Pull dma-mapping updates from Marek Szyprowski:
- swiotlb:
- new configuration option for the default pool size
(Jagadeesh Pagadala)
- reduce overhead for high watermark tracking (chenhuguanshen)
- minor code cleanups and improvements (Vova Sharaienko, Honglei Huang
and Marek Szyprowski)
- add proper tracking of the shared DMA state through direct, pool and
swiotlb paths (Aneesh Kumar K.V)
This is important for confidential-computing
* tag 'dma-mapping-7.3-2026-08-24' of git://git.kernel.org/pub/scm/linux/kernel/git/mszyprowski/linux:
dma/swiotlb: decouple high watermark tracking from CONFIG_DEBUG_FS
MAINTAINERS: update tree for DMA MAPPING HELPERS
dma/swiotlb: introduce Kconfig option for compile-time default pool size
dma-direct: Improve readability of the dma_direct_map_sg() for P2PDMA case
iommu/dma: simplify dma_iova_destroy() and drop the free_iova helper
dma-coherent: use KiB in DMA allocation logs
dma-coherent: fix spacing coding style issue
* tag 'dma-mapping-7.3-2026-08-24-2' of git://git.kernel.org/pub/scm/linux/kernel/git/mszyprowski/linux: (23 commits)
swiotlb: remove unused SWIOTLB_FORCE flag
dma: swiotlb: handle set_memory_decrypted() failures
dma: swiotlb: free dynamic pools from process context
dma-direct: rename ret to cpu_addr in alloc helpers
dma-direct: select DMA address encoding from __DMA_ATTR_ALLOC_CC_SHARED
dma-direct: set decrypted flag for remapped DMA allocations
dma-direct: make dma_direct_map_phys() honor DMA_ATTR_CC_SHARED
dma-direct: Move dma_direct_map_phys() to dma/direct.c
dma-direct: pass attrs to dma_capable() for DMA_ATTR_CC_SHARED checks
dma-mapping: make dma_pgprot() honor __DMA_ATTR_ALLOC_CC_SHARED
dma: swiotlb: track pool encryption state and honor DMA_ATTR_CC_SHARED
dma: swiotlb: pass mapping attributes by reference
dma-pool: track decrypted atomic pools and select them via attrs
dma-direct: use __DMA_ATTR_ALLOC_CC_SHARED in alloc/free paths
dma-mapping: Add internal shared allocation attribute
coco: arm64: s390: powerpc: Mark secure guests with CC_ATTR_GUEST_MEM_ENCRYPT
dma-direct: swiotlb: handle swiotlb alloc/free outside __dma_direct_alloc_pages
s390: Expose protected virtualization through cc_platform_has()
swiotlb: Preserve allocation virtual address for dynamic pools
dma: free atomic pool pages by physical address
...
ret in dma_direct_alloc() and dma_direct_alloc_pages() holds the returned
CPU mapping, not a generic return value. Rename it to cpu_addr and update
the remaining uses to match.
This makes the allocation paths easier to follow and keeps the local naming
consistent with what the variable actually represents.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Reviewed-by: Petr Tesarik <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Make the dma-direct helpers derive the DMA address encoding from
__DMA_ATTR_ALLOC_CC_SHARED instead of implicitly relying on
force_dma_unencrypted() inside phys_to_dma_direct()
Pass an explicit unencrypted/decrypted state into phys_to_dma_direct(),
make the alloc paths return DMA addresses that match the requested buffer
encryption state. Also only call dma_set_decrypted() when
__DMA_ATTR_ALLOC_CC_SHARED is actually set.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Reviewed-by: Mostafa Saleh <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Devices that are DMA non-coherent and require a remap were skipping
dma_set_decrypted(), leaving DMA buffers encrypted even when the device
requires unencrypted access. Move the call after the if (remap) branch
so that both the direct and remapped allocation paths correctly mark the
allocation as decrypted (or fail cleanly) before use.
Fix dma_direct_alloc() and dma_direct_free() to apply set_memory_*() to the
linear-map alias of the backing pages instead of the remapped CPU address.
Also disallow highmem pages for __DMA_ATTR_ALLOC_CC_SHARED, because highmem
buffers do not provide a usable linear-map address.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Teach dma_direct_map_phys() to select the DMA address encoding based on
DMA_ATTR_CC_SHARED.
Use phys_to_dma_unencrypted() for decrypted mappings and
phys_to_dma_encrypted() otherwise. If a device requires unencrypted DMA
but the source physical address is still encrypted, force the mapping
through swiotlb so the DMA address and backing memory attributes remain
consistent.
Update the arm64, x86, s390 and powerpc secure-guest setup to not use
swiotlb force option
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
[mszyprow: rebased onto latest changes in arch/arm64/mm/init.c]
Signed-off-by: Marek Szyprowski <[email protected]>
dma_direct_map_phys() is too large to benefit from being inlined. Move its
implementation to direct.c and leave the declaration in direct.h.
