| Message ID | 20221017195834.2295901-1-ricarkol@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: selftests: Add aarch64/page_fault_test | expand |
On Mon, Oct 17, 2022, Ricardo Koller wrote: > This series adds a new aarch64 selftest for testing stage 2 fault handling > for various combinations of guest accesses (e.g., write, S1PTW), backing > sources (e.g., anon), and types of faults (e.g., read on hugetlbfs with a > hole, write on a readonly memslot). Each test tries a different combination > and then checks that the access results in the right behavior (e.g., uffd > faults with the right address and write/read flag). Some interesting > combinations are: ... > Ricardo Koller (14): > KVM: selftests: Add a userfaultfd library > KVM: selftests: aarch64: Add virt_get_pte_hva() library function > KVM: selftests: Add missing close and munmap in > __vm_mem_region_delete() > KVM: selftests: aarch64: Construct DEFAULT_MAIR_EL1 using sysreg.h > macros > tools: Copy bitfield.h from the kernel sources > KVM: selftests: Stash backing_src_type in struct userspace_mem_region > KVM: selftests: Add vm->memslots[] and enum kvm_mem_region_type > KVM: selftests: Fix alignment in virt_arch_pgd_alloc() and > vm_vaddr_alloc() > KVM: selftests: Use the right memslot for code, page-tables, and data > allocations > KVM: selftests: aarch64: Add aarch64/page_fault_test > KVM: selftests: aarch64: Add userfaultfd tests into page_fault_test > KVM: selftests: aarch64: Add dirty logging tests into page_fault_test > KVM: selftests: aarch64: Add readonly memslot tests into > page_fault_test > KVM: selftests: aarch64: Add mix of tests into page_fault_test One alignment nit in the first patch, but definitely not worthy of a v11. If it gets fixed up when applying, yay. If not, no big deal.
On Mon, 17 Oct 2022 19:58:20 +0000, Ricardo Koller wrote: > This series adds a new aarch64 selftest for testing stage 2 fault handling > for various combinations of guest accesses (e.g., write, S1PTW), backing > sources (e.g., anon), and types of faults (e.g., read on hugetlbfs with a > hole, write on a readonly memslot). Each test tries a different combination > and then checks that the access results in the right behavior (e.g., uffd > faults with the right address and write/read flag). Some interesting > combinations are: > > [...] Applied to next, thanks! [01/14] KVM: selftests: Add a userfaultfd library commit: a93871d0ea9fd59fb5eb783619334183d7f07f51 [02/14] KVM: selftests: aarch64: Add virt_get_pte_hva() library function commit: 228f324dc718f702e8777164c4e2e7426824fb13 [03/14] KVM: selftests: Add missing close and munmap in __vm_mem_region_delete() commit: b6b03b86c0250a80b671313dbc0d7bcdbab78f41 [04/14] KVM: selftests: aarch64: Construct DEFAULT_MAIR_EL1 using sysreg.h macros commit: 41f5189ea9c08f7fc28340a7aefc93d0d2dcb769 [05/14] tools: Copy bitfield.h from the kernel sources commit: 590b949597b1e811d35df2f32021dd17d8e47f8c [06/14] KVM: selftests: Stash backing_src_type in struct userspace_mem_region commit: bd3ed7e1a47eb7b3838ca09439f1eb289ec3be1f [07/14] KVM: selftests: Add vm->memslots[] and enum kvm_mem_region_type commit: 290c5b54012b7f05e9c51af32d557574bf69a654 [08/14] KVM: selftests: Fix alignment in virt_arch_pgd_alloc() and vm_vaddr_alloc() commit: 5485e822e31a75dfac3713d94b6b22025d4895da [09/14] KVM: selftests: Use the right memslot for code, page-tables, and data allocations commit: 1446e331432d7f24ed56b870ad605a4345fee43f [10/14] KVM: selftests: aarch64: Add aarch64/page_fault_test commit: 35c5810157124cb71aaa939cd2d5508192714877 [11/14] KVM: selftests: aarch64: Add userfaultfd tests into page_fault_test commit: 3b1d915659c64dce079f4926a648f2271faea008 [12/14] KVM: selftests: aarch64: Add dirty logging tests into page_fault_test commit: a4edf25b3e25656c69cbc768d1c704868e4a616f [13/14] KVM: selftests: aarch64: Add readonly memslot tests into page_fault_test commit: 45acde40f538a30e759f3b3f4aa5089edf097b2f [14/14] KVM: selftests: aarch64: Add mix of tests into page_fault_test commit: ff2b5509e1d252cd18bb1430b5461d5044701559 Cheers, M.
