Linaro-mm-sig

linaro-mm-sig@lists.linaro.org

15 participants
3337 discussions

Re: [RFC v2 00/11] Add dmabuf read/write via io_uring

by Christian König

On 11/24/25 12:30, Pavel Begunkov wrote: > On 11/24/25 10:33, Christian König wrote: >> On 11/23/25 23:51, Pavel Begunkov wrote: >>> Picking up the work on supporting dmabuf in the read/write path. >> >> IIRC that work was completely stopped because it violated core dma_fence and DMA-buf rules and after some private discussion was considered not doable in general. >> >> Or am I mixing something up here? > > The time gap is purely due to me being busy. I wasn't CC'ed to those private > discussions you mentioned, but the v1 feedback was to use dynamic attachments > and avoid passing dma address arrays directly. > > https://lore.kernel.org/all/cover.1751035820.git.asml.silence@gmail.com/ > > I'm lost on what part is not doable. Can you elaborate on the core > dma-fence dma-buf rules? I most likely mixed that up, in other words that was a different discussion. When you use dma_fences to indicate async completion of events you need to be super duper careful that you only do this for in flight events, have the fence creation in the right order etc... For example once the fence is created you can't make any memory allocations any more, that's why we have this dance of reserving fence slots, creating the fence and then adding it. >> Since I don't see any dma_fence implementation at all that might actually be the case. > > See Patch 5, struct blk_mq_dma_fence. It's used in the move_notify > callback and is signaled when all inflight IO using the current > mapping are complete. All new IO requests will try to recreate the > mapping, and hence potentially wait with dma_resv_wait_timeout(). Without looking at the code that approach sounds more or less correct to me. >> On the other hand we have direct I/O from DMA-buf working for quite a while, just not upstream and without io_uring support. > > Have any reference? There is a WIP feature in AMDs GPU driver package for ROCm. But that can't be used as general purpose DMA-buf approach, because it makes use of internal knowledge about how the GPU driver is using the backing store. BTW when you use DMA addresses from DMA-buf always keep in mind that this memory can be written by others at the same time, e.g. you can't do things like compute a CRC first, then write to backing store and finally compare CRC. Regards, Christian.

