Linaro-mm-sig

linaro-mm-sig@lists.linaro.org

8 participants
3221 discussions

Re: [RFC v2 1/8] dma-buf: Add support for map/unmap APIs for interconnects

by Jason Gunthorpe

On Sun, Oct 26, 2025 at 09:44:13PM -0700, Vivek Kasireddy wrote: > For the map operation, the dma-buf core will create an xarray but > the exporter needs to populate it with the interconnect specific > addresses. And, similarly for unmap, the exporter is expected to > cleanup the individual entries of the xarray. I don't think we should limit this to xarrays, nor do I think it is a great datastructure for what is usually needed here.. I just posted the patches showing what iommufd needs, and it wants something like struct mapping { struct p2p_provider *provider; size_t nelms; struct phys_vec *phys; }; Which is not something that make sense as an xarray. I think the interconnect should have its own functions for map/unmap, ie instead of trying to have them as a commmon dma_buf_interconnect_ops do something like struct dma_buf_interconnect_ops { const char *name; bool (*supports_interconnects)(struct dma_buf_attachment *attach, const struct dma_buf_interconnect_match *, unsigned int num_ics); }; struct dma_buf_iov_interconnect_ops { struct dma_buf_interconnect_ops ic_ops; struct xx *(*map)(struct dma_buf_attachment *attach, unsigned int *bar_number, size_t *nelms); // No unmap for iov }; static inline struct xx *dma_buf_iov_map(struct dma_buf_attachment *attach, unsigned int *bar_number, size_t *nelms) { return container_of(attach->ic_ops, struct dma_buf_iov_interconnect_ops, ic_ops)->map( attach, bar_number, nelms)); } > +/** > + * dma_buf_attachment_is_dynamic - check if the importer can handle move_notify. > + * @attach: the attachment to check > + * > + * Returns true if a DMA-buf importer has indicated that it can handle dmabuf > + * location changes through the move_notify callback. > + */ > +static inline bool > +dma_buf_attachment_is_dynamic(struct dma_buf_attachment *attach) > +{ > + return !!attach->importer_ops; > +} Why is this in this patch? I also think this patch should be second in the series, it makes more sense to figure out how to attach with an interconnect then show how to map/unmap with that interconnect Like I'm not sure why this introduces allow_ic? Jason

1 month, 1 week

Re: [PATCH v5 9/9] vfio/pci: Add dma-buf export support for MMIO regions

by Jason Gunthorpe

On Sun, Oct 26, 2025 at 03:55:04PM +0800, Shuai Xue wrote: > > > 在 2025/10/22 20:50, Jason Gunthorpe 写道: > > On Mon, Oct 13, 2025 at 06:26:11PM +0300, Leon Romanovsky wrote: > > > From: Leon Romanovsky <leonro(a)nvidia.com> > > > > > > Add support for exporting PCI device MMIO regions through dma-buf, > > > enabling safe sharing of non-struct page memory with controlled > > > lifetime management. This allows RDMA and other subsystems to import > > > dma-buf FDs and build them into memory regions for PCI P2P operations. > > > > > > The implementation provides a revocable attachment mechanism using > > > dma-buf move operations. MMIO regions are normally pinned as BARs > > > don't change physical addresses, but access is revoked when the VFIO > > > device is closed or a PCI reset is issued. This ensures kernel > > > self-defense against potentially hostile userspace. > > > > Let's enhance this: > > > > Currently VFIO can take MMIO regions from the device's BAR and map > > them into a PFNMAP VMA with special PTEs. This mapping type ensures > > the memory cannot be used with things like pin_user_pages(), hmm, and > > so on. In practice only the user process CPU and KVM can safely make > > use of these VMA. When VFIO shuts down these VMAs are cleaned by > > unmap_mapping_range() to prevent any UAF of the MMIO beyond driver > > unbind. > > > > However, VFIO type 1 has an insecure behavior where it uses > > follow_pfnmap_*() to fish a MMIO PFN out of a VMA and program it back > > into the IOMMU. This has a long history of enabling P2P DMA inside > > VMs, but has serious lifetime problems by allowing a UAF of the MMIO > > after the VFIO driver has been unbound. > > Hi, Jason, > > Can you elaborate on this more? > > From my understanding of the VFIO type 1 implementation: > > - When a device is opened through VFIO type 1, it increments the > device->refcount > - During unbind, the driver waits for this refcount to drop to zero via > wait_for_completion(&device->comp) > - This should prevent the unbind() from completing while the device is > still in use > > Given this refcount mechanism, I do not figure out how the UAF can > occur. A second vfio device can be opened and then use follow_pfnmap_*() to read the first vfio device's PTEs. There is no relationship betweent the first and second VFIO devices, so once the first is unbound it sails through the device->comp while the second device retains the PFN in its type1 iommu_domain. Jason

