On Tue, 2012-05-01 at 13:47 +0300, Avi Kivity wrote: > On 05/01/2012 12:39 PM, Peter Zijlstra wrote: > > On Sun, 2012-04-29 at 15:23 +0300, Avi Kivity wrote: > > > On 04/27/2012 07:24 PM, Nikunj A. Dadhania wrote: > > > > flush_tlb_others_ipi depends on lot of statics in tlb.c. Replicated > > > > the flush_tlb_others_ipi as kvm_flush_tlb_others to further adapt to > > > > paravirtualization. > > > > > > > > Use the vcpu state information inside the kvm_flush_tlb_others to > > > > avoid sending ipi to pre-empted vcpus. > > > > > > > > * Do not send ipi's to offline vcpus and set flush_on_enter flag > > > > > > get_user_pages_fast() depends on the IPI to hold off page table teardown > > > while they are locklessly walked with interrupts disabled. If a vcpu > > > were to be preempted while in this critical section, another vcpu > > > tearing down page tables would go ahead and destroy them. when the > > > preempted vcpu resumes it then touches the freed pages. > > > > > > We could try to teach kvm and get_user_pages_fast() about this, but this > > > is intrusive. Another option is to replace the cpu_relax() loop with > > > something that sleeps and is then woken up by the TLB IPI handler if needed. > > > > I think something like > > > > select HAVE_RCU_TABLE_FREE if PARAVIRT > > > > or somesuch is just about all it takes. > > > > A slightly better option would be to wrap all that tlb_*table* goo into > > paravirt stuff and only do the RCU free when paravirt is indeed enabled, > > but other than that you're there. > > I infer from this that there is a cost involved with rcu freeing. Any > idea how much? No idea, so far that code has only been used on platforms that required it so they didn't have a choice in the matter. > Looks like this increases performance for the overcommitted case, and > also for the case where many vcpus are sleeping, while reducing > performance for the uncontended, high duty cycle case. Sounds backwards if you put it like that ;-) > > This should work because the preempted vcpu's RCU state would also be > > stalled and thus avoids the actual page-table from going away. > > It can be unstalled at any moment. But spin_lock_irq() > rcu_read_lock(). Right, but since gup_fast has IRQs disabled the RCU state machine (as driven by the tick) won't actually do anything until its done. To be clear, the case was where the gup_fast() performing vcpu was preempted in the middle of gup_fast(), on wakeup it would perform the TLB flush on the virt-enter hook, but meanwhile a sibling vcpu might have free'd the page-tables. By using call_rcu_sched() to free the page-tables you'd need to receive and process at least one tick on the woken up cpu after the freeing, but since the in-progress gup_fast() will have IRQs disabled this will be delayed. Anyway, I don't have any idea about the costs involved with HAVE_RCU_TABLE_FREE, but I don't think its much.. otherwise these other platforms (PPC,SPARC) wouldn't have used it, gup_fast() is a very specific case, whereas mmu-gather is something affecting pretty much all tasks. But mostly my comment was due to you saying modifying gup_fast() would be difficult.. I was thinking the one Kconfig line wasn't as onerous ;-) -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html