Skip to contentSkip to conversation

x86/mm: Don't try to change poison pages to uncacheable in a guest

PATCH In review Tony Luck 5 participants 19 messages May 5, 2020 – May 26, 2020 4 messages not in the archive

Subsystems (4)
lore
Share & tools
lore

Tony Luck

author maintainer
#0 lore raw

PATCH 1/1 x86/mm: Don't try to change poison pages to uncacheable in a guest

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

Fix is to check whether Linux is running as a guest. If it is, there is no point in trying to change the cache mode of the bad page. The VMM has taken the whole page away.

Fixes 284ce4011ba6 ("x86/memory_failure: Introduce {set, clear}_mce_nospec()")
---
 arch/x86/include/asm/set_memory.h | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/include/asm/set_memory.h +8 -0
index ec2c0a094b5d..e8b7e8a82d7a 100644
@@ -91,6 +91,14 @@ static inline int set_mce_nospec(unsigned long pfn)
 	unsigned long decoy_addr;
 	int rc;
 
+	/*
+	 * If we are running as a guest, then the hypervisor
+	 * will have removed all access from the page. So no
+	 * point trying to make it uncacheable
+	 */
+	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
+		return 0;
+
 	/*
 	 * Mark the linear address as UC to make sure we don't log more
 	 * errors because of speculative access to the page.
Signature · 1 line
2.21.1
maintainer
#1 lore raw
Quoted 8 lines · Tony Luck

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

Quoted 7 lines

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

Maybe I don't see it but how can clflush_cache_range_opt() call cpa_flush() ?

Quoted 4 lines

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

This is where you lost me - if the VMM unmaps the page during the first MCE, how can the guest even attempt to touch it and do this stack trace above?

/me is confused.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#2 lore raw

There is only one actual machine check. But the VMM simulates a second machine check to the guest when the guest tries to access the poisoned page.

The stack trace was from Jue. I didn’t try to check it. But it looked reasonable that Linux would flush the cache for a page that is transitioning from cacheable to uncacheable.

Sent from my iPhone

Quoted 38 lines · Borislav Petkov
Quoted 8 lines · Tony Luck

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

Quoted 7 lines

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

Maybe I don't see it but how can clflush_cache_range_opt() call cpa_flush() ?

Quoted 4 lines

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

This is where you lost me - if the VMM unmaps the page during the first MCE, how can the guest even attempt to touch it and do this stack trace above?

/me is confused.

Signature · 4 lines
Regards/Gruss,
   Boris.

https://people.kernel.org/tglx/notes-about-netiquette
maintainer
#3 lore raw
Quoted 3 lines · Tony

There is only one actual machine check. But the VMM simulates a second machine check to the guest when the guest tries to access the poisoned page.

If the VMM unmaps the bad page, why doesn't the guest get a #PF instead injected by the VMM instead of latter injecting a second #MCE?

If the guest tries to access an unmapped page, it should get a #PF, I'd expect.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#4 lore raw

But the guest isn’t likely to do the right thing with a page fault. The guest just accessed a page that it knows is poisoned (VMM just told it once that it was poisoned). There is no reason that the VMM should let the guest actually touch the poison a second time. But if the guest does, then the guest should get the expected response. I.e. another machine check.

Sent from my iPhone

Quoted 18 lines · Borislav Petkov
Quoted 3 lines · Tony

There is only one actual machine check. But the VMM simulates a second machine check to the guest when the guest tries to access the poisoned page.

If the VMM unmaps the bad page, why doesn't the guest get a #PF instead injected by the VMM instead of latter injecting a second #MCE?

If the guest tries to access an unmapped page, it should get a #PF, I'd expect.

Signature · 4 lines
Regards/Gruss,
   Boris.

https://people.kernel.org/tglx/notes-about-netiquette

A message that is not in the archive (sent off-list or not yet received): CAPcxDJ50pbuTbittyvPwKq1uUT8q8jJ+dHH8rCug8a1DDZXVYw@mail.gmail.com

A message that is not in the archive (sent off-list or not yet received): CAPcxDJ6f3pBpwiR9nvXN_g_HBa1RAMG+aOmgfXLFT6aZ9HQn3w@mail.gmail.com

maintainer
#7 lore raw

Hi,

lemme try to reply to three emails at once.

First of all, the two of you: pls do not top-post.

Quoted 6 lines · Tony

But the guest isn’t likely to do the right thing with a page fault. The guest just accessed a page that it knows is poisoned (VMM just told it once that it was poisoned). There is no reason that the VMM should let the guest actually touch the poison a second time. But if the guest does, then the guest should get the expected response. I.e. another machine check.

So Jue says below that the VMM has unmapped the guest page from the EPT. So the guest cannot touch the poison anymore.

How is then possible for the guest to touch it again if the page is not mapped anymore?

The guest should get a #PF when the page is unmapped and cause a new page to be mapped.

Quoted 1 line · Jue Wang

The stack is from guest MCE handler's processing of the first MCE injected.

Aha, so you've flipped the functions order in the trace. It all starts at

  set_mce_nospec(m->addr >> PAGE_SHIFT);

Now it makes sense.

Quoted 3 lines

Note before the first MCE is injected into guest (i.e., after the host MCE handler successfully finished MCE handling and notified VMM via SIGBUS with BUS_MCEERR_AR), VMM unmaps the guest page from EPT.

Ok, good.

Quoted 4 lines

The guest MCE handler finished the "normal" MCE handling and recovery (memory_failure() in mm/memory_failure.cc) successfully, it's the aftermath below leading to the stack trace:
https://github.com/torvalds/linux/blob/5a9ffb954a3933d7867f4341684a23e008d6839b/arch/x86/kernel/cpu/mce/core.c#L1101

Quoted 4 lines · Jue Wang

In other words, it's the *do_memory_failure -> set_mce_nospec* flow of guest MCE handler acting on the *first* MCE injection. As a result, the guest panics and resets *whenever* there is an MCE injected, even when the injected MCE is recoverable.

So IIUC that set_mce_nospec() thing should check whether m->addr is mapped first and only then mark it _uc and whatever monkey business it does. Not this blanket

if am I a guest?

test.