1 month, 1 week

Re: [PATCH] dma-fence: Fix safe access wrapper to call timeline name method

by Christian König

On 10/21/25 20:36, Tvrtko Ursulin wrote: > > On 21/10/2025 17:09, Akash Goel wrote: >> This commit fixes the wrapper function dma_fence_timeline_name(), that >> was added for safe access, to actually call the timeline name method of >> dma_fence_ops. >> >> Cc: <stable(a)vger.kernel.org> # v6.17+ >> Signed-off-by: Akash Goel <akash.goel(a)arm.com> > > Fixes: 506aa8b02a8d ("dma-fence: Add safe access helpers and document the rules") > > Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin(a)igalia.com> Good catch, Reviewed-by: Christian König <christian.koenig(a)amd.com> as well. Please ping me if you need somebody to push this to drm-misc-fixes. Thanks, Christian. > > Apologies for the copy and paste snafu. > > Regards, > > Tvrtko > >> --- >> drivers/dma-buf/dma-fence.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c >> index 3f78c56b58dc..39e6f93dc310 100644 >> --- a/drivers/dma-buf/dma-fence.c >> +++ b/drivers/dma-buf/dma-fence.c >> @@ -1141,7 +1141,7 @@ const char __rcu *dma_fence_timeline_name(struct dma_fence *fence) >> "RCU protection is required for safe access to returned string"); >> if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) >> - return fence->ops->get_driver_name(fence); >> + return fence->ops->get_timeline_name(fence); >> else >> return "signaled-timeline"; >> } >

