diff mbox

[RFC] x86/emulate: implement hvmemul_cmpxchg() with an actual CMPXCHG

Message ID 15d76365-71a4-5f83-9ab3-a5bcc2564910@bitdefender.com (mailing list archive)
State New, archived
Headers show

Commit Message

Razvan Cojocaru April 1, 2017, 4:56 p.m. UTC
On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>
>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>
>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>> would need to be collected to have a chance to understand what
>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>> address this was trying to act on).
>>>>>>
>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>> always LOCKed in my patch):
>>>>>
>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>> how this information helps understanding the MMIO emulation failure
>>>>> above.
>>>>
>>>> I've managed to obtain this log of emulation errors:
>>>> https://pastebin.com/Esy1SkHx 
>>>>
>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>> line correspond to CMXCHG_FAILED return codes.
>>>>
>>>> The very last line is a MMIO emulation failed.
>>>>
>>>> It's probably important that this happens with the model where
>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>> succeeds. The other model allows me to go further with the guest, but
>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>
>>> Interesting. You didn't clarify what the printed "offset" values are,
>>> and it doesn't look like these have any correlation with the underlying
>>> (guest) physical address, which we would also want to see. And then
>>> it strikes me as odd that in these last lines
>>>
>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>> 07 8b cb e8 da 4b ff ff 8b 45
>>>
>>> the instruction pointers and virtual addresses are different, but the
>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>> wonder whether there's an issue with us wrongly re-using previously
>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>> this guessing, by you checking the involved binary/ies.)
>>
>> Offset is the actual value of the "offset" parameter of
>> hvmemul_cmpxchg().
> 
> That's not very useful then, as for flat segments "offset" ==
> "virtual address" (i.e. you merely re-print in decimal what you've
> already printed in hex).

The attached patch (a combination of your patch and mine) produces the
following output when booting a Windows 7 32-bit guest with monitoring:
https://pastebin.com/ayiFmj1N

The failed MMIO emulation is caused by a mapping failure due to the
"!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
__hvm_copy() but it looks like that might not be the right way to use it.


Thanks,
Razvan

Comments

Jan Beulich April 3, 2017, 10:23 a.m. UTC | #1
>>> On 01.04.17 at 18:56, <rcojocaru@bitdefender.com> wrote:
> On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>>
>>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>>
>>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>>> would need to be collected to have a chance to understand what
>>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>>> address this was trying to act on).
>>>>>>>
>>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>>> always LOCKed in my patch):
>>>>>>
>>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>>> how this information helps understanding the MMIO emulation failure
>>>>>> above.
>>>>>
>>>>> I've managed to obtain this log of emulation errors:
>>>>> https://pastebin.com/Esy1SkHx 
>>>>>
>>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>>> line correspond to CMXCHG_FAILED return codes.
>>>>>
>>>>> The very last line is a MMIO emulation failed.
>>>>>
>>>>> It's probably important that this happens with the model where
>>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>>> succeeds. The other model allows me to go further with the guest, but
>>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>>
>>>> Interesting. You didn't clarify what the printed "offset" values are,
>>>> and it doesn't look like these have any correlation with the underlying
>>>> (guest) physical address, which we would also want to see. And then
>>>> it strikes me as odd that in these last lines
>>>>
>>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
> 
>>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>
>>>> the instruction pointers and virtual addresses are different, but the
>>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>>> wonder whether there's an issue with us wrongly re-using previously
>>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>>> this guessing, by you checking the involved binary/ies.)
>>>
>>> Offset is the actual value of the "offset" parameter of
>>> hvmemul_cmpxchg().
>> 
>> That's not very useful then, as for flat segments "offset" ==
>> "virtual address" (i.e. you merely re-print in decimal what you've
>> already printed in hex).
> 
> The attached patch (a combination of your patch and mine) produces the
> following output when booting a Windows 7 32-bit guest with monitoring:
> https://pastebin.com/ayiFmj1N 
> 
> The failed MMIO emulation is caused by a mapping failure due to the
> "!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
> being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
> __hvm_copy() but it looks like that might not be the right way to use it.

(XEN) [7] virtual address: 0x8276d09c, rc: 4
(XEN) gfn: 0x276d
(XEN) Dump follows for VCPU 7
(XEN) Mem event (RETRY) emulation failed: d4v7 32bit @ 0008:826a1c7c -> f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
(XEN) [9] virtual address: 0x8276d0a8, rc: 0
(XEN) gfn: 0xfed00
(XEN) !nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)
(XEN) hvmemul_vaddr_to_mfn() fail
(XEN) [7] virtual address: 0xffd080f0, rc: 1

It would help to know which of the earlier messages were also for
vCPU 7. In any event we again have the same suspicious pattern:
Virtual (and now also physical, assuming the gfn logged was that on
vCPU 7, not 9) address completely different, but ...

(XEN) MMIO emulation failed: d4v7 32bit @ 0008:8263bf3c -> f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41

... instruction bytes identical between the last mem event emulation
and the failed MMIO one. Relatively unlikely to be pure coincidence.

Jan
Razvan Cojocaru April 3, 2017, 6:20 p.m. UTC | #2
On 04/01/2017 07:56 PM, Razvan Cojocaru wrote:
> On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>>
>>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>>
>>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>>> would need to be collected to have a chance to understand what
>>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>>> address this was trying to act on).
>>>>>>>
>>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>>> always LOCKed in my patch):
>>>>>>
>>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>>> how this information helps understanding the MMIO emulation failure
>>>>>> above.
>>>>>
>>>>> I've managed to obtain this log of emulation errors:
>>>>> https://pastebin.com/Esy1SkHx 
>>>>>
>>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>>> line correspond to CMXCHG_FAILED return codes.
>>>>>
>>>>> The very last line is a MMIO emulation failed.
>>>>>
>>>>> It's probably important that this happens with the model where
>>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>>> succeeds. The other model allows me to go further with the guest, but
>>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>>
>>>> Interesting. You didn't clarify what the printed "offset" values are,
>>>> and it doesn't look like these have any correlation with the underlying
>>>> (guest) physical address, which we would also want to see. And then
>>>> it strikes me as odd that in these last lines
>>>>
>>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
>>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>
>>>> the instruction pointers and virtual addresses are different, but the
>>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>>> wonder whether there's an issue with us wrongly re-using previously
>>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>>> this guessing, by you checking the involved binary/ies.)
>>>
>>> Offset is the actual value of the "offset" parameter of
>>> hvmemul_cmpxchg().
>>
>> That's not very useful then, as for flat segments "offset" ==
>> "virtual address" (i.e. you merely re-print in decimal what you've
>> already printed in hex).
> 
> The attached patch (a combination of your patch and mine) produces the
> following output when booting a Windows 7 32-bit guest with monitoring:
> https://pastebin.com/ayiFmj1N
> 
> The failed MMIO emulation is caused by a mapping failure due to the
> "!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
> being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
> __hvm_copy() but it looks like that might not be the right way to use it.

