Re: [Xen-devel] PVHVM VCPU hotplug mechanism via ACPI hotplug doesn't work in Xen 4.[1, 2, 3]

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>]

On Tue, 7 May 2013, Konrad Rzeszutek Wilk wrote:
> On Tue, May 07, 2013 at 12:16:25AM +0000, Ross Philipson wrote:
> > > -----Original Message-----
> > > From: xen-devel-bounces@xxxxxxxxxxxxx [mailto:xen-devel-
> > > bounces@xxxxxxxxxxxxx] On Behalf Of Ross Philipson
> > > Sent: Monday, May 06, 2013 5:07 PM
> > > To: Konrad Rzeszutek Wilk; jinsong.liu@xxxxxxxxx; Stefano Stabellini;
> > > xen-devel@xxxxxxxxxxxxxxxxxxx
> > > Subject: Re: [Xen-devel] PVHVM VCPU hotplug mechanism via ACPI hotplug
> > > doesn't work in Xen 4.[1, 2, 3]
> > > 
> > > > -----Original Message-----
> > > > From: xen-devel-bounces@xxxxxxxxxxxxx [mailto:xen-devel-
> > > > bounces@xxxxxxxxxxxxx] On Behalf Of Konrad Rzeszutek Wilk
> > > > Sent: Monday, May 06, 2013 11:45 AM
> > > > To: jinsong.liu@xxxxxxxxx; Stefano Stabellini; xen-
> > > > devel@xxxxxxxxxxxxxxxxxxx
> > > > Subject: [Xen-devel] PVHVM VCPU hotplug mechanism via ACPI hotplug
> > > > doesn't work in Xen 4.[1, 2, 3]
> > > >
> > > > Which is probably the case b/c the Linux side implementation for
> > > > such simple operation as :
> > > >
> > > > echo 0 > /sys/devices/system/cpu/cpu1/online
> > > > echo 1 > /sys/devices/system/cpu/cpu1/online
> > > >
> > > > would have blown up so nobody tested it?
> > > >
> > > > Now that is fixed (v3.10 + http://lists.xen.org/archives/html/xen-
> > > > devel/2013-05/msg00503.html)
> > > > I tried to do 'xm vcpu-set latest X' for a PVHVM guest.
> > > >
> > > > The first iteration was using this simple guest config:
> > > > builder='hvm'
> > > > disk = [ 'file:/mnt/lab/latest/root_image.iso,hdc:cdrom,r']
> > > > memory = 2048
> > > > boot="d"
> > > > maxvcpus=4
> > > > vcpus=4
> > > > serial='pty'
> > > > vnclisten="0.0.0.0"
> > > > name="latest"
> > > > vif = [ 'mac=00:0F:4B:00:00:68, bridge=switch' ]
> > > >
> > > > And I tried simple combinations of 'x[m|l] vcpu-set latest 2|3|4' and
> > > > none of them worked.
> > > >
> > > > In Xen 4.1 (and Xen 4.3 if I use:device_model_version="qemu-xen-
> > > > traditional")
> > > > I saw that the qemu log does the right thing:
> > > > .. snip..
> > > > Remove vcpu 2
> > > > Remove vcpu 1
> > > >
> > > > and the guest's ACPI SCI is incrementing:
> > > > # cat /proc/interrupts  |grep acpi
> > > >    9:          1          0          0   IO-APIC-fasteoi   acpi
> > > > # cat /proc/interrupts  |grep acpi
> > > >    9:          2          0          0   IO-APIC-fasteoi   acpi
> > > >
> > > > But nothing looks to be happening. Where should I look?
> > > > The ACPI DSDT is a bit daunting. Has this ever worked in the past?
> > > > If so, what code ran to hotplug CPUs?
> > > 
> > > I am looking at qemu-traditional (I believe it is traditional
> > > here: http://xenbits.xen.org/gitweb/?p=qemu-xen-unstable.git;a=summary).