1 month, 1 week

Re: [RFC PATCH] dma-fence: Remove 64-bit flag

by Christian König

On 10/20/25 13:18, Matthew Brost wrote: > On Mon, Oct 20, 2025 at 10:16:23AM +0200, Philipp Stanner wrote: >> On Fri, 2025-10-17 at 14:28 -0700, Matthew Brost wrote: >>> On Fri, Oct 17, 2025 at 11:31:47AM +0200, Philipp Stanner wrote: >>>> It seems that DMA_FENCE_FLAG_SEQNO64_BIT has no real effects anymore, >>>> since seqno is a u64 everywhere. >>>> >>>> Remove the unneeded flag. >>>> >>>> Signed-off-by: Philipp Stanner <phasta(a)kernel.org> >>>> --- >>>> Seems to me that this flag doesn't really do anything anymore? >>>> >>>> I *suspect* that it could be that some drivers pass a u32 to >>>> dma_fence_init()? I guess they could be ported, couldn't they. >>>> >>> >>> Xe uses 32-bit hardware fence sequence numbers—see [1] and [2]. We could >>> switch to 64-bit hardware fence sequence numbers, but that would require >>> changes on the driver side. If you sent this to our CI, I’m fairly >>> certain we’d see a bunch of failures. I suspect this would also break >>> several other drivers. >> >> What exactly breaks? Help me out here; if you pass a u32 for a u64, > > Seqno wraps. > >> doesn't the C standard guarantee that the higher, unused 32 bits will >> be 0? > > return (int)(lower_32_bits(f1) - lower_32_bits(f2)) > 0; > > Look at the above logic. > > f1 = 0x0; > f2 = 0xffffffff; /* -1 */ > > The above statement will correctly return true. > > Compared to the below statement which returns false. > > return f1 > f2; > > We test seqno wraps in Xe by setting our initial seqno to -127, again if > you send this patch to our CI any test which sends more than 127 job on > queue will likely fail. Yeah, exactly that's why this flag is needed for quite a lot of things. Question is what is missing in the documentation to make that clear? Regards, Christian. > > Matt > >> >> Because the only thing the flag still does is do this lower_32 check in >> fence_is_later. >> >> P. >> >>> >>> As I mentioned, all Xe-supported platforms could be updated since their >>> rings support 64-bit store instructions. However, I suspect that very >>> old i915 platforms don’t support such instructions in the ring. I agree >>> this is a legacy issue, and we should probably use 64-bit sequence >>> numbers in Xe. But again, platforms and drivers that are decades old >>> might break as a result. >>> >>> Matt >>> >>> [1] https://elixir.bootlin.com/linux/v6.17.1/source/drivers/gpu/drm/xe/xe_hw_fe… >>> [2] https://elixir.bootlin.com/linux/v6.17.1/source/drivers/gpu/drm/xe/xe_hw_fe… >>> >>>> P. >>>> --- >>>> drivers/dma-buf/dma-fence.c | 3 +-- >>>> include/linux/dma-fence.h | 10 +--------- >>>> 2 files changed, 2 insertions(+), 11 deletions(-) >>>> >>>> diff --git a/drivers/dma-buf/dma-fence.c b/drivers/dma-buf/dma-fence.c >>>> index 3f78c56b58dc..24794c027813 100644 >>>> --- a/drivers/dma-buf/dma-fence.c >>>> +++ b/drivers/dma-buf/dma-fence.c >>>> @@ -1078,8 +1078,7 @@ void >>>> dma_fence_init64(struct dma_fence *fence, const struct dma_fence_ops *ops, >>>> spinlock_t *lock, u64 context, u64 seqno) >>>> { >>>> - __dma_fence_init(fence, ops, lock, context, seqno, >>>> - BIT(DMA_FENCE_FLAG_SEQNO64_BIT)); >>>> + __dma_fence_init(fence, ops, lock, context, seqno, 0); >>>> } >>>> EXPORT_SYMBOL(dma_fence_init64); >>>> >>>> diff --git a/include/linux/dma-fence.h b/include/linux/dma-fence.h >>>> index 64639e104110..4eca2db28625 100644 >>>> --- a/include/linux/dma-fence.h >>>> +++ b/include/linux/dma-fence.h >>>> @@ -98,7 +98,6 @@ struct dma_fence { >>>> }; >>>> >>>> enum dma_fence_flag_bits { >>>> - DMA_FENCE_FLAG_SEQNO64_BIT, >>>> DMA_FENCE_FLAG_SIGNALED_BIT, >>>> DMA_FENCE_FLAG_TIMESTAMP_BIT, >>>> DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, >>>> @@ -470,14 +469,7 @@ dma_fence_is_signaled(struct dma_fence *fence) >>>> */ >>>> static inline bool __dma_fence_is_later(struct dma_fence *fence, u64 f1, u64 f2) >>>> { >>>> - /* This is for backward compatibility with drivers which can only handle >>>> - * 32bit sequence numbers. Use a 64bit compare when the driver says to >>>> - * do so. >>>> - */ >>>> - if (test_bit(DMA_FENCE_FLAG_SEQNO64_BIT, &fence->flags)) >>>> - return f1 > f2; >>>> - >>>> - return (int)(lower_32_bits(f1) - lower_32_bits(f2)) > 0; >>>> + return f1 > f2; >>>> } >>>> >>>> /** >>>> -- >>>> 2.49.0 >>>> >>

