mm/hmm: fault non-owner device private entries
v3 cover mm/hmm: fault non-owner device private entries
Changes from v1 to v2:
Made code style changes suggested by Alistair Popple Added a self test to hmm-tests.c (Jason Gunthorpe)
Changes from v2 to v3:
Updated the change description for patch 1 to include an example user visible effect of the change.
Updated the change description for patch 2 to mention the deleted code. Added reviewed-by tags and updated the fixes reference.
Andrew's mm tree already has everything except the updated description for patch 2.
Ralph Campbell (2):
mm/hmm: fault non-owner device private entries mm/hmm: add a test for cross device private faults
mm/hmm.c | 19 ++++++++----------- tools/testing/selftests/vm/hmm-tests.c | 14 ++++++++++++-- 2 files changed, 20 insertions(+), 13 deletions(-)
Signature · 1 line
2.35.3
PATCH v3 1/2 mm/hmm: fault non-owner device private entries
If hmm_range_fault() is called with the HMM_PFN_REQ_FAULT flag and a device private PTE is found, the hmm_range::dev_private_owner page is used to determine if the device private page should not be faulted in. However, if the device private page is not owned by the caller, hmm_range_fault() returns an error instead of calling migrate_to_ram() to fault in the page.
For example, if a page is migrated to GPU private memory and a RDMA fault capable NIC tries to read the migrated page, without this patch it will get an error. With this patch, the page will be migrated back to system memory and the NIC will be able to read the data.
--- mm/hmm.c | 19 ++++++++----------- 1 file changed, 8 insertions(+), 11 deletions(-)
mm/hmm.c +8 -11
@@ -212,14 +212,6 @@ int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr, unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);#endif /* CONFIG_TRANSPARENT_HUGEPAGE */ -static inline bool hmm_is_device_private_entry (struct hmm_range *range, -swp_entry_t entry) -{ -return is_device_private_entry (entry) && -pfn_swap_entry_to_page (entry)->pgmap->owner == - range->dev_private_owner; -} -static inline unsigned long pte_to_hmm_pfn_flags (struct hmm_range *range,pte_t pte) {
@@ -252,10 +244,12 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,swp_entry_t entry =pte_to_swp_entry (pte);/* -* Never fault in device private pages, but just report -* the PFN even if not present. +* Don't fault in device private pages owned by the caller, +* just report the PFN. */ -if (hmm_is_device_private_entry (range, entry)) { +if (is_device_private_entry (entry) && +pfn_swap_entry_to_page (entry)->pgmap->owner == + range->dev_private_owner) { cpu_flags =HMM_PFN_VALID ;if (is_writable_device_private_entry (entry)) cpu_flags |=HMM_PFN_WRITE ;
@@ -273,6 +267,9 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,if (!non_swap_entry (entry)) goto fault; +if (is_device_private_entry (entry)) + goto fault; +if (is_device_exclusive_entry (entry)) goto fault;
Signature · 1 line
2.35.3
PATCH v3 2/2 mm/hmm: add a test for cross device private faults
In reply to Ralph CampbellAdd a simple test case for when hmm_range_fault() is called with the HMM_PFN_REQ_FAULT flag and a device private PTE is found for a device other than the hmm_range::dev_private_owner. This should cause the page to be faulted back to system memory from the other device and the PFN returned in the output array.
Also, remove a piece of code that unnecessarily unmaps part of the buffer.
--- tools/testing/selftests/vm/hmm-tests.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)
tools/testing/selftests/vm/hmm-tests.c +12 -2
@@ -1520,9 +1520,19 @@ TEST_F(hmm2, double_map)for (i =0 , ptr = buffer->mirror; i < size /sizeof (*ptr); ++i)ASSERT_EQ (ptr[i], i); -/* Punch a hole after the first page address. */ - ret =munmap (buffer->ptr + self->page_size, self->page_size); +/* Migrate pages to device 1 and try to read from device 0. */ + ret =hmm_dmirror_cmd (self->fd1,HMM_DMIRROR_MIGRATE , buffer, npages); +ASSERT_EQ (ret,0 ); +ASSERT_EQ (buffer->cpages, npages); + + ret =hmm_dmirror_cmd (self->fd0,HMM_DMIRROR_READ , buffer, npages);ASSERT_EQ (ret,0 ); +ASSERT_EQ (buffer->cpages, npages); +ASSERT_EQ (buffer->faults,1 ); + +/* Check what device 0 read. */ +for (i =0 , ptr = buffer->mirror; i < size /sizeof (*ptr); ++i) +ASSERT_EQ (ptr[i], i);hmm_buffer_free (buffer); }
Signature · 1 line
2.35.3
Andrew Morton
maintainerQuoted 2 lines · On Tue, 26 Jul 2022 17:08:35 -0700 Ralph Campbell <[email protected]> wrote:
Andrew's mm tree already has everything except the updated description for patch 2.
Thanks, I updated the changelog and the Link: tags in place.