No functional change in this patch
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Fold encrypted/decrypted pgprot selection into dma_pgprot() so callers
do not need to adjust the page protection separately.
Update dma_pgprot() to apply pgprot_decrypted() when DMA_ATTR_CC_SHARED or
__DMA_ATTR_ALLOC_CC_SHARED is set and pgprot_encrypted() otherwise Convert
the dma-direct mmap paths to pass DMA_ATTR_CC_SHARED instead of open-coding
force_dma_unencrypted() handling around dma_pgprot().
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Teach swiotlb to distinguish between encrypted and decrypted bounce
buffer pools, and make allocation and mapping paths select a pool whose
state matches the requested DMA attributes.
Add a cc_shared flag to io_tlb_mem, initialize it for the default and
restricted pools, and propagate __DMA_ATTR_ALLOC_CC_SHARED into swiotlb
pool allocation. Reject swiotlb alloc/map requests when the selected pool
does not match the required encrypted/decrypted state.
Also return DMA addresses with the matching phys_to_dma_{encrypted,
unencrypted} helper so the DMA address encoding stays consistent with the
chosen pool.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Teach the atomic DMA pool code to distinguish between encrypted and
unencrypted pools, and make pool allocation select the matching pool based
on DMA attributes.
Introduce a dma_gen_pool wrapper that records whether a pool is
unencrypted, initialize that state when the atomic pools are created, and
use it when expanding and resizing the pools. Update dma_alloc_from_pool()
to take attrs and skip pools whose encrypted state does not match
__DMA_ATTR_ALLOC_CC_SHARED. Update dma_free_from_pool() accordingly.
Also pass __DMA_ATTR_ALLOC_CC_SHARED from the swiotlb atomic allocation
path so decrypted swiotlb allocations are taken from the correct atomic
pool.
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Reviewed-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Propagate force_dma_unencrypted() into __DMA_ATTR_ALLOC_CC_SHARED in the
dma-direct allocation path and use the attribute to drive the related
decisions.
This updates dma_direct_alloc(), dma_direct_free(), and
dma_direct_alloc_pages() to fold the forced unencrypted case into attrs.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Reviewed-by: Petr Tesarik <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Move swiotlb allocation out of __dma_direct_alloc_pages() and handle it in
dma_direct_alloc() / dma_direct_alloc_pages().
This is needed for follow-up changes that simplify the handling of
memory encryption/decryption based on the DMA attribute flags.
swiotlb backing pages are already mapped decrypted by
swiotlb_update_mem_attributes() and rmem_swiotlb_device_init(), so
dma-direct should not call dma_set_decrypted() on allocation nor
dma_set_encrypted() on free for swiotlb-backed memory.
Update alloc/free paths to detect swiotlb-backed pages and skip
encrypt/decrypt transitions for those paths. Keep the existing highmem
rejection in dma_direct_alloc_pages() for swiotlb allocations.
Only for "restricted-dma-pool", we currently set `for_alloc = true`, while
rmem_swiotlb_device_init() decrypts the whole pool up front. This pool is
typically used together with "shared-dma-pool", where the shared region is
accessed after remap/ioremap and the returned address is suitable for
decrypted memory access. So existing code paths remain valid.
Reviewed-by: Jason Gunthorpe <[email protected]>
Tested-by: Jiri Pirko <[email protected]>
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Reviewed-by: Mostafa Saleh <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
dma_direct_alloc_pages() may satisfy atomic allocations from the coherent
atomic pools. The pool allocation is keyed by the virtual address stored in
the gen_pool, but the pages API returns only the backing struct page.
On architectures with CONFIG_DMA_DIRECT_REMAP, atomic pool chunks are added
to the gen_pool using their remapped virtual address.
dma_direct_free_pages() reconstructs a linear-map address with
page_address(page) and passes that to dma_free_from_pool(). That address
does not match the gen_pool virtual range, so the pool lookup can fail and
the code can fall through to freeing a pool-owned page through the normal
page allocator path.
Add a page-based pool free helper that looks up the owning pool chunk by
physical address, translates it back to the gen_pool virtual address, and
frees that address to the pool. Use it from dma_direct_free_pages() while
keeping the existing virtual-address helper for coherent allocation frees.
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
Commit 5b138c534f ("dma-direct: factor out a dma_direct_alloc_from_pool
helper") changed dma_direct_alloc_from_pool() to return the CPU address
from dma_alloc_from_pool(). That fits dma_direct_alloc(), but
dma_direct_alloc_pages() also uses the helper and expects a struct page *.
Fix this by making dma_direct_alloc_from_pool() return the struct page *
again, and pass the CPU address back through an out-parameter for the
dma_direct_alloc() caller.
Fixes: 5b138c534f ("dma-direct: factor out a dma_direct_alloc_from_pool helper")
Cc: [email protected]
Tested-by: Michael Kelley <[email protected]>
Tested-by: Mostafa Saleh <[email protected]>
Reviewed-by: Jason Gunthorpe <[email protected]>
Signed-off-by: Aneesh Kumar K.V (Arm) <[email protected]>
Reviewed-by: Mostafa Saleh <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Marek Szyprowski <[email protected]>
In dma_direct_map_sg(), the case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE
incorrectly used 'break' instead of falling through to MAP_NONE.