1 month, 1 week

[PATCH AUTOSEL 6.17] drm/xe: improve dma-resv handling for backup object

by Sasha Levin

From: Matthew Auld <matthew.auld(a)intel.com> [ Upstream commit edb1745fc618ba8ef63a45ce3ae60de1bdf29231 ] Since the dma-resv is shared we don't need to reserve and add a fence slot fence twice, plus no need to loop through the dependencies. Signed-off-by: Matthew Auld <matthew.auld(a)intel.com> Cc: Thomas Hellström <thomas.hellstrom(a)linux.intel.com> Cc: Matthew Brost <matthew.brost(a)intel.com> Reviewed-by: Jonathan Cavitt <jonathan.cavitt(a)intel.com> Reviewed-by: Thomas Hellström <thomas.hellstrom(a)linux.intel.com> Link: https://lore.kernel.org/r/20250829164715.720735-2-matthew.auld@intel.com Signed-off-by: Sasha Levin <sashal(a)kernel.org> --- LLM Generated explanations, may be completely bogus: YES Explanation - What it fixes - Removes redundant dma-resv operations when a backup BO shares the same reservation object as the original BO, preventing the same fence from being reserved/added twice to the same `dma_resv`. - Avoids scanning the same dependency set twice when source and destination BOs share the same `dma_resv`. - Why the change is correct - The backup object is created to share the parent’s reservation object, so a single reserve/add is sufficient: - The backup BO is initialized with the parent’s resv: `drivers/gpu/drm/xe/xe_bo.c:1309` (`xe_bo_init_locked(..., bo->ttm.base.resv, ...)`), ensuring `bo->ttm.base.resv == backup->ttm.base.resv`. - The patch adds an explicit invariant check to document and enforce this: `drivers/gpu/drm/xe/xe_bo.c:1225` (`xe_assert(xe, bo->ttm.base.resv == backup->ttm.base.resv)`). - With shared `dma_resv`, adding the same fence twice is at best redundant (wasting fence slots and memory) and at worst error-prone. Reserving fence slots only once and adding the fence once is the correct behavior. - Specific code changes and effects - Evict path (GPU migration copy case): - Before: reserves and adds fence on both `bo->ttm.base.resv` and `backup->ttm.base.resv`. - After: reserves and adds exactly once, guarded by the shared-resv assertion. - See single reserve and add: `drivers/gpu/drm/xe/xe_bo.c:1226` (reserve) and `drivers/gpu/drm/xe/xe_bo.c:1237` (add fence). This is the core fix; the removed second reserve/add on the backup is the redundant part eliminated. - Restore path (migration copy back): - Same simplification: reserve once, add once on the shared `dma_resv`. - See single reserve and add: `drivers/gpu/drm/xe/xe_bo.c:1375` (reserve) and `drivers/gpu/drm/xe/xe_bo.c:1387` (add fence). - Dependency handling in migrate: - Before: added deps for both src and dst based only on `src_bo != dst_bo`. - After: only add dst deps if the resv objects differ, avoiding double-walking the same `dma_resv`. - See updated condition: `drivers/gpu/drm/xe/xe_migrate.c:932` (`src_bo->ttm.base.resv != dst_bo->ttm.base.resv`). - User-visible impact without the patch - Duplicate `dma_resv_add_fence()` calls on the same reservation object can: - Consume extra shared-fence slots and memory. - Inflate dependency lists, causing unnecessary scheduler waits and overhead. - Increase failure likelihood of `dma_resv_reserve_fences()` under memory pressure. - These paths are exercised during suspend/resume flows of pinned VRAM BOs (evict/restore), so reliability and performance in power transitions can be affected. - Scope and risk - Small, focused changes localized to the Intel Xe driver migration/evict/restore paths: - Files: `drivers/gpu/drm/xe/xe_bo.c`, `drivers/gpu/drm/xe/xe_migrate.c`. - No API changes or architectural refactors; logic strictly reduces redundant operations. - The `xe_assert` acts as a safety net to catch unexpected non-shared `resv` usage; normal runtime behavior is unchanged when the invariant holds. - The CPU copy fallback paths are untouched. - Stable backport considerations - This is a clear correctness and robustness fix, not a feature. - Low regression risk if the stable branch also creates the backup BO with the parent’s `dma_resv` (as shown by the use of `xe_bo_init_locked(..., bo->ttm.base.resv, ...)` in `drivers/gpu/drm/xe/xe_bo.c:1309`). - If a stable branch diverges and the backup BO does not share the resv, this patch would need adjustment (i.e., keep dual reserve/add in that case). The added `xe_assert` helps surface such mismatches during testing. Conclusion: This commit fixes a real bug (duplicate fence reserve/add and duplicate dependency scanning on a shared `dma_resv`) with a minimal, well-scoped change. It aligns with stable rules (important bugfix, low risk, contained), so it should be backported. drivers/gpu/drm/xe/xe_bo.c | 13 +------------ drivers/gpu/drm/xe/xe_migrate.c | 2 +- 2 files changed, 2 insertions(+), 13 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index d07e23eb1a54d..5a61441d68af5 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -1242,14 +1242,11 @@ int xe_bo_evict_pinned(struct xe_bo *bo) else migrate = mem_type_to_migrate(xe, bo->ttm.resource->mem_type); + xe_assert(xe, bo->ttm.base.resv == backup->ttm.base.resv); ret = dma_resv_reserve_fences(bo->ttm.base.resv, 1); if (ret) goto out_backup; - ret = dma_resv_reserve_fences(backup->ttm.base.resv, 1); - if (ret) - goto out_backup; - fence = xe_migrate_copy(migrate, bo, backup, bo->ttm.resource, backup->ttm.resource, false); if (IS_ERR(fence)) { @@ -1259,8 +1256,6 @@ int xe_bo_evict_pinned(struct xe_bo *bo) dma_resv_add_fence(bo->ttm.base.resv, fence, DMA_RESV_USAGE_KERNEL); - dma_resv_add_fence(backup->ttm.base.resv, fence, - DMA_RESV_USAGE_KERNEL); dma_fence_put(fence); } else { ret = xe_bo_vmap(backup); @@ -1338,10 +1333,6 @@ int xe_bo_restore_pinned(struct xe_bo *bo) if (ret) goto out_unlock_bo; - ret = dma_resv_reserve_fences(backup->ttm.base.resv, 1); - if (ret) - goto out_unlock_bo; - fence = xe_migrate_copy(migrate, backup, bo, backup->ttm.resource, bo->ttm.resource, false); @@ -1352,8 +1343,6 @@ int xe_bo_restore_pinned(struct xe_bo *bo) dma_resv_add_fence(bo->ttm.base.resv, fence, DMA_RESV_USAGE_KERNEL); - dma_resv_add_fence(backup->ttm.base.resv, fence, - DMA_RESV_USAGE_KERNEL); dma_fence_put(fence); } else { ret = xe_bo_vmap(backup); diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c index 2a627ed64b8f8..ba9b8590eccb2 100644 --- a/drivers/gpu/drm/xe/xe_migrate.c +++ b/drivers/gpu/drm/xe/xe_migrate.c @@ -901,7 +901,7 @@ struct dma_fence *xe_migrate_copy(struct xe_migrate *m, if (!fence) { err = xe_sched_job_add_deps(job, src_bo->ttm.base.resv, DMA_RESV_USAGE_BOOKKEEP); - if (!err && src_bo != dst_bo) + if (!err && src_bo->ttm.base.resv != dst_bo->ttm.base.resv) err = xe_sched_job_add_deps(job, dst_bo->ttm.base.resv, DMA_RESV_USAGE_BOOKKEEP); if (err) -- 2.51.0

