Re: [RFC 1/3] [gdb] Call gdbarch_get_syscall_number less often

[Tom de Vries <tdevries@suse.de>]

On 11/21/23 23:17, John Baldwin wrote:
> On 11/21/23 1:19 PM, Simon Marchi wrote:
>> On 11/21/23 13:03, John Baldwin wrote:
>>> I think you can use linux_nat_target::set_syscall_catchpoint to do the
>>> invalidation.  From my reading it is called each time a syscall 
>>> catchpoint
>>> is added or removed, and needed is set to false if there are no more
>>> catchpoints (so syscall catches are disabled).   I think that means
>>> you could do something like:
>>>
>>> linux_nat_target::set_syscall_catchpoint ()
>>> {
>>>    if (!needed)
>>>      iterate_over_lwps (ptid_t (pid), [] (struct lwp_info *lp)
>>>                         {
>>>                           lp->syscall_number = -1;
>>>                         });
>>> }
>>
>> What about this corner case of a corner case:
>>
>>   1. you have a syscall catchpoint, you hit it (the syscall entry)
>>   2. you delete the syscall catchpoint
>>   3. you create a syscall catchpoint again
>>   4. you continue, expecting to hit the syscall exit
>>
>> Will step 2 cause set_syscall_catchpoint to be called, and will we
>> therefore lose the information that would have been useful at step 4?
> 
> Yes, but I think the generation counter idea is subject to the same issue
> unless in both cases you actually defer the geneation bump (or clearing
> syscall_number to -1) until you resume the lwp.  For that approach
> perhaps you could reset syscall_number to -1 in linux_resume_one_lwp_throw
> if catch_syscalls_enabled is zero?
> 

Thanks for the suggestion, but I went with a solution in 
linux_resume_one_lwp_throw ( 
https://sourceware.org/pipermail/gdb-patches/2023-November/204407.html ) .

AFAIU it should handle the corner case Simon mentions.

Thanks,
- Tom