On Mon, Oct 17, 2022 at 07:58:20PM +0000, Ricardo Koller wrote: > This series adds a new aarch64 selftest for testing stage 2 fault handling > for various combinations of guest accesses (e.g., write, S1PTW), backing > sources (e.g., anon), and types of faults (e.g., read on hugetlbfs with a > hole, write on a readonly memslot). Each test tries a different combination > and then checks that the access results in the right behavior (e.g., uffd > faults with the right address and write/read flag). Some interesting > combinations are: I'm seeing issues with the page_fault_test tests in both -next and mainline all the way back to v6.1 when they were introduced running on both the fast model and hardware. With -next the reports come back as: # selftests: kvm: page_fault_test # ==== Test Assertion Failure ==== # aarch64/page_fault_test.c:316: __a == __b # pid=851 tid=860 errno=0 - Success # 1 0x0000000000402253: uffd_generic_handler at page_fault_test.c:316 # 2 0x000000000040be07: uffd_handler_thread_fn at userfaultfd_util.c:97 # 3 0x0000ffff8b39edd7: ?? ??:0 # 4 0x0000ffff8b407e9b: ?? ??:0 # ASSERT_EQ(!!(flags & UFFD_PAGEFAULT_FLAG_WRITE), expect_write) failed. # !!(flags & UFFD_PAGEFAULT_FLAG_WRITE) is 0 # expect_write is 0x1 not ok 6 selftests: kvm: page_fault_test # exit=254 (addr2line seemed to be not doing much, I've not poked too hard at that). I've been unable to find any case where the program passes. Is this expected? Some random full runs on hardware: 4xA53: https://lava.sirena.org.uk/scheduler/job/244678 4xA72: https://lkft.validation.linaro.org/scheduler/job/6114427
Hi Mark, On Fri, Feb 3, 2023 at 6:49 AM Mark Brown <broonie@kernel.org> wrote: > > On Mon, Oct 17, 2022 at 07:58:20PM +0000, Ricardo Koller wrote: > > This series adds a new aarch64 selftest for testing stage 2 fault handling > > for various combinations of guest accesses (e.g., write, S1PTW), backing > > sources (e.g., anon), and types of faults (e.g., read on hugetlbfs with a > > hole, write on a readonly memslot). Each test tries a different combination > > and then checks that the access results in the right behavior (e.g., uffd > > faults with the right address and write/read flag). Some interesting > > combinations are: > > I'm seeing issues with the page_fault_test tests in both -next and > mainline all the way back to v6.1 when they were introduced running on > both the fast model and hardware. With -next the reports come back as: > > # selftests: kvm: page_fault_test > # ==== Test Assertion Failure ==== > # aarch64/page_fault_test.c:316: __a == __b > # pid=851 tid=860 errno=0 - Success > # 1 0x0000000000402253: uffd_generic_handler at page_fault_test.c:316 > # 2 0x000000000040be07: uffd_handler_thread_fn at userfaultfd_util.c:97 > # 3 0x0000ffff8b39edd7: ?? ??:0 > # 4 0x0000ffff8b407e9b: ?? ??:0 > # ASSERT_EQ(!!(flags & UFFD_PAGEFAULT_FLAG_WRITE), expect_write) failed. > # !!(flags & UFFD_PAGEFAULT_FLAG_WRITE) is 0 > # expect_write is 0x1 > not ok 6 selftests: kvm: page_fault_test # exit=254 > > (addr2line seemed to be not doing much, I've not poked too hard at > that). I've been unable to find any case where the program passes. > Is this expected? > > Some random full runs on hardware: That failure was fixed with this series: "KVM: selftests: aarch64: page_fault_test S1PTW related fixes" https://lore.kernel.org/kvmarm/20230127214353.245671-1-ricarkol@google.com/ which made it into kvmarm/fixes and should get into 6.2: https://lore.kernel.org/kvmarm/20230129190142.2481354-1-maz@kernel.org/ Note that the failing assert does not exist after the mentioned series: > # ASSERT_EQ(!!