As a result, segments traversing the host bridge skipped the required
dma_direct_map_phys() call entirely, leaving sg->dma_address
uninitialized and leading to DMA failures. Fix this by using
'fallthrough;'.
Fixes: a25e7962db ("PCI/P2PDMA: Refactor the p2pdma mapping helpers")
Reviewed-by: Logan Gunthorpe <[email protected]>
Signed-off-by: Li RongQing <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Calculate the correct physical address of the last byte of memory. Since
max_pfn is in fact "the PFN of the first page after the highest system RAM
in physical address space", the highest address that might be used for a
DMA buffer is one byte below max_pfn << PAGE_SHIFT.
This fix is unlikely to make any difference in practice. It's just that the
current formula is slightly confusing.
Signed-off-by: Petr Tesarik <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Currently, arch_sync_dma_for_cpu and arch_sync_dma_for_device
always wait for the completion of each DMA buffer. That is,
issuing the DMA sync and waiting for completion is done in a
single API call.
For scatter-gather lists with multiple entries, this means
issuing and waiting is repeated for each entry, which can hurt
performance. Architectures like ARM64 may be able to issue all
DMA sync operations for all entries first and then wait for
completion together.
To address this, arch_sync_dma_for_* now batches DMA operations
and performs a flush afterward. On ARM64, the flush is implemented
with a dsb instruction in arch_sync_dma_flush(). On other
architectures, arch_sync_dma_flush() is currently a nop.
Cc: Leon Romanovsky <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Marek Szyprowski <[email protected]>
Cc: Robin Murphy <[email protected]>
Cc: Ada Couprie Diaz <[email protected]>
Cc: Ard Biesheuvel <[email protected]>
Cc: Marc Zyngier <[email protected]>
Cc: Anshuman Khandual <[email protected]>
Cc: Ryan Roberts <[email protected]>
Cc: Suren Baghdasaryan <[email protected]>
Cc: Joerg Roedel <[email protected]>
Cc: Stefano Stabellini <[email protected]>
Cc: Oleksandr Tyshchenko <[email protected]>
Cc: Tangquan Zheng <[email protected]>
Reviewed-by: Juergen Gross <[email protected]> # drivers/xen/swiotlb-xen.c
Tested-by: Xueyuan Chen <[email protected]>
Signed-off-by: Barry Song <[email protected]>
Reviewed-by: Leon Romanovsky <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
This was done entirely with mindless brute force, using
git grep -l '\<k[vmz]*alloc_objs*(.*, GFP_KERNEL)' |
xargs sed -i 's/\(alloc_objs*(.*\), GFP_KERNEL)/\1)/'
to convert the new alloc_obj() users that had a simple GFP_KERNEL
argument to just drop that argument.
Note that due to the extreme simplicity of the scripting, any slightly
more complex cases spread over multiple lines would not be triggered:
they definitely exist, but this covers the vast bulk of the cases, and
the resulting diff is also then easier to check automatically.
For the same reason the 'flex' versions will be done as a separate
conversion.
Signed-off-by: Linus Torvalds <[email protected]>
This is the result of running the Coccinelle script from
scripts/coccinelle/api/kmalloc_objs.cocci. The script is designed to
avoid scalar types (which need careful case-by-case checking), and
instead replace kmalloc-family calls that allocate struct or union
object instances:
Single allocations: kmalloc(sizeof(TYPE), ...)
are replaced with: kmalloc_obj(TYPE, ...)
Array allocations: kmalloc_array(COUNT, sizeof(TYPE), ...)
are replaced with: kmalloc_objs(TYPE, COUNT, ...)
Flex array allocations: kmalloc(struct_size(PTR, FAM, COUNT), ...)
are replaced with: kmalloc_flex(*PTR, FAM, COUNT, ...)
(where TYPE may also be *VAR)
The resulting allocations no longer return "void *", instead returning
"TYPE *".
Signed-off-by: Kees Cook <[email protected]>
Pull VFIO updates from Alex Williamson:
- Move libvfio selftest artifacts in preparation of more tightly
coupled integration with KVM selftests (David Matlack)
- Fix comment typo in mtty driver (Chu Guangqing)
- Support for new hardware revision in the hisi_acc vfio-pci variant
driver where the migration registers can now be accessed via the PF.
