KFD user queue validation
Subsystems (4)
v2 cover KFD user queue validation
This patch series do additional queue buffers validation in the queue creation IOCTLS, fail the queue creation if buffers not mapped on the GPU with the expected size.
Ensure queue buffers residency by tracking the GPUVM virtual addresses for queue buffers to return error if the user tries to free and unmap them when the qeueu is active, or evict the queue if SVM memory is unmapped and freed from CPU.
Patch 1-2 is prepration work and general fix.
v2:
- patch 3/9, keep wptr_bo_gart in struct queue
Philip Yang (9):
drm/amdkfd: kfd_bo_mapped_dev support partition drm/amdkfd: amdkfd_free_gtt_mem clear the correct pointer drm/amdkfd: Refactor queue wptr_bo GART mapping drm/amdkfd: Validate user queue buffers drm/amdkfd: Ensure user queue buffers residency drm/amdkfd: Validate user queue svm memory residency drm/amdkfd: Validate user queue update drm/amdkfd: Store queue cwsr area size to node properties drm/amdkfd: Validate queue cwsr area and eop buffer size
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 6 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 6 + drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 61 +--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 19 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- .../amd/amdkfd/kfd_process_queue_manager.c | 79 +++- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 336 ++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 12 + drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 4 + 16 files changed, 489 insertions(+), 91 deletions(-)
Signature · 1 line
2.43.2
PATCH v2 1/9 drm/amdkfd: kfd_bo_mapped_dev support partition
Change amdgpu_amdkfd_bo_mapped_to_dev to use drm_priv as parameter instead of adev, to support spatial partition. This is only used by CRIU checkpoint restore now. No functional change.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 5 +++-- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- 3 files changed, 5 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -345,7 +345,7 @@ void amdgpu_amdkfd_ras_pasid_poison_consumption_handler(struct amdgpu_device *ad pasid_notify pasid_fn, void *data, uint32_t reset);bool amdgpu_amdkfd_is_fed (struct amdgpu_device *adev); -bool amdgpu_amdkfd_bo_mapped_to_dev (struct amdgpu_device *adev,struct kgd_mem *mem); +bool amdgpu_amdkfd_bo_mapped_to_dev (void *drm_priv,struct kgd_mem *mem);void amdgpu_amdkfd_block_mmu_notifications (void *p);int amdgpu_amdkfd_criu_resume (void *p); bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev,
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -2
@@ -3200,12 +3200,13 @@ int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,return 0 ; } -bool amdgpu_amdkfd_bo_mapped_to_dev (struct amdgpu_device *adev,struct kgd_mem *mem) +bool amdgpu_amdkfd_bo_mapped_to_dev (void *drm_priv,struct kgd_mem *mem) { +struct amdgpu_vm *vm =drm_priv_to_vm (drm_priv);struct kfd_mem_attachment *entry;list_for_each_entry (entry, &mem->attachments, list) { -if (entry->is_mapped && entry->adev == adev) +if (entry->is_mapped && entry->bo_va->base.vm == vm)return true; }return false;
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
@@ -1963,7 +1963,7 @@ static int criu_checkpoint_bos(struct kfd_process *p, bo_bucket->offset =amdgpu_bo_mmap_offset (dumper_bo);for (i =0 ; i < p->n_pdds; i++) { -if (amdgpu_amdkfd_bo_mapped_to_dev (p->pdds[i]->dev->adev, kgd_mem)) +if (amdgpu_amdkfd_bo_mapped_to_dev (p->pdds[i]->drm_priv, kgd_mem)) bo_priv->mapped_gpuids[dev_idx++] = p->pdds[i]->user_gpu_id; }
Signature · 1 line
2.43.2
PATCH v2 2/9 drm/amdkfd: amdkfd_free_gtt_mem clear the correct pointer
In reply to Philip YangPass pointer reference to amdgpu_bo_unref to clear the correct pointer, otherwise amdgpu_bo_unref clear the local variable, the original pointer not set to NULL, this could cause use-after-free bug.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 14 +++++++------- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 ++-- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c | 4 ++-- 8 files changed, 16 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +7 -7