Imagine a hypervisor which doesn't set X86_FEATURE_HYPERVISOR, i.e., CPUID(1)[EDX, bit 31]?

Thx.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#8 lore raw
Quoted 5 lines · Borislav Petkov

Hi,

lemme try to reply to three emails at once.

First of all, the two of you: pls do not top-post.

Sorry. Phone e-mail client is dumb.

Quoted 8 lines · Tony
Quoted 6 lines

But the guest isn’t likely to do the right thing with a page fault. The guest just accessed a page that it knows is poisoned (VMM just told it once that it was poisoned). There is no reason that the VMM should let the guest actually touch the poison a second time. But if the guest does, then the guest should get the expected response. I.e. another machine check.

Quoted 5 lines

So Jue says below that the VMM has unmapped the guest page from the EPT. So the guest cannot touch the poison anymore.

How is then possible for the guest to touch it again if the page is not mapped anymore?

The VMM wants to make sure that the guest can't touch the poison (this is important because not every touch of poison results in a recoverable machine check. If the guest's next touch is an unaligned access that crosses from the poison cache line to a non-poisoned line then h/w will signal a fatal machinecheck and the whole machine will go down).

Quoted 2 lines

The guest should get a #PF when the page is unmapped and cause a new page to be mapped.

The VMM gets the page fault (because the unmapping of the guest physical address is at the VMM EPT level). The VMM can't map a new page into that guest physical address because it has no way to replace the contents of the old page. The VMM could pass the #PF to the guest, but that would just confuse the guest (its page tables all say that the page is still valid). In this particular case the page is part of the 1:1 kernel map. So the kernel will OOPS (I think).

Quoted 22 lines · Jue Wang
Quoted 1 line

The stack is from guest MCE handler's processing of the first MCE injected.

Aha, so you've flipped the functions order in the trace. It all starts at

  set_mce_nospec(m->addr >> PAGE_SHIFT);

Now it makes sense.

Quoted 3 lines

Note before the first MCE is injected into guest (i.e., after the host MCE handler successfully finished MCE handling and notified VMM via SIGBUS with BUS_MCEERR_AR), VMM unmaps the guest page from EPT.

Ok, good.

Quoted 4 lines

The guest MCE handler finished the "normal" MCE handling and recovery (memory_failure() in mm/memory_failure.cc) successfully, it's the aftermath below leading to the stack trace:
https://github.com/torvalds/linux/blob/5a9ffb954a3933d7867f4341684a23e008d6839b/arch/x86/kernel/cpu/mce/core.c#L1101

On Sun, May 17, 2020 at 08:33:00AM -0700, Jue Wang wrote:

Quoted 8 lines
Quoted 4 lines

In other words, it's the *do_memory_failure -> set_mce_nospec* flow of guest MCE handler acting on the *first* MCE injection. As a result, the guest panics and resets *whenever* there is an MCE injected, even when the injected MCE is recoverable.

So IIUC that set_mce_nospec() thing should check whether m->addr is mapped first and only then mark it _uc and whatever monkey business it does. Not this blanket

PLease explain how a guest (that doesn't even know that it is a guest) is going to figure out that the EPT tables (that it has no way to access) have marked this page invalid in guest physical address space.

Quoted 6 lines

if am I a guest?

test.

Imagine a hypervisor which doesn't set X86_FEATURE_HYPERVISOR, i.e., CPUID(1)[EDX, bit 31]?

Guest is going to be screwed in this case.

-Tony

maintainer
#9 lore raw
Quoted 7 lines · Tony

The VMM gets the page fault (because the unmapping of the guest physical address is at the VMM EPT level). The VMM can't map a new page into that guest physical address because it has no way to replace the contents of the old page. The VMM could pass the #PF to the guest, but that would just confuse the guest (its page tables all say that the page is still valid). In this particular case the page is part of the 1:1 kernel map. So the kernel will OOPS (I think).

...

Quoted 3 lines

PLease explain how a guest (that doesn't even know that it is a guest) is going to figure out that the EPT tables (that it has no way to access) have marked this page invalid in guest physical address space.

So somewhere BUS_MCEERR_AR was mentioned. So I'm assuming the error severity was "action required". What does happen in the kernel, on baremetal, with an AR error in kernel space, i.e., kernel memory?

If we can't fixup the exception, we die.

So why should the guest behave any differently?

Now, if you want for the guest to be more "robust" and handle that thing, fine. But then you'd need an explicit way to tell the guest
kernel: "you've just had an MCE and I unmapped the page" so that the guest kernel can figure out what do to. Even if it means, to panic.

I.e., signal in an explicit way that EPT violation Jue is talking about in the other mail.

You can inject a #PF or better yet the *first* MCE which is being injected should say with a bit somehwere "I unmapped the address in m->addr". So that the guest kernel can handle that properly and know what *exactly* it is getting an MCE for.

What I don't like is the "am I running as a guest" check. Because someone else would come later and say, err, I'm not virtualizing this portion of MCA either, lemme add another "am I guest" check.

Sure, it is a lot easier but when stuff like that starts spreading around in the MCE code, then we can just as well disable MCE when virtualized altogether. It would be a lot easier for everybody.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#10 lore raw
Quoted 11 lines · Borislav Petkov
Quoted 7 lines · Tony

The VMM gets the page fault (because the unmapping of the guest physical address is at the VMM EPT level). The VMM can't map a new page into that guest physical address because it has no way to replace the contents of the old page. The VMM could pass the #PF to the guest, but that would just confuse the guest (its page tables all say that the page is still valid). In this particular case the page is part of the 1:1 kernel map. So the kernel will OOPS (I think).

...

Quoted 7 lines
Quoted 3 lines

PLease explain how a guest (that doesn't even know that it is a guest) is going to figure out that the EPT tables (that it has no way to access) have marked this page invalid in guest physical address space.

So somewhere BUS_MCEERR_AR was mentioned. So I'm assuming the error severity was "action required". What does happen in the kernel, on baremetal, with an AR error in kernel space, i.e., kernel memory?

Outside of the now infamous memcpy_mcsafe() any kernel consumption of poison results in a panic as the mce_severity() code will trip this case:

MCESEV(
PANIC, "Data load in unrecoverable area of kernel",
SER, MASK(MCI_STATUS_OVER|MCI_UC_SAR|MCI_ADDR|MCACOD, MCI_UC_SAR|MCI_ADDR|MCACOD_DATA), KERNEL
),

Quoted 3 lines

If we can't fixup the exception, we die.