Sorry to reply to this email instead of your original reply but I've
"left it" in my computer at work. Here's the last part of the log, with
the VCPU number logged for the GFN as well:

(XEN) [8] gfn: 0x2781
(XEN) [8] virtual address: 0x827810a8, rc: 0
(XEN) [8] gfn: 0x2781
(XEN) [8] virtual address: 0x827810a8, rc: 0
(XEN) [8] gfn: 0x2781
(XEN) [8] virtual address: 0x827810cc, rc: 0
(XEN) [8] gfn: 0x2781
(XEN) [8] virtual address: 0x8278109c, rc: 0
(XEN) [8] gfn: 0x2781
(XEN) [8] virtual address: 0x827810d0, rc: 0
(XEN) [11] gfn: 0x2781
(XEN) [8] gfn: 0x2781
(XEN) [11] virtual address: 0x8278109c, rc: 0
(XEN) [8] virtual address: 0x8278109c, rc: 4
(XEN) Dump follows for VCPU 8
(XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826b5c7c ->
f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
(XEN) [11] gfn: 0x2781
(XEN) [11] virtual address: 0x827810a8, rc: 0
(XEN) [11] gfn: 0x2781
(XEN) [11] virtual address: 0x827810a8, rc: 0
(XEN) [8] gfn: 0xfed00
(XEN) !page
(XEN) hvmemul_vaddr_to_mfn() fail
(XEN) [8] virtual address: 0xffd080f0, rc: 1
(XEN) MMIO emulation failed: d3v8 32bit @ 0008:8264ff3c -> f0 0f c1 08
85 c9 74 1f f6 c1 02 75 1a 41 8d 41

The code does look the same for the last two failures on VCPU 8, for
clearly different GFNs and virtual addresses. The first time it hits a
protected page, we try to emulate the instruction and it fails with
X86EMUL_CMPXCHG_FAILED (rc: 4). Then it somehow pops up again and this
time it's MMIO-emulated, and that fails as well (with UNHANDLEABLE,
since we can't seem to be able to map the memory).


Thanks,
Razvan
Razvan Cojocaru April 3, 2017, 6:36 p.m. UTC | #3
On 04/03/2017 09:20 PM, Razvan Cojocaru wrote:
> On 04/01/2017 07:56 PM, Razvan Cojocaru wrote:
>> On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>>>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>>>
>>>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>>>
>>>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>>>> would need to be collected to have a chance to understand what
>>>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>>>> address this was trying to act on).
>>>>>>>>
>>>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>>>> always LOCKed in my patch):
>>>>>>>
>>>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>>>> how this information helps understanding the MMIO emulation failure
>>>>>>> above.
>>>>>>
>>>>>> I've managed to obtain this log of emulation errors:
>>>>>> https://pastebin.com/Esy1SkHx 
>>>>>>
>>>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>>>> line correspond to CMXCHG_FAILED return codes.
>>>>>>
>>>>>> The very last line is a MMIO emulation failed.
>>>>>>
>>>>>> It's probably important that this happens with the model where
>>>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>>>> succeeds. The other model allows me to go further with the guest, but
>>>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>>>
>>>>> Interesting. You didn't clarify what the printed "offset" values are,
>>>>> and it doesn't look like these have any correlation with the underlying
>>>>> (guest) physical address, which we would also want to see. And then
>>>>> it strikes me as odd that in these last lines
>>>>>
>>>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
>>>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>
>>>>> the instruction pointers and virtual addresses are different, but the
>>>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>>>> wonder whether there's an issue with us wrongly re-using previously
>>>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>>>> this guessing, by you checking the involved binary/ies.)
>>>>
>>>> Offset is the actual value of the "offset" parameter of
>>>> hvmemul_cmpxchg().
>>>
>>> That's not very useful then, as for flat segments "offset" ==
>>> "virtual address" (i.e. you merely re-print in decimal what you've
>>> already printed in hex).
>>
>> The attached patch (a combination of your patch and mine) produces the
>> following output when booting a Windows 7 32-bit guest with monitoring:
>> https://pastebin.com/ayiFmj1N
>>
>> The failed MMIO emulation is caused by a mapping failure due to the
>> "!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
>> being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
>> __hvm_copy() but it looks like that might not be the right way to use it.
> 
> Sorry to reply to this email instead of your original reply but I've
> "left it" in my computer at work. Here's the last part of the log, with
> the VCPU number logged for the GFN as well:
> 
> (XEN) [8] gfn: 0x2781
> (XEN) [8] virtual address: 0x827810a8, rc: 0
> (XEN) [8] gfn: 0x2781
> (XEN) [8] virtual address: 0x827810a8, rc: 0
> (XEN) [8] gfn: 0x2781
> (XEN) [8] virtual address: 0x827810cc, rc: 0
> (XEN) [8] gfn: 0x2781
> (XEN) [8] virtual address: 0x8278109c, rc: 0
> (XEN) [8] gfn: 0x2781
> (XEN) [8] virtual address: 0x827810d0, rc: 0
> (XEN) [11] gfn: 0x2781
> (XEN) [8] gfn: 0x2781
> (XEN) [11] virtual address: 0x8278109c, rc: 0
> (XEN) [8] virtual address: 0x8278109c, rc: 4
> (XEN) Dump follows for VCPU 8
> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826b5c7c ->
> f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
> (XEN) [11] gfn: 0x2781
> (XEN) [11] virtual address: 0x827810a8, rc: 0
> (XEN) [11] gfn: 0x2781
> (XEN) [11] virtual address: 0x827810a8, rc: 0
> (XEN) [8] gfn: 0xfed00
> (XEN) !page
> (XEN) hvmemul_vaddr_to_mfn() fail
> (XEN) [8] virtual address: 0xffd080f0, rc: 1
> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8264ff3c -> f0 0f c1 08
> 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
> 
> The code does look the same for the last two failures on VCPU 8, for
> clearly different GFNs and virtual addresses. The first time it hits a
> protected page, we try to emulate the instruction and it fails with
> X86EMUL_CMPXCHG_FAILED (rc: 4). Then it somehow pops up again and this
> time it's MMIO-emulated, and that fails as well (with UNHANDLEABLE,
> since we can't seem to be able to map the memory).

Another log, making sure that the MMIO emulation fail really happens on
the same processor as the preceding (identical bytes) fail:

(XEN) [8] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f09c, rc: 0
(XEN) [8] virtual address: 0x8276f09c, rc: 4
(XEN) Dump follows for VCPU 8
(XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826a3861 ->
f0 0f ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
(XEN) [11] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f0d0, rc: 0
(XEN) [11] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f09c, rc: 0
(XEN) [11] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f0cc, rc: 0
(XEN) [11] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f09c, rc: 0
(XEN) [8] gfn: 0xfed00
(XEN) !page
(XEN) hvmemul_vaddr_to_mfn() fail
(XEN) [8] virtual address: 0xffd080f0, rc: 1
(XEN) Dump follows for VCPU 8
(XEN) MMIO emulation failed: d3v8 32bit @ 0008:8263df3c -> f0 0f ba 30
00 72 07 8b cb e8 da 4b ff ff 8b 45
(XEN) [11] gfn: 0x276f
(XEN) [11] virtual address: 0x8276f0d0, rc: 0


Thanks,
Razvan
Jan Beulich April 4, 2017, 9:07 a.m. UTC | #4
>>> On 03.04.17 at 20:36, <rcojocaru@bitdefender.com> wrote:
> Another log, making sure that the MMIO emulation fail really happens on
> the same processor as the preceding (identical bytes) fail:
> 
> (XEN) [8] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [8] virtual address: 0x8276f09c, rc: 4
> (XEN) Dump follows for VCPU 8
> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826a3861 ->
> f0 0f ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45

So what does your code do following this? Namely is there any exit
back to guest context prior to the retried operation succeeding (e.g.
are you allowing the instruction to be re-executed)? If there is, I'd
suspect you don't clean up some internal state and ...

> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f0d0, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f0cc, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [8] gfn: 0xfed00
> (XEN) !page
> (XEN) hvmemul_vaddr_to_mfn() fail
> (XEN) [8] virtual address: 0xffd080f0, rc: 1
> (XEN) Dump follows for VCPU 8
> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8263df3c -> f0 0f ba 30
> 00 72 07 8b cb e8 da 4b ff ff 8b 45

... this then happens in the context of an interrupt taken on
the instruction to be re-executed (i.e. before it actually gets
re-executed). "Normal" retry processing is not supposed to
reach guest context again, i.e. the retry is being performed
from (iirc) hvm_do_resume(). Hence it can re-use previously
collected information on the instruction and its operand(s).
Any exit to guest context, otoh, would require invalidation of
such internally cached state.

