On Mon, Dec 21, 2020 at 10:31:57AM -0800, Nadav Amit wrote:
On Dec 21, 2020, at 9:27 AM, Peter Xu peterx@redhat.com wrote:
Hi, Nadav,
On Sun, Dec 20, 2020 at 12:06:38AM -0800, Nadav Amit wrote:
[...]
So to correct myself, I think that what I really encountered was actually during MM_CP_UFFD_WP_RESOLVE (i.e., when the protection is removed). The problem was that in this case the “write”-bit was removed during unprotect. Sorry for the strange formatting to fit within 80 columns:
I assume I can ignore the race mentioned in the commit message but only refer to this one below. However I'm still confused. Please see below.
[ Start: PTE is writable ]
cpu0 cpu1 cpu2
[ Writable PTE cached in TLB ]Here cpu2 got writable pte in tlb. But why?
If below is an unprotect, it means it must have been protected once by userfaultfd, right? If so, the previous change_protection_range() which did the wr-protect should have done a tlb flush already before it returns (since pages>0 - we protected one pte at least). Then I can't see why cpu2 tlb has stall data.
Thanks, Peter. Just as you can munprotect() a region which was not protected before, you can ufff-unprotect a region that was not protected before. It might be that the user tried to unprotect a large region, which was partially protected and partially unprotected.
The selftest obviously blindly unprotect some regions to check for bugs.
So to your question - it was not write-protected (think about initial copy without write-protecting).
If that's the only case, how about we don't touch the ptes at all? Instead of playing with preserve_write, I'm thinking something like this right before ptep_modify_prot_start(), even for uffd_wp==true:
if (uffd_wp && pte_uffd_wp(old_pte)) { WARN_ON_ONCE(pte_write(old_pte)); continue; }
if (uffd_wp_resolve && !pte_uffd_wp(old_pte)) continue;
Then we can also avoid the heavy operations on changing ptes back and forth.
If I assume cpu2 doesn't have that cached tlb, then "write to old page" won't happen either, because cpu1/cpu2 will all go through the cow path and pgtable lock should serialize them.
userfaultfd_writeprotect() [ write-*unprotect* ] mwriteprotect_range() mmap_read_lock() change_protection()
change_protection_range() ... change_pte_range() [ *clear* “write”-bit ] [ defer TLB flushes] [ page-fault ] … wp_page_copy() cow_user_page() [ copy page ] [ write to old page ] … set_pte_at_notify()
[ End: cpu2 write not copied form old to new page. ]
Could you share how to reproduce the problem? I would be glad to give it a shot as well.
You can run the selftests/userfaultfd with my small patch [1]. I ran it with the following parameters: “ ./userfaultfd anon 100 100 “. I think that it is more easily reproducible with “mitigations=off idle=poll” as kernel parameters.
Thanks.
PS: Sorry to not have read the other series of yours. It seems to need some chunk of time so I postponed it a bit due to other things; but I'll read at least the fixes very soon.
Thanks again, I will post RFCv2 with some numbers soon.
I read the patch 1/3 of the series. Would it be better to post them separately just in case Andrew would like to pick them earlier?
Since you seem to be heavily working on uffd-wp - I do still have a few uffd-wp fixes locally even for anonymous. I think they're related to some corner cases like either thp or migration entry convertions, but anyway I'll see whether I should post them even earlier (I planned to add smap/pagemap support for uffd-wp so maybe I can even write some test case to verify some of them). Just a FYI...
Thanks,