In situations where no system memory is migrated to devmem, and in
upcoming patches where another GPU is performing the migration to
the newly allocated devmem buffer, there is nothing to ensure any
ongoing clear to the devmem allocation or async eviction from the
devmem allocation is complete.
Address that by passing a struct dma_fence down to the copy
functions, and ensure it is waited for before migration is marked
complete.
v3:
- New patch.
v4:
- Update the logic used for determining when to wait for the
pre_migrate_fence.
- Update the logic used for determining when to warn for the
pre_migrate_fence since the scheduler fences apparently
can signal out-of-order.
v5:
- Fix a UAF (CI)
- Remove references to source P2P migration (Himal)
- Put the pre_migrate_fence after migration.
v6:
- Pipeline the pre_migrate_fence dependency (Matt Brost)
Fixes: c5b3eb5a90 ("drm/xe: Add GPUSVM device memory copy vfunc functions")
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: <stable@vger.kernel.org> # v6.15+
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Acked-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> # For merging through drm-xe.
Link: https://patch.msgid.link/20251219113320.183860-4-thomas.hellstrom@linux.intel.com
(cherry picked from commit 16b5ad3195)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Patch series "mm: support device-private THP", v7.
This patch series introduces support for Transparent Huge Page (THP)
migration in zone device-private memory. The implementation enables
efficient migration of large folios between system memory and
device-private memory
Background
Current zone device-private memory implementation only supports PAGE_SIZE
granularity, leading to:
- Increased TLB pressure
- Inefficient migration between CPU and device memory
This series extends the existing zone device-private infrastructure to
support THP, leading to:
- Reduced page table overhead
- Improved memory bandwidth utilization
- Seamless fallback to base pages when needed
In my local testing (using lib/test_hmm) and a throughput test, the series
shows a 350% improvement in data transfer throughput and a 80% improvement
in latency
These patches build on the earlier posts by Ralph Campbell [1]
Two new flags are added in vma_migration to select and mark compound
pages. migrate_vma_setup(), migrate_vma_pages() and
migrate_vma_finalize() support migration of these pages when
MIGRATE_VMA_SELECT_COMPOUND is passed in as arguments.
The series also adds zone device awareness to (m)THP pages along with
fault handling of large zone device private pages. page vma walk and the
rmap code is also zone device aware. Support has also been added for
folios that might need to be split in the middle of migration (when the
src and dst do not agree on MIGRATE_PFN_COMPOUND), that occurs when src
side of the migration can migrate large pages, but the destination has not
been able to allocate large pages. The code supported and used
folio_split() when migrating THP pages, this is used when
MIGRATE_VMA_SELECT_COMPOUND is not passed as an argument to
migrate_vma_setup().
The test infrastructure lib/test_hmm.c has been enhanced to support THP
migration. A new ioctl to emulate failure of large page allocations has
been added to test the folio split code path. hmm-tests.c has new test
cases for huge page migration and to test the folio split path. A new
throughput test has been added as well.
The nouveau dmem code has been enhanced to use the new THP migration
capability.
mTHP support:
The patches hard code, HPAGE_PMD_NR in a few places, but the code has been
kept generic to support various order sizes. With additional refactoring
of the code support of different order sizes should be possible.
The future plan is to post enhancements to support mTHP with a rough
design as follows:
1. Add the notion of allowable thp orders to the HMM based test driver
2. For non PMD based THP paths in migrate_device.c, check to see if
a suitable order is found and supported by the driver
3. Iterate across orders to check the highest supported order for migration
4. Migrate and finalize
The mTHP patches can be built on top of this series, the key design
elements that need to be worked out are infrastructure and driver support
for multiple ordered pages and their migration.
HMM support for large folios was added in 10b9feee2d ("mm/hmm:
populate PFNs from PMD swap entry").
This patch (of 16)
Add routines to support allocation of large order zone device folios and
helper functions for zone device folios, to check if a folio is device
private and helpers for setting zone device data.
When large folios are used, the existing page_free() callback in pgmap is
called when the folio is freed, this is true for both PAGE_SIZE and higher
order pages.
Zone device private large folios do not support deferred split and scan
like normal THP folios.