Jan
Razvan Cojocaru April 4, 2017, 12:01 p.m. UTC | #5
On 04/04/2017 12:07 PM, Jan Beulich wrote:
>>>> On 03.04.17 at 20:36, <rcojocaru@bitdefender.com> wrote:
>> Another log, making sure that the MMIO emulation fail really happens on
>> the same processor as the preceding (identical bytes) fail:
>>
>> (XEN) [8] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [8] virtual address: 0x8276f09c, rc: 4
>> (XEN) Dump follows for VCPU 8
>> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826a3861 ->
>> f0 0f ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
> 
> So what does your code do following this? Namely is there any exit
> back to guest context prior to the retried operation succeeding (e.g.
> are you allowing the instruction to be re-executed)? If there is, I'd
> suspect you don't clean up some internal state and ...

That code is called from hvm_do_resume() ...

>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f0d0, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f0cc, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [8] gfn: 0xfed00
>> (XEN) !page
>> (XEN) hvmemul_vaddr_to_mfn() fail
>> (XEN) [8] virtual address: 0xffd080f0, rc: 1
>> (XEN) Dump follows for VCPU 8
>> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8263df3c -> f0 0f ba 30
>> 00 72 07 8b cb e8 da 4b ff ff 8b 45
> 
> ... this then happens in the context of an interrupt taken on
> the instruction to be re-executed (i.e. before it actually gets
> re-executed). "Normal" retry processing is not supposed to
> reach guest context again, i.e. the retry is being performed
> from (iirc) hvm_do_resume(). Hence it can re-use previously
> collected information on the instruction and its operand(s).
> Any exit to guest context, otoh, would require invalidation of
> such internally cached state.

... which does nothing if hvm_emulate_one_vm_event() fails. It doesn't
even check the return code. That's why I have added the do { } while (
rc == X86EMUL_RETRY ); loop around the code in
hvm_emulate_one_vm_event() and noticed that the BSODs appear much later
(and they're timeout BSODs, presumably because of a taking-too-long
RETRY loop), or the guest becomes unresponsive.

I'm not sure what cached state that would be, I'll take a closer look at
the code.

I'll retest as soon as possible (it's been a very busy time, sorry for
the late replies).


Thanks,
Razvan
Razvan Cojocaru April 8, 2017, 10:15 p.m. UTC | #6
On 04/03/2017 09:36 PM, Razvan Cojocaru wrote:
> On 04/03/2017 09:20 PM, Razvan Cojocaru wrote:
>> On 04/01/2017 07:56 PM, Razvan Cojocaru wrote:
>>> On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>>>>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>>>>
>>>>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>>>>
>>>>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>>>>> would need to be collected to have a chance to understand what
>>>>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>>>>> address this was trying to act on).
>>>>>>>>>
>>>>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>>>>> always LOCKed in my patch):
>>>>>>>>
>>>>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>>>>> how this information helps understanding the MMIO emulation failure
>>>>>>>> above.
>>>>>>>
>>>>>>> I've managed to obtain this log of emulation errors:
>>>>>>> https://pastebin.com/Esy1SkHx 
>>>>>>>
>>>>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>>>>> line correspond to CMXCHG_FAILED return codes.
>>>>>>>
>>>>>>> The very last line is a MMIO emulation failed.
>>>>>>>
>>>>>>> It's probably important that this happens with the model where
>>>>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>>>>> succeeds. The other model allows me to go further with the guest, but
>>>>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>>>>
>>>>>> Interesting. You didn't clarify what the printed "offset" values are,
>>>>>> and it doesn't look like these have any correlation with the underlying
>>>>>> (guest) physical address, which we would also want to see. And then
>>>>>> it strikes me as odd that in these last lines
>>>>>>
>>>>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
>>>>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>>>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>>>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>
>>>>>> the instruction pointers and virtual addresses are different, but the
>>>>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>>>>> wonder whether there's an issue with us wrongly re-using previously
>>>>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>>>>> this guessing, by you checking the involved binary/ies.)
>>>>>
>>>>> Offset is the actual value of the "offset" parameter of
>>>>> hvmemul_cmpxchg().
>>>>
>>>> That's not very useful then, as for flat segments "offset" ==
>>>> "virtual address" (i.e. you merely re-print in decimal what you've
>>>> already printed in hex).
>>>
>>> The attached patch (a combination of your patch and mine) produces the
>>> following output when booting a Windows 7 32-bit guest with monitoring:
>>> https://pastebin.com/ayiFmj1N
>>>
>>> The failed MMIO emulation is caused by a mapping failure due to the
>>> "!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
>>> being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
>>> __hvm_copy() but it looks like that might not be the right way to use it.
>>
>> Sorry to reply to this email instead of your original reply but I've
>> "left it" in my computer at work. Here's the last part of the log, with
>> the VCPU number logged for the GFN as well:
>>
>> (XEN) [8] gfn: 0x2781
>> (XEN) [8] virtual address: 0x827810a8, rc: 0
>> (XEN) [8] gfn: 0x2781
>> (XEN) [8] virtual address: 0x827810a8, rc: 0
>> (XEN) [8] gfn: 0x2781
>> (XEN) [8] virtual address: 0x827810cc, rc: 0
>> (XEN) [8] gfn: 0x2781
>> (XEN) [8] virtual address: 0x8278109c, rc: 0
>> (XEN) [8] gfn: 0x2781
>> (XEN) [8] virtual address: 0x827810d0, rc: 0
>> (XEN) [11] gfn: 0x2781
>> (XEN) [8] gfn: 0x2781
>> (XEN) [11] virtual address: 0x8278109c, rc: 0
>> (XEN) [8] virtual address: 0x8278109c, rc: 4
>> (XEN) Dump follows for VCPU 8
>> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826b5c7c ->
>> f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
>> (XEN) [11] gfn: 0x2781
>> (XEN) [11] virtual address: 0x827810a8, rc: 0
>> (XEN) [11] gfn: 0x2781
>> (XEN) [11] virtual address: 0x827810a8, rc: 0
>> (XEN) [8] gfn: 0xfed00
>> (XEN) !page
>> (XEN) hvmemul_vaddr_to_mfn() fail
>> (XEN) [8] virtual address: 0xffd080f0, rc: 1
>> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8264ff3c -> f0 0f c1 08
>> 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
>>
>> The code does look the same for the last two failures on VCPU 8, for
>> clearly different GFNs and virtual addresses. The first time it hits a
>> protected page, we try to emulate the instruction and it fails with
>> X86EMUL_CMPXCHG_FAILED (rc: 4). Then it somehow pops up again and this
>> time it's MMIO-emulated, and that fails as well (with UNHANDLEABLE,
>> since we can't seem to be able to map the memory).
> 
> Another log, making sure that the MMIO emulation fail really happens on
> the same processor as the preceding (identical bytes) fail:
> 
> (XEN) [8] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [8] virtual address: 0x8276f09c, rc: 4
> (XEN) Dump follows for VCPU 8
> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826a3861 ->
> f0 0f ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f0d0, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f0cc, rc: 0
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f09c, rc: 0
> (XEN) [8] gfn: 0xfed00
> (XEN) !page
> (XEN) hvmemul_vaddr_to_mfn() fail
> (XEN) [8] virtual address: 0xffd080f0, rc: 1
> (XEN) Dump follows for VCPU 8
> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8263df3c -> f0 0f ba 30
> 00 72 07 8b cb e8 da 4b ff ff 8b 45
> (XEN) [11] gfn: 0x276f
> (XEN) [11] virtual address: 0x8276f0d0, rc: 0

As you've suggested, this does indeed seem to happen because of the
current model of emulating because of vm_event replies: an instruction
needs to be emulated in hvm_do_resume(), but inside
hvm_emulate_one_vm_event(), hvm_emulate_one() returns RETRY, presumably
because of the new cmpxchg handler. This causes
hvm_emulate_one_vm_event() to simply return, doing nothing else. Then,
the guest resumes execution at the same place (since RIP has not been
modified), then hvm_do_resume() gets called again, except this time this
code:

 481     if ( !handle_hvm_io_completion(v) )
 482         return;

ends up trying to MMIO emulate the current instruction by calling
handle_mmio().