2 months, 2 weeks

Re: [RFC v2 00/11] Add dmabuf read/write via io_uring

by Christian König

On 11/23/25 23:51, Pavel Begunkov wrote: > Picking up the work on supporting dmabuf in the read/write path. IIRC that work was completely stopped because it violated core dma_fence and DMA-buf rules and after some private discussion was considered not doable in general. Or am I mixing something up here? Since I don't see any dma_fence implementation at all that might actually be the case. On the other hand we have direct I/O from DMA-buf working for quite a while, just not upstream and without io_uring support. Regards, Christian. > There > are two main changes. First, it doesn't pass a dma addresss directly by > rather wraps it into an opaque structure, which is extended and > understood by the target driver. > > The second big change is support for dynamic attachments, which added a > good part of complexity (see Patch 5). I kept the main machinery in nvme > at first, but move_notify can ask to kill the dma mapping asynchronously, > and any new IO would need to wait during submission, thus it was moved > to blk-mq. That also introduced an extra callback layer b/w driver and > blk-mq. > > There are some rough corners, and I'm not perfectly happy about the > complexity and layering. For v3 I'll try to move the waiting up in the > stack to io_uring wrapped into library helpers. > > For now, I'm interested what is the best way to test move_notify? And > how dma_resv_reserve_fences() errors should be handled in move_notify? > > The uapi didn't change, after registration it looks like a normal > io_uring registered buffer and can be used as such. Only non-vectored > fixed reads/writes are allowed. Pseudo code: > > // registration > reg_buf_idx = 0; > io_uring_update_buffer(ring, reg_buf_idx, { dma_buf_fd, file_fd }); > > // request creation > io_uring_prep_read_fixed(sqe, file_fd, buffer_offset, > buffer_size, file_offset, reg_buf_idx); > > And as previously, a good bunch of code was taken from Keith's series [1]. > > liburing based example: > > git: https://github.com/isilence/liburing.git dmabuf-rw > link: https://github.com/isilence/liburing/tree/dmabuf-rw > > [1] https://lore.kernel.org/io-uring/20220805162444.3985535-1-kbusch@fb.com/ > > Pavel Begunkov (11): > file: add callback for pre-mapping dmabuf > iov_iter: introduce iter type for pre-registered dma > block: move around bio flagging helpers > block: introduce dma token backed bio type > block: add infra to handle dmabuf tokens > nvme-pci: add support for dmabuf reggistration > nvme-pci: implement dma_token backed requests > io_uring/rsrc: add imu flags > io_uring/rsrc: extended reg buffer registration > io_uring/rsrc: add dmabuf-backed buffer registeration > io_uring/rsrc: implement dmabuf regbuf import > > block/Makefile | 1 + > block/bdev.c | 14 ++ > block/bio.c | 21 +++ > block/blk-merge.c | 23 +++ > block/blk-mq-dma-token.c | 236 +++++++++++++++++++++++++++++++ > block/blk-mq.c | 20 +++ > block/blk.h | 3 +- > block/fops.c | 3 + > drivers/nvme/host/pci.c | 217 ++++++++++++++++++++++++++++ > include/linux/bio.h | 49 ++++--- > include/linux/blk-mq-dma-token.h | 60 ++++++++ > include/linux/blk-mq.h | 21 +++ > include/linux/blk_types.h | 8 +- > include/linux/blkdev.h | 3 + > include/linux/dma_token.h | 35 +++++ > include/linux/fs.h | 4 + > include/linux/uio.h | 10 ++ > include/uapi/linux/io_uring.h | 13 +- > io_uring/rsrc.c | 201 +++++++++++++++++++++++--- > io_uring/rsrc.h | 23 ++- > io_uring/rw.c | 7 +- > lib/iov_iter.c | 30 +++- > 22 files changed, 948 insertions(+), 54 deletions(-) > create mode 100644 block/blk-mq-dma-token.c > create mode 100644 include/linux/blk-mq-dma-token.h > create mode 100644 include/linux/dma_token.h >