1 month, 1 week

Re: [PATCH v2] dma-buf: system_heap: use larger contiguous mappings instead of per-page mmap

by Maxime Ripard

On Tue, 21 Oct 2025 17:20:22 +1300, Barry Song wrote: > From: Barry Song <v-songbaohua(a)oppo.com> > > We can allocate high-order pages, but mapping them one by > one is inefficient. This patch changes the code to map > as large a chunk as possible. The code looks somewhat > > [ ... ] Reviewed-by: Maxime Ripard <mripard(a)kernel.org> Thanks! Maxime

1 month, 1 week

[PATCH v5 3/8] rust: drm: gem: Add raw_dma_resv() function

by Lyude Paul

For retrieving a pointer to the struct dma_resv for a given GEM object. We also introduce it in a new trait, BaseObjectPrivate, which we automatically implement for all gem objects and don't expose to users outside of the crate. Signed-off-by: Lyude Paul <lyude(a)redhat.com> --- rust/kernel/drm/gem/mod.rs | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/rust/kernel/drm/gem/mod.rs b/rust/kernel/drm/gem/mod.rs index 32bff2e8463f4..67813cfb0db42 100644 --- a/rust/kernel/drm/gem/mod.rs +++ b/rust/kernel/drm/gem/mod.rs @@ -200,6 +200,18 @@ fn create_mmap_offset(&self) -> Result<u64> { impl<T: IntoGEMObject> BaseObject for T {} +/// Crate-private base operations shared by all GEM object classes. +#[expect(unused)] +pub(crate) trait BaseObjectPrivate: IntoGEMObject { + /// Return a pointer to this object's dma_resv. + fn raw_dma_resv(&self) -> *mut bindings::dma_resv { + // SAFETY: `as_gem_obj()` always returns a valid pointer to the base DRM gem object + unsafe { (*self.as_raw()).resv } + } +} + +impl<T: IntoGEMObject> BaseObjectPrivate for T {} + /// A base GEM object. /// /// Invariants -- 2.51.0

