Re: [pushed+7.8] Re: [PATCH] Fix "attach" command vs user input race [Re: Regression for attach from stdin [Re: [pushed] Re: [PATCH v6 0/2] enable target-async by default]]

[Doug Evans <dje at google dot com>]

On Tue, Jul 29, 2014 at 2:48 PM, Doug Evans <dje@google.com> wrote:
> On Wed, Jul 9, 2014 at 10:09 AM, Pedro Alves <palves@redhat.com> wrote:
>> On 07/09/2014 05:37 PM, Doug Evans wrote:
>>
>>> spurious change
>>
>> Fixed.
>>
>>> I like this a lot better.  Thanks.
>>> The patch is ok with me, modulo removing the spurious change.
>>
>> Here's what I pushed to both master and gdb-7.8-branch.
>>
>> Thanks.
>>
>> -------------
>> From 1fe2833b6dd03602ba86aa334e81466ea9abe66a Mon Sep 17 00:00:00 2001
>> From: Pedro Alves <palves@redhat.com>
>> Date: Wed, 9 Jul 2014 17:52:58 +0100
>> Subject: [PATCH] Fix "attach" command vs user input race
>>
>> On async targets, a synchronous attach is done like this:
>>
>>    #1 - target_attach is called (PTRACE_ATTACH is issued)
>>    #2 - a continuation is installed
>>    #3 - we go back to the event loop
>>    #4 - target reports stop (SIGSTOP), event loop wakes up, and
>>         attach continuation is called
>>    #5 - among other things, the continuation calls
>>         target_terminal_inferior, which removes stdin from the event
>>         loop
>>
>> Note that in #3, GDB is still processing user input.  If the user is
>> fast enough, e.g., with something like:
>>
>>   echo -e "attach PID\nset xxx=1" | gdb
>>
>> ... then the "set" command is processed before the attach completes.
>>
>> We get worse behavior even, if input is a tty and therefore
>> readline/editing is enabled, with e.g.,:
>>
>>  (gdb) attach PID\nset xxx=1
>>
>> we then crash readline/gdb, with:
>>
>>  Attaching to program: attach-wait-input, process 14537
>>  readline: readline_callback_read_char() called with no handler!
>>  Aborted
>>  $
>>
>> Fix this by calling target_terminal_inferior before #3 above.
>>
>> The test covers both scenarios by running with editing/readline forced
>> to both on and off.
>>
>> gdb/
>> 2014-07-09  Pedro Alves  <palves@redhat.com>
>>
>>         * infcmd.c (attach_command_post_wait): Don't call
>>         target_terminal_inferior here.
>>         (attach_command): Call it here instead.
>>
>> gdb/testsuite/
>> 2014-07-09  Pedro Alves  <palves@redhat.com>
>>
>>         * gdb.base/attach-wait-input.exp: New file.
>>         * gdb.base/attach-wait-input.c: New file.
>
> Hi.
>
> Is this TODO still needed after this patch?
>
> infcmd.c:
>
> /*
>  * TODO:
>  * Should save/restore the tty state since it might be that the
>  * program to be debugged was started on this tty and it wants
>  * the tty in some state other than what we want.  If it's running
>  * on another terminal or without a terminal, then saving and
>  * restoring the tty state is a harmless no-op.
>  * This only needs to be done if we are attaching to a process.
>  */

A related issue (or the same one if one prefers):

post_create_inferior does this:

  /* Be sure we own the terminal in case write operations are performed.  */
  target_terminal_ours ();

but post_create_inferior is called *after* target_post_attach
in attach_command_post_wait:

  /* Take any necessary post-attaching actions for this platform.  */
  target_post_attach (ptid_get_pid (inferior_ptid));

  post_create_inferior (&current_target, from_tty);

What if target_post_attach does some writes?
Seems like it can, e.g., some of the ptrace checking stuff may print a warning.
Plus attach_command_post_wait calls some other stuff before
post_create_inferior that could cause some writes to the terminal.

Question: Is there a specific terminal state that needs to be in
effect when attach_command_post_wait returns?
Or can we just call target_terminal_ours at its start?
[and leave it to other code to switch back to target_terminal_inferior
as needed - e.g. proceed calls resume which will call
target_terminal_inferior]