2 months, 2 weeks

Re: [PATCH v3 27/28] drm/amdgpu: get rid of amdgpu_ttm_clear_buffer

by Christian König

On 11/21/25 11:12, Pierre-Eric Pelloux-Prayer wrote: > It's doing the same thing as amdgpu_fill_buffer(src_data=0), so drop it. > > The only caveat is that amdgpu_res_cleared() return value is only valid > right after allocation. > > --- > v2: introduce new "bool consider_clear_status" arg > --- > > Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer(a)amd.com> It would be better to have that ealier in the patch set, but I guess that gives you rebasing problems? Christian. > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 16 ++-- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 90 +++++----------------- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 7 +- > 3 files changed, 33 insertions(+), 80 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > index 7d8d70135cc2..dccc31d0128e 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > @@ -725,13 +725,17 @@ int amdgpu_bo_create(struct amdgpu_device *adev, > bo->tbo.resource->mem_type == TTM_PL_VRAM) { > struct dma_fence *fence; > > - r = amdgpu_ttm_clear_buffer(adev, bo, bo->tbo.base.resv, &fence); > + r = amdgpu_fill_buffer(adev, amdgpu_ttm_next_clear_entity(adev), > + bo, 0, NULL, &fence, > + true, AMDGPU_KERNEL_JOB_ID_TTM_CLEAR_BUFFER); > if (unlikely(r)) > goto fail_unreserve; > > - dma_resv_add_fence(bo->tbo.base.resv, fence, > - DMA_RESV_USAGE_KERNEL); > - dma_fence_put(fence); > + if (fence) { > + dma_resv_add_fence(bo->tbo.base.resv, fence, > + DMA_RESV_USAGE_KERNEL); > + dma_fence_put(fence); > + } > } > if (!bp->resv) > amdgpu_bo_unreserve(bo); > @@ -1323,8 +1327,8 @@ void amdgpu_bo_release_notify(struct ttm_buffer_object *bo) > goto out; > > r = amdgpu_fill_buffer(adev, amdgpu_ttm_next_clear_entity(adev), > - abo, 0, &bo->base._resv, > - &fence, AMDGPU_KERNEL_JOB_ID_CLEAR_ON_RELEASE); > + abo, 0, &bo->base._resv, &fence, > + false, AMDGPU_KERNEL_JOB_ID_CLEAR_ON_RELEASE); > if (WARN_ON(r)) > goto out; > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > index 39cfe2dbdf03..c65c411ce26e 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > @@ -459,7 +459,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo, > > r = amdgpu_fill_buffer(adev, entity, > abo, 0, NULL, &wipe_fence, > - AMDGPU_KERNEL_JOB_ID_MOVE_BLIT); > + false, AMDGPU_KERNEL_JOB_ID_MOVE_BLIT); > if (r) { > goto error; > } else if (wipe_fence) { > @@ -2459,79 +2459,28 @@ static int amdgpu_ttm_fill_mem(struct amdgpu_device *adev, > } > > /** > - * amdgpu_ttm_clear_buffer - clear memory buffers > + * amdgpu_fill_buffer - fill a buffer with a given value > * @adev: amdgpu device object > - * @bo: amdgpu buffer object > - * @resv: reservation object > - * @fence: dma_fence associated with the operation > + * @entity: optional entity to use. If NULL, the clearing entities will be > + * used to load-balance the partial clears > + * @bo: the bo to fill > + * @src_data: the value to set > + * @resv: fences contained in this reservation will be used as dependencies. > + * @out_fence: the fence from the last clear will be stored here. It might be > + * NULL if no job was run. > + * @dependency: optional input dependency fence. > + * @consider_clear_status: true if