Link: https://lkml.kernel.org/r/20251001065707.920170-1-balbirs@nvidia.com
Link: https://lkml.kernel.org/r/20251001065707.920170-2-balbirs@nvidia.com
Link: https://lore.kernel.org/linux-mm/20201106005147.20113-1-rcampbell@nvidia.com/ [1]
Signed-off-by: Balbir Singh <balbirs@nvidia.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: Ying Huang <ying.huang@linux.alibaba.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Nico Pache <npache@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Lyude Paul <lyude@redhat.com>
Cc: Danilo Krummrich <dakr@kernel.org>
Cc: David Airlie <airlied@gmail.com>
Cc: Simona Vetter <simona@ffwll.ch>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Mika Penttilä <mpenttil@redhat.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Francois Dugast <francois.dugast@intel.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Felix Kuehling <Felix.Kuehling@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
If the order is greater than zero, allocate a folio when populating the
RAM PFNs instead of allocating individual pages one after the other. For
example if 2MB folios are used instead of 4KB pages, this reduces the
number of calls to the allocation API by 512.
v2:
- Use page order instead of extra argument (Matthew Brost)
- Allocate with folio_alloc() (Matthew Brost)
- Loop for mpages and free_pages based on order (Matthew Brost)
v3:
- Fix loops in drm_pagemap_migrate_populate_ram_pfn() (Matthew Brost)
v4:
- Use folio_trylock(), set local variable to NULL (Matthew Brost)
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Acked-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Acked-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://lore.kernel.org/r/20250805140028.599361-5-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
If the page is part of a folio, DMA map the whole folio at once instead of
mapping individual pages one after the other. For example if 2MB folios
are used instead of 4KB pages, this reduces the number of DMA mappings by
512.
The folio order (and consequently, the size) is persisted in the struct
drm_pagemap_device_addr to be available at the time of unmapping.
v2:
- Initialize order variable (Matthew Brost)
- Set proto and dir for completeness (Matthew Brost)
- Do not populate drm_pagemap_addr, document it (Matthew Brost)
- Add and use macro NR_PAGES(order) (Matthew Brost)
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Acked-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://lore.kernel.org/r/20250805140028.599361-4-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
This struct embeds more information than just the DMA address. This will
help later to support folio orders greater than zero. At this point, there
is no functional change as the only struct member used is addr.
In Xe, adapt to the new drm_gpusvm_devmem_ops type signatures using struct
drm_pagemap_addr, as well as the internal xe SVM functions implementing
those operations. The use of this struct is propagated to xe_migrate as it
makes indexed accesses to the next DMA address but they are no longer
contiguous.
v2:
- Rename drm_pagemap_device_addr to drm_pagemap_addr (Matthew Brost)
- Squash with patch for Xe (Matthew Brost)
- Set proto and dir for completeness (Matthew Brost)
- Assess DMA map protocol (Matthew Brost)
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Acked-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://lore.kernel.org/r/20250805140028.599361-3-francois.dugast@intel.com
Signed-off-by: Francois Dugast <francois.dugast@intel.com>
The migration functionality and track-keeping of per-pagemap VRAM
mapped to the CPU mm is not per GPU_vm, but rather per pagemap.
This is also reflected by the functions not needing the drm_gpusvm
structures. So move to drm_pagemap.
With this, drm_gpusvm shouldn't really access the page zone-device-data
since its meaning is internal to drm_pagemap. Currently it's used to
reject mapping ranges backed by multiple drm_pagemap allocations.
For now, make the zone-device-data a void pointer.
Alter the interface of drm_gpusvm_migrate_to_devmem() to ensure we don't
pass a gpusvm pointer.
Rename CONFIG_DRM_XE_DEVMEM_MIRROR to CONFIG_DRM_XE_PAGEMAP.
Matt is listed as author of this commit since he wrote most of the code,
and it makes sense to retain his git authorship.
Thomas mostly moved the code around.
v3:
- Kerneldoc fixes (CI)
- Don't update documentation about how the drm_pagemap
migration should be interpreted until upcoming
patches where the functionality is implemented.
(Matt Brost)
v4:
- More kerneldoc fixes around timeslice_ms
(Himal Ghimiray, Matt Brost)
v6:
- Fix an uninitialized pagemap pointer (CI)
Co-developed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20250619134035.170086-2-thomas.hellstrom@linux.intel.com