So why should the guest behave any differently?

We don't see this particular problem on baremetal because a CLFLUSH instruction isn't *consuming* data. It's just evicting things from the cache to memory. So we reference the virtual address, which works fine on baremetal because the kernel 1:1 map is still active. But in the guest case the guest physical address has gone away. So we trap to the VMM.

Quoted 5 lines

Now, if you want for the guest to be more "robust" and handle that thing, fine. But then you'd need an explicit way to tell the guest
kernel: "you've just had an MCE and I unmapped the page" so that the guest kernel can figure out what do to. Even if it means, to panic.

Quoted 7 lines

I.e., signal in an explicit way that EPT violation Jue is talking about in the other mail.

You can inject a #PF or better yet the *first* MCE which is being injected should say with a bit somehwere "I unmapped the address in m->addr". So that the guest kernel can handle that properly and know what *exactly* it is getting an MCE for.

That question only makes any sense if you know you are running as a guest and that someone else has unmapped the address. It's a meaningless question to ask if you are running bare metal. So we'd still have a check for FEATURE_HYPERVISOR

Quoted 7 lines

What I don't like is the "am I running as a guest" check. Because someone else would come later and say, err, I'm not virtualizing this portion of MCA either, lemme add another "am I guest" check.

Sure, it is a lot easier but when stuff like that starts spreading around in the MCE code, then we can just as well disable MCE when virtualized altogether. It would be a lot easier for everybody.

Maybe it isn't pretty. But I don't see another practical solution.

The VMM is doing exactly the right thing here. It should not trust that the guest will behave and not touch the poison location again. If/when the guest does touch the poison, the right action is for the VMM to fake a new machine check to the guest.

Theoretlcally the VMM could decode the instruction that the guest was trying to use on the poison page and decide "oh, this is that weird case in Linux where it's just trying to CLFLUSH the page. I'll just step the return IP past the CLFLUSH and let the guest continue".

But that doesn't sound at all reasonable to me (especially as the next step is to realize that Linux is going to repeat that for every cache line in the page, so you also want to VMM to fudge the register contents to skip to the end of the loop and avoid another 63 VMEXITs).

N.B. Linux wants to switch the page to uncacheable so that in the persistant memory case the filesytem code can continue to access the other "blocks" in the page, rather than lose all of them. That's futile in the case where the VMM took the whole 4K away. Maybe Dan needs to think about the guest case too.

-Tony

#11 lore raw

On Mon, May 18, 2020 at 11:26 AM Luck, Tony <[email protected]> wrote:
[..]

Quoted 5 lines

N.B. Linux wants to switch the page to uncacheable so that in the persistant memory case the filesytem code can continue to access the other "blocks" in the page, rather than lose all of them. That's futile in the case where the VMM took the whole 4K away. Maybe Dan needs to think about the guest case too.

I think increasing the blast-radius to 4K is the best we can do without a paravirt mechanism to coordinate errors. There's also the existing problem that set_mce_nospec() fails on pmem due to the memtype lookup.

#12 lore raw
In reply to Luck, Tony
Quoted 11 lines · Tony

Maybe it isn't pretty. But I don't see another practical solution.

The VMM is doing exactly the right thing here. It should not trust that the guest will behave and not touch the poison location again. If/when the guest does touch the poison, the right action is for the VMM to fake a new machine check to the guest.

Theoretlcally the VMM could decode the instruction that the guest was trying to use on the poison page and decide "oh, this is that weird case in Linux where it's just trying to CLFLUSH the page. I'll just step the return IP past the CLFLUSH and let the guest continue".

That's actually doable in KVM, e.g. a hack could be done in <10 lines of code. A proper fix that integrates with KVM's emulator would be substantially more code and effort though.

Quoted 4 lines

But that doesn't sound at all reasonable to me (especially as the next step is to realize that Linux is going to repeat that for every cache line in the page, so you also want to VMM to fudge the register contents to skip to the end of the loop and avoid another 63 VMEXITs).

Eh, 63 VM-Exits is peanuts in the grand scheme. Even with the host-side gup() that's probably less than 50us.

Quoted 5 lines

N.B. Linux wants to switch the page to uncacheable so that in the persistant memory case the filesytem code can continue to access the other "blocks" in the page, rather than lose all of them. That's futile in the case where the VMM took the whole 4K away. Maybe Dan needs to think about the guest case too.

This is where I'm unclear as to the guest behavior. Is it doing *just* CLFLUSH, or is it doing CLFLUSH followed by other accesses to the poisoned page? If it's the former, then it's probably worth at least exploring a KVM fix. If it's the latter, then yeah, emulating CLFLUSH for a poisoned #MC is pointless. I assume it's the latter since the goal is to recover data?

Oh, and FWIW, the guest won't actually get UC for that page.

maintainer
#13 lore raw
In reply to Luck, Tony
Quoted 4 lines · Tony

That question only makes any sense if you know you are running as a guest and that someone else has unmapped the address. It's a meaningless question to ask if you are running bare metal. So we'd still have a check for FEATURE_HYPERVISOR

Maybe I'm not making myself clear enough here: I'm talking about using a *special* MCE signature which says "your mapping has disappeared from underneath you." Maybe a bit in MCi_MISC which the hw doesn't use. Or some other deprecated bit, whatever.

If that signature is unique you won't have to check for hypervisor - you *know* it comes from it.

Because the hypervisor would be telling the guest: "I have removed the page from under you, you should act accordingly" with that signature. Vs the kernel going with the unspecific "am I running as a guest"?

See the huge difference?

Quoted 1 line

Maybe it isn't pretty. But I don't see another practical solution.

See above. Below too. I actually got two.

Quoted 4 lines

The VMM is doing exactly the right thing here. It should not trust that the guest will behave and not touch the poison location again. If/when the guest does touch the poison, the right action is for the VMM to fake a new machine check to the guest.

Yes, and that new machine check should tell the guest: "do not CLFLUSH this address - I've unmapped it and you don't have to do anything." Basically what your hypervisor check does but *actually* stating why it raised the second MCE.

