gdb: Coalesce/aggregate (async) vCont packets/actions
authorPedro Alves <palves@redhat.com>
Wed, 26 Oct 2016 10:08:28 +0000 (11:08 +0100)
committerPedro Alves <palves@redhat.com>
Wed, 26 Oct 2016 15:22:50 +0000 (16:22 +0100)
commit85ad3aaf403d2104c82010494d3d4a93a36e2e6f
treeffe792616bfdd77915a3c5246e43e9bbdaf7a875
parent5a04c4cf5df6d13596e79e7b84520cbe245a5a4d
gdb: Coalesce/aggregate (async) vCont packets/actions

Currently, with "maint set target-non-stop on", that is, when gdb
connects with the non-stop/asynchronous variant of the remote
protocol, even with "set non-stop off", GDB always sends one vCont
packet per thread resumed.  This patch makes GDB aggregate and
coalesce vCont packets, so we send vCont packets like "vCont;s:p1.1;c"
in non-stop mode too.

Basically, this is done by:

  - Adding a new target method target_commit_resume that is called
    after calling target_resume one or more times.  When resuming a
    batch of threads, we'll only call target_commit_resume once after
    calling target_resume for all threads.

  - Making the remote target defer sending the actual vCont packet to
    target_commit_resume.

Special care must be taken to avoid sending a vCont action with a
"wildcard" thread-id (all threads of process / all threads) when that
would resume threads/processes that should not be resumed.  See
remote_commit_resume comments for details.

Unlike all-stop's remote_resume implementation, this handles the case
of too many actions resulting in a too-big vCont packet, by flushing
the vCont packet and starting a new one.

E.g., imagining that the "c" action in:

  vCont;s:1;c

overflows the packet buffer, we split the actions like:

  vCont;s:1
  vCont;c

Tested on x86_64 Fedora 20, with and without "maint set
target-non-stop on".

Also tested with a hack that makes remote_commit_resume flush the vCont
packet after every action appended (which caught a few bugs).

gdb/ChangeLog:
2016-10-26  Pedro Alves  <palves@redhat.com>

* inferior.h (ALL_NON_EXITED_INFERIORS): New macro.
* infrun.c (do_target_resume): Call target_commit_resume.
(proceed): Defer target_commit_resume while looping over threads,
resuming them.  Call target_commit_resume at the end.
* record-btrace.c (record_btrace_commit_resume): New function.
(init_record_btrace_ops): Install it as to_commit_resume method.
* record-full.c (record_full_commit_resume): New function.
(record_full_wait_1): Call the beneath target's to_commit_resume
method.
(init_record_full_ops): Install record_full_commit_resume as
to_commit_resume method.
* remote.c (struct private_thread_info) <last_resume_step,
last_resume_sig, vcont_resumed>: New fields.
(remote_add_thread): Set the new thread's vcont_resumed flag.
(demand_private_info): Delete.
(get_private_info_thread, get_private_info_ptid): New functions.
(remote_update_thread_list): Adjust.
(process_initial_stop_replies): Clear the thread's vcont_resumed
flag.
(remote_resume): If connected in non-stop mode, record the resume
request and return early.
(struct private_inferior): New.
(struct vcont_builder): New.
(vcont_builder_restart, vcont_builder_flush)
(vcont_builder_push_action): New functions.
(MAX_ACTION_SIZE): New macro.
(remote_commit_resume): New function.
(thread_pending_fork_status, is_pending_fork_parent_thread): New
functions.
(check_pending_event_prevents_wildcard_vcont_callback)
(check_pending_events_prevent_wildcard_vcont): New functions.
(process_stop_reply): Adjust.  Clear the thread's vcont_resumed
flag.
(init_remote_ops): Install remote_commit_resume.
* target-delegates.c: Regenerate.
* target.c (defer_target_commit_resume): New global.
(target_commit_resume, make_cleanup_defer_target_commit_resume):
New functions.
* target.h (struct target_ops) <to_commit_resume>: New field.
(target_resume): Update comments.
(target_commit_resume): New declaration.
gdb/ChangeLog
gdb/inferior.h
gdb/infrun.c
gdb/record-btrace.c
gdb/record-full.c
gdb/remote.c
gdb/target-delegates.c
gdb/target.c
gdb/target.h