Re: [RFC 0/3] Use reinsert breakpoint for vCont;s

[Antoine Tremblay <antoine.tremblay@ericsson.com>]

Yao Qi writes:

> Nowadays, reinsert breakpoint is used in GDBserver to step over a
> breakpoint.  I want to use it to handle vCont;s.  The motivation
> of this work is to exercise software single step in GDBserver side.
> In the past two weeks, I am fixing various test fails, but still
> can't fix all of them.  I want to post something here, and hope
> people can help me on this area.
>
> Suppose GDB is able to send vCont;s to GDBserver using software
> single step (done by patch 3), what should GDBserver do?  It call
> function single_step if lwp->resume->kind is resume_step.  See
> patch 2.
>
> With this change, reinsert breakpoint is used for two purposes,
> 1) step over GDBserver breakpoint, 2) handle vCont;s.  Here are some
> facts or assumptions in my mind,
>
>  - reinsert breakpoints can be inserted for both step over and
>    vCont;s together.  GDBserver should finish all step-overs
>    before resuming the threads, see scenario b) below,
>  - GDB doesn't send more than one vCont s actions in one vCont
>    packet, although RSP doc doesn't say this.
>
> It is straightforward to insert reinsert breakpoints for vCont;s,
> but I am not sure when to delete them.  Here are some scenarios,
>
> a) vCont;s thread A, and vCont;c thread B.  Thread A hits the reinsert
>    breakpoints, and GDBserver can remove them.  What is the proper
>    place to remove them?
>
> b) vCont;s thread A, and vCont;c thread B.  Thread B hits breakpoints
>    (not reinsert), do we remove reinsert breakpoints?  My answer is
>    no.  In the following step-over, reinsert breakpoints for step-over
>    are deleted, but reinsert breakpoints for vCont;s (thread A) are still
>    there.
>
> c) vCont;s thread A, and vCont;c thread B.  Thread B hits the reinsert
>    breakpoints (for thread A vCont;s), do we remove reinsert breakpoints?
>    I think no, we can just step over it for thread B.
>
> d) vCont;s thread A, and vCont;c thread B.  A signal arrives, do we remove
>    reinsert breakpoints?  Yes, I think so.
>
> IMO, b) requires reinsert breakpoint thread specific, so that we can delete
> reinsert breakpoints for step-over of thread B, but keep reinsert breakpoints
> for vCont;s of thread A.  That is what patch 1 does.
>
> I tried different ways to remove reinsert breakpoints in GDBserver, but still
> can't fix fails in gdb.threads/schedlock.exp, that the program gets SIGILL or
> SIGSEGV.  These fails can't happen in every run, and they are disappeared
> when I turn on debugging output in GDBserver.  I suspect they are about the
> improper management to reinsert breakpoints.
>

I'm not sure at the moment but this makes me think of an issue I sent
here: https://www.sourceware.org/ml/gdb/2015-11/msg00030.html

Actually thanks to a good discussion with LTTng maintainer, (Mathieu
Desnoyer) last week I got the hint that this could be due to the
situation described here:
https://community.arm.com/groups/processors/blog/2010/02/17/caches-and-self-modifying-code

Short story, the instruction cache could be out of sync after we've
removed a software breakpoint and the processor would execute the
breakpoint instruction causing a SIGILL rather then the original memory.

The solution however, calling __clear_cache(...) requires that we're in
the same process for the memory addresses to be valid, I think, so I'm
not sure how to fix this yet...

I'm still early in my investigation but it may be something to consider.

> (gdb) PASS: gdb.threads/schedlock.exp: schedlock=off: cmd=next: call_function=1: next to increment (5)
> next^M
> 78          while (*myp > 0)^M
> (gdb) next^M
> ^M
> Thread 1 "schedlock" received signal SIGILL, Illegal instruction.^M
> [Switching to Thread 3797.3797]^M
> 0x000087f8 in thread_function (arg=0x0) at /home/yao/SourceCode/gnu/gdb/git/gdb/testsuite/gdb.threads/schedlock.c:78^M
> 78          while (*myp > 0)^M
> (gdb) FAIL: gdb.threads/schedlock.exp: schedlock=off: cmd=next: call_function=1: next to increment (6)
>
> Any ideas on the overall design, how to handle vCont;s in GDBserver using
> software single step? or is it a completely wrong thing to handle vCont;s
> using software single step?
>
> *** BLURB HERE ***
>
> Yao Qi (3):
>   make reinsert breakpoint thread specific
>   use reinsert breakpoint for vCont;s
>   [GDBserver] Support vCont s and S actions with software single step
>
>  gdb/gdbserver/linux-low.c | 37 ++++++++++++++++++++++++++++++++-----
>  gdb/gdbserver/mem-break.c | 29 ++++++++++++++++++++++++++---
>  gdb/gdbserver/mem-break.h | 13 +++++++++----
>  gdb/gdbserver/server.c    | 13 ++++++++-----
>  4 files changed, 75 insertions(+), 17 deletions(-)