Quoted 4 lines

Theoretlcally the VMM could decode the instruction that the guest was trying to use on the poison page and decide "oh, this is that weird case in Linux where it's just trying to CLFLUSH the page. I'll just step the return IP past the CLFLUSH and let the guest continue".

... or not inject the second MCE at all.

That would be fixing it in the HV.

Because there's this other way to look at it and come to think of it, fixing this in the HV makes a lot more sense. Why?

Well, let me elaborate:

The hypervisor just removed that page under the guest's feet and if that hypervisor wants to support unenlightened guests which cannot even deal with pages disappearing from under their feet, then that hypervisor better not inject that second MCE.

Why would it even inject the MCE - what can the guest even do about it? Exactly *nothing*. The page is unmapped and gone, the guest cannot salvage any information anymore from it.

And yes, the hypervisor has *all* the information, it knows which page it just removed so if the guest tries to access memory which HV just poisoned and is within the range which was covered by that page, then it should *not* inject that MCE. The guest can't handle it and why would it inject it - it is an access to a poisoned page which the HV *knows* it won't succeed so why bother?

The HV simply returns without injecting the MCE and so on, until the 4K page's end. It simply ignores guest accesses to the poisoned page. Without any guest changes.

Quoted 5 lines

N.B. Linux wants to switch the page to uncacheable so that in the persistant memory case the filesytem code can continue to access the other "blocks" in the page, rather than lose all of them. That's futile in the case where the VMM took the whole 4K away. Maybe Dan needs to think about the guest case too.

Yes, if the 4K page just went away, marking it UC doesn't make any sense.

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#14 lore raw

PATCH v2 1/1 x86/mm: Change so poison pages are either unmapped or marked uncacheable

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

Fix is to check the scope of the poison by checking the MCi_MISC register. If the entire page is affected, then unmap the page. If only part of the page is affected, then mark the page as uncacheable.

This assumes that VMMs will do the logical thing and pass in the "whole page scope" via the MCi_MISC register (since they unmapped the entire page).

Fixes 284ce4011ba6 ("x86/memory_failure: Introduce {set, clear}_mce_nospec()")

---

Quoted 4 lines · Boris

Maybe I'm not making myself clear enough here: I'm talking about using a *special* MCE signature which says "your mapping has disappeared from underneath you." Maybe a bit in MCi_MISC which the hw doesn't use. Or some other deprecated bit, whatever.

My reactions (in order)
1) How can I do that for 10 year old h/w?
2) Oh wait, Boris is talking about the VMM faking a signature to the guest. They can mess with whatever bits they want.
3) Unused/deprecated bits are in short supply. Taking one over and getting all the VMM vendors to agree to this "standard" is going to be a huge effort.
4) Could we do something inside the existing architecture?
5) MCi_MISC uses low order bits to indicate the "blast radius" of an error
6) Asks Jue "Could you set those MCi_MISC bits to say the whole page is gone?
7) Jue: We already do that!
8) Me: writes this patch

 arch/x86/include/asm/set_memory.h | 19 +++++++++++++------
 arch/x86/kernel/cpu/mce/core.c    | 11 +++++++++--
 include/linux/set_memory.h        |  2 +-
 3 files changed, 23 insertions(+), 9 deletions(-)
arch/x86/include/asm/set_memory.h +13 -6
index ec2c0a094b5d..5948218f35c5 100644
@@ -86,28 +86,35 @@ int set_direct_map_default_noflush(struct page *page);
 extern int kernel_set_to_readonly;
 
 #ifdef CONFIG_X86_64
-static inline int set_mce_nospec(unsigned long pfn)
+/*
+ * Prevent speculative access to the page by either unmapping
+ * it (if we do not require access to any part of the page) or
+ * marking it uncacheable (if we want to try to retrieve data
+ * from non-poisoned lines in the page).
+ */
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	unsigned long decoy_addr;
 	int rc;
 
 	/*
-	 * Mark the linear address as UC to make sure we don't log more
-	 * errors because of speculative access to the page.
 	 * We would like to just call:
-	 *      set_memory_uc((unsigned long)pfn_to_kaddr(pfn), 1);
+	 *      set_memory_XX((unsigned long)pfn_to_kaddr(pfn), 1);
 	 * but doing that would radically increase the odds of a
 	 * speculative access to the poison page because we'd have
 	 * the virtual address of the kernel 1:1 mapping sitting
 	 * around in registers.
 	 * Instead we get tricky.  We create a non-canonical address
 	 * that looks just like the one we want, but has bit 63 flipped.
-	 * This relies on set_memory_uc() properly sanitizing any __pa()
+	 * This relies on set_memory_XX() properly sanitizing any __pa()
 	 * results with __PHYSICAL_MASK or PTE_PFN_MASK.
 	 */
 	decoy_addr = (pfn << PAGE_SHIFT) + (PAGE_OFFSET ^ BIT(63));
 
-	rc = set_memory_uc(decoy_addr, 1);
+	if (unmap)
+		rc = set_memory_np(decoy_addr, 1);
+	else
+		rc = set_memory_uc(decoy_addr, 1);
 	if (rc)
 		pr_warn("Could not invalidate pfn=0x%lx from 1:1 map\n", pfn);
 	return rc;
arch/x86/kernel/cpu/mce/core.c +9 -2
index 54165f3569e8..c1a480a27164 100644
@@ -529,6 +529,13 @@ bool mce_is_memory_error(struct mce *m)
 }
 EXPORT_SYMBOL_GPL(mce_is_memory_error);
 
+static bool whole_page(struct mce *m)
+{
+	if (!mca_cfg.ser || !(m->status & MCI_STATUS_MISCV))
+		return true;
+	return MCI_MISC_ADDR_LSB(m->misc) >= PAGE_SHIFT;
+}
+
 bool mce_is_correctable(struct mce *m)
 {
 	if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED)
@@ -600,7 +607,7 @@ static int uc_decode_notifier(struct notifier_block *nb, unsigned long val,
 
 	pfn = mce->addr >> PAGE_SHIFT;
 	if (!memory_failure(pfn, 0))
-		set_mce_nospec(pfn);
+		set_mce_nospec(pfn, whole_page(mce));
 
 	return NOTIFY_OK;
 }
@@ -1098,7 +1105,7 @@ static int do_memory_failure(struct mce *m)
 	if (ret)
 		pr_err("Memory error not recovered");
 	else
-		set_mce_nospec(m->addr >> PAGE_SHIFT);
+		set_mce_nospec(m->addr >> PAGE_SHIFT, whole_page(m));
 	return ret;
 }
 
include/linux/set_memory.h +1 -1
index 86281ac7c305..860e0f843c12 100644
@@ -26,7 +26,7 @@ static inline int set_direct_map_default_noflush(struct page *page)
 #endif
 
