simplify the implementation of targets, this patch makes it so that
when calling certain target methods, the contract between the core and
the targets guarantees that commit_resumed_state is false. This way,
the target doesn't need two paths, one for commit_resumed_state ==
true and one for commit_resumed_state == false. It can just assert
that commit_resumed_state is false and work with that assumption.
This also helps catch places where we forgot to disable
commit_resumed_state before calling the method, which represents a
probable optimization opportunity. The commit adds assertions in the
target method wrappers (target_resume and friends) to have some
confidence that this contract between the core and the targets is
respected.
The scoped_disable_commit_resumed type is used to disable the commit
resumed state of all process targets on construction, and selectively
re-enable it on destruction (see below for criteria). Note that it
only sets the process_stratum_target::commit_resumed_state flag. A
subsequent call to maybe_call_commit_resumed_all_targets is necessary
to call the commit_resumed method on all target stacks with process
targets that got their commit_resumed_state flag turned back on. This
separation is because we don't want to call the commit_resumed methods
in scoped_disable_commit_resumed's destructor, as they may throw.
On destruction, commit-resumed is not re-enabled for a given target
if:
1. this target has no threads resumed, or
2. this target has at least one resumed thread with a pending status
known to the core (saved in thread_info::suspend::waitstatus).
The first point is not technically necessary, because a proper
commit_resumed implementation would be a no-op if the target has no
resumed threads. But since we have a flag do to a quick check, it
shouldn't hurt.
The second point is more important: together with the
scoped_disable_commit_resumed instance added in fetch_inferior_event,
it makes it so the "Resuming with pending events" described above is
handled efficiently. Here's what happens in that case:
1. The user types "continue".
2. Upon destruction, the scoped_disable_commit_resumed in the
`proceed` function does not enable commit-resumed, as it sees some
threads have pending statuses.
3. fetch_inferior_event is called to handle another event, the
breakpoint hit evaluates to false, and that thread is resumed.
Because there are still more threads with pending statuses, the
destructor of scoped_disable_commit_resumed in
fetch_inferior_event still doesn't enable commit-resumed.
4. Rinse and repeat step 3, until the last pending status is handled
by fetch_inferior_event. In that case,
scoped_disable_commit_resumed's destructor sees there are no more
threads with pending statues, so it asks the target to commit
resumed threads.
This allows us to avoid all unnecessary back and forths, there is a
single commit_resumed call once all pending statuses are processed.
This change required remote_target::remote_stop_ns to learn how to
handle stopping threads that were resumed but pending vCont. The
simplest example where that happens is when using the remote target in
all-stop, but with "maint set target-non-stop on", to force it to
operate in non-stop mode under the hood. If two threads hit a
breakpoint at the same time, GDB will receive two stop replies. It
will present the stop for one thread and save the other one in
thread_info::suspend::waitstatus.
Before this patch, when doing "continue", GDB first resumes the thread
without a pending status:
Sending packet: $vCont;c:p172651.172676#f3
It then consumes the pending status in the next fetch_inferior_event
call:
[infrun] do_target_wait_1: Using pending wait status status->kind = stopped, signal = GDB_SIGNAL_TRAP for Thread 1517137.1517137.
[infrun] target_wait (-1.0.0, status) =
[infrun] 1517137.1517137.0 [Thread 1517137.1517137],
[infrun] status->kind = stopped, signal = GDB_SIGNAL_TRAP
It then realizes it needs to stop all threads to present the stop, so
stops the thread it just resumed:
[infrun] stop_all_threads: Thread 1517137.1517137 not executing
[infrun] stop_all_threads: Thread 1517137.1517174 executing, need stop
remote_stop called
Sending packet: $vCont;t:p172651.172676#04
This is an unnecessary resume/stop. With this patch, we don't commit
resumed threads after proceeding, because of the pending status:
[infrun] maybe_commit_resumed_all_process_targets: not requesting commit-resumed for target extended-remote, a thread has a pending waitstatus
When GDB handles the pending status and stop_all_threads runs, we stop a
resumed but pending vCont thread:
remote_stop_ns: Enqueueing phony stop reply for thread pending vCont-resume (1520940, 1520976, 0)
That thread was never actually resumed on the remote stub / gdbserver,
so we shouldn't send a packet to the remote side asking to stop the
thread.
Note that there are paths that resume the target and then do a
synchronous blocking wait, in sort of nested event loop, via
wait_sync_command_done. For example, inferior function calls, or any
run control command issued from a breakpoint command list. We handle
that making wait_sync_command_one a "sync" point -- force forward
progress, or IOW, force-enable commit-resumed state.
gdb/ChangeLog:
yyyy-mm-dd Simon Marchi <simon.marchi@efficios.com>
Pedro Alves <pedro@palves.net>
* infcmd.c (run_command_1, attach_command, detach_command)
(interrupt_target_1): Use scoped_disable_commit_resumed.