(flags & UFFD_PAGEFAULT_FLAG_WRITE), expect_write) failed. > > 4xA53: https://lava.sirena.org.uk/scheduler/job/244678 > 4xA72: https://lkft.validation.linaro.org/scheduler/job/6114427 Thanks, Ricardo
On Fri, Feb 03, 2023 at 07:36:37AM -0800, Ricardo Koller wrote: > On Fri, Feb 3, 2023 at 6:49 AM Mark Brown <broonie@kernel.org> wrote: > > On Mon, Oct 17, 2022 at 07:58:20PM +0000, Ricardo Koller wrote: > > # ==== Test Assertion Failure ==== > > # aarch64/page_fault_test.c:316: __a == __b > > # pid=851 tid=860 errno=0 - Success > > # ASSERT_EQ(!!(flags & UFFD_PAGEFAULT_FLAG_WRITE), expect_write) failed. > > # !!(flags & UFFD_PAGEFAULT_FLAG_WRITE) is 0 > > # expect_write is 0x1 > That failure was fixed with this series: > "KVM: selftests: aarch64: page_fault_test S1PTW related fixes" > https://lore.kernel.org/kvmarm/20230127214353.245671-1-ricarkol@google.com/ > which made it into kvmarm/fixes and should get into 6.2: > https://lore.kernel.org/kvmarm/20230129190142.2481354-1-maz@kernel.org/ > Note that the failing assert does not exist after the mentioned series: > > # ASSERT_EQ(!!(flags & UFFD_PAGEFAULT_FLAG_WRITE), expect_write) failed. Ah, good. That's not made it into -next yet unfortunately so far as I can see but hopefully it'll turn up shortly and everything will start passing. Thanks!
This series adds a new aarch64 selftest for testing stage 2 fault handling for various combinations of guest accesses (e.g., write, S1PTW), backing sources (e.g., anon), and types of faults (e.g., read on hugetlbfs with a hole, write on a readonly memslot). Each test tries a different combination and then checks that the access results in the right behavior (e.g., uffd faults with the right address and write/read flag). Some interesting combinations are: - loading an instruction leads to a stage 1 page-table-walk that misses on stage 2 because the backing memslot for the page table it not in host memory (a hole was punched right there) and the fault is handled using userfaultfd. The expected behavior is that this leads to a userfaultfd fault marked as a write. See commit c4ad98e4b72c ("KVM: arm64: Assume write fault on S1PTW permission fault on instruction fetch") for why that's a write. - a cas (compare-and-swap) on a readonly memslot leads to a failed vcpu run. - write-faulting on a memslot that's marked for userfaultfd handling and dirty logging should result in a uffd fault and having the respective bit set in the dirty log. The first 8 commits of this series add library support. The first one adds a new userfaultfd library. Commits 2-5 add some misc library changes that will be used by the new test, like a library function to get the GPA of a PTE. Commits 6-8 breaks the implicit assumption that code and page tables live in memslot memslots should allocators use. This is then used by the new test to place the page tables in a specific memslot. The last 5 commits add the new selftest, one type of test at a time. It