region reported as cleared by amdgpu_res_cleared() > + * are skipped. > + * @k_job_id: trace id > * > - * Clear the memory buffer resource. > - * > - * Returns: > - * 0 for success or a negative error code on failure. > */ > -int amdgpu_ttm_clear_buffer(struct amdgpu_device *adev, > - struct amdgpu_bo *bo, > - struct dma_resv *resv, > - struct dma_fence **fence) > -{ > - struct amdgpu_ttm_buffer_entity *entity; > - struct amdgpu_res_cursor cursor; > - u64 addr; > - int r = 0; > - > - if (!adev->mman.buffer_funcs_enabled) > - return -EINVAL; > - > - if (!fence) > - return -EINVAL; > - entity = &adev->mman.clear_entities[0]; > - *fence = dma_fence_get_stub(); > - > - amdgpu_res_first(bo->tbo.resource, 0, amdgpu_bo_size(bo), &cursor); > - > - mutex_lock(&entity->lock); > - while (cursor.remaining) { > - struct dma_fence *next = NULL; > - u64 size; > - > - if (amdgpu_res_cleared(&cursor)) { > - amdgpu_res_next(&cursor, cursor.size); > - continue; > - } > - > - /* Never clear more than 256MiB at once to avoid timeouts */ > - size = min(cursor.size, 256ULL << 20); > - > - r = amdgpu_ttm_map_buffer(adev, entity, > - &bo->tbo, bo->tbo.resource, &cursor, > - 1, false, false, &size, &addr); > - if (r) > - goto err; > - > - r = amdgpu_ttm_fill_mem(adev, entity, 0, addr, size, resv, > - &next, true, > - AMDGPU_KERNEL_JOB_ID_TTM_CLEAR_BUFFER); > - if (r) > - goto err; > - > - dma_fence_put(*fence); > - *fence = next; > - > - amdgpu_res_next(&cursor, size); > - } > -err: > - mutex_unlock(&entity->lock); > - > - return r; > -} > - > int amdgpu_fill_buffer(struct amdgpu_device *adev, > struct amdgpu_ttm_buffer_entity *entity, > struct amdgpu_bo *bo, > uint32_t src_data, > struct dma_resv *resv, > - struct dma_fence **f, > + struct dma_fence **out_fence, > + bool consider_clear_status, > u64 k_job_id) > { > struct dma_fence *fence = NULL; > @@ -2551,6 +2500,11 @@ int amdgpu_fill_buffer(struct amdgpu_device *adev, > struct dma_fence *next; > uint64_t cur_size, to; > > + if (consider_clear_status && amdgpu_res_cleared(&dst)) { > + amdgpu_res_next(&dst, dst.size); > + continue; > + } > + > /* Never fill more than 256MiB at once to avoid timeouts */ > cur_size = min(dst.size, 256ULL << 20); > > @@ -2574,9 +2528,7 @@ int amdgpu_fill_buffer(struct amdgpu_device *adev, > } > error: > mutex_unlock(&entity->lock); > - if (f) > - *f = dma_fence_get(fence); > - dma_fence_put(fence); > + *out_fence = fence; > return r; > } > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > index 653a4d17543e..f3bdbcec9afc 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > @@ -181,16 +181,13 @@ int amdgpu_copy_buffer(struct amdgpu_device *adev, > struct dma_resv *resv, > struct dma_fence **fence, > bool vm_needs_flush, uint32_t copy_flags); > -int amdgpu_ttm_clear_buffer(struct amdgpu_device *adev, > - struct amdgpu_bo *bo, > - struct dma_resv *resv, > - struct dma_fence **fence); > int amdgpu_fill_buffer(struct amdgpu_device *adev, > struct amdgpu_ttm_buffer_entity *entity, > struct amdgpu_bo *bo, > uint32_t src_data, > struct dma_resv *resv, > - struct dma_fence **f, > + struct dma_fence **out_fence, > + bool consider_clear_status, > u64 k_job_id); > struct amdgpu_ttm_buffer_entity *amdgpu_ttm_next_clear_entity(struct amdgpu_device *adev); >