 #ifndef set_mce_nospec
-static inline int set_mce_nospec(unsigned long pfn)
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	return 0;
 }
Signature · 1 line
2.21.1
merge bot Applied
#15 lore raw

[tip: ras/core] x86/{mce,mm}: Change so poison pages are either unmapped or marked uncacheable

The following commit has been merged into the ras/core branch of tip:

The following commit has been merged into the ras/core branch of tip:

Commit-ID:     3cb1ada80fe29e2fa022b5f20370b65718e0a744
Gitweb:        https://git.kernel.org/tip/3cb1ada80fe29e2fa022b5f20370b65718e0a744
Author:        Tony Luck <[email protected]>
AuthorDate:    Wed, 20 May 2020 09:35:46 -07:00
Committer:     Borislav Petkov <[email protected]>
CommitterDate: Mon, 25 May 2020 12:46:40 +02:00

x86/{mce,mm}: Change so poison pages are either unmapped or marked uncacheable

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

Fix is to check the scope of the poison by checking the MCi_MISC register. If the entire page is affected, then unmap the page. If only part of the page is affected, then mark the page as uncacheable.

This assumes that VMMs will do the logical thing and pass in the "whole page scope" via the MCi_MISC register (since they unmapped the entire page).

Fixes 284ce4011ba6 ("x86/memory_failure: Introduce {set, clear}_mce_nospec()")
---
 arch/x86/include/asm/set_memory.h | 19 +++++++++++++------
 arch/x86/kernel/cpu/mce/core.c    | 11 +++++++++--
 include/linux/set_memory.h        |  2 +-
 3 files changed, 23 insertions(+), 9 deletions(-)
arch/x86/include/asm/set_memory.h +13 -6
index ec2c0a0..5948218 100644
@@ -86,28 +86,35 @@ int set_direct_map_default_noflush(struct page *page);
 extern int kernel_set_to_readonly;
 
 #ifdef CONFIG_X86_64
-static inline int set_mce_nospec(unsigned long pfn)
+/*
+ * Prevent speculative access to the page by either unmapping
+ * it (if we do not require access to any part of the page) or
+ * marking it uncacheable (if we want to try to retrieve data
+ * from non-poisoned lines in the page).
+ */
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	unsigned long decoy_addr;
 	int rc;
 
 	/*
-	 * Mark the linear address as UC to make sure we don't log more
-	 * errors because of speculative access to the page.
 	 * We would like to just call:
-	 *      set_memory_uc((unsigned long)pfn_to_kaddr(pfn), 1);
+	 *      set_memory_XX((unsigned long)pfn_to_kaddr(pfn), 1);
 	 * but doing that would radically increase the odds of a
 	 * speculative access to the poison page because we'd have
 	 * the virtual address of the kernel 1:1 mapping sitting
 	 * around in registers.
 	 * Instead we get tricky.  We create a non-canonical address
 	 * that looks just like the one we want, but has bit 63 flipped.
-	 * This relies on set_memory_uc() properly sanitizing any __pa()
+	 * This relies on set_memory_XX() properly sanitizing any __pa()
 	 * results with __PHYSICAL_MASK or PTE_PFN_MASK.
 	 */
 	decoy_addr = (pfn << PAGE_SHIFT) + (PAGE_OFFSET ^ BIT(63));
 
-	rc = set_memory_uc(decoy_addr, 1);
+	if (unmap)
+		rc = set_memory_np(decoy_addr, 1);
+	else
+		rc = set_memory_uc(decoy_addr, 1);
 	if (rc)
 		pr_warn("Could not invalidate pfn=0x%lx from 1:1 map\n", pfn);
 	return rc;
arch/x86/kernel/cpu/mce/core.c +9 -2
index 02e1f16..e35aece 100644
@@ -518,6 +518,13 @@ bool mce_is_memory_error(struct mce *m)
 }
 EXPORT_SYMBOL_GPL(mce_is_memory_error);
 
+static bool whole_page(struct mce *m)
+{
+	if (!mca_cfg.ser || !(m->status & MCI_STATUS_MISCV))
+		return true;
+	return MCI_MISC_ADDR_LSB(m->misc) >= PAGE_SHIFT;
+}
+
 bool mce_is_correctable(struct mce *m)
 {
 	if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED)
@@ -571,7 +578,7 @@ static int uc_decode_notifier(struct notifier_block *nb, unsigned long val,
 
 	pfn = mce->addr >> PAGE_SHIFT;
 	if (!memory_failure(pfn, 0)) {
-		set_mce_nospec(pfn);
+		set_mce_nospec(pfn, whole_page(mce));
 		mce->kflags |= MCE_HANDLED_UC;
 	}
 
@@ -1069,7 +1076,7 @@ static int do_memory_failure(struct mce *m)
 	if (ret)
 		pr_err("Memory error not recovered");
 	else
-		set_mce_nospec(m->addr >> PAGE_SHIFT);
+		set_mce_nospec(m->addr >> PAGE_SHIFT, whole_page(m));
 	return ret;
 }
 
include/linux/set_memory.h +1 -1
index 86281ac..860e0f8 100644
@@ -26,7 +26,7 @@ static inline int set_direct_map_default_noflush(struct page *page)
 #endif
 
 #ifndef set_mce_nospec
-static inline int set_mce_nospec(unsigned long pfn)
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	return 0;
 }
maintainer Tested-by Jue Wang
#16 lore raw

Re: [tip: ras/core] x86/{mce,mm}: Change so poison pages are either unmapped or marked uncacheable

Quoted 8 lines · On Mon, May 25, 2020 at 11:00:03AM -0000, tip-bot2 for Tony Luck wrote:

The following commit has been merged into the ras/core branch of tip:

Commit-ID:     3cb1ada80fe29e2fa022b5f20370b65718e0a744
Gitweb:        https://git.kernel.org/tip/3cb1ada80fe29e2fa022b5f20370b65718e0a744
Author:        Tony Luck <[email protected]>
AuthorDate:    Wed, 20 May 2020 09:35:46 -07:00
Committer:     Borislav Petkov <[email protected]>
CommitterDate: Mon, 25 May 2020 12:46:40 +02:00

Ok, I had to change this one due to other pending changes in tip:x86/entry. The new version below.

