Linux: Access memory even if threads are running

Linux: Access memory even if threads are running

Currently, on GNU/Linux, if you try to access memory and you have a
running thread selected, GDB fails the memory accesses, like:

 (gdb) c&
 Continuing.
 (gdb) p global_var
 Cannot access memory at address 0x555555558010

Or:

 (gdb) b main
 Breakpoint 2 at 0x55555555524d: file access-mem-running.c, line 59.
 Warning:
 Cannot insert breakpoint 2.
 Cannot access memory at address 0x55555555524d

This patch removes this limitation.  It teaches the native Linux
target to read/write memory even if the target is running.  And it
does this without temporarily stopping threads.  We now get:

 (gdb) c&
 Continuing.
 (gdb) p global_var
 $1 = 123
 (gdb) b main
 Breakpoint 2 at 0x555555555259: file access-mem-running.c, line 62.

(The scenarios above work correctly with current GDBserver, because
GDBserver temporarily stops all threads in the process whenever GDB
wants to access memory (see prepare_to_access_memory /
done_accessing_memory).  Freezing the whole process makes sense when
we need to be sure that we have a consistent view of memory and don't
race with the inferior changing it at the same time as GDB is
accessing it.  But I think that's a too-heavy hammer for the default
behavior.  I think that ideally, whether to stop all threads or not
should be policy decided by gdb core, probably best implemented by
exposing something like gdbserver's prepare_to_access_memory /
done_accessing_memory to gdb core.)

Currently, if we're accessing (reading/writing) just a few bytes, then
the Linux native backend does not try accessing memory via
/proc/<pid>/mem and goes straight to ptrace
PTRACE_PEEKTEXT/PTRACE_POKETEXT.  However, ptrace always fails when
the ptracee is running.  So the first step is to prefer
/proc/<pid>/mem even for small accesses.  Without further changes
however, that may cause a performance regression, due to constantly
opening and closing /proc/<pid>/mem for each memory access.  So the
next step is to keep the /proc/<pid>/mem file open across memory
accesses.  If we have this, then it doesn't make sense anymore to even
have the ptrace fallback, so the patch disables it.

I've made it such that GDB only ever has one /proc/<pid>/mem file open
at any time.  As long as a memory access hits the same inferior
process as the previous access, then we reuse the previously open
file.  If however, we access memory of a different process, then we
close the previous file and open a new one for the new process.

If we wanted, we could keep one /proc/<pid>/mem file open per
inferior, and never close them (unless the inferior exits or execs).
However, having seen bfd patches recently about hitting too many open
file descriptors, I kept the logic to have only one file open tops.
Also, we need to handle memory accesses for processes for which we
don't have an inferior object, for when we need to detach a
fork-child, and we'd probaly want to handle caching the open file for
that scenario (no inferior for process) too, which would probably end
up meaning caching for last non-inferior process, which is very much
what I'm proposing anyhow.  So always having one file open likely ends
up a smaller patch.

The next step is handling the case of GDB reading/writing memory
through a thread that is running and exits.  The access should not
result in a user-visible failure if the inferior/process is still
alive.

Once we manage to open a /proc/<lwpid>/mem file, then that file is
usable for memory accesses even if the corresponding lwp exits and is
reaped.  I double checked that trying to open the same
/proc/<lwpid>/mem path again fails because the lwp is really gone so
there's no /proc/<lwpid>/ entry on the filesystem anymore, but the
previously open file remains usable.  It's only when the whole process
execs that we need to reopen a new file.

When the kernel destroys the whole address space, i.e., when the
process exits or execs, the reads/writes fail with 0 aka EOF, in which
case there's nothing else to do than returning a memory access
failure.  Note this means that when we get an exec event, we need to
reopen the file, to access the process's new address space.

If we need to open (or reopen) the /proc/<pid>/mem file, and the LWP
we're opening it for exits before we open it and before we reap the
LWP (i.e., the LWP is zombie), the open fails with EACCES.  The patch
handles this by just looking for another thread until it finds one
that we can open a /proc/<pid>/mem successfully for.

If we need to open (or reopen) the /proc/<pid>/mem file, and the LWP
we're opening has exited and we already reaped it, which is the case
if the selected thread is in THREAD_EXIT state, the open fails with
ENOENT.  The patch handles this the same way as a zombie race
(EACCES), instead of checking upfront whether we're accessing a
known-exited thread, because that would result in more complicated
code, because we also need to handle accessing lwps that are not
listed in the core thread list, and it's the core thread list that
records the THREAD_EXIT state.

The patch includes two testcases:

#1 - gdb.base/access-mem-running.exp

  This is the conceptually simplest - it is single-threaded, and has
  GDB read and write memory while the program is running.  It also
  tests setting a breakpoint while the program is running, and checks
  that the breakpoint is hit immediately.

#2 - gdb.threads/access-mem-running-thread-exit.exp

  This one is more elaborate, as it continuously spawns short-lived
  threads in order to exercise accessing memory just while threads are
  exiting.  It also spawns two different processes and alternates
  accessing memory between the two processes to exercise the reopening
  the /proc file frequently.  This also ends up exercising GDB reading
  from an exited thread frequently.  I confirmed by putting abort()
  calls in the EACCES/ENOENT paths added by the patch that we do hit
  all of them frequently with the testcase.  It also exits the
  process's main thread (i.e., the main thread becomes zombie), to
  make sure accessing memory in such a corner-case scenario works now
  and in the future.

The tests fail on GNU/Linux native before the code changes, and pass
after.  They pass against current GDBserver, again because GDBserver
supports memory access even if all threads are running, by
transparently pausing the whole process.

gdb/ChangeLog:
yyyy-mm-dd  Pedro Alves  <pedro@palves.net>

	PR mi/15729
	PR gdb/13463
	* linux-nat.c (linux_nat_target::detach): Close the
	/proc/<pid>/mem file if it was open for this process.
	(linux_handle_extended_wait) <PTRACE_EVENT_EXEC>: Close the
	/proc/<pid>/mem file if it was open for this process.
	(linux_nat_target::mourn_inferior): Close the /proc/<pid>/mem file
	if it was open for this process.
	(linux_nat_target::xfer_partial): Adjust.  Do not fall back to
	inf_ptrace_target::xfer_partial for memory accesses.
	(last_proc_mem_file): New.
	(maybe_close_proc_mem_file): New.
	(linux_proc_xfer_memory_partial_pid): New, with bits factored out
	from linux_proc_xfer_partial.
	(linux_proc_xfer_partial): Delete.
	(linux_proc_xfer_memory_partial): New.

gdb/testsuite/ChangeLog
yyyy-mm-dd  Pedro Alves  <pedro@palves.net>

	PR mi/15729
	PR gdb/13463
	* gdb.base/access-mem-running.c: New.
	* gdb.base/access-mem-running.exp: New.
	* gdb.threads/access-mem-running-thread-exit.c: New.
	* gdb.threads/access-mem-running-thread-exit.exp: New.

Change-Id: Ib3c082528872662a3fc0ca9b31c34d4876c874c9