2 months, 3 weeks

Re: [PATCH 6/9] iommufd: Have pfn_reader process DMABUF iopt_pages

by Jason Gunthorpe

On Thu, Nov 20, 2025 at 08:04:37AM +0000, Tian, Kevin wrote: > > From: Jason Gunthorpe <jgg(a)nvidia.com> > > Sent: Saturday, November 8, 2025 12:50 AM > > + > > +static int pfn_reader_fill_dmabuf(struct pfn_reader_dmabuf *dmabuf, > > + struct pfn_batch *batch, > > + unsigned long start_index, > > + unsigned long last_index) > > +{ > > + unsigned long start = dmabuf->start_offset + start_index * PAGE_SIZE; > > + > > + /* > > + * This works in PAGE_SIZE indexes, if the dmabuf is sliced and > > + * starts/ends at a sub page offset then the batch to domain code will > > + * adjust it. > > + */ > > dmabuf->start_offset comes from pages->dmabuf.start, which is initialized as: > > pages->dmabuf.start = start - start_byte; > > so it's always page-aligned. Where is the sub-page offset coming from? I need to go over this again to check it, this sub-page stuff is a bit convoluted. start_offset should include the sub page offset here.. > > @@ -1687,6 +1737,12 @@ static void __iopt_area_unfill_domain(struct > > iopt_area *area, > > > > lockdep_assert_held(&pages->mutex); > > > > + if (iopt_is_dmabuf(pages)) { > > + iopt_area_unmap_domain_range(area, domain, start_index, > > + last_index); > > + return; > > + } > > + > > this belongs to patch3? This is part of programming the domain with the dmabuf, the patch3 was about the revoke which is a slightly different topic though they are both similar. Thanks, Jason

2 months, 3 weeks

Re: [PATCH 2/9] iommufd: Add DMABUF to iopt_pages

by Jason Gunthorpe

On Thu, Nov 20, 2025 at 07:55:04AM +0000, Tian, Kevin wrote: > > From: Jason Gunthorpe <jgg(a)nvidia.com> > > Sent: Saturday, November 8, 2025 12:50 AM > > > > > > @@ -2031,7 +2155,10 @@ int iopt_pages_rw_access(struct iopt_pages > > *pages, unsigned long start_byte, > > if ((flags & IOMMUFD_ACCESS_RW_WRITE) && !pages->writable) > > return -EPERM; > > > > - if (pages->type == IOPT_ADDRESS_FILE) > > + if (iopt_is_dmabuf(pages)) > > + return -EINVAL; > > + > > probably also add helpers for other types, e.g.: > > iopt_is_user() > iopt_is_memfd() The helper was to integrate the IS_ENABLED() check for DMABUF, there are not so many others uses, I think leave it to not bloat the patch. > > + if (pages->type != IOPT_ADDRESS_USER) > > return iopt_pages_rw_slow(pages, start_index, last_index, > > start_byte % PAGE_SIZE, data, > > length, > > flags); > > -- > > then the following WARN_ON() becomes useless: > > if (IS_ENABLED(CONFIG_IOMMUFD_TEST) && > WARN_ON(pages->type != IOPT_ADDRESS_USER)) > return -EINVAL; Yep Thanks, Jason