1 month, 2 weeks

[PATCH v5 2/8] rust: helpers: Add bindings/wrappers for dma_resv_lock

by Lyude Paul

From: Asahi Lina <lina(a)asahilina.net> This is just for basic usage in the DRM shmem abstractions for implied locking, not intended as a full DMA Reservation abstraction yet. Signed-off-by: Asahi Lina <lina(a)asahilina.net> Signed-off-by: Daniel Almeida <daniel.almeida(a)collabora.com> Reviewed-by: Alice Ryhl <aliceryhl(a)google.com> Signed-off-by: Lyude Paul <lyude(a)redhat.com> --- rust/bindings/bindings_helper.h | 1 + rust/helpers/dma-resv.c | 13 +++++++++++++ rust/helpers/helpers.c | 1 + 3 files changed, 15 insertions(+) create mode 100644 rust/helpers/dma-resv.c diff --git a/rust/bindings/bindings_helper.h b/rust/bindings/bindings_helper.h index 2e43c66635a2c..07f79e125c329 100644 --- a/rust/bindings/bindings_helper.h +++ b/rust/bindings/bindings_helper.h @@ -48,6 +48,7 @@ #include <linux/cpumask.h> #include <linux/cred.h> #include <linux/debugfs.h> +#include <linux/dma-resv.h> #include <linux/device/faux.h> #include <linux/dma-direction.h> #include <linux/dma-mapping.h> diff --git a/rust/helpers/dma-resv.c b/rust/helpers/dma-resv.c new file mode 100644 index 0000000000000..05501cb814513 --- /dev/null +++ b/rust/helpers/dma-resv.c @@ -0,0 +1,13 @@ +// SPDX-License-Identifier: GPL-2.0 + +#include <linux/dma-resv.h> + +int rust_helper_dma_resv_lock(struct dma_resv *obj, struct ww_acquire_ctx *ctx) +{ + return dma_resv_lock(obj, ctx); +} + +void rust_helper_dma_resv_unlock(struct dma_resv *obj) +{ + dma_resv_unlock(obj); +} diff --git a/rust/helpers/helpers.c b/rust/helpers/helpers.c index 551da6c9b5064..36d40f911345c 100644 --- a/rust/helpers/helpers.c +++ b/rust/helpers/helpers.c @@ -25,6 +25,7 @@ #include "cred.c" #include "device.c" #include "dma.c" +#include "dma-resv.c" #include "drm.c" #include "err.c" #include "irq.c" -- 2.51.0