> > > 
> > > The code for raising the SCI for processor hotplug doesn't
> > > look quite right to me. This is the code I see in
> > > qemu_cpu_add_remove:
> > > 
> > >     if (gpe_state.gpe0_en[0] & 4) {
> > >         qemu_set_irq(sci_irq, 1);
> > >         qemu_set_irq(sci_irq, 0);
> > >     }
> > > 
> > > I would expect the code to actually set the GPE status bit for
> > > the GPE register in question (in this case bit 2 for _L02), maybe:
> > > 
> > >     if ((gpe_state.gpe0_en[0] & 4)&&
> > >         ((gpe_state.gpe0_sts[0] & 4) == 0)) {
> > >         gpe_state.gpe0_sts[0] &= 4;
> > >         qemu_irq_raise(sci_irq);
> > >     }
> > > 
> > > I also don't understand why it is raising the SCI as level and
> > > then edge. That GPE is defined in the DSDT as level:
> > > 
> > >     /* Define GPE control method '_L02'. */
> > >     push_block("Scope", "\\_GPE");
> > >     push_block("Method", "_L02");
> > >     stmt("Return", "\\_SB.PRSC()");
> > > 
> > > Maybe look at how the PCI hotplug GPE's occur (ACPI_PHP_GPE_BIT).
> > > Without
> > > setting a status register the SCI may just be going in the bit bucket.
> > > 
> > > Anyway hope this is helpful at all...
> > 
> > Never mind, I missed the calls above to enable_processor
> > and disable_processor which seem to be doing the right
> > thing (long day on a plane). My sample code above had a
> > type-o but g->gpe0_sts[0] |= 4; looks right.
> > 
> > I still don't understand why it raises the SCI level then edge though.
> 
> Neither do I but it actually looks to work. The problem I had was that
> the $@$@#( generic hotplug code has a race, which means that udev
> never ends up onlining the CPU (https://lkml.org/lkml/2012/4/30/198)
> So it looks like the CPU nevers gets onlined - and it did not until
> I manually wrote a script to online the CPUs.
> 
> Now it can online, albeit I am hitting another bug <sigh>:
> 
> [   41.465296] CPU 1 got hotplugged
> [   41.469607] installing Xen timer for CPU 1
> [   41.475935] SMP alternatives: lockdep: fixing up alternatives
> [   41.483014] SMP alternatives: switching to SMP code
> [   41.499193] smpboot: Booting Node 0 Processor 1 APIC 0x4
> [   41.517099] ------------[ cut here ]------------
> [   41.519532] kernel BUG at 
> /home/konrad/ssd/konrad/linux/arch/x86/xen/time.c:337!
> [   41.519532] invalid opcode: 0000 [#1] SMP
> [   41.519532] Modules linked in: dm_multipath dm_mod iscsi_boot_sysfs 
> iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi libcrc32c crc32c sg 
> sr_mod cdrom ata_generic crc32c_intel ata_piix libata scsi_mod xen_blkfront 
> xen_netfront fb_sys_fops sysimgblt sysfillrect syscopyarea xen_kbdfront xenfs 
> xen_privcmd
> [   41.519532] CPU 1
> [   41.519532] Pid: 0, comm: swapper/1 Not tainted 
> 3.9.0upstream-00022-g49c1bf1-dirty #3 Xen HVM domU
> [   41.519532] RIP: 0010:[<ffffffff81044f5c>]  [<ffffffff81044f5c>] 
> xen_vcpuop_set_mode+0x3c/0x80
> [   41.519532] RSP: 0000:ffff880074467db8  EFLAGS: 00010082
> [   41.519532] RAX: ffffffffffffffea RBX: ffff880074a2be40 RCX: 
> 0000000000000001