2 months, 3 weeks

Re: [PATCH v3 09/28] drm/amdgpu: pass the entity to use to ttm public functions

by Christian König

On 11/21/25 11:12, Pierre-Eric Pelloux-Prayer wrote: > This way the caller can select the one it wants to use. > > Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer(a)amd.com> I'm wondering if it wouldn't make sense to put a pointer to adev into each amdgpu_ttm_buffer_entity. But that is maybe something for another patch. For now: Reviewed-by: Christian König <christian.koenig(a)amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c | 3 +- > drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 4 +-- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 34 +++++++++---------- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 16 +++++---- > drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 3 +- > 5 files changed, 32 insertions(+), 28 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c > index 3636b757c974..a050167e76a4 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c > @@ -37,7 +37,8 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size, > > stime = ktime_get(); > for (i = 0; i < n; i++) { > - r = amdgpu_copy_buffer(adev, saddr, daddr, size, NULL, &fence, > + r = amdgpu_copy_buffer(adev, &adev->mman.default_entity, > + saddr, daddr, size, NULL, &fence, > false, 0); > if (r) > goto exit_do_move; > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > index 926a3f09a776..858eb9fa061b 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c > @@ -1322,8 +1322,8 @@ void amdgpu_bo_release_notify(struct ttm_buffer_object *bo) > if (r) > goto out; > > - r = amdgpu_fill_buffer(abo, 0, &bo->base._resv, &fence, true, > - AMDGPU_KERNEL_JOB_ID_CLEAR_ON_RELEASE); > + r = amdgpu_fill_buffer(&adev->mman.clear_entity, abo, 0, &bo->base._resv, > + &fence, AMDGPU_KERNEL_JOB_ID_CLEAR_ON_RELEASE); > if (WARN_ON(r)) > goto out; > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > index 3d850893b97f..1d3afad885da 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > @@ -359,7 +359,7 @@ static int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev, > write_compress_disable)); > } > > - r = amdgpu_copy_buffer(adev, from, to, cur_size, resv, > + r = amdgpu_copy_buffer(adev, entity, from, to, cur_size, resv, > &next, true, copy_flags); > if (r) > goto error; > @@ -414,8 +414,9 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo, > (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) { > struct dma_fence *wipe_fence = NULL; > > - r = amdgpu_fill_buffer(abo, 0, NULL, &wipe_fence, > - false, AMDGPU_KERNEL_JOB_ID_MOVE_BLIT); > + r = amdgpu_fill_buffer(&adev->mman.move_entity, > + abo, 0, NULL, &wipe_fence, > + AMDGPU_KERNEL_JOB_ID_MOVE_BLIT); > if (r) { > goto error; > } else if (wipe_fence) { > @@ -2258,7 +2259,9 @@ static int amdgpu_ttm_prepare_job(struct amdgpu_device *adev, > DMA_RESV_USAGE_BOOKKEEP); > } > > -int amdgpu_copy_buffer(struct amdgpu_device *adev, uint64_t src_offset, > +int amdgpu_copy_buffer(struct amdgpu_device *adev, > + struct amdgpu_ttm_buffer_entity *entity, > + uint64_t src_offset, > uint64_t dst_offset, uint32_t byte_count, > struct dma_resv *resv, > struct dma_fence **fence, > @@ -2282,7 +2285,7 @@ int amdgpu_copy_buffer(struct amdgpu_device *adev, uint64_t src_offset, > max_bytes = adev->mman.buffer_funcs->copy_max_bytes; > num_loops = DIV_ROUND_UP(byte_count, max_bytes); > num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8); > - r = amdgpu_ttm_prepare_job(adev, &adev->mman.move_entity, num_dw, > + r = amdgpu_ttm_prepare_job(adev, entity, num_dw, > resv, vm_needs_flush, &job, > AMDGPU_KERNEL_JOB_ID_TTM_COPY_BUFFER); > if (r) > @@ -2411,22 +2414,18 @@ int amdgpu_ttm_clear_buffer(struct amdgpu_bo *bo, > return r; > } > > -int amdgpu_fill_buffer(struct amdgpu_bo *bo, > - uint32_t src_data, > - struct dma_resv *resv, > - struct dma_fence **f, > - bool delayed, > - u64 k_job_id) > +int amdgpu_fill_buffer(struct amdgpu_ttm_buffer_entity *entity, > + struct amdgpu_bo *bo, > + uint32_t src_data, > + struct dma_resv *resv, > + struct dma_fence **f, > + u64 k_job_id) > { > struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); > - struct amdgpu_ttm_buffer_entity *entity; > struct dma_fence *fence = NULL; > struct amdgpu_res_cursor dst; > int r; > > - entity = delayed ? &adev->mman.clear_entity : > - &adev->mman.move_entity; > - > if (!adev->mman.buffer_funcs_enabled) { > dev_err(adev->dev, > "Trying to clear memory with ring turned off.\n"); > @@ -2443,13 +2442,14 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo, > /* Never fill more than 256MiB at once to avoid timeouts */ > cur_size = min(dst.size, 256ULL << 20); > > - r = amdgpu_ttm_map_buffer(adev, &adev->mman.default_entity, > + r = amdgpu_ttm_map_buffer(adev, entity, > &bo->tbo, bo->tbo.resource, &dst, > 1, false, &cur_size, &to); > if (r) > goto error; > > - r = amdgpu_ttm_fill_mem(adev, entity, src_data, to, cur_size, resv, > + r = amdgpu_ttm_fill_mem(adev, entity, > + src_data, to, cur_size, resv, > &next, true, k_job_id); > if (r) > goto error; > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > index 41bbc25680a2..9288599c9c46 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h > @@ -167,7 +167,9 @@ int amdgpu_ttm_init(struct amdgpu_device *adev); > void amdgpu_ttm_fini(struct amdgpu_device *adev); > void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, > bool enable); > -int amdgpu_copy_buffer(struct amdgpu_device *adev, uint64_t src_offset, > +int amdgpu_copy_buffer(struct amdgpu_device *adev, > + struct amdgpu_ttm_buffer_entity *entity, > + uint64_t src_offset, > uint64_t dst_offset, uint32_t byte_count, > struct dma_resv *resv, > struct dma_fence **fence, > @@ -175,12 +177,12 @@ int amdgpu_copy_buffer(struct amdgpu_device *adev, uint64_t src_offset, > int amdgpu_ttm_clear_buffer(struct amdgpu_bo *bo, > struct dma_resv *resv, > struct dma_fence **fence); > -int amdgpu_fill_buffer(struct amdgpu_bo *bo, > - uint32_t src_data, > - struct dma_resv *resv, > - struct dma_fence **fence, > - bool delayed, > - u64 k_job_id); > +int amdgpu_fill_buffer(struct amdgpu_ttm_buffer_entity *entity, > + struct amdgpu_bo *bo, > + uint32_t src_data, > + struct dma_resv *resv, > + struct dma_fence **f, > + u64 k_job_id); > > int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo); > void amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo); > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > index ade1d4068d29..9c76f1ba0e55 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > @@ -157,7 +157,8 @@ svm_migrate_copy_memory_gart(struct amdgpu_device *adev, dma_addr_t *sys, > goto out_unlock; > } > > - r = amdgpu_copy_buffer(adev, gart_s, gart_d, size * PAGE_SIZE, > + r = amdgpu_copy_buffer(adev, entity, > + gart_s, gart_d, size * PAGE_SIZE, > NULL, &next, true, 0); > if (r) { > dev_err(adev->dev, "fail %d to copy memory\n", r);

