On Sat, Oct 09, 2010 at 08:48:15PM +0200, Avi Kivity wrote: > On 10/07/2010 08:03 PM, Gleb Natapov wrote: > >> >> > >> >Host side keeps track of outstanding apfs and will not send apf for the > >> >same phys address twice. It will halt vcpu instead. > >> > >> What about different pages, running the scheduler code? > >> > >We can get couple of nested apfs, just like we can get nested > >interrupts. Since scheduler disables preemption second apf will halt. > > How much is a couple? > > Consider: > > SIGSTOP > Entire process swapped out > SIGCONT > > We can get APF's on the current code, the scheduler code, the stack, > any debugging code in between (e.g. ftrace), and the page tables for > all of these. > Lets count them all. Suppose guest is in a userspace process code and guest memory is completely swapped out. Guest starts to run and faults in userspace. Apf is queued but can't be delivered due to faults in idt and exception stack. All of them will be taken synchronously due to event pending check. After apf is delivered any fault in apf code will be takes synchronously since interrupt are disabled. Just before calling schedule() interrupts are enabled, so next pf that will happen during call to schedule() will be taken asynchronously. Which will cause another call to schedule() at which point vcpu will be halted since two apfs happened at the same address. So I counted two of them. -- Gleb. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>