Skip to contentSkip to conversation

x86/sev: Split up runtime #VC handler for correct state tracking

bug report Dan Carpenter 2 participants 7 messages Aug 4, 2021 – Aug 6, 2021 KVM

lore
Share & tools
lore
#0 lore raw

Hello Joerg Roedel,

The patch be1a5408868a: "x86/sev: Split up runtime #VC handler for correct state tracking" from Jun 18, 2021, leads to the following static checker warning:

	arch/x86/kernel/kvm.c:153 kvm_async_pf_task_wait_schedule()
	warn: sleeping in atomic context

arch/x86/kernel/sev.c
1416 * Handle #DB before calling into !noinstr code to avoid recursive #DB.
1417 */
1418 if (vc_is_db(error_code)) {
1419 exc_debug(regs);
1420 return;
1421 }
1422

  1423          irq_state = irqentry_nmi_enter(regs);
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

preempt disabled inside irqentry_nmi_enter().

  1424  
  1425          instrumentation_begin();
  1426  
  1427          if (!vc_raw_handle_exception(regs, error_code)) {
                     ^^^^^^^^^^^^^^^^^^^^^^^^

These sleeping in atomic static checker warnings come with a lot of caveats because the call tree is very long and it's easy to have false positives.

--> vc_raw_handle_exception()

    --> vc_forward_exception()
        --> exc_page_fault()

Page faults always sleep right?

  1428                  /* Show some debug info */
  1429                  show_regs(regs);
  1430  
  1431                  /* Ask hypervisor to sev_es_terminate */
  1432                  sev_es_terminate(GHCB_SEV_ES_REASON_GENERAL_REQUEST);
  1433  
  1434                  /* If that fails and we get here - just panic */
  1435                  panic("Returned from Terminate-Request to Hypervisor\n");
  1436          }
  1437  
  1438          instrumentation_end();
  1439          irqentry_nmi_exit(regs, irq_state);
  1440  }

regards,
dan carpenter

#1 lore raw

Hi Dan,

Quoted 9 lines · Dan Carpenter

These sleeping in atomic static checker warnings come with a lot of caveats because the call tree is very long and it's easy to have false positives.

--> vc_raw_handle_exception()

    --> vc_forward_exception()
        --> exc_page_fault()

Page faults always sleep right?

No, page faults do no always sleep, only when IO needs to be done to fulfill the page fault request. In this case, the page-fault handler will never sleep, because it is called with preemption disabled. The page-fault handler can detect this and just do nothing. The #VC handler will return for re-fault in this case.

Hope that helps,

Joerg

#2 lore raw
Quoted 11 lines · Joerg Roedel

Hi Dan,

Quoted 8 lines · Dan Carpenter

These sleeping in atomic static checker warnings come with a lot of caveats because the call tree is very long and it's easy to have false positives.

--> vc_raw_handle_exception()

    --> vc_forward_exception()
        --> exc_page_fault()
Quoted 7 lines
Quoted 1 line

Page faults always sleep right?

No, page faults do no always sleep, only when IO needs to be done to fulfill the page fault request. In this case, the page-fault handler will never sleep, because it is called with preemption disabled. The page-fault handler can detect this and just do nothing. The #VC handler will return for re-fault in this case.

Hm... Ok. Let give you the rest of the call tree then because I'm not seeing where it checks preempt count.

exc_page_fault() <-- called with preempt disabled
--> kvm_handle_async_pf()

    --> __kvm_handle_async_pf()
        --> kvm_async_pf_task_wait_schedule() calls schedule().

regards,
dan carpenter

#3 lore raw
Quoted 2 lines · Dan Carpenter

Hm... Ok. Let give you the rest of the call tree then because I'm not seeing where it checks preempt count.

The check is in faulthandler_disabled(), it checks for in_atomic().

Regards,

Joerg

#4 lore raw
Quoted 6 lines · Joerg Roedel
Quoted 2 lines · Dan Carpenter

Hm... Ok. Let give you the rest of the call tree then because I'm not seeing where it checks preempt count.

The check is in faulthandler_disabled(), it checks for in_atomic().

That check is later after the caller tree that I pasted as already called schedule().

regards,
dan carpenter

#5 lore raw
In reply to Dan Carpenter
Quoted 4 lines · Dan Carpenter

exc_page_fault() <-- called with preempt disabled
--> kvm_handle_async_pf()

    --> __kvm_handle_async_pf()
        --> kvm_async_pf_task_wait_schedule() calls schedule().

This call path can not be taken in the page-fault handler when called from the #VC handler. To take this path the host needs to inject an async page-fault, especially setting async pf flags, without injecting a page-fault exception on its own ... and when the #VC handler is running. KVM is not doing that.

Okay, the hypervisor can be malicious, but otherwise this can't happen. To mitigate a malicious hypervisor threat here it might be a solution to not call the page-fault handler directly from the #VC handler and let it re-fault after the #VC handler returned.

Regards,

Joerg

#6 lore raw
Quoted 12 lines · Joerg Roedel
Quoted 4 lines · Dan Carpenter

exc_page_fault() <-- called with preempt disabled
--> kvm_handle_async_pf()

    --> __kvm_handle_async_pf()
        --> kvm_async_pf_task_wait_schedule() calls schedule().

This call path can not be taken in the page-fault handler when called from the #VC handler. To take this path the host needs to inject an async page-fault, especially setting async pf flags, without injecting a page-fault exception on its own ... and when the #VC handler is running. KVM is not doing that.

Quoted 5 lines

Okay, the hypervisor can be malicious, but otherwise this can't happen. To mitigate a malicious hypervisor threat here it might be a solution to not call the page-fault handler directly from the #VC handler and let it re-fault after the #VC handler returned.

Thanks for taking a look at this.

Also it turns out that my check wasn't taking in_atomic() into consideration either so I've added that.

regards,
dan carpenter