I've had hvm_emulate_one_vm_event() return what hvm_emulate_one()
returns, and loop while ( rc == RETRY ) in hvm_do_resume() - this is
functionally equivalent to what I've been doing before by wrapping the
hvm_emulate_one_vm_event() code in a do {} while ( rc == RETRY ); loop
as previously discussed. However, as stated before, at a much later
point I still get rare BSODs, or the guest becomes unresponsive. Digging
a bit more, I've found that what seems to be happening is that emulating
some instructions returns EXCEPTION at some point (for instructions
using the new cmpxchg handler, this can happen if paging_gva_to_gfn()
returns INVALID_GFN and neither PFEC_page_paged nor PFEC_page_shared are
set). Since this causes hvm_emulate_one_vm_event() to
hvm_inject_event(), I'm not sure that it does not mess with the
interrupt part of hvm_do_resume(). This will need more testing to figure
out exactly what's going wrong.


Thanks,
Razvan
Razvan Cojocaru April 9, 2017, 11:03 a.m. UTC | #7
On 04/09/2017 01:15 AM, Razvan Cojocaru wrote:
> I'm not sure that it does not mess with the
> interrupt part of hvm_do_resume().

Actually I've now printk()ed
v->arch.hvm_vcpu.inject_event.vector on the EXCEPTION case of
hvm_emulate_one_vm_event(), and it's
always -1 in my tests, so the interrupt theory doesn't check out. I'll
have to think about another scenario to test.


Thanks,
Razvan
Jan Beulich April 10, 2017, 10:18 a.m. UTC | #8
>>> On 09.04.17 at 00:15, <rcojocaru@bitdefender.com> wrote:
> On 04/03/2017 09:36 PM, Razvan Cojocaru wrote:
>> On 04/03/2017 09:20 PM, Razvan Cojocaru wrote:
>>> On 04/01/2017 07:56 PM, Razvan Cojocaru wrote:
>>>> On 03/31/2017 06:04 PM, Jan Beulich wrote:
>>>>>>>> On 31.03.17 at 17:01, <rcojocaru@bitdefender.com> wrote:
>>>>>> On 03/31/2017 05:46 PM, Jan Beulich wrote:
>>>>>>>>>> On 31.03.17 at 11:56, <rcojocaru@bitdefender.com> wrote:
>>>>>>>> On 03/31/2017 10:34 AM, Jan Beulich wrote:
>>>>>>>>>>>> On 31.03.17 at 08:17, <rcojocaru@bitdefender.com> wrote:
>>>>>>>>>> On 03/30/2017 06:47 PM, Jan Beulich wrote:
>>>>>>>>>>>> Speaking of emulated MMIO, I've got this when the guest was crashing
>>>>>>>>>>>> immediately (pre RETRY loop):
>>>>>>>>>>>>
>>>>>>>>>>>>  MMIO emulation failed: d3v8 32bit @ 0008:82679f3c -> f0 0f ba 30 00 72
>>>>>>>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>>>>>
>>>>>>>>>>> That's a BTR, which we should be emulating fine. More information
>>>>>>>>>>> would need to be collected to have a chance to understand what
>>>>>>>>>>> might be going one (first of all the virtual and physical memory
>>>>>>>>>>> address this was trying to act on).
>>>>>>>>>>
>>>>>>>>>> Right, the BTR part should be fine, but I think the LOCK part is what's
>>>>>>>>>> causing the issue. I've done a few more test runs to see what return
>>>>>>>>>> RETRY (dumping the instruction with an "(r)" prefix to distinguish from
>>>>>>>>>> the UNHANDLEABLE dump), and a couple of instructions return RETRY (BTR
>>>>>>>>>> and XADD, both LOCK-prefixed, which means they now involve CMPXCHG
>>>>>>>>>> handler, which presumably now fails - possibly simply because it's
>>>>>>>>>> always LOCKed in my patch):
>>>>>>>>>
>>>>>>>>> Well, all of that looks to be expected behavior. I'm afraid I don't see
>>>>>>>>> how this information helps understanding the MMIO emulation failure
>>>>>>>>> above.
>>>>>>>>
>>>>>>>> I've managed to obtain this log of emulation errors:
>>>>>>>> https://pastebin.com/Esy1SkHx 
>>>>>>>>
>>>>>>>> The "virtual address" lines that are not followed by any "Mem event"
>>>>>>>> line correspond to CMXCHG_FAILED return codes.
>>>>>>>>
>>>>>>>> The very last line is a MMIO emulation failed.
>>>>>>>>
>>>>>>>> It's probably important that this happens with the model where
>>>>>>>> hvm_emulate_one_vm_event() does _not_ re-try the emulation until it
>>>>>>>> succeeds. The other model allows me to go further with the guest, but
>>>>>>>> eventually I get timeout-related BSODs or the guest becomes unresponsive.
>>>>>>>
>>>>>>> Interesting. You didn't clarify what the printed "offset" values are,
>>>>>>> and it doesn't look like these have any correlation with the underlying
>>>>>>> (guest) physical address, which we would also want to see. And then
>>>>>>> it strikes me as odd that in these last lines
>>>>>>>
>>>>>>> (XEN) Mem event (RETRY) emulation failed: d5v8 32bit @ 0008:826bb861 -> f0 0f 
> 
>>>>>> ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>> (XEN) virtual address: 0xffd080f0, offset: 4291854576
>>>>>>> (XEN) MMIO emulation failed: d5v8 32bit @ 0008:82655f3c -> f0 0f ba 30 00 72 
>>>>>> 07 8b cb e8 da 4b ff ff 8b 45
>>>>>>>
>>>>>>> the instruction pointers and virtual addresses are different, but the
>>>>>>> code bytes are exactly the same. This doesn't seem very likely, so I
>>>>>>> wonder whether there's an issue with us wrongly re-using previously
>>>>>>> fetched insn bytes. (Of course I'd be happy to be proven wrong with
>>>>>>> this guessing, by you checking the involved binary/ies.)
>>>>>>
>>>>>> Offset is the actual value of the "offset" parameter of
>>>>>> hvmemul_cmpxchg().
>>>>>
>>>>> That's not very useful then, as for flat segments "offset" ==
>>>>> "virtual address" (i.e. you merely re-print in decimal what you've
>>>>> already printed in hex).
>>>>
>>>> The attached patch (a combination of your patch and mine) produces the
>>>> following output when booting a Windows 7 32-bit guest with monitoring:
>>>> https://pastebin.com/ayiFmj1N 
>>>>
>>>> The failed MMIO emulation is caused by a mapping failure due to the
>>>> "!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)" condition
>>>> being true in hvmemul_vaddr_to_mfn(). I've ripped that off from
>>>> __hvm_copy() but it looks like that might not be the right way to use it.
>>>
>>> Sorry to reply to this email instead of your original reply but I've
>>> "left it" in my computer at work. Here's the last part of the log, with
>>> the VCPU number logged for the GFN as well:
>>>
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [8] virtual address: 0x827810a8, rc: 0
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [8] virtual address: 0x827810a8, rc: 0
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [8] virtual address: 0x827810cc, rc: 0
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [8] virtual address: 0x8278109c, rc: 0
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [8] virtual address: 0x827810d0, rc: 0
>>> (XEN) [11] gfn: 0x2781
>>> (XEN) [8] gfn: 0x2781
>>> (XEN) [11] virtual address: 0x8278109c, rc: 0
>>> (XEN) [8] virtual address: 0x8278109c, rc: 4
>>> (XEN) Dump follows for VCPU 8
>>> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826b5c7c ->
>>> f0 0f c1 08 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
>>> (XEN) [11] gfn: 0x2781
>>> (XEN) [11] virtual address: 0x827810a8, rc: 0
>>> (XEN) [11] gfn: 0x2781
>>> (XEN) [11] virtual address: 0x827810a8, rc: 0
>>> (XEN) [8] gfn: 0xfed00
>>> (XEN) !page
>>> (XEN) hvmemul_vaddr_to_mfn() fail
>>> (XEN) [8] virtual address: 0xffd080f0, rc: 1
>>> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8264ff3c -> f0 0f c1 08
>>> 85 c9 74 1f f6 c1 02 75 1a 41 8d 41
>>>
>>> The code does look the same for the last two failures on VCPU 8, for
>>> clearly different GFNs and virtual addresses. The first time it hits a
>>> protected page, we try to emulate the instruction and it fails with
>>> X86EMUL_CMPXCHG_FAILED (rc: 4). Then it somehow pops up again and this
>>> time it's MMIO-emulated, and that fails as well (with UNHANDLEABLE,
>>> since we can't seem to be able to map the memory).
>> 
>> Another log, making sure that the MMIO emulation fail really happens on
>> the same processor as the preceding (identical bytes) fail:
>> 
>> (XEN) [8] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [8] virtual address: 0x8276f09c, rc: 4
>> (XEN) Dump follows for VCPU 8
>> (XEN) Mem event (RETRY) emulation failed: d3v8 32bit @ 0008:826a3861 ->
>> f0 0f ba 30 00 72 07 8b cb e8 da 4b ff ff 8b 45
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f0d0, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f0cc, rc: 0
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f09c, rc: 0
>> (XEN) [8] gfn: 0xfed00
>> (XEN) !page
>> (XEN) hvmemul_vaddr_to_mfn() fail
>> (XEN) [8] virtual address: 0xffd080f0, rc: 1
>> (XEN) Dump follows for VCPU 8
>> (XEN) MMIO emulation failed: d3v8 32bit @ 0008:8263df3c -> f0 0f ba 30
>> 00 72 07 8b cb e8 da 4b ff ff 8b 45
>> (XEN) [11] gfn: 0x276f
>> (XEN) [11] virtual address: 0x8276f0d0, rc: 0
> 
> As you've suggested, this does indeed seem to happen because of the
> current model of emulating because of vm_event replies: an instruction
> needs to be emulated in hvm_do_resume(), but inside
> hvm_emulate_one_vm_event(), hvm_emulate_one() returns RETRY, presumably
> because of the new cmpxchg handler. This causes
> hvm_emulate_one_vm_event() to simply return, doing nothing else. Then,
> the guest resumes execution at the same place (since RIP has not been
> modified), then hvm_do_resume() gets called again, except this time this
> code:
> 
>  481     if ( !handle_hvm_io_completion(v) )
>  482         return;
> 
> ends up trying to MMIO emulate the current instruction by calling
> handle_mmio().
> 
> I've had hvm_emulate_one_vm_event() return what hvm_emulate_one()
> returns, and loop while ( rc == RETRY ) in hvm_do_resume() - this is
> functionally equivalent to what I've been doing before by wrapping the
> hvm_emulate_one_vm_event() code in a do {} while ( rc == RETRY ); loop
> as previously discussed. However, as stated before, at a much later
> point I still get rare BSODs, or the guest becomes unresponsive. Digging
> a bit more, I've found that what seems to be happening is that emulating
> some instructions returns EXCEPTION at some point (for instructions
> using the new cmpxchg handler, this can happen if paging_gva_to_gfn()
> returns INVALID_GFN and neither PFEC_page_paged nor PFEC_page_shared are
> set).

