> On Dec 23, 2020, at 6:00 PM, Andrea Arcangeli <aarcange@xxxxxxxxxx> wrote: > > On Wed, Dec 23, 2020 at 05:21:43PM -0800, Andy Lutomirski wrote: >> I don’t love this as a long term fix. AFAICT we can have mm_tlb_flush_pending set for quite a while — mprotect seems like it can wait in IO while splitting a huge page, for example. That gives us a window in which every write fault turns into a TLB flush. > > mprotect can't run concurrently with a page fault in the first place. > > One other near zero cost improvement easy to add if this would be "if > (vma->vm_flags & (VM_SOFTDIRTY|VM_UFFD_WP))" and it could be made > conditional to the two config options too. > > Still I don't mind doing it in some other way, uffd-wp has much easier > time doing it in another way in fact. > > Whatever performs better is fine, but queuing up pending invalidate > ranges don't look very attractive since it'd be a fixed cost that we'd > always have to pay even when there's no fault (and there can't be any > fault at least for mprotect). I think there are other cases in which Andy’s concern is relevant (MADV_PAGEOUT). Perhaps holding some small bitmap based on part of the deferred flushed pages (e.g., bits 12-17 of the address or some other kind of a single hash-function bloom-filter) would be more performant to avoid (most) unnecessary TLB flushes. It will be cleared before a TLB flush and set while holding the PTL. Checking if a flush is needed, under the PTL, would require a single memory access (although potentially cache miss). It will however require one atomic operation for each page-table whose PTEs’ flushes are deferred - in contrast to the current scheme which requires two atomic operations for the *entire* operation.
