This ports the switching from page to folio in add_to_swap_cache().
We saw multiple different hangs on mmap_lock, where the task holding the
lock was livelocked spinning in this stack:
+-> __x64_sys_ioctl
+-> kvm_vcpu_ioctl
+-> kvm_arch_vcpu_ioctl_run
+-> vcpu_run
+-> vcpu_enter_guest
+-> kvm_mmu_page_fault
+-> kvm_tdp_page_fault
+-> kvm_faultin_pfn
+-> __kvm_faultin_pfn
+-> hva_to_pfn
+-> get_user_pages_unlocked
+-> get_user_pages_unlocked
+-> mmap_read_lock # 1
+-> __get_user_pages_locked # 2
+-> for-loop # taken once
+-> __get_user_pages
+-> retry-loop # constantly spinning
+-> faultin_page # return 0 to trigger retry
+-> handle_mm_fault
+-> __handle_mm_fault
+-> handle_pte_fault
+-> do_swap_page
+-> lookup_swap_cache # returns non-NULL
+-> if (swapcache)
+-> if (!folio_test_swapcache || page_private(page) !=
entry.val)
+-> goto out_page
+-> return 0
That can be due to an inconsistency in swapcache flag setting/reading,
one can see that PageSwapCache reads the flag from folio, but
SetPageSwapCache/ClearPageSwapCache instead affect the flag from page.
After applying those patches SetPageSwapCache/ClearPageSwapCache become
unused, thus all paths seek this flag from folio now. With it I don't
see any hangs on mmap_lock anymore (on the same test setup).
https://virtuozzo.atlassian.net/browse/PSBM-153264
Signed-off-by: Pavel Tikhomirov <ptikhomi...@virtuozzo.com>
Matthew Wilcox (Oracle) (2):
mm/swap: convert __read_swap_cache_async() to use a folio
mm/swap: convert add_to_swap_cache() to take a folio
mm/shmem.c | 2 +-
mm/swap.h | 4 +--
mm/swap_state.c | 71 +++++++++++++++++++++++++------------------------
3 files changed, 39 insertions(+), 38 deletions(-)
--
2.43.0
_______________________________________________
Devel mailing list
Devel@openvz.org
https://lists.openvz.org/mailman/listinfo/devel