When enabled for this support, the full BAR can be exposed to the
user (Longfang Liu)
- Fix vfio cdev support for VF token passing, using the correct size
for the kernel structure, thereby actually allowing userspace to
provide a non-zero UUID token. Also set the match token callback for
the hisi_acc, fixing VF token support for this this vfio-pci variant
driver (Raghavendra Rao Ananta)
- Introduce internal callbacks on vfio devices to simplify and
consolidate duplicate code for generating VFIO_DEVICE_GET_REGION_INFO
data, removing various ioctl intercepts with a more structured
solution (Jason Gunthorpe)
- Introduce dma-buf support for vfio-pci devices, allowing MMIO regions
to be exposed through dma-buf objects with lifecycle managed through
move operations. This enables low-level interactions such as a
vfio-pci based SPDK drivers interacting directly with dma-buf capable
RDMA devices to enable peer-to-peer operations. IOMMUFD is also now
able to build upon this support to fill a long standing feature gap
versus the legacy vfio type1 IOMMU backend with an implementation of
P2P support for VM use cases that better manages the lifecycle of the
P2P mapping (Leon Romanovsky, Jason Gunthorpe, Vivek Kasireddy)
- Convert eventfd triggering for error and request signals to use RCU
mechanisms in order to avoid a 3-way lockdep reported deadlock issue
(Alex Williamson)
- Fix a 32-bit overflow introduced via dma-buf support manifesting with
large DMA buffers (Alex Mastro)
- Convert nvgrace-gpu vfio-pci variant driver to insert mappings on
fault rather than at mmap time. This conversion serves both to make
use of huge PFNMAPs but also to both avoid corrected RAS events
during reset by now being subject to vfio-pci-core's use of
unmap_mapping_range(), and to enable a device readiness test after
reset (Ankit Agrawal)
- Refactoring of vfio selftests to support multi-device tests and split
code to provide better separation between IOMMU and device objects.
This work also enables a new test suite addition to measure parallel
device initialization latency (David Matlack)
* tag 'vfio-v6.19-rc1' of https://github.com/awilliam/linux-vfio: (65 commits)
vfio: selftests: Add vfio_pci_device_init_perf_test
vfio: selftests: Eliminate INVALID_IOVA
vfio: selftests: Split libvfio.h into separate header files
vfio: selftests: Move vfio_selftests_*() helpers into libvfio.c
vfio: selftests: Rename vfio_util.h to libvfio.h
vfio: selftests: Stop passing device for IOMMU operations
vfio: selftests: Move IOVA allocator into iova_allocator.c
vfio: selftests: Move IOMMU library code into iommu.c
vfio: selftests: Rename struct vfio_dma_region to dma_region
vfio: selftests: Upgrade driver logging to dev_err()
vfio: selftests: Prefix logs with device BDF where relevant
vfio: selftests: Eliminate overly chatty logging
vfio: selftests: Support multiple devices in the same container/iommufd
vfio: selftests: Introduce struct iommu
vfio: selftests: Rename struct vfio_iommu_mode to iommu_mode
vfio: selftests: Allow passing multiple BDFs on the command line
vfio: selftests: Split run.sh into separate scripts
vfio: selftests: Move run.sh into scripts directory
vfio/nvgrace-gpu: wait for the GPU mem to be ready
vfio/nvgrace-gpu: Inform devmem unmapped after reset
...
Prior to commit a25e7962db ("PCI/P2PDMA: Refactor the p2pdma mapping
helpers"), P2P segments were mapped using the pci_p2pdma_map_segment()
helper. This helper was responsible for populating sg->dma_address,
marking the bus address, and also setting sg_dma_len(sg).
The refactor[1] removed this helper and moved the mapping logic directly
into the callers. While iommu_dma_map_sg() was correctly updated to set
the length in the new flow, it was missed in dma_direct_map_sg().
Thus, in dma_direct_map_sg(), the PCI_P2PDMA_MAP_BUS_ADDR case sets the
dma_address and marks the segment, but immediately executes 'continue',
which causes the loop to skip the standard assignment logic at the end:
sg_dma_len(sg) = sg->length;
As a result, when CONFIG_NEED_SG_DMA_LENGTH is enabled, the dma_length
field remains uninitialized (zero) for P2P bus address mappings. This
breaks upper-layer drivers (for e.g. RDMA/IB) that rely on sg_dma_len()
to determine the transfer size.
Fix this by explicitly setting the DMA length in the
PCI_P2PDMA_MAP_BUS_ADDR case before continuing to the next scatterlist
entry.
Fixes: a25e7962db ("PCI/P2PDMA: Refactor the p2pdma mapping helpers")
Reported-by: Jacob Moroni <[email protected]>
Signed-off-by: Pranjal Shrivastava <[email protected]>
[1]
https://lore.kernel.org/all/ac14a0e94355bf898de65d023ccf8a2ad22a3ece.1746424934.git.leon@kernel.org/
Reviewed-by: Logan Gunthorpe <[email protected]>
Reviewed-by: Leon Romanovsky <[email protected]>
Reviewed-by: Shivaji Kant <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Update the pci_p2pdma_bus_addr_map() function to take a direct pointer
to the p2pdma_provider structure instead of the pci_p2pdma_map_state.
This simplifies the API by removing the need for callers to extract
the provider from the state structure.
