On Thu, Mar 12, 2015 at 12:59 PM, Dr. David Alan Gilbert <dgilbert@xxxxxxxxxx> wrote: > * Andrey Korolyov (andrey@xxxxxxx) wrote: >> On Wed, Mar 11, 2015 at 10:59 PM, Dr. David Alan Gilbert >> <dgilbert@xxxxxxxxxx> wrote: >> > * Andrey Korolyov (andrey@xxxxxxx) wrote: >> >> On Wed, Mar 11, 2015 at 10:33 PM, Dr. David Alan Gilbert >> >> <dgilbert@xxxxxxxxxx> wrote: >> >> > * Kevin O'Connor (kevin@xxxxxxxxxxxx) wrote: >> >> >> On Wed, Mar 11, 2015 at 02:45:31PM -0400, Kevin O'Connor wrote: >> >> >> > On Wed, Mar 11, 2015 at 02:40:39PM -0400, Kevin O'Connor wrote: >> >> >> > > For what it's worth, I can't seem to trigger the problem if I move the >> >> >> > > cmos read above the SIPI/LAPIC code (see patch below). >> >> >> > >> >> >> > Ugh! >> >> >> > >> >> >> > That's a seabios bug. Main processor modifies the rtc index >> >> >> > (rtc_read()) while APs try to clear the NMI bit by modifying the rtc >> >> >> > index (romlayout.S:transition32). >> >> >> > >> >> >> > I'll put together a fix. >> >> >> >> >> >> The seabios patch below resolves the issue for me. >> >> > >> >> > Thanks! Looks good here. >> >> > >> >> > Andrey, Paolo, Bandan: Does it fix it for you as well? >> >> > >> >> >> >> Thanks Kevin, Dave, >> >> >> >> I`m afraid that I`m hitting something different not only because >> >> different suberror code but also because of mine version of seabios - >> >> I am using 1.7.5 and corresponding code in the proposed patch looks >> >> different - there is no smp-related code patch is about of. Those >> >> mentioned devices went to production successfully and I`m afraid I >> >> cannot afford playing on them anymore, even if I re-trigger the issue >> >> with patched 1.8.1-rc, there is no way to switch to a different kernel >> >> and retest due to specific conditions of this production suite. I`ve >> >> ordered a pair of new shoes^W 2620v2-s which should arrive to me next >> > >> > Well I was testing on a pair of 'E5-2620 v2'; but as you saw my test case >> > was pretty simple. If you can suggest any flags I should add etc to the >> > test I'd be happy to give it a go. >> > >> > Dave >> >> Here is mine launch string: >> >> qemu-system-x86_64 -enable-kvm -name vmtest -S -machine >> pc-i440fx-2.1,accel=kvm,usb=off -cpu SandyBridge,+kvm_pv_eoi -m 512 >> -realtime mlock=off -smp 12,sockets=1,cores=12,threads=12 -numa >> node,nodeid=0,cpus=0-11,mem=512 -nographic -no-user-config -nodefaults >> -device sga -rtc base=utc,driftfix=slew -global >> kvm-pit.lost_tick_policy=discard -no-hpet -no-shutdown -global >> PIIX4_PM.disable_s3=1 -global PIIX4_PM.disable_s4=1 -boot strict=on >> -device nec-usb-xhci,id=usb,bus=pci.0,addr=0x4 -device >> virtio-serial-pci,id=virtio-serial0,bus=pci.0,addr=0x5 -m >> 512,slots=31,maxmem=16384M -object >> memory-backend-ram,id=mem0,size=512M -device >> pc-dimm,id=dimm0,node=0,memdev=mem0 >> >> I omitted disk backend in this example, but there is a chance that my >> problem is not reproducible without some calls made explicitly by a >> bootloader (not sure what to say for mid-runtime failures). > > It seems to survive OK: Thanks David, I`ll go through test sequence and report. Unfortunately my orchestration does not have even a hundred millisecond precision for libvirt events, so I can`t