first adds core tests, then uffd, then dirty logging, then readonly memslots tests, and finally combinations of the previous ones (like uffd and dirty logging at the same time). v9 -> v10: https://lore.kernel.org/kvmarm/20221011010628.1734342-1-ricarkol@google.com/ - collected r-b's from Andrew - fixed indentation in several places (mainly alignment of params) [Sean] - renamed args in uffd_setup [Sean] v8 -> v9: https://lore.kernel.org/kvmarm/20220922031857.2588688-1-ricarkol@google.com/ - removed check before trying madvise(MADV_DONTNEED) on anon hugetlb. [Sean] - renamed punch_hole_in_memslot() [Sean] - changed the comment describing the accesses on "holes" [Sean] - collectd r-b's from Sean v7 -> v8: https://lore.kernel.org/kvmarm/20220920042551.3154283-1-ricarkol@google.com/ - applied Sean's suggestion of adding a fourth region: DATA, and renamed the old DATA one to DATA_TEST. [Sean] - removed some unneeded code in page_fault_test.c. [Andrew] - removed r-b's from Andrew and Oliver in commit "KVM: selftests: Use the right memslot for code, page-tables, and data allocations", as the commit changed quite a bit (again). Thanks for the reviews! would it have been OK to keep it? not sure how strict we all are about collecting r-b's on a commit that changed. Ricardo Koller (14): KVM: selftests: Add a userfaultfd library KVM: selftests: aarch64: Add virt_get_pte_hva() library function KVM: selftests: Add missing close and munmap in __vm_mem_region_delete() KVM: selftests: aarch64: Construct DEFAULT_MAIR_EL1 using sysreg.h macros tools: Copy bitfield.h from the kernel sources KVM: selftests: Stash backing_src_type in struct userspace_mem_region KVM: selftests: Add vm->memslots[] and enum kvm_mem_region_type KVM: selftests: Fix alignment in virt_arch_pgd_alloc() and vm_vaddr_alloc() KVM: selftests: Use the right memslot for code, page-tables, and data allocations KVM: selftests: aarch64: Add aarch64/page_fault_test KVM: selftests: aarch64: Add userfaultfd tests into page_fault_test KVM: selftests: aarch64: Add dirty logging tests into page_fault_test KVM: selftests: aarch64: Add readonly memslot tests into page_fault_test KVM: selftests: aarch64: Add mix of tests into page_fault_test tools/include/linux/bitfield.h | 176 +++ tools/testing/selftests/kvm/.gitignore | 1 + tools/testing/selftests/kvm/Makefile | 2 + .../selftests/kvm/aarch64/page_fault_test.c | 1112 +++++++++++++++++ .../selftests/kvm/demand_paging_test.c | 228 +--- .../selftests/kvm/include/aarch64/processor.h | 35 +- .../selftests/kvm/include/kvm_util_base.h | 31 +- .../selftests/kvm/include/userfaultfd_util.h | 45 + .../selftests/kvm/lib/aarch64/processor.c | 48 +- tools/testing/selftests/kvm/lib/elf.c | 3 +- tools/testing/selftests/kvm/lib/kvm_util.c | 82 +- .../selftests/kvm/lib/riscv/processor.c | 29 +- .../selftests/kvm/lib/s390x/processor.c | 8 +- .../selftests/kvm/lib/userfaultfd_util.c | 186 +++ .../selftests/kvm/lib/x86_64/processor.c | 13 +- 15 files changed, 1723 insertions(+), 276 deletions(-) create mode 100644 tools/include/linux/bitfield.h create mode 100644 tools/testing/selftests/kvm/aarch64/page_fault_test.c create mode 100644 tools/testing/selftests/kvm/include/userfaultfd_util.h create mode 100644 tools/testing/selftests/kvm/lib/userfaultfd_util.c