Apparently in line with your later reply - if you got INVALID_GFN
from paging_gva_to_gfn(), the earlier read of the same memory
should have failed already (unless something plays with the P2M
behind your back).

Jan

> Since this causes hvm_emulate_one_vm_event() to
> hvm_inject_event(), I'm not sure that it does not mess with the
> interrupt part of hvm_do_resume(). This will need more testing to figure
> out exactly what's going wrong.
> 
> 
> Thanks,
> Razvan
diff mbox

Patch

diff --git a/xen/arch/x86/hvm/emulate.c b/xen/arch/x86/hvm/emulate.c
index 2d92957..f6244af 100644
--- a/xen/arch/x86/hvm/emulate.c
+++ b/xen/arch/x86/hvm/emulate.c
@@ -20,6 +20,7 @@ 
 #include <asm/hvm/emulate.h>
 #include <asm/hvm/hvm.h>
 #include <asm/hvm/ioreq.h>
+#include <asm/hvm/nestedhvm.h>
 #include <asm/hvm/trace.h>
 #include <asm/hvm/support.h>
 #include <asm/hvm/svm/svm.h>
@@ -1029,6 +1030,86 @@  static int hvmemul_wbinvd_discard(
     return X86EMUL_OKAY;
 }
 
+static int hvmemul_vaddr_to_mfn(
+    unsigned long addr,
+    mfn_t *mfn,
+    uint32_t pfec,
+    struct x86_emulate_ctxt *ctxt)
+{
+    paddr_t gpa = addr & ~PAGE_MASK;
+    struct page_info *page;
+    p2m_type_t p2mt;
+    unsigned long gfn;
+    struct vcpu *curr = current;
+    struct hvm_emulate_ctxt *hvmemul_ctxt =
+        container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
+
+    gfn = paging_gva_to_gfn(curr, addr, &pfec);
+
+    printk("gfn: 0x%lx\n", gfn);
+
+    if ( gfn == gfn_x(INVALID_GFN) )
+    {
+        pagefault_info_t pfinfo = {};
+
+        if ( ( pfec & PFEC_page_paged ) || ( pfec & PFEC_page_shared ) )
+            return X86EMUL_RETRY;
+
+        pfinfo.linear = addr;
+        pfinfo.ec = pfec;
+
+        x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
+        return X86EMUL_EXCEPTION;
+    }
+
+    gpa |= (paddr_t)gfn << PAGE_SHIFT;
+
+    /*
+     * No need to do the P2M lookup for internally handled MMIO, benefiting
+     * - 32-bit WinXP (& older Windows) on AMD CPUs for LAPIC accesses,
+     * - newer Windows (like Server 2012) for HPET accesses.
+     */
+    if ( !nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa) )
+    {
+        printk("!nestedhvm_vcpu_in_guestmode(curr) && hvm_mmio_internal(gpa)\n");
+        return X86EMUL_UNHANDLEABLE;
+    }
+
+    page = get_page_from_gfn(curr->domain, gfn, &p2mt, P2M_UNSHARE);
+
+    if ( !page )
+    {
+        printk("!page\n");
+        return X86EMUL_UNHANDLEABLE;
+    }
+
+    if ( p2m_is_paging(p2mt) )
+    {
+        put_page(page);
+        p2m_mem_paging_populate(curr->domain, gfn);
+        return X86EMUL_RETRY;
+    }
+
+    if ( p2m_is_shared(p2mt) )
+    {
+        put_page(page);
+        return X86EMUL_RETRY;
+    }
+
+    if ( p2m_is_grant(p2mt) )
+    {
+        put_page(page);
+        printk("p2m_is_grant(p2mt)\n");
+        return X86EMUL_UNHANDLEABLE;
+    }
+
+    *mfn = _mfn(page_to_mfn(page));
+
+    put_page(page);
+
+    return X86EMUL_OKAY;
+}
+
 static int hvmemul_cmpxchg(
     enum x86_segment seg,
     unsigned long offset,
@@ -1037,8 +1118,98 @@  static int hvmemul_cmpxchg(
     unsigned int bytes,
     struct x86_emulate_ctxt *ctxt)
 {
-    /* Fix this in case the guest is really relying on r-m-w atomicity. */
-    return hvmemul_write(seg, offset, p_new, bytes, ctxt);
+    unsigned long addr = 0, reps = 1;
+    int rc = X86EMUL_OKAY;
+    unsigned long old = 0, new = 0;
+    uint32_t pfec = PFEC_page_present | PFEC_write_access;
+    struct hvm_emulate_ctxt *hvmemul_ctxt =
+        container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
+    mfn_t mfn[2];
+    void *map = NULL;
+    struct domain *currd = current->domain;
+    unsigned long ret;
+
+    if ( is_x86_system_segment(seg) )
+        pfec |= PFEC_implicit;
+    else if ( hvmemul_ctxt->seg_reg[x86_seg_ss].attr.fields.dpl == 3 )
+        pfec |= PFEC_user_mode;
+
+    rc = hvmemul_virtual_to_linear(
+        seg, offset, bytes, &reps, hvm_access_write, hvmemul_ctxt, &addr);
+
+    if ( rc != X86EMUL_OKAY || !bytes )
+    {
+        printk("rc != X86EMUL_OKAY || !bytes\n");
+        goto out;
+    }
+
+    rc = hvmemul_vaddr_to_mfn(addr, &mfn[0], pfec, ctxt);
+
+    if ( rc != X86EMUL_OKAY )
+    {
+        printk("hvmemul_vaddr_to_mfn() fail\n");
+        goto out;
+    }
+
+    if ( likely(((addr + bytes - 1) & PAGE_MASK) == (addr & PAGE_MASK)) )
+    {
+        /* Whole write fits on a single page. */
+        mfn[1] = INVALID_MFN;
+        map = map_domain_page(mfn[0]);
+    }
+    else
+    {
+        rc = hvmemul_vaddr_to_mfn((addr + bytes - 1) & PAGE_MASK, &mfn[1],
+                                  pfec, ctxt);
+        if ( rc != X86EMUL_OKAY )
+        {
+            printk("hvmemul_vaddr_to_mfn(mfn[1]) fail\n");
+            goto out;
+        }
+
+        map = vmap(mfn, 2);
+    }
+
+    if ( !map )
+    {
+        printk("!map\n");
+        return X86EMUL_UNHANDLEABLE;
+    }
+
+    map += (addr & ~PAGE_MASK);
+
+    memcpy(&old, p_old, bytes);
+    memcpy(&new, p_new, bytes);
+
+    ret = __cmpxchg(map, old, new, bytes);
+
+    if ( ret != old )
+    {
+        memcpy(p_old, &ret, bytes);
+        rc = X86EMUL_CMPXCHG_FAILED;
+    }
+
+    paging_mark_dirty(currd, mfn[0]);
+
+    if ( unlikely(mfn_valid(mfn[1])) )
+    {
+        paging_mark_dirty(currd, mfn[1]);
+        vunmap((void *)((unsigned long)map & PAGE_MASK));
+    }
+    else
+        unmap_domain_page(map);
+
+out:
+    if ( rc != X86EMUL_OKAY )
+    {
+        if ( rc != X86EMUL_CMPXCHG_FAILED )
+            rc = X86EMUL_UNHANDLEABLE;
+    }
+
+    printk("[%d] virtual address: 0x%lx, rc: %d\n",
+           current->vcpu_id, addr, rc);
+
+    return rc;
 }
 
