This is the mail archive of the
mailing list for the GDB project.
[rfc] Options for "info mappings" etc. (Re: [PATCH] Implement new `info core mappings' command)
- From: "Ulrich Weigand" <uweigand at de dot ibm dot com>
- To: gdb-patches at sourceware dot org
- Cc: jan dot kratochvil at redhat dot com, pedro at codesourcery dot com, sergiodj at redhat dot com
- Date: Wed, 23 Nov 2011 17:32:08 +0100 (CET)
- Subject: [rfc] Options for "info mappings" etc. (Re: [PATCH] Implement new `info core mappings' command)
Sergio Durigan Junior <email@example.com> writes:
> > Ok, so I would like to ask for another round of review of this patch
> > then. I am aware of Ulrich's work towards a new `info mappings'
> > command, but since I've got no response from him yet, I decided to
> > continue my efforts here.
First of all, sorry for the late reply. I ran into some issues with
implementing the proposed "info mappings" and then got side-tracked ...
Anyway, here's my current thoughts on the issues and some options on
how to move forward.
Today we have the following set of commands:
This uses the target_find_memory_regions callback, which is implemented
on Linux native by reading /proc/.../maps. My pending patch would have
implemented this callback on remote as well, via a new packet type
qXfer:process-map:read, implemented on Linux gdbserver by reading
- info proc ... (in particular, info proc mappings)
This is implemented on Linux native target only, where it reads from
/proc directly. Note that:
* this reads a variety of other files beyond /proc/.../maps
* it can actually read from a completely different process
(info proc takes an optional PID argument)
(There is also a separate implementation of info proc for procfs
targets, i.e. Solaris, Irix, Tru64).
- info core ... (info core mappings and info core exe)
This would be implemented by Sergio's patch, and available for
core file targets only (any platform). It would support only
"exe" and "mappings", and only the current core file.
This state of affairs, even assuming both my and Sergio's pending
patch sets were applied, has a couple of drawbacks:
- there is still no "info proc mappings" like command for the remote
target, even if the remote side runs gdbserver on Linux
- there is still code duplication between the "info proc mappings"
and target_find_memory_regions implementations in linux-nat.c
To fix these, I had been thinking along the lines of implementing
the following set of features ("Option 1"):
a) Extend target_find_memory_regions to provide mapped file names
This requires extending the implementation in linux-nat.c, and
updating other existing implementations and users.
b) Implement target_find_memory_regions for remote targets
This would use xfer with a new object type TARGET_OBJECT_PROCESS_MAP,
implemented via a packet type qXfer:process-map:read providing a new
XML formatted <process-map>. Implementation on gdbserver in my patch.
c) Implement target_find_memory_regions in corefile.c
This would use techniques similar to Sergio's current patch.
d) Implement new "info mappings" in core GDB
This would be independent of existing "info proc" commands. It would
be implemented across all targets, and simply call into the (newly
extended) target_find_memory_regions to get its data.
This would fix the first of the two problems mentioned above, in that
we now also have a working "info mappings" with remote gdbserver targets.
However, we still have code duplication; in fact the duplication is now
even user-visible in the sense that we now have a generic "info mappings"
in addition to the Linux-specific "info proc mappings".
Another drawback is that we do not have anything like Sergio's proposed
"info core exe" command; nor do have anything like "info proc ..." for
any other the *other* commands except "mappings" for remote targets.
Finally, as a (very minor) drawback: in non-XML builds of GDB, the
"info mappings" command would not work with remote targets.
I have been thinking about ways to address these, and come up with one
that would basically export arbitrary /proc files via xfer ("Option 2"):
a) Implement TARGET_OBJECT_PROC xfer object.
This would use the name of the /proc file as "Annex" (e.g. "maps" /
"exe" / "cmdline" ...). On Linux native targets, this can be directly
implemented via reads from /proc. For remote targets, this would be
implemented via a new qXfer:proc:read packet which simply returns raw
contents of the requested /proc file (no XML format required). For
core file targets, we could synthesize Linux /proc/.../maps and
/proc/.../exe contents via something like Sergio's patch.
b) Implement gdbarch target_find_memory_regions fallback
This can be implemented on Linux targets (linux-tdep.c) via reading
TARGET_OBJECT_PROC annex "maps", which would then automatically work
for native, remote, and core file targets. The implementation in
linux-nat.c would then be superfluous.
c) Implement generic "info proc ..." command
This would call out to a gdbarch architecture-specific implementation.
(We'd need to take care that on procfs targets, we still use the
original implementation in procfs.c instead of the generic one.)
A Linux (linux-tdep.c) implementation of that callback would then use
TARGET_OBJECT_PROC xfers (and thus work native, remote, and core).
This would make the linux-nat.c implementation superfluous.
This fixes all the drawbacks mentioned with Option 1 above: there are
no new commands, no more code duplication, we support "info proc exe"
for core files, and we don't even require XML.
However, there are still problems with this option:
- "info proc mappings" and "info proc exe" would work on core files,
but only *Linux* core files -- at least until other targets implement
equivalent support. Sergio's "info core ..." would work anywhere.
- "info proc ..." would lose the possibility to query properties of
*other* processes except the current one (i.e. the "info proc PID"
variant could no longer be supported)
The first problem doesn't look really serious to me: if we wouldn't
support "info proc" on a native target, it doesn't seem important
to support it on a core file produced on that target (in particular
if the information we can synthesize is rather sparse anyway).
The second problem also may not be really serious any more: with
current GDB, the user could always use multi-inferior support to
(temporarily) attach to the process they want to see info about,
instead of specifying the PID in the command. However, this would
indeed reflect a UI change ...
There is a variant of Option 2 that would actually solve that latter
problem as well: we might encode a target PID in the TARGET_OBJECT_PROC
request, e.g. by using "PID/maps" instead of plain "maps" as the Annex.
(As minor variations, we could keep TARGET_OBJECT_PROC as-is and add
a new TARGET_OBJECT_REMOTE_PROC that takes the PID argument.)
The only drawback of this method seems to be that it would introduce
a somewhat "alien" remote protocol packet type: all the xfer commands
usually refer to the current inferior, not some random other process ...
I'd really appreciate thoughts / suggestions on whether the above
drawbacks are acceptable, and which option we should proceed with.
Dr. Ulrich Weigand
GNU Toolchain for Linux on System z and Cell BE