Can you guys run this branch to make sure it still works as expected?

https://git.kernel.org/pub/scm/linux/kernel/git/bp/bp.git/log/?h=tip-ras-core

Thx.

---
From 4d37444a762f4c35289ac86fe880e018731701f9 Mon Sep 17 00:00:00 2001
From: Tony Luck <[email protected]>
Date: Wed, 20 May 2020 09:35:46 -0700
Subject: [PATCH] x86/{mce,mm}: Unmap the entire page if the whole page is
affected and poisoned

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

Fix is to check the scope of the poison by checking the MCi_MISC register. If the entire page is affected, then unmap the page. If only part of the page is affected, then mark the page as uncacheable.

This assumes that VMMs will do the logical thing and pass in the "whole page scope" via the MCi_MISC register (since they unmapped the entire page).

[ bp: Adjust to x86/entry changes. ]

Fixes 284ce4011ba6 ("x86/memory_failure: Introduce {set, clear}_mce_nospec()")
---
 arch/x86/include/asm/set_memory.h | 19 +++++++++++++------
 arch/x86/kernel/cpu/mce/core.c    | 18 ++++++++++++++----
 include/linux/sched.h             |  4 +++-
 include/linux/set_memory.h        |  2 +-
 4 files changed, 31 insertions(+), 12 deletions(-)
arch/x86/include/asm/set_memory.h +13 -6
index ec2c0a094b5d..5948218f35c5 100644
@@ -86,28 +86,35 @@ int set_direct_map_default_noflush(struct page *page);
 extern int kernel_set_to_readonly;
 
 #ifdef CONFIG_X86_64
-static inline int set_mce_nospec(unsigned long pfn)
+/*
+ * Prevent speculative access to the page by either unmapping
+ * it (if we do not require access to any part of the page) or
+ * marking it uncacheable (if we want to try to retrieve data
+ * from non-poisoned lines in the page).
+ */
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	unsigned long decoy_addr;
 	int rc;
 
 	/*
-	 * Mark the linear address as UC to make sure we don't log more
-	 * errors because of speculative access to the page.
 	 * We would like to just call:
-	 *      set_memory_uc((unsigned long)pfn_to_kaddr(pfn), 1);
+	 *      set_memory_XX((unsigned long)pfn_to_kaddr(pfn), 1);
 	 * but doing that would radically increase the odds of a
 	 * speculative access to the poison page because we'd have
 	 * the virtual address of the kernel 1:1 mapping sitting
 	 * around in registers.
 	 * Instead we get tricky.  We create a non-canonical address
 	 * that looks just like the one we want, but has bit 63 flipped.
-	 * This relies on set_memory_uc() properly sanitizing any __pa()
+	 * This relies on set_memory_XX() properly sanitizing any __pa()
 	 * results with __PHYSICAL_MASK or PTE_PFN_MASK.
 	 */
 	decoy_addr = (pfn << PAGE_SHIFT) + (PAGE_OFFSET ^ BIT(63));
 
-	rc = set_memory_uc(decoy_addr, 1);
+	if (unmap)
+		rc = set_memory_np(decoy_addr, 1);
+	else
+		rc = set_memory_uc(decoy_addr, 1);
 	if (rc)
 		pr_warn("Could not invalidate pfn=0x%lx from 1:1 map\n", pfn);
 	return rc;
arch/x86/kernel/cpu/mce/core.c +14 -4
index ffee8a2f435d..753bc7731f12 100644
@@ -520,6 +520,14 @@ bool mce_is_memory_error(struct mce *m)
 }
 EXPORT_SYMBOL_GPL(mce_is_memory_error);
 
+static bool whole_page(struct mce *m)
+{
+	if (!mca_cfg.ser || !(m->status & MCI_STATUS_MISCV))
+		return true;
+
+	return MCI_MISC_ADDR_LSB(m->misc) >= PAGE_SHIFT;
+}
+
 bool mce_is_correctable(struct mce *m)
 {
 	if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED)
@@ -573,7 +581,7 @@ static int uc_decode_notifier(struct notifier_block *nb, unsigned long val,
 
 	pfn = mce->addr >> PAGE_SHIFT;
 	if (!memory_failure(pfn, 0)) {
-		set_mce_nospec(pfn);
+		set_mce_nospec(pfn, whole_page(mce));
 		mce->kflags |= MCE_HANDLED_UC;
 	}
 
@@ -1173,11 +1181,12 @@ static void kill_me_maybe(struct callback_head *cb)
 	int flags = MF_ACTION_REQUIRED;
 
 	pr_err("Uncorrected hardware memory error in user-access at %llx", p->mce_addr);
-	if (!(p->mce_status & MCG_STATUS_RIPV))
+
+	if (!p->mce_ripv)
 		flags |= MF_MUST_KILL;
 
 	if (!memory_failure(p->mce_addr >> PAGE_SHIFT, flags)) {
-		set_mce_nospec(p->mce_addr >> PAGE_SHIFT);
+		set_mce_nospec(p->mce_addr >> PAGE_SHIFT, p->mce_whole_page);
 		return;
 	}
 
@@ -1331,7 +1340,8 @@ void noinstr do_machine_check(struct pt_regs *regs)
 		BUG_ON(!on_thread_stack() || !user_mode(regs));
 
 		current->mce_addr = m.addr;
-		current->mce_status = m.mcgstatus;
+		current->mce_ripv = !!(m.mcgstatus & MCG_STATUS_RIPV);
+		current->mce_whole_page = whole_page(&m);
 		current->mce_kill_me.func = kill_me_maybe;
 		if (kill_it)
 			current->mce_kill_me.func = kill_me_now;
include/linux/sched.h +3 -1
index 1d68ee36c583..6293fc2f3fc0 100644
@@ -1304,7 +1304,9 @@ struct task_struct {
 
