----- On Jan 12, 2022, at 4:00 PM, Florian Weimer fw@xxxxxxxxxxxxx wrote: > * Mathieu Desnoyers: > >> ----- On Jan 12, 2022, at 11:00 AM, Florian Weimer fw@xxxxxxxxxxxxx wrote: >> >>> * Peter Zijlstra: >>> >>>> On Wed, Jan 12, 2022 at 04:16:36PM +0100, Florian Weimer wrote: >>>> >>>>> You could perhaps push a signal frame onto the stack. It's going to >>>>> be expensive, but it's already in the context switch path, so maybe it >>>>> does not matter. >>>> >>>> Please no! Signals are a trainwreck that need change (see the whole >>>> AVX-512 / AMX saga), we shouldn't use more of that just cause. >>> >>> If it's a signal, it should be modeled as such. I think it's pretty >>> close to a synchronous signal. > > (an asynchronous signal) > >> Florian, just to validate here: is your argument about AVX-512/AMX or about >> rseq abort-at-ip ? > > rseq abort-at-ip. I wonder if it is possible to use regular stack > unwinding (through the signal frame) to figure out where the abort > happened, and use the existing cleanup handler functionality in GCC. > (Although -fnon-call-exceptions is not quite up to this, but in theory > we would have to fix this for POSIX asynchronous cancellation/Ada > asynchronous transfer of control support anyway.) OK, so if we take x86-64 as an example, the abort would do: /* Skip redzone and sigframe */ regs->sp -= 128 + sizeof(struct rt_sigframe); It's unclear to me what should be the size of rt_sigframe here. Should it include floating point state as well ? Then at regs->sp, we can store a signal frame. The interesting bits would go into (struct rt_sigframe __user *)->uc which is a struct sigcontext. It would contain all register state at the point where the critical section was aborted. The only register state we really care about is sigcontext "rip", given that the userspace abort handler can save all other relevant registers by itself. Then should we populate the other struct rt_sigframe fields or just zero them ? I wonder what the meaning of "pretcode" and "struct siginfo" would be in the context of an rseq abort ? Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