2 months, 3 weeks

Re: [PATCH v3 06/28] drm/amdgpu: add amdgpu_ttm_job_submit helper

by Christian König

On 11/21/25 11:12, Pierre-Eric Pelloux-Prayer wrote: > Deduplicate the IB padding code and will also be used > later to check locking. > > Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer(a)amd.com> Reviewed-by: Christian König <christian.koenig(a)amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 34 ++++++++++++------------- > 1 file changed, 16 insertions(+), 18 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > index 17e1892c44a2..be1232b2d55e 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c > @@ -162,6 +162,18 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo, > *placement = abo->placement; > } > > +static struct dma_fence * > +amdgpu_ttm_job_submit(struct amdgpu_device *adev, struct amdgpu_job *job, u32 num_dw) > +{ > + struct amdgpu_ring *ring; > + > + ring = adev->mman.buffer_funcs_ring; > + amdgpu_ring_pad_ib(ring, &job->ibs[0]); > + WARN_ON(job->ibs[0].length_dw > num_dw); > + > + return amdgpu_job_submit(job); > +} > + > /** > * amdgpu_ttm_map_buffer - Map memory into the GART windows > * @adev: the device being used > @@ -185,7 +197,6 @@ static int amdgpu_ttm_map_buffer(struct amdgpu_device *adev, > { > unsigned int offset, num_pages, num_dw, num_bytes; > uint64_t src_addr, dst_addr; > - struct amdgpu_ring *ring; > struct amdgpu_job *job; > void *cpu_addr; > uint64_t flags; > @@ -240,10 +251,6 @@ static int amdgpu_ttm_map_buffer(struct amdgpu_device *adev, > amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, > dst_addr, num_bytes, 0); > > - ring = adev->mman.buffer_funcs_ring; > - amdgpu_ring_pad_ib(ring, &job->ibs[0]); > - WARN_ON(job->ibs[0].length_dw > num_dw); > - > flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem); > if (tmz) > flags |= AMDGPU_PTE_TMZ; > @@ -261,7 +268,7 @@ static int amdgpu_ttm_map_buffer(struct amdgpu_device *adev, > amdgpu_gart_map_vram_range(adev, pa, 0, num_pages, flags, cpu_addr); > } > > - dma_fence_put(amdgpu_job_submit(job)); > + dma_fence_put(amdgpu_ttm_job_submit(adev, job, num_dw)); > return 0; > } > > @@ -1497,10 +1504,7 @@ static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo, > amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr, > PAGE_SIZE, 0); > > - amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]); > - WARN_ON(job->ibs[0].length_dw > num_dw); > - > - fence = amdgpu_job_submit(job); > + fence = amdgpu_ttm_job_submit(adev, job, num_dw); > > if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout)) > r = -ETIMEDOUT; > @@ -2285,11 +2289,9 @@ int amdgpu_copy_buffer(struct amdgpu_device *adev, uint64_t src_offset, > byte_count -= cur_size_in_bytes; > } > > - amdgpu_ring_pad_ib(ring, &job->ibs[0]); > - WARN_ON(job->ibs[0].length_dw > num_dw); > - *fence = amdgpu_job_submit(job); > if (r) > goto error_free; > + *fence = amdgpu_ttm_job_submit(adev, job, num_dw); > > return r; > > @@ -2307,7 +2309,6 @@ static int amdgpu_ttm_fill_mem(struct amdgpu_device *adev, uint32_t src_data, > u64 k_job_id) > { > unsigned int num_loops, num_dw; > - struct amdgpu_ring *ring; > struct amdgpu_job *job; > uint32_t max_bytes; > unsigned int i; > @@ -2331,10 +2332,7 @@ static int amdgpu_ttm_fill_mem(struct amdgpu_device *adev, uint32_t src_data, > byte_count -= cur_size; > } > > - ring = adev->mman.buffer_funcs_ring; > - amdgpu_ring_pad_ib(ring, &job->ibs[0]); > - WARN_ON(job->ibs[0].length_dw > num_dw); > - *fence = amdgpu_job_submit(job); > + *fence = amdgpu_ttm_job_submit(adev, job, num_dw); > return 0; > } >