The change updates all callers across the kernel (block layer, IOMMU,
DMA direct, and HMM) to pass the provider pointer directly, making
the code more explicit and reducing unnecessary indirection. This
also removes the runtime warning check since callers now have direct
control over which provider they use.
Tested-by: Alex Mastro <[email protected]>
Tested-by: Nicolin Chen <[email protected]>
Signed-off-by: Leon Romanovsky <[email protected]>
Acked-by: Ankit Agrawal <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Alex Williamson <[email protected]>
Introduce new DMA mapping functions dma_map_phys() and dma_unmap_phys()
that operate directly on physical addresses instead of page+offset
parameters. This provides a more efficient interface for drivers that
already have physical addresses available.
The new functions are implemented as the primary mapping layer, with
the existing dma_map_page_attrs()/dma_map_resource() and
dma_unmap_page_attrs()/dma_unmap_resource() functions converted to simple
wrappers around the phys-based implementations.
In case dma_map_page_attrs(), the struct page is converted to physical
address with help of page_to_phys() function and dma_map_resource()
provides physical address as is together with addition of DMA_ATTR_MMIO
attribute.
The old page-based API is preserved in mapping.c to ensure that existing
code won't be affected by changing EXPORT_SYMBOL to EXPORT_SYMBOL_GPL
variant for dma_*map_phys().
Reviewed-by: Jason Gunthorpe <[email protected]>
Reviewed-by: Keith Busch <[email protected]>
Signed-off-by: Leon Romanovsky <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/54cc52af91777906bbe4a386113437ba0bcfba9c.1757423202.git.leonro@nvidia.com
Convert the DMA direct mapping functions to accept physical addresses
directly instead of page+offset parameters. The functions were already
operating on physical addresses internally, so this change eliminates
the redundant page-to-physical conversion at the API boundary.
The functions dma_direct_map_page() and dma_direct_unmap_page() are
renamed to dma_direct_map_phys() and dma_direct_unmap_phys() respectively,
with their calling convention changed from (struct page *page,
unsigned long offset) to (phys_addr_t phys).
Architecture-specific functions arch_dma_map_page_direct() and
arch_dma_unmap_page_direct() are similarly renamed to
arch_dma_map_phys_direct() and arch_dma_unmap_phys_direct().
The is_pci_p2pdma_page() checks are replaced with DMA_ATTR_MMIO checks
to allow integration with dma_direct_map_resource and dma_direct_map_phys()
is extended to support MMIO path either.
Reviewed-by: Jason Gunthorpe <[email protected]>
Signed-off-by: Leon Romanovsky <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/bb15a22f76dc2e26683333ff54e789606cfbfcf0.1757423202.git.leonro@nvidia.com
Convert a goto-based loop to a while() loop. To allow the simplification,
return early when allocation from CMA is successful. As a bonus, this early
return avoids a repeated dma_coherent_ok() check.
No functional change.
Signed-off-by: Petr Tesarik <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
To support the upcoming non-scatterlist mapping helpers, we need to go
back to have them called outside of the DMA API. Thus move them out of
dma-map-ops.h, which is only for DMA API implementations to pci-p2pdma.h,
which is for driver use.
Note that the core helper is still not exported as the mapping is
expected to be done only by very highlevel subsystem code at least for
now.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Logan Gunthorpe <[email protected]>
Acked-by: Bjorn Helgaas <[email protected]>
Tested-by: Jens Axboe <[email protected]>
Reviewed-by: Luis Chamberlain <[email protected]>
Reviewed-by: Lu Baolu <[email protected]>
Signed-off-by: Leon Romanovsky <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
The current scheme with a single helper to determine the P2P status
and map a scatterlist segment force users to always use the map_sg
helper to DMA map, which we're trying to get away from because they
are very cache inefficient.
Refactor the code so that there is a single helper that checks the P2P
state for a page, including the result that it is not a P2P page to
simplify the callers, and a second one to perform the address translation
for a bus mapped P2P transfer that does not depend on the scatterlist
structure.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Logan Gunthorpe <[email protected]>
Acked-by: Bjorn Helgaas <[email protected]>
Tested-by: Jens Axboe <[email protected]>
Reviewed-by: Luis Chamberlain <[email protected]>
Reviewed-by: Lu Baolu <[email protected]>
Signed-off-by: Leon Romanovsky <[email protected]>
Signed-off-by: Marek Szyprowski <[email protected]>
When the CMA allocation succeeds but isn't addressable, its buffer has
already been released and the page is set to NULL. So later when the
normal page allocation succeeds but isn't addressable, __free_pages()
can be used to free that normal page rather than using
dma_free_contiguous that does extra checks that are not needed.
Signed-off-by: Chen Yu <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
The hardware DMA limit might not be power of 2. When RAM range starts
above 0, say 4GB, DMA limit of 30 bits should end at 5GB. A single high
bit can not encode this limit.
Use a plain address for the DMA zone limit instead.
Since the DMA zone can now potentially span beyond 4GB physical limit of
DMA32, make sure to use DMA zone for GFP_DMA32 allocations in that case.
