Re: [rfc, ping] Remote "info proc" and core file generation

[Pedro Alves <alves dot ped at gmail dot com>]

On 01/05/2012 03:17 PM, Ulrich Weigand wrote:
> Hello,
>
> given the problems with my latest attempt to access /proc remotely via
> generic file access routines documented here:
> http://sourceware.org/ml/gdb-patches/2011-12/msg00782.html
>
> I would like to go back to my earlier approach using TARGET_INFO_PROC:
> http://sourceware.org/ml/gdb-patches/2011-12/msg00007.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00008.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00009.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00010.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00011.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00014.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00015.html
> http://sourceware.org/ml/gdb-patches/2011-12/msg00016.html
>
> In the meantime, I've got approval for the doc and bfd parts, and
> Joel has regression-tested the patches on a procfs target (Irix).
>
> So the only thing that stops this patch series from going in as-is
> is consensus that TARGET_INFO_PROC is the right abstraction level.
>
> Given the experiments I did in the meantime (see above), I'd now
> argue that this *is* the proper level of abstraction:
>
> - TARGET_INFO_PROC allows the *contents* of Linux /proc files to
>    be passed through unchanged, so we don't have to define our own
>    formats (and keep updating them) -- the one drawback is that the
>    contents are obviously Linux-specific, but that's OK as long as
>    the target objects are only used in linux-tdep code.
>
> - At the same time, *access* to those contents is abstracted.  This
>    means we do *not* have to know exactly where on the target the
>    /proc files are found: e.g. in the classic remote target, the GDB
>    host side does not even know the PID of the inferior process on
>    the target.  (Another possibility might be a Linux kernel remote
>    target that operates via hardware debugging or in-kernel debugging
>    and still provides access to Linux processes: such remote stubs
>    could also implement TARGET_INFO_PROC, even if they may not
>    provide general access to the file system.)
>
> Pedro, you had been raising concerns about this initially.  Did you
> have a chance to look at the discussion refered to at the top of
> this mail?

I've replied now.  Sorry for the delay...

> Do you still feel that TARGET_INFO_PROC is inappropiate?

I still do.  :-(

- there's the issue I raised about needing to cache the object across
  the whole transfer, lest the file disappears
  of changes behind your feet.  Easy to fix, though.

- I don't see the advantage over separate target objects for each
  proc/... subtype.  We already have a mechanism to report back
  some object is not supports (simply don't support
  the qXfer:object:read packet).  Why use the annex instead?
  An e.g., getting at the current process'es executable is something
  that would be useful for target_pid_to_exec_file, for attach.
  If we're having a specific packet for that, shouldn't it be a "top-level"
  packet, rather than buried in TARGET_INFO_PROC?

- if GDB already needs to know what target it is talking to (for the
  gdbarch methods), then I don't see what gain do we have from half
  an abstraction -- this is what leads me to consider instead reading
  from the target filesystem.