1 month, 2 weeks

Re: [PATCH v5 2/2] accel: Add Arm Ethos-U NPU driver

by Rob Herring

On Tue, Oct 21, 2025 at 4:43 PM Matthew Brost <matthew.brost(a)intel.com> wrote: > > On Sat, Oct 18, 2025 at 12:42:30AM -0700, Matthew Brost wrote: > > On Fri, Oct 17, 2025 at 11:43:51PM -0700, Matthew Brost wrote: > > > On Fri, Oct 17, 2025 at 10:37:46AM -0500, Rob Herring wrote: > > > > On Thu, Oct 16, 2025 at 11:25:34PM -0700, Matthew Brost wrote: > > > > > On Thu, Oct 16, 2025 at 04:06:05PM -0500, Rob Herring (Arm) wrote: > > > > > > Add a driver for Arm Ethos-U65/U85 NPUs. The Ethos-U NPU has a > > > > > > relatively simple interface with single command stream to describe > > > > > > buffers, operation settings, and network operations. It supports up to 8 > > > > > > memory regions (though no h/w bounds on a region). The Ethos NPUs > > > > > > are designed to use an SRAM for scratch memory. Region 2 is reserved > > > > > > for SRAM (like the downstream driver stack and compiler). Userspace > > > > > > doesn't need access to the SRAM. > > > > > > > > Thanks for the review. > > > > > > > > [...] > > > > > > > > > > +static struct dma_fence *ethosu_job_run(struct drm_sched_job *sched_job) > > > > > > +{ > > > > > > + struct ethosu_job *job = to_ethosu_job(sched_job); > > > > > > + struct ethosu_device *dev = job->dev; > > > > > > + struct dma_fence *fence = NULL; > > > > > > + int ret; > > > > > > + > > > > > > + if (unlikely(job->base.s_fence->finished.error)) > > > > > > + return NULL; > > > > > > + > > > > > > + fence = ethosu_fence_create(dev); > > > > > > > > > > Another reclaim issue: ethosu_fence_create allocates memory using > > > > > GFP_KERNEL. Since we're already in the DMA fence signaling path > > > > > (reclaim), this can lead to a deadlock. > > > > > > > > > > Without too much thought, you likely want to move this allocation to > > > > > ethosu_job_do_push, but before taking dev->sched_lock or calling > > > > > drm_sched_job_arm. > > > > > > > > > > We really should fix the DRM scheduler work queue to be tainted with > > > > > reclaim. If I recall correctly, we'd need to update the work queue > > > > > layer. Let me look into that—I've seen this type of bug several times, > > > > > and lockdep should be able to catch it. > > > > > > > > Likely the rocket driver suffers from the same issues... > > > > > > > > > > I am not surprised by this statement. > > > > > > > > > > > > > > + if (IS_ERR(fence)) > > > > > > + return fence; > > > > > > + > > > > > > + if (job->done_fence) > > > > > > + dma_fence_put(job->done_fence); > > > > > > + job->done_fence = dma_fence_get(fence); > > > > > > + > > > > > > + ret = pm_runtime_get_sync(dev->base.dev); > > > > > > > > > > I haven't looked at your PM design, but this generally looks quite > > > > > dangerous with respect to reclaim. For example, if your PM resume paths > > > > > allocate memory or take locks that allocate memory underneath, you're > > > > > likely to run into issues. > > > > > > > > > > A better approach would be to attach a PM reference to your job upon > > > > > creation and release it upon job destruction. That would be safer and > > > > > save you headaches in the long run. > > > > > > > > Our PM is nothing more than clock enable/disable and register init. > > > > > > > > If the runtime PM API doesn't work and needs special driver wrappers, > > > > then I'm inclined to just not use it and manage clocks directly (as > > > > that's all it is doing). > > > > > > > > > > Yes, then you’re probably fine. More complex drivers can do all sorts of > > > things during a PM wake, which is why PM wakes should generally be the > > > outermost layer. I still suggest, to future-proof your code, that you > > > move the PM reference to an outer layer. > > > > > > > Also, taking a PM reference in a function call — as opposed to tying it > > to a object's lifetime — is risky. It can quickly lead to imbalances in > > PM references if things go sideways or function calls become unbalanced. > > Depending on how your driver uses the DRM scheduler, this seems like a > > real possibility. > > > > Matt > > > > > > > > > > > > This is what we do in Xe [1] [2]. > > > > > > > > > > Also, in general, this driver has been reviewed (RB’d), but it's not > > > > > great that I spotted numerous issues within just five minutes. I suggest > > > > > taking a step back and thoroughly evaluating everything this driver is > > > > > doing. > > > > > > > > Well, if it is hard to get simple drivers right, then it's a problem > > > > with the subsystem APIs IMO. > > > > > > > > > > Yes, agreed. We should have assertions and lockdep annotations in place > > > to catch driver-side misuses. This is the second driver I’ve randomly > > > looked at over the past year that has broken DMA fencing and reclaim > > > rules. I’ll take an action item to fix this in the DRM scheduler, but > > > I’m afraid I’ll likely break multiple drivers in the process as misuess > > > / lockdep will complain. > > I've posted a series [1] for the DRM scheduler which will complain about the > things I've pointed out here. Thanks. I ran v6 with them and no lockdep splats. Rob

1 month, 2 weeks

[PATCH v5 0/9] vfio/pci: Allow MMIO regions to be exported through dma-buf

by Leon Romanovsky

Changelog: v5: * Rebased on top of v6.18-rc1. * Added more validation logic to make sure that DMA-BUF length doesn't overflow in various scenarios. * Hide kernel config from the users. * Fixed type conversion issue. DMA ranges are exposed with u64 length, but DMA-BUF uses "unsigned int" as a length for SG entries. * Added check to prevent from VFIO drivers which reports BAR size different from PCI, do not use DMA-BUF functionality. v4: https://lore.kernel.org/all/cover.1759070796.git.leon@kernel.org * Split pcim_p2pdma_provider() to two functions, one that initializes array of providers and another to return right provider pointer. v3: https://lore.kernel.org/all/cover.1758804980.git.leon@kernel.org * Changed pcim_p2pdma_enable() to be pcim_p2pdma_provider(). * Cache provider in vfio_pci_dma_buf struct instead of BAR index. * Removed misleading comment from pcim_p2pdma_provider(). * Moved MMIO check to be in pcim_p2pdma_provider(). v2: https://lore.kernel.org/all/cover.1757589589.git.leon@kernel.org/ * Added extra patch which adds new CONFIG, so next patches can reuse * it. * Squashed "PCI/P2PDMA: Remove redundant bus_offset from map state" into the other patch. * Fixed revoke calls to be aligned with true->false semantics. * Extended p2pdma_providers to be per-BAR and not global to whole * device. * Fixed possible race between dmabuf states and revoke. * Moved revoke to PCI BAR zap block. v1: https://lore.kernel.org/all/cover.1754311439.git.leon@kernel.org * Changed commit messages. * Reused DMA_ATTR_MMIO attribute. * Returned support for multiple DMA ranges per-dMABUF. v0: https://lore.kernel.org/all/cover.1753274085.git.leonro@nvidia.com --------------------------------------------------------------------------- Based on "[PATCH v6 00/16] dma-mapping: migrate to physical address-based API" https://lore.kernel.org/all/cover.1757423202.git.leonro@nvidia.com/ series. --------------------------------------------------------------------------- This series extends the VFIO PCI subsystem to support exporting MMIO regions from PCI device BARs as dma-buf objects, enabling safe sharing of non-struct page memory with controlled lifetime management. This allows RDMA and other subsystems to import dma-buf FDs and build them into memory regions for PCI P2P operations. The series supports a use case for SPDK where a NVMe device will be owned by SPDK through VFIO but interacting with a RDMA device. The RDMA device may directly access the NVMe CMB or directly manipulate the NVMe device's doorbell using PCI P2P. However, as a general mechanism, it can support many other scenarios with VFIO. This dmabuf approach can be usable by iommufd as well for generic and safe P2P mappings. In addition to the SPDK use-case mentioned above, the capability added in this patch series can also be useful when a buffer (located in device memory such as VRAM) needs to be shared between any two dGPU devices or instances (assuming one of them is bound to VFIO PCI) as long as they are P2P DMA compatible. The implementation provides a revocable attachment mechanism using dma-buf move operations. MMIO regions are normally pinned as BARs don't change physical addresses, but access is revoked when the VFIO device is closed or a PCI reset is issued. This ensures kernel self-defense against potentially hostile userspace. The series includes significant refactoring of the PCI P2PDMA subsystem to separate core P2P functionality from memory allocation features, making it more modular and suitable for VFIO use cases that don't need struct page support. ----------------------------------------------------------------------- The series is based originally on https://lore.kernel.org/all/20250307052248.405803-1-vivek.kasireddy@intel.c… but heavily rewritten to be based on DMA physical API. ----------------------------------------------------------------------- The WIP branch can be found here: https://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma.git/log/?h=… Thanks Leon Romanovsky (7): PCI/P2PDMA: Separate the mmap() support from the core logic PCI/P2PDMA: Simplify bus address mapping API PCI/P2PDMA: Refactor to separate core P2P functionality from memory allocation PCI/P2PDMA: Export pci_p2pdma_map_type() function types: move phys_vec definition to common header vfio/pci: Enable peer-to-peer DMA transactions by default vfio/pci: Add dma-buf export support for MMIO regions Vivek Kasireddy (2): vfio: Export vfio device get and put registration helpers vfio/pci: Share the core device pointer while invoking feature functions block/blk-mq-dma.c | 7 +- drivers/iommu/dma-iommu.c | 4 +- drivers/pci/p2pdma.c | 175 ++++++++--- drivers/vfio/pci/Kconfig | 3 + drivers/vfio/pci/Makefile | 2 + drivers/vfio/pci/vfio_pci_config.c | 22 +- drivers/vfio/pci/vfio_pci_core.c | 63 ++-- drivers/vfio/pci/vfio_pci_dmabuf.c | 446 +++++++++++++++++++++++++++++ drivers/vfio/pci/vfio_pci_priv.h | 23 ++ drivers/vfio/vfio_main.c | 2 + include/linux/pci-p2pdma.h | 120 +++++--- include/linux/types.h | 5 + include/linux/vfio.h | 2 + include/linux/vfio_pci_core.h | 1 + include/uapi/linux/vfio.h | 25 ++ kernel/dma/direct.c | 4 +- mm/hmm.c | 2 +- 17 files changed, 785 insertions(+), 121 deletions(-) create mode 100644 drivers/vfio/pci/vfio_pci_dmabuf.c -- 2.51.0

1 month, 2 weeks

Jump to page:

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

Linaro-mm-sig