Signed-off-by: Catalin Marinas <[email protected]>
Co-developed-by: Baruch Siach <[email protected]>
Signed-off-by: Baruch Siach <[email protected]>
Reviewed-by: Catalin Marinas <[email protected]>
Reviewed-by: Petr Tesarik <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
With CONFIG_SWIOTLB_DYNAMIC enabled, each round-trip map/unmap pair
in the swiotlb results in 6 calls to swiotlb_find_pool(). In multiple
places, the pool is found and used in one function, and then must
be found again in the next function that is called because only the
tlb_addr is passed as an argument. These are the six call sites:
dma_direct_map_page:
1. swiotlb_map -> swiotlb_tbl_map_single -> swiotlb_bounce
dma_direct_unmap_page:
2. dma_direct_sync_single_for_cpu -> is_swiotlb_buffer
3. dma_direct_sync_single_for_cpu -> swiotlb_sync_single_for_cpu ->
swiotlb_bounce
4. is_swiotlb_buffer
5. swiotlb_tbl_unmap_single -> swiotlb_del_transient
6. swiotlb_tbl_unmap_single -> swiotlb_release_slots
Reduce the number of calls by finding the pool at a higher level, and
passing it as an argument instead of searching again. A key change is
for is_swiotlb_buffer() to return a pool pointer instead of a boolean,
and then pass this pool pointer to subsequent swiotlb functions.
There are 9 occurrences of is_swiotlb_buffer() used to test if a buffer
is a swiotlb buffer before calling a swiotlb function. To reduce code
duplication in getting the pool pointer and passing it as an argument,
introduce inline wrappers for this pattern. The generated code is
essentially unchanged.
Since is_swiotlb_buffer() no longer returns a boolean, rename some
functions to reflect the change:
* swiotlb_find_pool() becomes __swiotlb_find_pool()
* is_swiotlb_buffer() becomes swiotlb_find_pool()
* is_xen_swiotlb_buffer() becomes xen_swiotlb_find_pool()
With these changes, a round-trip map/unmap pair requires only 2 pool
lookups (listed using the new names and wrappers):
dma_direct_unmap_page:
1. dma_direct_sync_single_for_cpu -> swiotlb_find_pool
2. swiotlb_tbl_unmap_single -> swiotlb_find_pool
These changes come from noticing the inefficiencies in a code review,
not from performance measurements. With CONFIG_SWIOTLB_DYNAMIC,
__swiotlb_find_pool() is not trivial, and it uses an RCU read lock,
so avoiding the redundant calls helps performance in a hot path.
When CONFIG_SWIOTLB_DYNAMIC is *not* set, the code size reduction
is minimal and the perf benefits are likely negligible, but no
harm is done.
No functional change is intended.
Signed-off-by: Michael Kelley <[email protected]>
Reviewed-by: Petr Tesarik <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
On TDX it is possible for the untrusted host to cause
set_memory_encrypted() or set_memory_decrypted() to fail such that an
error is returned and the resulting memory is shared. Callers need to
take care to handle these errors to avoid returning decrypted (shared)
memory to the page allocator, which could lead to functional or security
issues.
DMA could free decrypted/shared pages if dma_set_decrypted() fails. This
should be a rare case. Just leak the pages in this case instead of
freeing them.
Signed-off-by: Rick Edgecombe <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
A bus_dma_region necessarily stores both CPU and DMA base addresses for
a range, so there's no need to also store the difference between them.
Signed-off-by: Robin Murphy <[email protected]>
Acked-by: Rob Herring <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
There is an unusual case that the range map covers right up to the top
of system RAM, but leaves a hole somewhere lower down. Then it prevents
the nvme device dma mapping in the checking path of phys_to_dma() and
causes the hangs at boot.
E.g. On an Armv8 Ampere server, the dsdt ACPI table is:
Method (_DMA, 0, Serialized) // _DMA: Direct Memory Access
{
Name (RBUF, ResourceTemplate ()
{
QWordMemory (ResourceConsumer, PosDecode, MinFixed,
MaxFixed, Cacheable, ReadWrite,
0x0000000000000000, // Granularity
0x0000000000000000, // Range Minimum
0x00000000FFFFFFFF, // Range Maximum
0x0000000000000000, // Translation Offset
0x0000000100000000, // Length
,, , AddressRangeMemory, TypeStatic)
QWordMemory (ResourceConsumer, PosDecode, MinFixed,
MaxFixed, Cacheable, ReadWrite,
0x0000000000000000, // Granularity
0x0000006010200000, // Range Minimum
0x000000602FFFFFFF, // Range Maximum
0x0000000000000000, // Translation Offset
0x000000001FE00000, // Length
,, , AddressRangeMemory, TypeStatic)
QWordMemory (ResourceConsumer, PosDecode, MinFixed,
MaxFixed, Cacheable, ReadWrite,
0x0000000000000000, // Granularity
0x00000060F0000000, // Range Minimum
0x00000060FFFFFFFF, // Range Maximum
0x0000000000000000, // Translation Offset
0x0000000010000000, // Length
,, , AddressRangeMemory, TypeStatic)
QWordMemory (ResourceConsumer, PosDecode, MinFixed,
MaxFixed, Cacheable, ReadWrite,
0x0000000000000000, // Granularity
0x0000007000000000, // Range Minimum
0x000003FFFFFFFFFF, // Range Maximum
0x0000000000000000, // Translation Offset
0x0000039000000000, // Length
,, , AddressRangeMemory, TypeStatic)
})
But the System RAM ranges are:
cat /proc/iomem |grep -i ram
90000000-91ffffff : System RAM
92900000-fffbffff : System RAM
880000000-fffffffff : System RAM
8800000000-bff5990fff : System RAM
bff59d0000-bff5a4ffff : System RAM
bff8000000-bfffffffff : System RAM
So some RAM ranges are out of dma_range_map.