> [   41.519532] RDX: 0000000000000000 RSI: 0000000000000001 RDI: 
> 0000000000000009
> [   41.519532] RBP: ffff880074467db8 R08: 0000000000000001 R09: 
> 0000000000000000
> [   41.519532] R10: 0000000000000008 R11: 0000000000000001 R12: 
> 0000000000000001
> [   41.519532] R13: 0000000000000082 R14: 0000000000000000 R15: 
> 0000000000000082
> [   41.519532] FS:  0000000000000000(0000) GS:ffff880074a20000(0000) 
> knlGS:0000000000000000
> [   41.519532] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [   41.519532] CR2: 0000000000000000 CR3: 0000000001c0c000 CR4: 
> 00000000000406e0
> [   41.519532] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 
> 0000000000000000
> [   41.519532] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 
> 0000000000000400
> [   41.519532] Process swapper/1 (pid: 0, threadinfo ffff880074466000, task 
> ffff880074464300)
> [   41.519532] Stack:
> [   41.519532]  ffff880074467dd8 ffffffff810e8385 ffff880074a2be40 
> ffff880074a2be40
> [   41.519532]  ffff880074467df8 ffffffff810e83e6 ffff880074467df8 
> 0000000000000000
> [   41.519532]  ffff880074467e28 ffffffff810e84b7 ffffffff810e8ae4 
> ffff880074a2be40
> [   41.519532] Call Trace:
> [   41.519532]  [<ffffffff810e8385>] clockevents_set_mode+0x25/0x70
> [   41.519532]  [<ffffffff810e83e6>] clockevents_shutdown+0x16/0x30
> [   41.519532]  [<ffffffff810e84b7>] clockevents_exchange_device+0xb7/0x110
> [   41.519532]  [<ffffffff810e8ae4>] ? tick_notify+0x114/0x420
> [   41.519532]  [<ffffffff810e8b99>] tick_notify+0x1c9/0x420
> [   41.519532]  [<ffffffff810e8221>] ? clockevents_register_device+0x31/0x170
> [   41.519532]  [<ffffffff8169dd9d>] notifier_call_chain+0x4d/0x70
> [   41.519532]  [<ffffffff810be851>] raw_notifier_call_chain+0x11/0x20
> [   41.519532]  [<ffffffff810e82d0>] clockevents_register_device+0xe0/0x170
> [   41.519532]  [<ffffffff8104502c>] xen_setup_cpu_clockevents+0x2c/0x50
> [   41.519532]  [<ffffffff81045066>] xen_hvm_setup_cpu_clockevents+0x16/0x20
> [   41.519532]  [<ffffffff8168d48f>] start_secondary+0x1ea/0x1f9
> [   41.519532] Code: 73 2d 48 63 c9 bf 09 00 00 00 31 d2 48 89 ce e8 bb c3 fb 
> ff 85 c0 75 13 bf 07 00 00 00 48 89 ce 31 d2 e8 a8 c3 fb ff 85 c0 74 09 <0f> 
> 0b eb fe 83 ff 03 74 02 c9 c3 bf 07 00 00 00 48 63 f1 31 d2 
> [   41.519532] RIP  [<ffffffff81044f5c>] xen_vcpuop_set_mode+0x3c/0x80
> [   41.519532]  RSP <ffff880074467db8>
> [   41.519532] ---[ end trace a182694869545b1a ]---
> [   41.519532] Kernel panic - not syncing: Attempted to kill the idle task!
> 
> Thought this is using xend, as you cannot in xl have maxvcpus != vcpus.
> 
> For reference this is the guest config:
> 
> builder='hvm'
> device_model_version = 'qemu-xen-traditional'
> disk = [ 'file:/mnt/lab/latest/root_image.iso,hdc:cdrom,r']
> memory = 2048
> boot="d"
> maxvcpus=32
> vcpus=1
> serial='pty'
> vnclisten="0.0.0.0"
> name="latest"
> vif = [ 'mac=00:0F:4B:00:00:68, bridge=switch' ]
> 

This needs more investigating but it is possible that cpu hotplug is
broken in xl. This should be considered a regression.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel