Re: S3 resume issue in xstate_init

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 17/08/2021 13:21, Jan Beulich wrote:
> On 17.08.2021 13:44, Marek Marczykowski-Górecki wrote:
>> On Tue, Aug 17, 2021 at 12:14:36PM +0100, Andrew Cooper wrote:
>>> On 17/08/2021 12:02, Marek Marczykowski-Górecki wrote:
>>>> On Tue, Aug 17, 2021 at 03:25:21AM +0200, Marek Marczykowski-Górecki wrote:
>>>>> Hi,
>>>>>
>>>>> I've got another S3 issue:
>>>>>
>>>>> (XEN) Preparing system for ACPI S3 state.
>>>>> (XEN) Disabling non-boot CPUs ...
>>>>> (XEN) Broke affinity for IRQ1, new: ffff
>>>>> (XEN) Broke affinity for IRQ16, new: ffff
>>>>> (XEN) Broke affinity for IRQ9, new: ffff
>>>>> (XEN) Broke affinity for IRQ139, new: ffff
>>>>> (XEN) Broke affinity for IRQ8, new: ffff
>>>>> (XEN) Broke affinity for IRQ14, new: ffff
>>>>> (XEN) Broke affinity for IRQ20, new: ffff
>>>>> (XEN) Broke affinity for IRQ137, new: ffff
>>>>> (XEN) Broke affinity for IRQ138, new: ffff
>>>>> (XEN) Entering ACPI S3 state.
>>>>> (XEN) mce_intel.c:773: MCA Capability: firstbank 0, extended MCE MSR 0, 
>>>>> BCAST, CMCI
>>>>> (XEN) CPU0 CMCI LVT vector (0xf1) already installed
>>>>> (XEN) Finishing wakeup from ACPI S3 state.
>>>>> (XEN) microcode: CPU0 updated from revision 0xca to 0xea, date = 
>>>>> 2021-01-05
>>>>> (XEN) xstate: size: 0x440 (uncompressed 0x440) and states: 0x1f
>>>>> (XEN) Enabling non-boot CPUs  ...
>>>>> (XEN) xstate: size: 0x440 (uncompressed 0x240) and states: 0x1f
>>>>> (XEN) Xen BUG at xstate.c:673
>>>>> (XEN) ----[ Xen-4.16-unstable  x86_64  debug=y  Not tainted ]----
>>>>> (XEN) CPU:    1
>>>>> (XEN) RIP:    e008:[<ffff82d040350ee4>] xstate_init+0x24b/0x2ff
>>>>> (XEN) RFLAGS: 0000000000010087   CONTEXT: hypervisor
>>>>> (XEN) rax: 0000000000000240   rbx: 000000000000001f   rcx: 
>>>>> 0000000000000440
>>>>> (XEN) rdx: 0000000000000001   rsi: 000000000000000a   rdi: 
>>>>> 000000000000001f
>>>>> (XEN) rbp: ffff83025dc9fd38   rsp: ffff83025dc9fd20   r8:  
>>>>> 0000000000000001
>>>>> (XEN) r9:  ffff83025dc9fc88   r10: 0000000000000001   r11: 
>>>>> 0000000000000001
>>>>> (XEN) r12: ffff83025dc9fd80   r13: 000000000000001f   r14: 
>>>>> 0000000000000001
>>>>> (XEN) r15: 0000000000000000   cr0: 000000008005003b   cr4: 
>>>>> 00000000003526e0
>>>>> (XEN) cr3: 0000000049656000   cr2: 0000000000000000
>>>>> (XEN) fsb: 0000000000000000   gsb: 0000000000000000   gss: 
>>>>> 0000000000000000
>>>>> (XEN) ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: e008
>>>>> (XEN) Xen code around <ffff82d040350ee4> (xstate_init+0x24b/0x2ff):
>>>>> (XEN)  ff e9 a2 00 00 00 0f 0b <0f> 0b 89 f8 89 f1 0f a2 89 f2 4c 8b 0d 
>>>>> cb b4 0f
>>>>> (XEN) Xen stack trace from rsp=ffff83025dc9fd20:
>>>>> (XEN)    0000000000000240 ffff83025dc9fd80 0000000000000001 
>>>>> ffff83025dc9fd70
>>>>> (XEN)    ffff82d04027e7a1 000000004035a7f1 7ffafbbf01100800 
>>>>> 00000000bfebfbff
>>>>> (XEN)    0000000000000001 00000000000000c8 ffff83025dc9feb8 
>>>>> ffff82d0402e43ce
>>>>> (XEN)    000000160a9e0106 bfebfbff80000008 2c1008007ffaf3bf 
>>>>> 0000000f00000121
>>>>> (XEN)    00000000029c6fbf 0000000000000100 000000009c002e00 
>>>>> 02afcd7f00000000
>>>>> (XEN)    756e654700000000 6c65746e49656e69 65746e4904b21920 
>>>>> 726f43202952286c
>>>>> (XEN)    376920294d542865 432048303537382d 322e322040205550 
>>>>> 000000007a484730
>>>>> (XEN)    ffff830000000000 ffff83025dc9fe18 00002400402e8e0b 
>>>>> 000000085dc9fe30
>>>>> (XEN)    00000002402e9f21 0000000000000001 ffffffff00000000 
>>>>> ffff82d0402e0040
>>>>> (XEN)    00000000003526e0 ffff83025dc9fe68 ffff82d04027bd15 
>>>>> 0000000000000001
>>>>> (XEN)    ffff8302590a0000 0000000000000000 00000000000000c8 
>>>>> 0000000000000001
>>>>> (XEN)    0000000000000001 ffff83025dc9feb8 ffff82d0402e32b7 
>>>>> 0000000000000001
>>>>> (XEN)    0000000000000001 00000000000000c8 0000000000000001 
>>>>> ffff83025dc9fee8
>>>>> (XEN)    ffff82d04030e401 0000000000000001 0000000000000000 
>>>>> 0000000000000000
>>>>> (XEN)    0000000000000000 0000000000000000 ffff82d040200122 
>>>>> 0800002000000002
>>>>> (XEN)    0100000400010000 0000002000000000 2000000000100000 
>>>>> 0000001000000000
>>>>> (XEN)    2000000000000000 0000000029000000 0000008000000000 
>>>>> 00110000a0000000
>>>>> (XEN)    8000000080000000 4000000000000008 0000100000000000 
>>>>> 0200000040000080
>>>>> (XEN)    0004000000000000 0000010000000002 0400002030000000 
>>>>> 0000000060000000
>>>>> (XEN)    0400001000010000 0000000010000000 0000004010000000 
>>>>> 0000000000000000
>>>>> (XEN) Xen call trace:
>>>>> (XEN)    [<ffff82d040350ee4>] R xstate_init+0x24b/0x2ff
>>>>> (XEN)    [<ffff82d04027e7a1>] F identify_cpu+0x318/0x4af
>>>>> (XEN)    [<ffff82d0402e43ce>] F recheck_cpu_features+0x1f/0x72
>>>>> (XEN)    [<ffff82d04030e401>] F start_secondary+0x255/0x38a
>>>>> (XEN)    [<ffff82d040200122>] F __high_start+0x82/0x91
>>>>> (XEN) 
>>>>> (XEN) 
>>>>> (XEN) ****************************************
>>>>> (XEN) Panic on CPU 1:
>>>>> (XEN) Xen BUG at xstate.c:673
>>>>> (XEN) ****************************************
>>>>> (XEN) 
>>>>> (XEN) Reboot in five seconds...
>>>>>
>>>>> This is with added debug patch:
>>>>>
>>>>> diff --git a/xen/arch/x86/xstate.c b/xen/arch/x86/xstate.c
>>>>> index 6aaf9a2f1546..7873a21b356a 100644
>>>>> --- a/xen/arch/x86/xstate.c
>>>>> +++ b/xen/arch/x86/xstate.c
>>>>> @@ -668,6 +668,8 @@ void xstate_init(struct cpuinfo_x86 *c)
>>>>>      else
>>>>>      {
>>>>>          BUG_ON(xfeature_mask != feature_mask);
>>>>> +        printk("xstate: size: %#x (uncompressed %#x) and states: 
>>>>> %#"PRIx64"\n",
>>>>> +               xsave_cntxt_size, hw_uncompressed_size(feature_mask), 
>>>>> feature_mask);
>>>>>          BUG_ON(xsave_cntxt_size != hw_uncompressed_size(feature_mask));
>>>>>      }
>>>>>  
>>>>>
>>>>> As can be seen above - the xsave size differs between BSP and other
>>>>> CPU(s) - likely because of (not) loaded ucode update there.
>>>>> I guess it's a matter of moving ucode loading somewhere else, right?
>>>> Few more data points:
>>>>
>>>> 1. The CPU is i7-8750H (family 6, model 158, stepping 10).
>>>> 2. I do have "smt=off" on the Xen cmdline, if that matters.
>>> As a datapoint, it would be interesting to confirm what the behaviour is
>>> with SMT enabled.
>>>
>>> I'd expect it not to make a difference, because smt=off is a purely Xen
>>> construct and doesn't change the hardware configuration.
>> Uhm, changing to smt=on actually _did_ change it. Now it doesn't crash!
>>
>> Let me add CPU number to the above printk - is smp_processor_id() the
>> thing I want?
>> With that, I get:
>> https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgist.github.com%2Fmarmarek%2Fae604a1e5cf49639a1eec9e220c037ca&amp;data=04%7C01%7CAndrew.Cooper3%40citrix.com%7Cf7c03eb0fa0341ac60bf08d96179964f%7C335836de42ef43a2b145348c2ee9ca5b%7C0%7C0%7C637647997122346495%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=YZg2191dKLlppcbdCBwTnXuk08Vq37pcHOftyjrGO3Q%3D&amp;reserved=0
>> Note that at boot all CPUs reports 0x440 (but only later are parked).
> And for a feature mask of 0x1f only 0x440 can possibly be correct.