@@ -364,15 +364,15 @@ int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,return r; } -void amdgpu_amdkfd_free_gtt_mem (struct amdgpu_device *adev,void *mem_obj) +void amdgpu_amdkfd_free_gtt_mem (struct amdgpu_device *adev,void **mem_obj) { -struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj; +struct amdgpu_bo **bo = (struct amdgpu_bo **) mem_obj; -amdgpu_bo_reserve (bo, true); -amdgpu_bo_kunmap (bo); -amdgpu_bo_unpin (bo); -amdgpu_bo_unreserve (bo); -amdgpu_bo_unref (&(bo)); +amdgpu_bo_reserve (*bo, true); +amdgpu_bo_kunmap (*bo); +amdgpu_bo_unpin (*bo); +amdgpu_bo_unreserve (*bo); +amdgpu_bo_unref (bo); } int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -235,7 +235,7 @@ int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,int amdgpu_amdkfd_alloc_gtt_mem (struct amdgpu_device *adev,size_t size,void **mem_obj,uint64_t *gpu_addr,void **cpu_ptr,bool mqd_gfx9); -void amdgpu_amdkfd_free_gtt_mem (struct amdgpu_device *adev,void *mem_obj); +void amdgpu_amdkfd_free_gtt_mem (struct amdgpu_device *adev,void **mem_obj);int amdgpu_amdkfd_alloc_gws (struct amdgpu_device *adev,size_t size,void **mem_obj);void amdgpu_amdkfd_free_gws (struct amdgpu_device *adev,void *mem_obj);
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -1
@@ -423,7 +423,7 @@ static int kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, err_create_queue:if (wptr_bo) -amdgpu_amdkfd_free_gtt_mem (dev->adev, wptr_bo); +amdgpu_amdkfd_free_gtt_mem (dev->adev, (void **)&wptr_bo); err_wptr_map_gart: err_bind_process: err_pdd:
drivers/gpu/drm/amd/amdkfd/kfd_device.c +2 -2
@@ -907,7 +907,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, kfd_doorbell_error:kfd_gtt_sa_fini (kfd); kfd_gtt_sa_init_error: -amdgpu_amdkfd_free_gtt_mem (kfd->adev, kfd->gtt_mem); +amdgpu_amdkfd_free_gtt_mem (kfd->adev, &kfd->gtt_mem); alloc_gtt_mem_failure: dev_err(kfd_device, "device %x:%x NOT added due to errors\n",
@@ -925,7 +925,7 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd)kfd_doorbell_fini (kfd);ida_destroy (&kfd->doorbell_ida);kfd_gtt_sa_fini (kfd); -amdgpu_amdkfd_free_gtt_mem (kfd->adev, kfd->gtt_mem); +amdgpu_amdkfd_free_gtt_mem (kfd->adev, &kfd->gtt_mem); }kfree (kfd);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
@@ -2621,7 +2621,7 @@ static void deallocate_hiq_sdma_mqd(struct kfd_node *dev, {WARN (!mqd,"No hiq sdma mqd trunk to free" ); -amdgpu_amdkfd_free_gtt_mem (dev->adev, mqd->gtt_mem); +amdgpu_amdkfd_free_gtt_mem (dev->adev, &mqd->gtt_mem); }void device_queue_manager_uninit (struct device_queue_manager *dqm)
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c +1 -1
@@ -225,7 +225,7 @@ void kfd_free_mqd_cp(struct mqd_manager *mm, void *mqd, struct kfd_mem_obj *mqd_mem_obj) {if (mqd_mem_obj->gtt_mem) { -amdgpu_amdkfd_free_gtt_mem (mm->dev->adev, mqd_mem_obj->gtt_mem); +amdgpu_amdkfd_free_gtt_mem (mm->dev->adev, &mqd_mem_obj->gtt_mem);kfree (mqd_mem_obj); }else {kfd_gtt_sa_free (mm->dev, mqd_mem_obj);
drivers/gpu/drm/amd/amdkfd/kfd_process.c +1 -1
@@ -1048,7 +1048,7 @@ static void kfd_process_destroy_pdds(struct kfd_process *p)if (pdd->dev->kfd->shared_resources.enable_mes)amdgpu_amdkfd_free_gtt_mem (pdd->dev->adev, - pdd->proc_ctx_bo); + &pdd->proc_ctx_bo);/* * before destroying pdd, make sure to report availability * for auto suspend
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -2
@@ -204,9 +204,9 @@ static void pqm_clean_queue_resource(struct process_queue_manager *pqm, }if (dev->kfd->shared_resources.enable_mes) { -amdgpu_amdkfd_free_gtt_mem (dev->adev, pqn->q->gang_ctx_bo); +amdgpu_amdkfd_free_gtt_mem (dev->adev, &pqn->q->gang_ctx_bo);if (pqn->q->wptr_bo) -amdgpu_amdkfd_free_gtt_mem (dev->adev, pqn->q->wptr_bo); +amdgpu_amdkfd_free_gtt_mem (dev->adev, (void **)&pqn->q->wptr_bo); } }
Signature · 1 line
2.43.2
PATCH v2 3/9 drm/amdkfd: Refactor queue wptr_bo GART mapping
In reply to Philip YangAdd helper function kfd_queue_acquire_buffers to get queue wptr_bo reference from queue write_ptr if it is mapped to the KFD node with expected size.
