Re: [PATCH] gdb.trace: Remove struct tracepoint_action_ops.

["Marcin Kościelnicki" <koriakin@0x04.net>]

On 25/01/16 12:53, Pedro Alves wrote:
> On 01/23/2016 07:31 PM, Marcin Kościelnicki wrote:
>> The struct tracepoint_action has an ops field, pointing to
>> a tracepoint_action_ops structure, containing send and download ops.
>> However, this field is only present when compiled in gdbserver, and not
>> when compiled in IPA.  When gdbserver is downloading tracepoint actions
>> to IPA, it skips offsetof(struct tracepoint_action, type) bytes from
>> its struct tracepoint_action, to get to the part that corresponds to
>> IPA's struct tracepoint_action.
>>
>> Unfortunately, this fails badly on ILP32 platforms where alignof(long long)
>> == 8.  Consider struct collect_memory_action layout in gdbserver:
>>
>> 0-3: base.ops
>> 4: base.type
>> 8-15: addr
>> 16-23: len
>> 24-27: basereg
>> sizeof == 32
>>
>> and its layout in IPA:
>>
>> 0: base.type
>> 8-15: addr
>> 16-23: len
>> 24-27: basereg
>> sizeof == 32
>>
>> When gdbserver tries to download it to IPA, it skips the first 4 bytes
>> (base.ops), figuring the rest will match what IPA expects - which is
>> not true, since addr is aligned to 8 bytes and will be at a different
>> relative position to base.type.
>>
>> The problem went unnoticed on the currently supported platforms, since
>> aarch64 and x86_64 have ops aligned to 8 bytes, and i386 has only 4-byte
>> alignment for long long.
>>
>> There are a few possible ways around this problem.  I decided on removing
>> ops altogether, since they can be easily inlined in their (only) places
>> of use - in fact allowing us share the code between 'L' and 'R'.  Any
>> approach where struct tracepoint_action is different between IPA and
>> gdbserver is just asking for trouble.
>>
>> Found on s390.  Tested on x86_64, s390, s390x.
>
> Hmm, this is essentially the same as:
>
>   https://sourceware.org/ml/gdb-patches/2015-03/msg00995.html
>
> Right?
>
> Seems that other patch inlines things a bit less though, which offhand
> looks preferable.  WDYT?
>
> Not sure what happened to that series.  I thought most of it (if not all)
> had been approved already.
>
> Thanks,
> Pedro Alves
>

Huh, I didn't know about that patch series.  Good to know, since I was 
going to try doing ppc tracepoints next, and had no idea that has 
already been attempted.  What happened to that patchset/author?  Kind of 
strange to abandon mostly-finished code when there's a $3k bounty on it.

The other patch looks fine too, I have no preference here.

Marcin Kościelnicki