tell if the immediate start-up failures happened before bootloader execution or during it, all I have for those is a less-than-two-second interval between actual pass of a launch command and paused state event. QEMU logging also does not give me timestamps for an emulation errors even with appropriate timestamp arg. > > while true; do (sleep 1; echo -e '\001cc\n'; sleep 5; echo -e 'q\n')|/opt/qemu-try-world3/bin/qemu-system-x86_64 -enable-kvm -name vmtest -S -machine pc-i440fx-2.1,accel=kvm,usb=off -cpu SandyBridge,+kvm_pv_eoi -m 512 -realtime mlock=off -smp 12,sockets=1,cores=12,threads=12 -numa node,nodeid=0,cpus=0-11,mem=512 -nographic -no-user-config -device sga -rtc base=utc,driftfix=slew -global kvm-pit.lost_tick_policy=discard -no-hpet -no-shutdown -global PIIX4_PM.disable_s3=1 -global PIIX4_PM.disable_s4=1 -boot strict=on -device nec-usb-xhci,id=usb,bus=pci.0,addr=0x4 -device virtio-serial-pci,id=virtio-serial0,bus=pci.0,addr=0x5 -m 512,slots=31,maxmem=16384M -object memory-backend-ram,id=mem0,size=512M -device pc-dimm,id=dimm0,node=0,memdev=mem0 ~/pi.vfd 2>&1 | tee /tmp/qemu.op; grep "internal error" /tmp/qemu.op -q && break; done > > Dave > >> >> > >> >> Monday, so I`ll be able to test a) against 1.8.0-release, b) against >> >> patched bios code, c) reproduce initial error on master/3.19 (may be >> >> I`ll take them before weekend by going into this computer shop in >> >> person). Until then, I have a very deep feeling that mine issue is not >> >> there :) Also I became very curious on how a lack of IDT feature may >> >> completely eliminate the issue appearance for me, the only possible >> >> explanation is a clock-related race which is kinda stupid suggestion >> >> and unlikely to exist in nature. >> >> >> >> Thanks again for everyone for throughout testing and ideas! >> >> >> >> > >> >> >> -Kevin >> >> >> >> >> >> >> >> >> --- a/src/romlayout.S >> >> >> +++ b/src/romlayout.S >> >> >> @@ -22,7 +22,8 @@ >> >> >> // %edx = return location (in 32bit mode) >> >> >> // Clobbers: ecx, flags, segment registers, cr0, idt/gdt >> >> >> DECLFUNC transition32 >> >> >> -transition32_for_smi: >> >> >> +transition32_nmi_off: >> >> >> + // transition32 when NMI and A20 are already initialized >> >> >> movl %eax, %ecx >> >> >> jmp 1f >> >> >> transition32: >> >> >> @@ -205,7 +206,7 @@ __farcall16: >> >> >> entry_smi: >> >> >> // Transition to 32bit mode. >> >> >> movl $1f + BUILD_BIOS_ADDR, %edx >> >> >> - jmp transition32_for_smi >> >> >> + jmp transition32_nmi_off >> >> >> .code32 >> >> >> 1: movl $BUILD_SMM_ADDR + 0x8000, %esp >> >> >> calll _cfunc32flat_handle_smi - BUILD_BIOS_ADDR >> >> >> @@ -216,8 +217,10 @@ entry_smi: >> >> >> DECLFUNC entry_smp >> >> >> entry_smp: >> >> >> // Transition to 32bit mode. >> >> >> + cli >> >> >> + cld >> >> >> movl $2f + BUILD_BIOS_ADDR, %edx >> >> >> - jmp transition32 >> >> >> + jmp transition32_nmi_off >> >> >> .code32 >> >> >> // Acquire lock and take ownership of shared stack >> >> >> 1: rep ; nop >> >> > -- >> >> > Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK >> > -- >> > Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK > -- > Dr. David Alan Gilbert / dgilbert@xxxxxxxxxx / Manchester, UK -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html