 #ifdef CONFIG_X86_MCE
 	u64				mce_addr;
-	u64				mce_status;
+	__u64				mce_ripv : 1,
+					mce_whole_page : 1,
+					__mce_reserved : 62;
 	struct callback_head		mce_kill_me;
 #endif
 
include/linux/set_memory.h +1 -1
index 86281ac7c305..860e0f843c12 100644
@@ -26,7 +26,7 @@ static inline int set_direct_map_default_noflush(struct page *page)
 #endif
 
 #ifndef set_mce_nospec
-static inline int set_mce_nospec(unsigned long pfn)
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	return 0;
 }
Signature · 7 lines
2.21.0

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
author maintainer
#17 lore raw

RE: [tip: ras/core] x86/{mce,mm}: Change so poison pages are either unmapped or marked uncacheable

Quoted 6 lines

Ok, I had to change this one due to other pending changes in tip:x86/entry. The new version below.

Can you guys run this branch to make sure it still works as expected?

https://git.kernel.org/pub/scm/linux/kernel/git/bp/bp.git/log/?h=tip-ras-core

Tested the native case. We correctly try to set the page uncacheable because the scope of the error is a cache line.

I don't have the right setup to test the virtualization case. Maybe Jue can test again?

-Tony

A message that is not in the archive (sent off-list or not yet received): CAPcxDJ5arJojbY4pzOvYh=waSPd3X_JJb1_PSuzd+jQ0qbvFsA@mail.gmail.com

A message that is not in the archive (sent off-list or not yet received): CAPcxDJ54EgX-SaDV=Lm+a2-43O68LhomyYfYdCDz38HGJCkh7g@mail.gmail.com

maintainer
#20 lore raw

Re: [tip: ras/core] x86/{mce,mm}: Change so poison pages are either unmapped or marked uncacheable

Quoted 9 lines · Jue Wang
Quoted 4 lines · Jue Wang

I tried to test this but my guest image build setup was not able to build from kernel/git/bp/bp.git tip-ras-core branch. It appeared there was some bindeb-pkg issue.

The bindeb-pkg issue is resolved and I tested the following branch in KVM guest and the injected MCE is recovered.
https://git.kernel.org/pub/scm/linux/kernel/git/bp/bp.git/log/?h=tip-ras-core

Thanks to both of you!

Signature · 4 lines
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette
merge bot Applied
#21 lore raw

[tip: ras/core] x86/{mce,mm}: Unmap the entire page if the whole page is affected and poisoned

In reply to Luck, Tony
The following commit has been merged into the ras/core branch of tip:

The following commit has been merged into the ras/core branch of tip:

Commit-ID:     be69f6c5cd38c457c22f6e718077f6524437369d
Gitweb:        https://git.kernel.org/tip/be69f6c5cd38c457c22f6e718077f6524437369d
Author:        Tony Luck <[email protected]>
AuthorDate:    Wed, 20 May 2020 09:35:46 -07:00
Committer:     Borislav Petkov <[email protected]>
CommitterDate: Mon, 25 May 2020 22:37:41 +02:00

x86/{mce,mm}: Unmap the entire page if the whole page is affected and poisoned

An interesting thing happened when a guest Linux instance took a machine check. The VMM unmapped the bad page from guest physical space and passed the machine check to the guest.

Linux took all the normal actions to offline the page from the process that was using it. But then guest Linux crashed because it said there was a second machine check inside the kernel with this stack trace:

do_memory_failure
set_mce_nospec
set_memory_uc
_set_memory_uc
change_page_attr_set_clr
cpa_flush
clflush_cache_range_opt

This was odd, because a CLFLUSH instruction shouldn't raise a machine check (it isn't consuming the data). Further investigation showed that the VMM had passed in another machine check because is appeared that the guest was accessing the bad page.

Fix is to check the scope of the poison by checking the MCi_MISC register. If the entire page is affected, then unmap the page. If only part of the page is affected, then mark the page as uncacheable.

This assumes that VMMs will do the logical thing and pass in the "whole page scope" via the MCi_MISC register (since they unmapped the entire page).

[ bp: Adjust to x86/entry changes. ]

Fixes 284ce4011ba6 ("x86/memory_failure: Introduce {set, clear}_mce_nospec()")
---
 arch/x86/include/asm/set_memory.h | 19 +++++++++++++------
 arch/x86/kernel/cpu/mce/core.c    | 18 ++++++++++++++----
 include/linux/sched.h             |  4 +++-
 include/linux/set_memory.h        |  2 +-
 4 files changed, 31 insertions(+), 12 deletions(-)
arch/x86/include/asm/set_memory.h +13 -6
index ec2c0a0..5948218 100644
@@ -86,28 +86,35 @@ int set_direct_map_default_noflush(struct page *page);
 extern int kernel_set_to_readonly;
 
 #ifdef CONFIG_X86_64
-static inline int set_mce_nospec(unsigned long pfn)
+/*
+ * Prevent speculative access to the page by either unmapping
+ * it (if we do not require access to any part of the page) or
+ * marking it uncacheable (if we want to try to retrieve data
+ * from non-poisoned lines in the page).
+ */
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	unsigned long decoy_addr;
 	int rc;
 
 	/*
-	 * Mark the linear address as UC to make sure we don't log more
-	 * errors because of speculative access to the page.
 	 * We would like to just call:
-	 *      set_memory_uc((unsigned long)pfn_to_kaddr(pfn), 1);
+	 *      set_memory_XX((unsigned long)pfn_to_kaddr(pfn), 1);
 	 * but doing that would radically increase the odds of a
 	 * speculative access to the poison page because we'd have
 	 * the virtual address of the kernel 1:1 mapping sitting
 	 * around in registers.
 	 * Instead we get tricky.  We create a non-canonical address
 	 * that looks just like the one we want, but has bit 63 flipped.
-	 * This relies on set_memory_uc() properly sanitizing any __pa()
+	 * This relies on set_memory_XX() properly sanitizing any __pa()
 	 * results with __PHYSICAL_MASK or PTE_PFN_MASK.
 	 */
 	decoy_addr = (pfn << PAGE_SHIFT) + (PAGE_OFFSET ^ BIT(63));
 
