gdb/linux-nat: Fix mem access ptrace fallback (PR threads/31579)

gdb/linux-nat: Fix mem access ptrace fallback (PR threads/31579)

Old RHEL systems have a kernel that does not support writing memory
via /proc/pid/mem.  On such systems, we fallback to accessing memory
via ptrace.  That has a few downsides described in the "Accessing
inferior memory" section at the top of linux-nat.c.

The target_xfer interface for memory access uses inferior_ptid as
sideband argument to indicate which process to access.  Memory access
is process-wide, it is not thread-specific, so inferior_ptid is
sometimes pointed at a process-wide ptid_t for the memory access
(i.e., a ptid that looks like {pid, 0, 0}).  That is the case for any
code that uses scoped_restore_current_inferior_for_memory, for
example.

That is what causes the issue described in PR 31579, where thread_db
calls into the debugger to read memory, which reaches our
ps_xfer_memory function, which does:

  static ps_err_e
  ps_xfer_memory (const struct ps_prochandle *ph, psaddr_t addr,
		  gdb_byte *buf, size_t len, int write)
  {
    scoped_restore_current_inferior_for_memory save_inferior (ph->thread->inf);

    ...
      ret = target_read_memory (core_addr, buf, len);
    ...
  }

If linux_nat_target::xfer_partial falls back to inf_ptrace_target with
a pid-ptid, then the ptrace code will do the ptrace call targeting
pid, the leader LWP.  That may fail with ESRCH if the leader is
currently running, or zombie.  That is the case in the scenario in
question, because thread_db is consulted for an event of a non-leader
thread, before we've stopped the whole process.

Fix this by having the ptrace fallback code try to find a stopped LWP
to use with ptrace.

I chose to handle this in the linux-nat target instead of in common
code because (global) memory is a process-wide property, and this
avoids having to teach all the code paths that use
scoped_restore_current_inferior_for_memory to find some stopped thread
to access memory through, which is a ptrace quirk.  That is
effectively what we used to do before we started relying on writable
/proc/pid/mem.  I'd rather not go back there.

To trigger this on modern kernels you have to hack linux-nat.c to
force the ptrace fallback code, like so:

 --- a/gdb/linux-nat.c
 +++ b/gdb/linux-nat.c
 @@ -3921,7 +3921,7 @@ linux_nat_target::xfer_partial (enum target_object object,
	  poke would incorrectly write memory to the post-exec address
	  space, while the core was trying to write to the pre-exec
	  address space.  */
 -      if (proc_mem_file_is_writable ())
 +      if (0 && proc_mem_file_is_writable ())

With that hack, I was able to confirm that the fix fixes hundreds of
testsuite failures.  Compared to a test run with pristine master, the
hack above + this commit's fix shows that some non-stop-related tests
fail, but that is expected, because those are tests that need to
access memory while the program is running.  (I made no effort to
temporarily pause an lwp if no ptrace-stopped lwp is found.)

Change-Id: I24a4f558e248aff7bc7c514a88c698f379f23180
Tested-By: Hannes Domani <ssbssa@yahoo.de>
Approved-By: Andrew Burgess <aburgess@redhat.com>