On Friday 14 November 2014 06:12 AM, David Gibson wrote: > On Thu, Nov 13, 2014 at 08:06:55PM +0530, Aravinda Prasad wrote: >> >> >> On Thursday 13 November 2014 06:14 PM, David Gibson wrote: >>> On Thu, Nov 13, 2014 at 05:18:16PM +0530, Aravinda Prasad wrote: >>>> On Thursday 13 November 2014 04:02 PM, David Gibson wrote: >>>>> On Thu, Nov 13, 2014 at 11:28:30AM +0530, Aravinda Prasad wrote: >>> [snip] >>>>>>>>> Having to retry the hcall from here seems very awkward. This is a >>>>>>>>> private hcall, so you can define it to do whatever retries are >>>>>>>>> necessary internally (and I don't think your current implementation >>>>>>>>> can fail anyway). >>>>>>>> >>>>>>>> Retrying is required in the cases when multi-processors experience >>>>>>>> machine check at or about the same time. As per PAPR, subsequent >>>>>>>> processors should serialize and wait for the first processor to issue >>>>>>>> the ibm,nmi-interlock call. The second processor retries if the first >>>>>>>> processor which received a machine check is still reading the error log >>>>>>>> and is yet to issue ibm,nmi-interlock call. >>>>>>> >>>>>>> Hmm.. ok. But I don't see any mechanism in the patches by which >>>>>>> H_REPORT_MC_ERR will report failure if another CPU has an MC in >>>>>>> progress. >>>>>> >>>>>> h_report_mc_err returns 0 if another VCPU is processing machine check >>>>>> and in that case we retry. h_report_mc_err returns error log address if >>>>>> no other VCPU is processing machine check. >>>>> >>>>> Uh.. how? I'm only seeing one return statement in the implementation >>>>> in 3/4. >>>> >>>> This part is in 4/4 which handles ibm,nmi-interlock call in >>>> h_report_mc_err() >>>> >>>> + if (mc_in_progress == 1) { >>>> + return 0; >>>> + } >>> >>> Ah, right, missed the change to h_report_mc_err() in the later patch. >>> >>>>>>>> Retrying cannot be done internally in h_report_mc_err hcall: only one >>>>>>>> thread can succeed entering qemu upon parallel hcall and hence retrying >>>>>>>> inside the hcall will not allow the ibm,nmi-interlock from first CPU to >>>>>>>> succeed. >>>>>>> >>>>>>> It's possible, but would require some fiddling inside the h_call to >>>>>>> unlock and wait for the other CPUs to finish, so yes, it might be more >>>>>>> trouble than it's worth. >>>>>>> >>>>>>>> >>>>>>>>> >>>>>>>>>> + mtsprg 2,4 >>>>>>>>> >>>>>>>>> Um.. doesn't this clobber the value of r3 you saved in SPRG2 just >>>>>>>>> above. >>>>>>>> >>>>>>>> The r3 saved in SPRG2 is moved to rtas area in the private hcall and >>>>>>>> hence it is fine to clobber r3 here >>>>>>> >>>>>>> Ok, if you're going to do some magic register saving inside the HCALL, >>>>>>> why not do the SRR[01] and CR restoration inside there as well. >>>>>> >>>>>> SRR0/1 is clobbered while returning from HCALL and hence cannot be >>>>>> restored in HCALL. For CR, we need to do the restoration here as we >>>>>> clobber CR after returning from HCALL (the instruction checking the >>>>>> return value of hcall clobbers CR). >>>>> >>>>> Hrm. AFAICT SRR0/1 shouldn't be clobbered when returning from an >>>> >>>> As hcall is an interrupt, SRR0 is set to nip and SRR1 to msr just before >>>> executing rfid. >>> >>> AFAICT the return path from the hypervisor - including for hcalls - >>> uses HSSR0/1 and hrfid, so ordinary SRR0/SRR1 should be ok. >> >> I see SRR0 and SRR1 clobbered when the HCALL from guest returns. >> Previous discussions on this is in the link below: >> >> http://lists.nongnu.org/archive/html/qemu-devel/2014-09/msg01148.html > > Hrm. Well, I guess if it happened it happened, but Alex's explanation > for why doesn't make sense to me. > > Did you execute cpu_synchronize_state() *before* attempting to set > SRR0/1 in the hcall?
Yes I did. > >> Further I searched QEMU source code but could not find whether it is >> using rfid/hrfid. However, ISA for sc instruction mentions that SRR0 and >> SRR1 are modified. > > Well of course it isn't in the qemu source, the low-level return to > guest is within the host kernel, specifically fast_guest_return in > arch/powerpc/kvm/book3s_hv_rmhandlers.S which uses hrfid. > > If I'm reading the ISA correctly then yes, SRR0/1 are clobbered on > entry, but that's on *entry* so can be overwritten by the hcall > handler itself. Hmm.. ok. I need to take a look into it in detail. > > -- Regards, Aravinda