-	rc = set_memory_uc(decoy_addr, 1);
+	if (unmap)
+		rc = set_memory_np(decoy_addr, 1);
+	else
+		rc = set_memory_uc(decoy_addr, 1);
 	if (rc)
 		pr_warn("Could not invalidate pfn=0x%lx from 1:1 map\n", pfn);
 	return rc;
arch/x86/kernel/cpu/mce/core.c +14 -4
index ffee8a2..753bc77 100644
@@ -520,6 +520,14 @@ bool mce_is_memory_error(struct mce *m)
 }
 EXPORT_SYMBOL_GPL(mce_is_memory_error);
 
+static bool whole_page(struct mce *m)
+{
+	if (!mca_cfg.ser || !(m->status & MCI_STATUS_MISCV))
+		return true;
+
+	return MCI_MISC_ADDR_LSB(m->misc) >= PAGE_SHIFT;
+}
+
 bool mce_is_correctable(struct mce *m)
 {
 	if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED)
@@ -573,7 +581,7 @@ static int uc_decode_notifier(struct notifier_block *nb, unsigned long val,
 
 	pfn = mce->addr >> PAGE_SHIFT;
 	if (!memory_failure(pfn, 0)) {
-		set_mce_nospec(pfn);
+		set_mce_nospec(pfn, whole_page(mce));
 		mce->kflags |= MCE_HANDLED_UC;
 	}
 
@@ -1173,11 +1181,12 @@ static void kill_me_maybe(struct callback_head *cb)
 	int flags = MF_ACTION_REQUIRED;
 
 	pr_err("Uncorrected hardware memory error in user-access at %llx", p->mce_addr);
-	if (!(p->mce_status & MCG_STATUS_RIPV))
+
+	if (!p->mce_ripv)
 		flags |= MF_MUST_KILL;
 
 	if (!memory_failure(p->mce_addr >> PAGE_SHIFT, flags)) {
-		set_mce_nospec(p->mce_addr >> PAGE_SHIFT);
+		set_mce_nospec(p->mce_addr >> PAGE_SHIFT, p->mce_whole_page);
 		return;
 	}
 
@@ -1331,7 +1340,8 @@ void noinstr do_machine_check(struct pt_regs *regs)
 		BUG_ON(!on_thread_stack() || !user_mode(regs));
 
 		current->mce_addr = m.addr;
-		current->mce_status = m.mcgstatus;
+		current->mce_ripv = !!(m.mcgstatus & MCG_STATUS_RIPV);
+		current->mce_whole_page = whole_page(&m);
 		current->mce_kill_me.func = kill_me_maybe;
 		if (kill_it)
 			current->mce_kill_me.func = kill_me_now;
include/linux/sched.h +3 -1
index 1d68ee3..6293fc2 100644
@@ -1304,7 +1304,9 @@ struct task_struct {
 
 #ifdef CONFIG_X86_MCE
 	u64				mce_addr;
-	u64				mce_status;
+	__u64				mce_ripv : 1,
+					mce_whole_page : 1,
+					__mce_reserved : 62;
 	struct callback_head		mce_kill_me;
 #endif
 
include/linux/set_memory.h +1 -1
index 86281ac..860e0f8 100644
@@ -26,7 +26,7 @@ static inline int set_direct_map_default_noflush(struct page *page)
 #endif
 
 #ifndef set_mce_nospec
-static inline int set_mce_nospec(unsigned long pfn)
+static inline int set_mce_nospec(unsigned long pfn, bool unmap)
 {
 	return 0;
 }
#22 lore raw
In reply to Borislav Petkov
Quoted 23 lines · Borislav Petkov
Quoted 7 lines · Tony

The VMM gets the page fault (because the unmapping of the guest physical address is at the VMM EPT level). The VMM can't map a new page into that guest physical address because it has no way to replace the contents of the old page. The VMM could pass the #PF to the guest, but that would just confuse the guest (its page tables all say that the page is still valid). In this particular case the page is part of the 1:1 kernel map. So the kernel will OOPS (I think).

...

Quoted 3 lines

PLease explain how a guest (that doesn't even know that it is a guest) is going to figure out that the EPT tables (that it has no way to access) have marked this page invalid in guest physical address space.

So somewhere BUS_MCEERR_AR was mentioned. So I'm assuming the error severity was "action required". What does happen in the kernel, on baremetal, with an AR error in kernel space, i.e., kernel memory?

If we can't fixup the exception, we die.

So why should the guest behave any differently?

Quoted 7 lines

Now, if you want for the guest to be more "robust" and handle that thing, fine. But then you'd need an explicit way to tell the guest
kernel: "you've just had an MCE and I unmapped the page" so that the guest kernel can figure out what do to. Even if it means, to panic.

I.e., signal in an explicit way that EPT violation Jue is talking about in the other mail.

Well, technically the CLFUSH thing is a KVM emulation bug, but it sounds like that's a moot point since the pmem-enabled guest will make real accesses to the poisoned page shortly thereafter. E.g. teaching KVM to eat the -EHWPOISON on CLFLUSH would only postpone the guest's death.

As for how the second #MC occurs, on the EPT violation, KVM does a gup() to translate the virtual address to a pfn (KVM maintains a simple GPA->HVA lookup). gup() returns -EHWPOISON for the poisoned page, which KVM redirects into a BUS_MCEERR_AR. The userspace VMM, e.g. Qemu, sees the BUS_MCEERR_AR and sends it back into the guest as a virtual #MC.

Quoted 12 lines

You can inject a #PF or better yet the *first* MCE which is being injected should say with a bit somehwere "I unmapped the address in m->addr". So that the guest kernel can handle that properly and know what *exactly* it is getting an MCE for.

What I don't like is the "am I running as a guest" check. Because someone else would come later and say, err, I'm not virtualizing this portion of MCA either, lemme add another "am I guest" check.

Sure, it is a lot easier but when stuff like that starts spreading around in the MCE code, then we can just as well disable MCE when virtualized altogether. It would be a lot easier for everybody.