 static int hvmemul_validate(
@@ -1961,59 +2132,64 @@  int hvm_emulate_one_mmio(unsigned long mfn, unsigned long gla)
 void hvm_emulate_one_vm_event(enum emul_kind kind, unsigned int trapnr,
     unsigned int errcode)
 {
-    struct hvm_emulate_ctxt ctx = {{ 0 }};
-    int rc;
+    int rc = X86EMUL_OKAY;
 
-    hvm_emulate_init_once(&ctx, NULL, guest_cpu_user_regs());
+    /* do { */
+        struct hvm_emulate_ctxt ctx = {{ 0 }};
 
-    switch ( kind )
-    {
-    case EMUL_KIND_NOWRITE:
-        rc = _hvm_emulate_one(&ctx, &hvm_emulate_ops_no_write);
-        break;
-    case EMUL_KIND_SET_CONTEXT_INSN: {
-        struct vcpu *curr = current;
-        struct hvm_vcpu_io *vio = &curr->arch.hvm_vcpu.hvm_io;
+        hvm_emulate_init_once(&ctx, NULL, guest_cpu_user_regs());
 
-        BUILD_BUG_ON(sizeof(vio->mmio_insn) !=
-                     sizeof(curr->arch.vm_event->emul.insn.data));
-        ASSERT(!vio->mmio_insn_bytes);
+        switch ( kind )
+        {
+        case EMUL_KIND_NOWRITE:
+            rc = _hvm_emulate_one(&ctx, &hvm_emulate_ops_no_write);
+            break;
+        case EMUL_KIND_SET_CONTEXT_INSN: {
+            struct vcpu *curr = current;
+            struct hvm_vcpu_io *vio = &curr->arch.hvm_vcpu.hvm_io;
+
+            BUILD_BUG_ON(sizeof(vio->mmio_insn) !=
+                         sizeof(curr->arch.vm_event->emul.insn.data));
+            ASSERT(!vio->mmio_insn_bytes);
+
+            /*
+             * Stash insn buffer into mmio buffer here instead of ctx
+             * to avoid having to add more logic to hvm_emulate_one.
+             */
+            vio->mmio_insn_bytes = sizeof(vio->mmio_insn);
+            memcpy(vio->mmio_insn, curr->arch.vm_event->emul.insn.data,
+                   vio->mmio_insn_bytes);
+        }
+        /* Fall-through */
+        default:
+            ctx.set_context = (kind == EMUL_KIND_SET_CONTEXT_DATA);
+            rc = hvm_emulate_one(&ctx);
+        }
 
-        /*
-         * Stash insn buffer into mmio buffer here instead of ctx
-         * to avoid having to add more logic to hvm_emulate_one.
-         */
-        vio->mmio_insn_bytes = sizeof(vio->mmio_insn);
-        memcpy(vio->mmio_insn, curr->arch.vm_event->emul.insn.data,
-               vio->mmio_insn_bytes);
-    }
-    /* Fall-through */
-    default:
-        ctx.set_context = (kind == EMUL_KIND_SET_CONTEXT_DATA);
-        rc = hvm_emulate_one(&ctx);
-    }
+        if ( rc != X86EMUL_OKAY )
+        {
+            printk("Dump follows for VCPU %d\n", current->vcpu_id);
+        }
 
-    switch ( rc )
-    {
-    case X86EMUL_RETRY:
-        /*
-         * This function is called when handling an EPT-related vm_event
-         * reply. As such, nothing else needs to be done here, since simply
-         * returning makes the current instruction cause a page fault again,
-         * consistent with X86EMUL_RETRY.
-         */
-        return;
-    case X86EMUL_UNHANDLEABLE:
-        hvm_dump_emulation_state(XENLOG_G_DEBUG "Mem event", &ctx);
-        hvm_inject_hw_exception(trapnr, errcode);
-        break;
-    case X86EMUL_EXCEPTION:
-        if ( ctx.ctxt.event_pending )
-            hvm_inject_event(&ctx.ctxt.event);
-        break;
-    }
+        switch ( rc )
+        {
+        case X86EMUL_RETRY:
+            /* break; */
+            hvm_dump_emulation_state(XENLOG_G_DEBUG "Mem event (RETRY)", &ctx);
+            return;
+        case X86EMUL_UNHANDLEABLE:
+            hvm_dump_emulation_state(XENLOG_G_DEBUG "Mem event", &ctx);
+            hvm_inject_hw_exception(trapnr, errcode);
+            break;
+        case X86EMUL_EXCEPTION:
+            if ( ctx.ctxt.event_pending )
+                hvm_inject_event(&ctx.ctxt.event);
+            break;
+        }
+
+        hvm_emulate_writeback(&ctx);
 