Fix it by checking whether each of the system RAM resources can be
properly encompassed within the dma_range_map.
Signed-off-by: Jia He <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Log a warning once when dma_alloc_coherent fails because the platform
does not support coherent allocations at all.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Robin Murphy <[email protected]>
Reviewed-by: Greg Ungerer <[email protected]>
Tested-by: Greg Ungerer <[email protected]>
The logic in dma_direct_alloc when to use the atomic pool vs remapping
grew a bit unreadable. Consolidate it into a single check, and clean
up the set_uncached vs remap logic a bit as well.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Robin Murphy <[email protected]>
Reviewed-by: Greg Ungerer <[email protected]>
Tested-by: Greg Ungerer <[email protected]>
Instead of using arch_dma_alloc if none of the generic coherent
allocators are used, require the architectures to explicitly opt into
providing it. This will used to deal with the case of m68knommu and
coldfire where we can't do any coherent allocations whatsoever, and
also makes it clear that arch_dma_alloc is a last resort.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Robin Murphy <[email protected]>
Reviewed-by: Greg Ungerer <[email protected]>
Tested-by: Greg Ungerer <[email protected]>
Try to allocate a transient memory pool if no suitable slots can be found
and the respective SWIOTLB is allowed to grow. The transient pool is just
enough big for this one bounce buffer. It is inserted into a per-device
list of transient memory pools, and it is freed again when the bounce
buffer is unmapped.
Transient memory pools are kept in an RCU list. A memory barrier is
required after adding a new entry, because any address within a transient
buffer must be immediately recognized as belonging to the SWIOTLB, even if
it is passed to another CPU.
Deletion does not require any synchronization beyond RCU ordering
guarantees. After a buffer is unmapped, its physical addresses may no
longer be passed to the DMA API, so the memory range of the corresponding
stale entry in the RCU list never matches. If the memory range gets
allocated again, then it happens only after a RCU quiescent state.
Since bounce buffers can now be allocated from different pools, add a
parameter to swiotlb_alloc_pool() to let the caller know which memory pool
is used. Add swiotlb_find_pool() to find the memory pool corresponding to
an address. This function is now also used by is_swiotlb_buffer(), because
a simple boundary check is no longer sufficient.
The logic in swiotlb_alloc_tlb() is taken from __dma_direct_alloc_pages(),
simplified and enhanced to use coherent memory pools if needed.
Note that this is not the most efficient way to provide a bounce buffer,
but when a DMA buffer can't be mapped, something may (and will) actually
break. At that point it is better to make an allocation, even if it may be
an expensive operation.
Signed-off-by: Petr Tesarik <[email protected]>
Reviewed-by: Greg Kroah-Hartman <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Since both callers of dma_direct_optimal_gfp_mask() pass
dev->coherent_dma_mask as the second argument, it is better to
remove that parameter altogether.
Not only is reducing number of parameters good for readability, but
the new function signature is also more logical: The optimal flags
depend only on data contained in struct device.
While touching this code, let's also rename phys_mask to phys_limit
in dma_direct_alloc_from_pool(), because it is indeed a limit.