* infrun.c (do_target_resume): Remove
target_commit_resume call.
(commit_resume_all_targets): Remove.
(maybe_set_commit_resumed_all_targets): New.
(maybe_call_commit_resumed_all_targets): New.
(enable_commit_resumed): New.
(scoped_disable_commit_resumed::scoped_disable_commit_resumed)
(scoped_disable_commit_resumed::~scoped_disable_commit_resumed)
(scoped_disable_commit_resumed::reset)
(scoped_disable_commit_resumed::reset_and_commit)
(scoped_enable_commit_resumed::scoped_enable_commit_resumed)
(scoped_enable_commit_resumed::~scoped_enable_commit_resumed):
New.
(proceed): Use scoped_disable_commit_resumed and
maybe_call_commit_resumed_all_targets.
(fetch_inferior_event): Use scoped_disable_commit_resumed.
* infrun.h (struct scoped_disable_commit_resumed): New.
(maybe_call_commit_resumed_all_process_targets): New.
(struct scoped_enable_commit_resumed): New.
* mi/mi-main.c (exec_continue): Use scoped_disable_commit_resumed.
* process-stratum-target.h (class process_stratum_target):
<commit_resumed_state>: New.
* record-full.c (record_full_wait_1): Change commit_resumed_state
around calling commit_resumed.
* remote.c (class remote_target) <commit_resume>: Rename to...
<commit_resumed>: ... this.
(struct stop_reply): Move up.
(remote_target::commit_resume): Rename to...
(remote_target::commit_resumed): ... this. Check if there is any
thread pending vCont resume.
(remote_target::remote_stop_ns): Generate stop replies for resumed
but pending vCont threads.
(remote_target::wait_ns): Add gdb_assert.
* target-delegates.c: Regenerate.
* target.c (target_wait, target_resume): Assert that the current
process_stratum target isn't in commit-resumed state.
(defer_target_commit_resume): Remove.
(target_commit_resume): Remove.
(target_commit_resumed): New.
(make_scoped_defer_target_commit_resume): Remove.
(target_stop): Assert that the current process_stratum target
isn't in commit-resumed state.
* target.h (struct target_ops) <commit_resume>: Rename to ...
<commit_resumed>: ... this.
(target_commit_resume): Remove.
(target_commit_resumed): New.
(make_scoped_defer_target_commit_resume): Remove.
* top.c (wait_sync_command_done): Use
scoped_enable_commit_resumed.
[1] https://github.com/ROCm-Developer-Tools/ROCgdb/
[2] https://github.com/ROCm-Developer-Tools/ROCdbgapi
Change-Id: I836135531a29214b21695736deb0a81acf8cf566
Diffstat (limited to 'gdb/target.h')
-rw-r--r-- | gdb/target.h | 31 |
1 files changed, 13 insertions, 18 deletions
diff --git a/gdb/target.h b/gdb/target.h index 5d14edb..0aef372 100644 --- a/gdb/target.h +++ b/gdb/target.h @@ -485,8 +485,15 @@ struct target_ops int TARGET_DEBUG_PRINTER (target_debug_print_step), enum gdb_signal) TARGET_DEFAULT_NORETURN (noprocess ()); - virtual void commit_resume () + + /* Ensure that all resumed threads are committed to the target. + + See the description of + process_stratum_target::commit_resumed_state for more + details. */ + virtual void commit_resumed () TARGET_DEFAULT_IGNORE (); + /* See target_wait's description. Note that implementations of this method must not assume that inferior_ptid on entry is pointing at the thread or inferior that ends up reporting an @@ -1463,23 +1470,11 @@ extern void target_disconnect (const char *, int); target_commit_resume below. */ extern void target_resume (ptid_t ptid, int step, enum gdb_signal signal); -/* Commit a series of resumption requests previously prepared with - target_resume calls. - - GDB always calls target_commit_resume after calling target_resume - one or more times. A target may thus use this method in - coordination with the target_resume method to batch target-side - resumption requests. In that case, the target doesn't actually - resume in its target_resume implementation. Instead, it prepares - the resumption in target_resume, and defers the actual resumption - to target_commit_resume. E.g., the remote target uses this to - coalesce multiple resumption requests in a single vCont packet. */ -extern void target_commit_resume (); - -/* Setup to defer target_commit_resume calls, and reactivate - target_commit_resume on destruction, if it was previously - active. */ -extern scoped_restore_tmpl<int> make_scoped_defer_target_commit_resume (); +/* Ensure that all resumed threads are committed to the target. + + See the description of process_stratum_target::commit_resumed_state + for more details. */ +extern void target_commit_resumed (); /* For target_read_memory see target/target.h. */ |