On Dec 23, 2020, at 7:34 PM, Yu Zhao yuzhao@google.com wrote:
On Wed, Dec 23, 2020 at 07:09:10PM -0800, Nadav Amit wrote:
On Dec 23, 2020, at 6:00 PM, Andrea Arcangeli aarcange@redhat.com 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).
That patch only demonstrate a rough idea and I should have been elaborate: if we ever decide to go that direction, we only need to worry about "jumping through hoops", because the final patch (set) I have in mind would not only have the build time optimization Andrea suggested but also include runtime optimizations like skipping do_swap_page() path and (!PageAnon() || page_mapcount > 1). Rest assured, the performance impact on do_wp_page() from occasionally an additional TLB flush on top of a page copy is negligible.
I agree with you to a certain extent, since there is anyhow another TLB flush in this path when the PTE is set after copying.
Yet, I think that having a combined and efficient central mechanism for pending TLB flushes is important even for robustness: to prevent the development of new independent deferred flushing schemes. I specifically do not like tlb_flush_batched which every time that I look at gets me confused. For example the following code completely confuses me:
void flush_tlb_batched_pending(struct mm_struct *mm) { if (data_race(mm->tlb_flush_batched)) { flush_tlb_mm(mm);
/* * Do not allow the compiler to re-order the clearing of * tlb_flush_batched before the tlb is flushed. */ barrier(); mm->tlb_flush_batched = false; } }
… and then I ask myself (no answer):
1. What prevents concurrent flush_tlb_batched_pending() which is called by madvise_free_pte_range(), for instance from madvise_free_pte_range(), from clearing new deferred flush indication that was just set by set_tlb_ubc_flush_pending()? Can it cause a missed TLB flush later?
2. Why the write to tlb_flush_batched is not done with WRITE_ONCE()?
3. Should we have smp_wmb() instead of barrier()? (probably the barrier() is not needed at all since flush_tlb_mm() serializes if a flush is needed).
4. Why do we need 2 deferred TLB flushing mechanisms?