2 months, 3 weeks

Re: [PATCH v9 10/11] vfio/pci: Add dma-buf export support for MMIO regions

by Leon Romanovsky

On Thu, Nov 20, 2025 at 05:04:13PM -0700, Alex Williamson wrote: > On Thu, 20 Nov 2025 11:28:29 +0200 > Leon Romanovsky <leon(a)kernel.org> wrote: > > diff --git a/drivers/vfio/pci/vfio_pci_core.c b/drivers/vfio/pci/vfio_pci_core.c > > index 142b84b3f225..51a3bcc26f8b 100644 > > --- a/drivers/vfio/pci/vfio_pci_core.c > > +++ b/drivers/vfio/pci/vfio_pci_core.c > ... > > @@ -2487,8 +2500,11 @@ static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set, > > > > err_undo: > > list_for_each_entry_from_reverse(vdev, &dev_set->device_list, > > - vdev.dev_set_list) > > + vdev.dev_set_list) { > > + if (__vfio_pci_memory_enabled(vdev)) > > + vfio_pci_dma_buf_move(vdev, false); > > up_write(&vdev->memory_lock); > > + } > > I ran into a bug here. In the hot reset path we can have dev_sets > where one or more devices are not opened by the user. The vconfig > buffer for the device is established on open. However: > > bool __vfio_pci_memory_enabled(struct vfio_pci_core_device *vdev) > { > struct pci_dev *pdev = vdev->pdev; > u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]); > ... > > Leads to a NULL pointer dereference. > > I think the most straightforward fix is simply to test the open_count > on the vfio_device, which is also protected by the dev_set->lock that > we already hold here: > > --- a/drivers/vfio/pci/vfio_pci_core.c > +++ b/drivers/vfio/pci/vfio_pci_core.c > @@ -2501,7 +2501,7 @@ static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set, > err_undo: > list_for_each_entry_from_reverse(vdev, &dev_set->device_list, > vdev.dev_set_list) { > - if (__vfio_pci_memory_enabled(vdev)) > + if (vdev->vdev.open_count && __vfio_pci_memory_enabled(vdev)) > vfio_pci_dma_buf_move(vdev, false); > up_write(&vdev->memory_lock); > } > > Any other suggestions? This should be the only reset path with this > nuance of affecting non-opened devices. Thanks, It seems right to me. Thanks > > Alex

2 months, 3 weeks

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

by Sumit Semwal

Hi Barry, On Fri, 21 Nov 2025 at 06:54, Barry Song <21cnbao(a)gmail.com> wrote: > > Hi Sumit, > > > > > Using the micro-benchmark below, we see that mmap becomes > > 3.5X faster: > > > Marcin pointed out to me off-tree that it is actually 35x faster, > not 3.5x faster. Sorry for my poor math. I assume you can fix this > when merging it? Sure, I corrected this, and is merged to drm-misc-next Thanks, Sumit. > > > > > W/ patch: > > > > ~ # ./a.out > > mmap 512MB took 200266.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 198151.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 197069.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 196781.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 198102.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 195552.000 us, verify OK > > > > W/o patch: > > > > ~ # ./a.out > > mmap 512MB took 6987470.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 6970739.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 6984383.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 6971311.000 us, verify OK > > ~ # ./a.out > > mmap 512MB took 6991680.000 us, verify OK > > > Thanks > Barry

2 months, 3 weeks

Re: [PATCH v9 10/11] vfio/pci: Add dma-buf export support for MMIO regions

by Jason Gunthorpe

On Thu, Nov 20, 2025 at 05:04:13PM -0700, Alex Williamson wrote: > @@ -2501,7 +2501,7 @@ static int vfio_pci_dev_set_hot_reset(struct vfio_device_set *dev_set, > err_undo: > list_for_each_entry_from_reverse(vdev, &dev_set->device_list, > vdev.dev_set_list) { > - if (__vfio_pci_memory_enabled(vdev)) > + if (vdev->vdev.open_count && __vfio_pci_memory_enabled(vdev)) > vfio_pci_dma_buf_move(vdev, false); > up_write(&vdev->memory_lock); > } > > Any other suggestions? This should be the only reset path with this > nuance of affecting non-opened devices. Thanks, Seems reasonable, but should it be in __vfio_pci_memory_enabled() just to be robust? Jason

2 months, 3 weeks

Jump to page:

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

Linaro-mm-sig