-    hvm_emulate_writeback(&ctx);
+    /* } while( rc == X86EMUL_RETRY ); */
 }
 
 void hvm_emulate_init_once(
diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
index 4dbd24f..a67cd55 100644
--- a/xen/arch/x86/mm.c
+++ b/xen/arch/x86/mm.c
@@ -5176,16 +5176,17 @@  static int ptwr_emulated_read(
 
 static int ptwr_emulated_update(
     unsigned long addr,
-    paddr_t old,
-    paddr_t val,
+    intpte_t *p_old,
+    intpte_t val,
     unsigned int bytes,
-    unsigned int do_cmpxchg,
     struct ptwr_emulate_ctxt *ptwr_ctxt)
 {
     unsigned long mfn;
     unsigned long unaligned_addr = addr;
     struct page_info *page;
     l1_pgentry_t pte, ol1e, nl1e, *pl1e;
+    intpte_t old = p_old ? *p_old : 0;
+    unsigned int offset = 0;
     struct vcpu *v = current;
     struct domain *d = v->domain;
     int ret;
@@ -5199,28 +5200,30 @@  static int ptwr_emulated_update(
     }
 
     /* Turn a sub-word access into a full-word access. */
-    if ( bytes != sizeof(paddr_t) )
+    if ( bytes != sizeof(val) )
     {
-        paddr_t      full;
-        unsigned int rc, offset = addr & (sizeof(paddr_t)-1);
+        intpte_t full;
+        unsigned int rc;
+
+        offset = addr & (sizeof(full) - 1);
 
         /* Align address; read full word. */
-        addr &= ~(sizeof(paddr_t)-1);
-        if ( (rc = copy_from_user(&full, (void *)addr, sizeof(paddr_t))) != 0 )
+        addr &= ~(sizeof(full) - 1);
+        if ( (rc = copy_from_user(&full, (void *)addr, sizeof(full))) != 0 )
         {
             x86_emul_pagefault(0, /* Read fault. */
-                               addr + sizeof(paddr_t) - rc,
+                               addr + sizeof(full) - rc,
                                &ptwr_ctxt->ctxt);
             return X86EMUL_EXCEPTION;
         }
         /* Mask out bits provided by caller. */
-        full &= ~((((paddr_t)1 << (bytes*8)) - 1) << (offset*8));
+        full &= ~((((intpte_t)1 << (bytes * 8)) - 1) << (offset * 8));
         /* Shift the caller value and OR in the missing bits. */
-        val  &= (((paddr_t)1 << (bytes*8)) - 1);
+        val  &= (((intpte_t)1 << (bytes * 8)) - 1);
         val <<= (offset)*8;
         val  |= full;
         /* Also fill in missing parts of the cmpxchg old value. */
-        old  &= (((paddr_t)1 << (bytes*8)) - 1);
+        old  &= (((intpte_t)1 << (bytes * 8)) - 1);
         old <<= (offset)*8;
         old  |= full;
     }
@@ -5242,7 +5245,7 @@  static int ptwr_emulated_update(
     {
     default:
         if ( is_pv_32bit_domain(d) && (bytes == 4) && (unaligned_addr & 4) &&
-             !do_cmpxchg && (l1e_get_flags(nl1e) & _PAGE_PRESENT) )
+             !p_old && (l1e_get_flags(nl1e) & _PAGE_PRESENT) )
         {
             /*
              * If this is an upper-half write to a PAE PTE then we assume that
@@ -5273,21 +5276,26 @@  static int ptwr_emulated_update(
     /* Checked successfully: do the update (write or cmpxchg). */
     pl1e = map_domain_page(_mfn(mfn));
     pl1e = (l1_pgentry_t *)((unsigned long)pl1e + (addr & ~PAGE_MASK));
-    if ( do_cmpxchg )
+    if ( p_old )
     {
-        int okay;
-        intpte_t t = old;
         ol1e = l1e_from_intpte(old);
 
-        okay = paging_cmpxchg_guest_entry(v, &l1e_get_intpte(*pl1e),
-                                          &t, l1e_get_intpte(nl1e), _mfn(mfn));
-        okay = (okay && t == old);
+        if ( !paging_cmpxchg_guest_entry(v, &l1e_get_intpte(*pl1e),
+                                         &old, l1e_get_intpte(nl1e), _mfn(mfn)) )
+            ret = X86EMUL_UNHANDLEABLE;
+        else if ( l1e_get_intpte(ol1e) == old )
+            ret = X86EMUL_OKAY;
+        else
+        {
+            *p_old = old >> (offset * 8);
+            ret = X86EMUL_CMPXCHG_FAILED;
+        }
 
-        if ( !okay )
+        if ( ret != X86EMUL_OKAY )
         {
             unmap_domain_page(pl1e);
             put_page_from_l1e(nl1e, d);
-            return X86EMUL_RETRY;
+            return ret;
         }
     }
     else
@@ -5314,9 +5322,9 @@  static int ptwr_emulated_write(
     unsigned int bytes,
     struct x86_emulate_ctxt *ctxt)
 {
-    paddr_t val = 0;
+    intpte_t val = 0;
 
-    if ( (bytes > sizeof(paddr_t)) || (bytes & (bytes - 1)) || !bytes )
+    if ( (bytes > sizeof(val)) || (bytes & (bytes - 1)) || !bytes )
     {
         MEM_LOG("ptwr_emulate: bad write size (addr=%lx, bytes=%u)",
                 offset, bytes);
@@ -5325,9 +5333,9 @@  static int ptwr_emulated_write(
 
     memcpy(&val, p_data, bytes);
 
-    return ptwr_emulated_update(
-        offset, 0, val, bytes, 0,
-        container_of(ctxt, struct ptwr_emulate_ctxt, ctxt));
+    return ptwr_emulated_update(offset, NULL, val, bytes,
+                                container_of(ctxt, struct ptwr_emulate_ctxt,
+                                             ctxt));
 }
 
 static int ptwr_emulated_cmpxchg(
@@ -5338,21 +5346,20 @@  static int ptwr_emulated_cmpxchg(
     unsigned int bytes,
     struct x86_emulate_ctxt *ctxt)
 {
-    paddr_t old = 0, new = 0;
+    intpte_t new = 0;
 
-    if ( (bytes > sizeof(paddr_t)) || (bytes & (bytes -1)) )
+    if ( (bytes > sizeof(new)) || (bytes & (bytes -1)) )
     {
         MEM_LOG("ptwr_emulate: bad cmpxchg size (addr=%lx, bytes=%u)",
                 offset, bytes);
         return X86EMUL_UNHANDLEABLE;
     }
 
-    memcpy(&old, p_old, bytes);
     memcpy(&new, p_new, bytes);
 
-    return ptwr_emulated_update(
-        offset, old, new, bytes, 1,
-        container_of(ctxt, struct ptwr_emulate_ctxt, ctxt));
+    return ptwr_emulated_update(offset, p_old, new, bytes,
+                                container_of(ctxt, struct ptwr_emulate_ctxt,
+                                             ctxt));
 }
 
 static int pv_emul_is_mem_write(const struct x86_emulate_state *state,
diff --git a/xen/arch/x86/mm/shadow/common.c b/xen/arch/x86/mm/shadow/common.c
index d93f2ab..06dc9f6 100644
--- a/xen/arch/x86/mm/shadow/common.c
+++ b/xen/arch/x86/mm/shadow/common.c
@@ -285,7 +285,7 @@  hvm_emulate_cmpxchg(enum x86_segment seg,
     struct sh_emulate_ctxt *sh_ctxt =
         container_of(ctxt, struct sh_emulate_ctxt, ctxt);
     struct vcpu *v = current;
-    unsigned long addr, old, new;
+    unsigned long addr, new = 0;
     int rc;
 
     if ( bytes > sizeof(long) )
@@ -296,12 +296,10 @@  hvm_emulate_cmpxchg(enum x86_segment seg,
     if ( rc )
         return rc;
 
