On Thu, 03 May 2012 15:47:26 +0300 Avi Kivity <avi@xxxxxxxxxx> wrote: > On 05/03/2012 03:29 PM, Peter Zijlstra wrote: > > On Thu, 2012-05-03 at 21:22 +0900, Takuya Yoshikawa wrote: > > > Although the real use case is out of this RFC patch, we are now discussing > > > a case in which we may hold a spin_lock for long time, ms order, depending > > > on workload; and in that case, other threads -- VCPU threads -- should be > > > given higher priority for that problematic lock. > > > > Firstly, if you can hold a lock that long, it shouldn't be a spinlock, > > In fact with your mm preemptibility work it can be made into a mutex, if > the entire mmu notifier path can be done in task context. However it > ends up a strange mutex - you can sleep while holding it but you may not > allocate, because you might recurse into an mmu notifier again. > > Most uses of the lock only involve tweaking some bits though. I might find a real way to go. After your "mmu_lock -- TLB-flush" decoupling, we can change the current get_dirty work flow like this: for ... { take mmu_lock for 4K*8 gfns { // with 4KB dirty_bitmap_buffer xchg dirty bits // 64/32 gfns at once write protect them } release mmu_lock copy_to_user } TLB flush This reduces the size of dirty_bitmap_buffer and does not hold mmu_lock so long. I should have think of a way not to hold the spin_lock so long as Peter said. My lack of thinking might be the real problem. Thanks, Takuya -- 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