Re: [PATCH 0/8] gdb/s390: Add regular and fast tracepoint support.

[Marcin KoÅcielnicki <koriakin at 0x04 dot net>]

A few updates on the status...

On 24/01/16 13:12, Marcin KoÅcielnicki wrote:
> This patchset adds support for regular and fast tracepoints on
> s390-ibm-linux-gnu and s390x-ibm-linux-gnu.  It depends on the following
> yet-unlanded patches:
>
> - https://sourceware.org/ml/gdb-patches/2016-01/msg00597.html (for 31-bit
>    fast tracepoint support)
> - https://sourceware.org/ml/gdb-patches/2016-01/msg00596.html (just
>    a testsuite fix)

These two have landed by now.

> - https://sourceware.org/ml/gdb-patches/2016-01/msg00510.html (fixes
>    setting a breakpoint and a tracepoint on the same location)

Still waiting for review.
> Patches 1-5 add working regular tracepoint support, patches 6-7 add
> working fast tracepoint support on top of that, and patch 8 makes fast
> tracepoints use compiled agent expressions.

I have updated the first 5 patches and hope to get them pushed first; 
there are outstanding issues on #6 and #8 (doc comments, indents, 
reading auxv from /proc) which I will fix soon (when I'm done with issue 
(3) below).
> Patches 1 and 7 are in target-independent code, the remaining ones
> are in s390-specific code.
>
> It has been tested on s390-ibm-linux-gnu and s390x-ibm-linux-gnu on
> a z13 machine (without vector extensions).  The parts handling vector
> extensions are untested.  It has also been regression-tested on x86_64.
>
> There are 2 test failures still left in gdb.trace:
>
> - gdb.trace/unavailable.exp: print derived_whole has 4 failures on 64-bit
>    only.  The same failure happens on x86_64, so it's unlikely to be
>    s390-specific.
> - gdb.trace/mi-tsv-changed.exp: create delete modify: tvariable $tvar3
>    modified is an intermittent failure.  It seems to be a race of some
>    kind - in the failing runs, gdb shows the same messages, but in
>    different order.  I strongly suspect this is a target-independent
>    issues that only happens on s390 due to timing factors.
>
> There are also three issues affecting s390 tracepoints that don't show
> up in the testsuite:
>
> 1. Target independent: tfile format doesn't contain target information
>     (tdesc).  While this affects all platforms with multiple tdescs
>     (eg. x86_64 is unable to pull AVX registers from tfile for that
>     reason), it horribly breaks 31-bit s390 with high GPRs - it has
>     a completely different GPR layout from plain 31-bit s390, so the
>     collected registers will be garbled.  I suppose the proper fix to
>     that would be to add tdesc information to tfile format.  Unfortunately,
>     I don't see a way to extend it in a backwards-compatible way.

This is fixed already.

> 2. Target independent: 32-bit (or 31-bit for us) IPA cannot be used with
>     64-bit gdbserver, due to communication involving lots of structs with
>     pointer types.  Fixing that would be quite involved, but possible
>     (I don't suppose we have to maintain compatibility between IPA/gdbserver
>     from different gdb versions?).
>
> 3. s390 specific: 31-bit gdbserver doesn't know about high GPRs, and
>     cannot collect them if they're in use.  Seems fixable with average
>     effort.  Unfortunately, fixing that will break tfile, unless (1)
>     is fixed first...

Since (1) is fixed, I'm going to work on that one now.

> These three interact in bad ways, summarised below:
>
> - 64-bit linux, 64-bit gdbserver, 64-bit target: works OK, but you won't be
>    able to see VX registers in tfile.

Works OK now, including VX in tfile.
> - 64-bit linux, 64-bit gdbserver, 31-bit target:
>
>    - no fast tracepoint support due to (2)
>    - tfile completely broken due to (1)

tfile should work now.
> - 64-bit linux, 31-bit gdbserver, 31-bit target:
>
>    - works OK, but you won't be able to see VX registers or high GPRs at all,
>      due to (3)
>    - if (3) were fixed, tfile will be completely broken due to (1)

Second line no longer applies.
> - 31-bit linux, 31-bit gdbserver, 31-bit target: works OK
>
> In summary, there's no way at all to use fast tracepoints on 31-bit target
> if what you want to collect involves high GPRs.  While there's support for
> it in the IPA, it's currently disabled, to match what gdbserver supports.
>
> Fast tracepoint support assumes z900+ CPU.  The g5/g6 CPUs only have jump
> instuctions with +-64kiB of range, making them pretty much useless for
> our purposes, so not much loss here.