Signed-off-by: Petr Tesarik <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Pull dma-mapping updates from Christoph Hellwig:
- convert arm32 to the common dma-direct code (Arnd Bergmann, Robin
Murphy, Christoph Hellwig)
- restructure the PCIe peer to peer mapping support (Logan Gunthorpe)
- allow the IOMMU code to communicate an optional DMA mapping length
and use that in scsi and libata (John Garry)
- split the global swiotlb lock (Tianyu Lan)
- various fixes and cleanup (Chao Gao, Dan Carpenter, Dongli Zhang,
Lukas Bulwahn, Robin Murphy)
* tag 'dma-mapping-5.20-2022-08-06' of git://git.infradead.org/users/hch/dma-mapping: (45 commits)
swiotlb: fix passing local variable to debugfs_create_ulong()
dma-mapping: reformat comment to suppress htmldoc warning
PCI/P2PDMA: Remove pci_p2pdma_[un]map_sg()
RDMA/rw: drop pci_p2pdma_[un]map_sg()
RDMA/core: introduce ib_dma_pci_p2p_dma_supported()
nvme-pci: convert to using dma_map_sgtable()
nvme-pci: check DMA ops when indicating support for PCI P2PDMA
iommu/dma: support PCI P2PDMA pages in dma-iommu map_sg
iommu: Explicitly skip bus address marked segments in __iommu_map_sg()
dma-mapping: add flags to dma_map_ops to indicate PCI P2PDMA support
dma-direct: support PCI P2PDMA pages in dma-direct map_sg
dma-mapping: allow EREMOTEIO return code for P2PDMA transfers
PCI/P2PDMA: Introduce helpers for dma_map_sg implementations
PCI/P2PDMA: Attempt to set map_type if it has not been set
lib/scatterlist: add flag for indicating P2PDMA segments in an SGL
swiotlb: clean up some coding style and minor issues
dma-mapping: update comment after dmabounce removal
scsi: sd: Add a comment about limiting max_sectors to shost optimal limit
ata: libata-scsi: cap ata_device->max_sectors according to shost->max_sectors
scsi: scsi_transport_sas: cap shost opt_sectors according to DMA optimal limit
...
Add PCI P2PDMA support for dma_direct_map_sg() so that it can map
PCI P2PDMA pages directly without a hack in the callers. This allows
for heterogeneous SGLs that contain both P2PDMA and regular pages.
A P2PDMA page may have three possible outcomes when being mapped:
1) If the data path between the two devices doesn't go through the
root port, then it should be mapped with a PCI bus address
2) If the data path goes through the host bridge, it should be mapped
normally, as though it were a CPU physical address
3) It is not possible for the two devices to communicate and thus
the mapping operation should fail (and it will return -EREMOTEIO).
SGL segments that contain PCI bus addresses are marked with
sg_dma_mark_pci_p2pdma() and are ignored when unmapped.
P2PDMA mappings are also failed if swiotlb needs to be used on the
mapping.
Signed-off-by: Logan Gunthorpe <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
The third parameter of dma_set_encrypted() is a size in bytes rather than
the number of pages.
Fixes: 4d0564785b ("dma-direct: factor out dma_set_{de,en}crypted helpers")
Signed-off-by: Dexuan Cui <[email protected]>
Reviewed-by: Robin Murphy <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
The original x86 sev_alloc() only called set_memory_decrypted() on
memory returned by alloc_pages_node(), so the page order calculation
fell out of that logic. However, the common dma-direct code has several
potential allocators, not all of which are guaranteed to round up the
underlying allocation to a power-of-two size, so carrying over that
calculation for the encryption/decryption size was a mistake. Fix it by
rounding to a *number* of pages, rather than an order.
Until recently there was an even worse interaction with DMA_DIRECT_REMAP
where we could have ended up decrypting part of the next adjacent
vmalloc area, only averted by no architecture actually supporting both
configs at once. Don't ask how I found that one out...
Fixes: c10f07aa27 ("dma/direct: Handle force decryption for DMA coherent buffers in common code")
Signed-off-by: Robin Murphy <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Acked-by: David Rientjes <[email protected]>
When dma_direct_alloc_pages encounters a highmem page it just gives up
currently. But what we really should do is to try memory using the
page allocator instead - without this platforms with a global highmem
CMA pool will fail all dma_alloc_pages allocations.
Fixes: efa70f2fdc ("dma-mapping: add a new dma_alloc_pages API")
Reported-by: Mark O'Neill <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
pgprot_decrypted is used by AMD SME systems to allow access to memory
that was set to not encrypted using set_memory_decrypted. That only
happens for dma-direct memory as the IOMMU solves the addressing
challenges for the encryption bit using its own remapping.
Move the pgprot_decrypted call out of dma_pgprot which is also used
by the IOMMU mappings and into dma-direct so that it is only used with
memory that was set decrypted.
Fixes: f5ff79fddf ("dma-mapping: remove CONFIG_DMA_REMAP")
Reported-by: Alex Xu (Hello71) <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Tested-by: Alex Xu (Hello71) <[email protected]>
CONFIG_DMA_REMAP is used to build a few helpers around the core
vmalloc code, and to use them in case there is a highmem page in
dma-direct, and to make dma coherent allocations be able to use
non-contiguous pages allocations for DMA allocations in the dma-iommu
layer.
Right now it needs to be explicitly selected by architectures, and
is only done so by architectures that require remapping to deal
with devices that are not DMA coherent. Make it unconditional for
builds with CONFIG_MMU as it is very little extra code, but makes
it much more likely that large DMA allocations succeed on x86.
This fixes hot plugging a NVMe thunderbolt SSD for me, which tries
to allocate a 1MB buffer that is otherwise hard to obtain due to
memory fragmentation on a heavily used laptop.
Signed-off-by: Christoph Hellwig <[email protected]>
Reviewed-by: Robin Murphy <[email protected]>
Linus Torvalds