As a complete tangent, we really want an mpx=off option (or extend
cpuid=no-mpx to clobber bnd{regs,csr} too).  For systems not wanting to
use MPX in the first place - and I'd expect QubesOS falls into this
category - this will increase performance by not partitioning the
physical register file, as well as reducing the xstate size for client
parts.

> I'm kind of guessing that set_xcr0() mistakenly skips the actual XCR0
> write, due to the cached value matching the to-be-written one, yet
> the cache having gone stale across S3.

This is a rats nest, and area for concern, but ...

>  I think this is to be expected
> for previously parked CPUs, as those don't have their per-CPU data
> de-allocated (and hence also not re-setup, and thus also not starting
> out as zero).

... we don't deallocate regular APs either, so I don't see why the smt=
setting would make a difference in this case.

To be clear - I think your guess about set_xcr0() skipping the write is
correct, because 0x240 is correct for xcr0=X87, but I don't see why smt=
makes a difference.

>  I guess an easy fix would be to write 0 to
> this_cpu(xcr0) directly early in xstate_init(), maybe in an "else"
> to the early "if ( bsp )".
>
> I'm not sure though this would be a good fix longer term, as there
> might easily be other similar issues elsewhere. IOW we may need to
> see whether per-CPU data initialization wouldn't want changing.

We've got other registers too, like MSR_TSC_AUX, but I don't think we
want to be doing anything as drastic as changing how the initialisation
works.

The S3 path needs to explicitly write every register we do lazy context
switching of.

~Andrew