On 26.08.22 17:55, Peter Xu wrote:
On Fri, Aug 26, 2022 at 04:47:22PM +0200, David Hildenbrand wrote:
To me anon exclusive only shows this mm exclusively owns this page. I didn't quickly figure out why that requires different handling on tlb flushs. Did I perhaps miss something?
GUP-fast is the magic bit, we have to make sure that we won't see new GUP pins, thus the TLB flush.
include/linux/mm.h:gup_must_unshare() contains documentation.
Hmm.. Shouldn't ptep_get_and_clear() (e.g., xchg() on x86_64) already guarantees that no other process/thread will see this pte anymore afterwards?
You could have a GUP-fast thread that just looked up the PTE and is going to pin the page afterwards, after the ptep_get_and_clear() returned. You'll have to wait until that thread finished.
Another user that relies on this interaction between GUP-fast and TLB flushing is for example mm/ksm.c:write_protect_page()
There is a comment in there explaining the interaction a bit more detailed.
Maybe we'll be able to handle this differently in the future (maybe once this turns out to be an actual performance problem). Unfortunately, mm->write_protect_seq isn't easily usable because we'd need have to make sure we're the exclusive writer.
For now, it's not too complicated. For PTEs: * try_to_migrate_one() already uses ptep_clear_flush(). * try_to_unmap_one() already conditionally used ptep_clear_flush(). * migrate_vma_collect_pmd() was the one case that didn't use it already (and I wonder why it's different than try_to_migrate_one()).