-    old = new = 0;
-    memcpy(&old, p_old, bytes);
     memcpy(&new, p_new, bytes);
 
     return v->arch.paging.mode->shadow.x86_emulate_cmpxchg(
-               v, addr, old, new, bytes, sh_ctxt);
+               v, addr, p_old, new, bytes, sh_ctxt);
 }
 
 static const struct x86_emulate_ops hvm_shadow_emulator_ops = {
diff --git a/xen/arch/x86/mm/shadow/multi.c b/xen/arch/x86/mm/shadow/multi.c
index 4798c93..d8270db 100644
--- a/xen/arch/x86/mm/shadow/multi.c
+++ b/xen/arch/x86/mm/shadow/multi.c
@@ -4783,11 +4783,11 @@  sh_x86_emulate_write(struct vcpu *v, unsigned long vaddr, void *src,
 
 static int
 sh_x86_emulate_cmpxchg(struct vcpu *v, unsigned long vaddr,
-                        unsigned long old, unsigned long new,
-                        unsigned int bytes, struct sh_emulate_ctxt *sh_ctxt)
+                       unsigned long *p_old, unsigned long new,
+                       unsigned int bytes, struct sh_emulate_ctxt *sh_ctxt)
 {
     void *addr;
-    unsigned long prev;
+    unsigned long prev, old = *p_old;
     int rv = X86EMUL_OKAY;
 
     /* Unaligned writes are only acceptable on HVM */
@@ -4811,7 +4811,10 @@  sh_x86_emulate_cmpxchg(struct vcpu *v, unsigned long vaddr,
     }
 
     if ( prev != old )
-        rv = X86EMUL_RETRY;
+    {
+        *p_old = prev;
+        rv = X86EMUL_CMPXCHG_FAILED;
+    }
 
     SHADOW_DEBUG(EMULATE, "va %#lx was %#lx expected %#lx"
                   " wanted %#lx now %#lx bytes %u\n",
diff --git a/xen/arch/x86/x86_emulate/x86_emulate.c b/xen/arch/x86/x86_emulate/x86_emulate.c
index bb67be6..444c84c 100644
--- a/xen/arch/x86/x86_emulate/x86_emulate.c
+++ b/xen/arch/x86/x86_emulate/x86_emulate.c
@@ -1862,6 +1862,9 @@  protmode_load_seg(
 
         default:
             return rc;
+
+        case X86EMUL_CMPXCHG_FAILED:
+            return X86EMUL_RETRY;
         }
 
         /* Force the Accessed flag in our local copy. */
@@ -6680,6 +6683,7 @@  x86_emulate(
         break;
 
     case X86EMUL_OPC(0x0f, 0xb0): case X86EMUL_OPC(0x0f, 0xb1): /* cmpxchg */
+        fail_if(!ops->cmpxchg);
         /* Save real source value, then compare EAX against destination. */
         src.orig_val = src.val;
         src.val = _regs.r(ax);
@@ -6688,8 +6692,17 @@  x86_emulate(
         if ( _regs.eflags & X86_EFLAGS_ZF )
         {
             /* Success: write back to memory. */
-            dst.val = src.orig_val;
+            dst.val = src.val;
+            rc = ops->cmpxchg(dst.mem.seg, dst.mem.off, &dst.val,
+                              &src.orig_val, dst.bytes, ctxt);
+            if ( rc == X86EMUL_CMPXCHG_FAILED )
+            {
+               _regs.eflags &= ~X86_EFLAGS_ZF;
+               rc = X86EMUL_OKAY;
+            }
         }
+        if ( _regs.eflags & X86_EFLAGS_ZF )
+            dst.type = OP_NONE;
         else
         {
             /* Failure: write the value we saw to EAX. */
@@ -6994,6 +7007,7 @@  x86_emulate(
 
         if ( memcmp(old, aux, op_bytes) )
         {
+        cmpxchgNb_failed:
             /* Expected != actual: store actual to rDX:rAX and clear ZF. */
             _regs.r(ax) = !(rex_prefix & REX_W) ? old->u32[0] : old->u64[0];
             _regs.r(dx) = !(rex_prefix & REX_W) ? old->u32[1] : old->u64[1];
@@ -7003,7 +7017,7 @@  x86_emulate(
         {
             /*
              * Expected == actual: Get proposed value, attempt atomic cmpxchg
-             * and set ZF.
+             * and set ZF if successful.
              */
             if ( !(rex_prefix & REX_W) )
             {
@@ -7016,10 +7030,20 @@  x86_emulate(
                 aux->u64[1] = _regs.r(cx);
             }
 
-            if ( (rc = ops->cmpxchg(ea.mem.seg, ea.mem.off, old, aux,
-                                    op_bytes, ctxt)) != X86EMUL_OKAY )
+            switch ( rc = ops->cmpxchg(ea.mem.seg, ea.mem.off, old, aux,
+                                       op_bytes, ctxt) )
+            {
+            case X86EMUL_OKAY:
+                _regs.eflags |= X86_EFLAGS_ZF;
+                break;
+
+            case X86EMUL_CMPXCHG_FAILED:
+                rc = X86EMUL_OKAY;
+                goto cmpxchgNb_failed;
+
+            default:
                 goto done;
-            _regs.eflags |= X86_EFLAGS_ZF;
+            }
         }
         break;
     }
@@ -7942,6 +7966,8 @@  x86_emulate(
             rc = ops->cmpxchg(
                 dst.mem.seg, dst.mem.off, &dst.orig_val,
                 &dst.val, dst.bytes, ctxt);
+            if ( rc == X86EMUL_CMPXCHG_FAILED )
+                rc = X86EMUL_RETRY;
         }
         else
         {
diff --git a/xen/arch/x86/x86_emulate/x86_emulate.h b/xen/arch/x86/x86_emulate/x86_emulate.h
index 9c5fcde..6a8d6a0 100644
--- a/xen/arch/x86/x86_emulate/x86_emulate.h
+++ b/xen/arch/x86/x86_emulate/x86_emulate.h
@@ -153,6 +153,8 @@  struct x86_emul_fpu_aux {
 #define X86EMUL_EXCEPTION      2
  /* Retry the emulation for some reason. No state modified. */
 #define X86EMUL_RETRY          3
+ /* (cmpxchg accessor): CMPXCHG failed. */
+#define X86EMUL_CMPXCHG_FAILED 4
  /*
   * Operation fully done by one of the hooks:
   * - validate(): operation completed (except common insn retire logic)
@@ -160,7 +162,7 @@  struct x86_emul_fpu_aux {
   * - read_io() / write_io(): bypass GPR update (non-string insns only)
   * Undefined behavior when used anywhere else.
   */
-#define X86EMUL_DONE           4
+#define X86EMUL_DONE           5
 
 /* FPU sub-types which may be requested via ->get_fpu(). */
 enum x86_emulate_fpu_type {
@@ -250,6 +252,8 @@  struct x86_emulate_ops
     /*
      * cmpxchg: Emulate an atomic (LOCKed) CMPXCHG operation.
      *  @p_old: [IN ] Pointer to value expected to be current at @addr.
+     *          [OUT] Pointer to value found at @addr (may always be
+     *                updated, meaningful for X86EMUL_CMPXCHG_FAILED only).
      *  @p_new: [IN ] Pointer to value to write to @addr.
      *  @bytes: [IN ] Operation size (up to 8 (x86/32) or 16 (x86/64) bytes).
      */
diff --git a/xen/include/asm-x86/paging.h b/xen/include/asm-x86/paging.h
index f262c9e..3efcf6d 100644
--- a/xen/include/asm-x86/paging.h
+++ b/xen/include/asm-x86/paging.h
@@ -89,7 +89,7 @@  struct shadow_paging_mode {
                                             void *src, u32 bytes,
                                             struct sh_emulate_ctxt *sh_ctxt);
     int           (*x86_emulate_cmpxchg   )(struct vcpu *v, unsigned long va,
-                                            unsigned long old, 
+                                            unsigned long *old,
                                             unsigned long new,
                                             unsigned int bytes,
                                             struct sh_emulate_ctxt *sh_ctxt);