Re: [PATCH] Only construct ucontext for SA_SIGINFO signal handlers

[Jon TURNEY <jon.turney@dronecode.org.uk>]

On 03/04/2015 15:08, Corinna Vinschen wrote:
> On Apr  3 13:51, Jon TURNEY wrote:
>> On 03/04/2015 13:17, Corinna Vinschen wrote:
>>> On Apr  3 13:18, Corinna Vinschen wrote:
>>>> On Apr  2 20:30, Jon TURNEY wrote:
>>>>
>>>>>         sigset_t this_oldmask = set_process_mask_delta ();
>>>>> -      thiscontext.uc_sigmask = this_oldmask;
>>>>> +      context.uc_sigmask = this_oldmask;
>>>>           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>>>>
>>>> This call to set_process_mask_delta() should occur before constructing
>>>> the context, so that filling in uc_sigmask can be moved into the above
>>>> `'if' branch.
>>
>> Ok, I will move it.
>>
>>>> On second thought, isn't this slightly wrong anyway?  Shouldn't that be
>>>>
>>>>           context.uc_sigmask = _my_tls.sigmask;
>>>> 	 context.uc_mcontext.oldmask = this_oldmask;
>>
>> As I wrote elsewhere:  You'll have to help me understand what the difference
>> in meaning between ucontext_t.uc_sigmask and ucontext_t.uc_mcontext.oldmask
>> is.
>>
>> I don't see how the value of _my_tls.sigmask has any meaning at that point
>> in the code.
>
> Ok, I had a look into the Linux source and searched the web, and here's
> the problem.
>
> One is that sigset_t on Linux is not just a 32 or 64 bit bitmask anymore,
> but an array of ulong's used as a rather big sigmask.
>
> OTOH, mcontext_t::oldmask is only the size of "unsigned long".  In fact,
> as it turns out by inspecting the Linux kernel, oldmask is nothing else
> than the first bits of uc_sigmask which fit into an unsigned long.  And
> in the net I found that oldmask is just the old representation of
> sigset_t, before the Linux kernel allowed more signals than fit into
> a bitmask of unsigned long size.  In fact, it's only for backward compat,
> but unused these days.
>
> Given that, setting context.uc_sigmask to this_oldmask is apparently
> the right thing to do.  For emulating backward compat (which we don't
> need, but it also doesn't hurt), we could set oldmask to the same
> value:
>
>    context.uc_sigmask = context.uc_mcontext.oldmask = this_oldmask;
>

Thank you very much for researching this.  I tried but wasn't able to 
discover anything much.

What you suggest seems right, so I'll make an updated patch including that.