Messages in this thread |  | | Date | Wed, 14 Dec 2022 12:22:49 +0100 | Subject | Re: [PATCH mm-unstable RFC 00/26] mm: support __HAVE_ARCH_PTE_SWP_EXCLUSIVE on all architectures with swap PTEs | From | David Hildenbrand <> |
| |
On 06.12.22 15:47, David Hildenbrand wrote: > This is the follow-up on [1]: > [PATCH v2 0/8] mm: COW fixes part 3: reliable GUP R/W FOLL_GET of > anonymous pages > > After we implemented __HAVE_ARCH_PTE_SWP_EXCLUSIVE on most prominent > enterprise architectures, implement __HAVE_ARCH_PTE_SWP_EXCLUSIVE on all > remaining architectures that support swap PTEs. > > This makes sure that exclusive anonymous pages will stay exclusive, even > after they were swapped out -- for example, making GUP R/W FOLL_GET of > anonymous pages reliable. Details can be found in [1]. > > This primarily fixes remaining known O_DIRECT memory corruptions that can > happen on concurrent swapout, whereby we can lose DMA reads to a page > (modifying the user page by writing to it). > > To verify, there are two test cases (requiring swap space, obviously): > (1) The O_DIRECT+swapout test case [2] from Andrea. This test case tries > triggering a race condition. > (2) My vmsplice() test case [3] that tries to detect if the exclusive > marker was lost during swapout, not relying on a race condition. > > > For example, on 32bit x86 (with and without PAE), my test case fails > without these patches: > $ ./test_swp_exclusive > FAIL: page was replaced during COW > But succeeds with these patches: > $ ./test_swp_exclusive > PASS: page was not replaced during COW > > > Why implement __HAVE_ARCH_PTE_SWP_EXCLUSIVE for all architectures, even > the ones where swap support might be in a questionable state? This is the > first step towards removing "readable_exclusive" migration entries, and > instead using pte_swp_exclusive() also with (readable) migration entries > instead (as suggested by Peter). The only missing piece for that is > supporting pmd_swp_exclusive() on relevant architectures with THP > migration support. > > As all relevant architectures now implement __HAVE_ARCH_PTE_SWP_EXCLUSIVE,, > we can drop __HAVE_ARCH_PTE_SWP_EXCLUSIVE in the last patch. > > > RFC because some of the swap PTE layouts are really tricky and I really > need some feedback related to deciphering these layouts and "using yet > unused PTE bits in swap PTEs". I tried cross-compiling all relevant setups > (phew, I might only miss some power/nohash variants), but only tested on > x86 so far.
As I was messing with sparc64 either way and got debian to boot under QEMU, I verified that the sparc64 change also seems to work as expected (under sun4u).
-- Thanks,
David / dhildenb
|  |