Add wptr_bo to structure queue_properties because structure queue is allocated after queue buffers are validated, then we can remove wptr_bo parameter from pqm_create_queue.
Rename structure queue wptr_bo_gart to hold wptr_bo reference for GART mapping and umapping. Move MES wptr_bo_gart mapping to init_user_queue, the same location with queue ctx_bo GART mapping.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 56 +++--------------- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 6 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 13 +++-- .../amd/amdkfd/kfd_process_queue_manager.c | 45 +++++++++++---- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 57 +++++++++++++++++++ 7 files changed, 116 insertions(+), 68 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -322,7 +322,7 @@ int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem, void **kptr, uint64_t *size);void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel (struct kgd_mem *mem); -int amdgpu_amdkfd_map_gtt_bo_to_gart (struct amdgpu_bo *bo); +int amdgpu_amdkfd_map_gtt_bo_to_gart (struct amdgpu_bo *bo,struct amdgpu_bo **bo_gart);int amdgpu_amdkfd_gpuvm_restore_process_bos (void *process_info,struct dma_fence __rcu **ef);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -2
@@ -2226,11 +2226,12 @@ int amdgpu_amdkfd_gpuvm_sync_memory(/** * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count * @bo: Buffer object to be mapped +* @bo_gart: Return bo reference * * Before return, bo reference count is incremented. To release the reference and unpin/ * unmap the BO, call amdgpu_amdkfd_free_gtt_mem. */ -int amdgpu_amdkfd_map_gtt_bo_to_gart (struct amdgpu_bo *bo) +int amdgpu_amdkfd_map_gtt_bo_to_gart (struct amdgpu_bo *bo,struct amdgpu_bo **bo_gart) {int ret;
@@ -2257,7 +2258,7 @@ int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo)amdgpu_bo_unreserve (bo); - bo =amdgpu_bo_ref (bo); + *bo_gart =amdgpu_bo_ref (bo);return 0 ;
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +9 -47
86 lines not on this page · raw message
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -4
@@ -208,10 +208,8 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, queue_input.mqd_addr = q->gart_mqd_addr; queue_input.wptr_addr = (uint64_t )q->properties.write_ptr; -if (q->wptr_bo) { - wptr_addr_off = (uint64_t )q->properties.write_ptr & (PAGE_SIZE -1 ); - queue_input.wptr_mc_addr =amdgpu_bo_gpu_offset (q->wptr_bo) + wptr_addr_off; - } + wptr_addr_off = (uint64_t )q->properties.write_ptr & (PAGE_SIZE -1 ); + queue_input.wptr_mc_addr =amdgpu_bo_gpu_offset (q->properties.wptr_bo) + wptr_addr_off; queue_input.is_kfd_process =1 ; queue_input.is_aql_queue = (q->properties.format ==KFD_QUEUE_FORMAT_AQL );
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +9 -4
@@ -494,8 +494,8 @@ struct queue_properties {uint64_t queue_size;uint32_t priority;uint32_t queue_percent; -uint32_t *read_ptr; -uint32_t *write_ptr; +void __user *read_ptr; +void __user *write_ptr;void __iomem *doorbell_ptr;uint32_t doorbell_off;bool is_interop;
@@ -522,6 +522,8 @@ struct queue_properties {uint64_t tba_addr;uint64_t tma_addr;uint64_t exception_status; + +struct amdgpu_bo *wptr_bo; };#define QUEUE_IS_ACTIVE (q) ((q).queue_size > 0 && \
@@ -604,7 +606,7 @@ struct queue {uint64_t gang_ctx_gpu_addr;void *gang_ctx_cpu_ptr; -struct amdgpu_bo *wptr_bo; +struct amdgpu_bo *wptr_bo_gart; };enum KFD_MQD_TYPE {
@@ -1284,6 +1286,10 @@ int init_queue(struct queue **q, const struct queue_properties *properties);void uninit_queue (struct queue *q);void print_queue_properties (struct queue_properties *q);void print_queue (struct queue *q); +int kfd_queue_buffer_get (struct amdgpu_vm *vm,void __user *addr,struct amdgpu_bo **pbo, +u64 expected_size); +int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties); +int kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties);struct mqd_manager *mqd_manager_init_cik (enum KFD_MQD_TYPE type,struct kfd_node *dev);
@@ -1320,7 +1326,6 @@ int pqm_create_queue(struct process_queue_manager *pqm,struct file *f, struct queue_properties *properties, unsigned int *qid, - struct amdgpu_bo *wptr_bo, const struct kfd_criu_queue_priv_data *q_data, const void *restore_mqd, const void *restore_ctl_stack,
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +35 -10
107 lines not on this page · raw message
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +57 -0
60 lines not on this page · raw message
Signature · 1 line
2.43.2
PATCH v2 4/9 drm/amdkfd: Validate user queue buffers
In reply to Philip YangFind user queue rptr, ring buf, eop buffer and cwsr area BOs, and check BOs are mapped on the GPU with correct size and take the BO reference.
--- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 4 +++ drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 38 ++++++++++++++++++++++++-- 2 files changed, 40 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4 -0
@@ -524,6 +524,10 @@ struct queue_properties {uint64_t exception_status;struct amdgpu_bo *wptr_bo; +struct amdgpu_bo *rptr_bo; +struct amdgpu_bo *ring_bo; +struct amdgpu_bo *eop_buf_bo; +struct amdgpu_bo *cwsr_bo; };#define QUEUE_IS_ACTIVE (q) ((q).queue_size > 0 && \
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +36 -2
@@ -97,7 +97,8 @@ int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_if (!mapping) goto out_err; -if (user_addr != mapping->start || user_addr + size -1 != mapping->last) { +if (user_addr != mapping->start || + (size !=0 && user_addr + size -1 != mapping->last)) {pr_debug ("expected size 0x%llx not equal to mapping addr 0x%llx size 0x%llx\n" , expected_size, mapping->start <<AMDGPU_GPU_PAGE_SHIFT , (mapping->last - mapping->start +1 ) <<AMDGPU_GPU_PAGE_SHIFT );
@@ -124,18 +125,51 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope err =kfd_queue_buffer_get (vm, properties->write_ptr, &properties->wptr_bo,PAGE_SIZE );if (err) + goto out_err_unreserve; + + err =kfd_queue_buffer_get (vm, properties->read_ptr, &properties->rptr_bo,PAGE_SIZE ); +if (err) + goto out_err_unreserve; + + err =kfd_queue_buffer_get (vm, (void *)properties->queue_address, + &properties->ring_bo, properties->queue_size); +if (err) + goto out_err_unreserve; + +/* only compute queue requires EOP buffer and CWSR area */ +if (properties->type !=KFD_QUEUE_TYPE_COMPUTE ) goto out_unreserve; +/* EOP buffer is not required for all ASICs */ +if (properties->eop_ring_buffer_address) { + err =kfd_queue_buffer_get (vm, (void *)properties->eop_ring_buffer_address, + &properties->eop_buf_bo, + properties->eop_ring_buffer_size); +if (err) + goto out_err_unreserve; + } + + err =kfd_queue_buffer_get (vm, (void *)properties->ctx_save_restore_area_address, + &properties->cwsr_bo,0 ); +if (err) + goto out_err_unreserve; + +out_unreserve:amdgpu_bo_unreserve (vm->root.bo);return 0 ; -out_unreserve: +out_err_unreserve:amdgpu_bo_unreserve (vm->root.bo); +kfd_queue_release_buffers (pdd, properties);return err; }int kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) {amdgpu_bo_unref (&properties->wptr_bo); +amdgpu_bo_unref (&properties->rptr_bo); +amdgpu_bo_unref (&properties->ring_bo); +amdgpu_bo_unref (&properties->eop_buf_bo); +amdgpu_bo_unref (&properties->cwsr_bo);return 0 ; }
Signature · 1 line
2.43.2
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
#regzbot introduced: 68e599db7a549f010a329515f3508d8a8c3467a4 #regzbot monitor: https://bugs.debian.org/1093124
Hello,
Quoted 6 lines · Philip Yang
Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and check BOs are mapped on the GPU with correct size and take the BO reference.
Signed-off-by Philip Yang <[email protected]>Reviewed-by Felix Kuehling <[email protected]>
This change made it into v6.12-rc1 as 68e599db7a54 ("drm/amdkfd:
Validate user queue buffers"). A Debian user (Dieter Faulbaum, on Cc) reported that this change introduced a regression using a gfx803 device resulting in a HSA exception when e.g. darktable is used. I didn't even try to understand the problem, but maybe one of you have an idea about the issue?!
See more details in the Debian bug at https://bugs.debian.org/1093124
Best regards
Uwe
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
On 2025-02-12 17:42, Uwe Kleine-König wrote:
#regzbot introduced: 68e599db7a549f010a329515f3508d8a8c3467a4 #regzbot monitor: https://bugs.debian.org/1093124 Hello, On Thu, Jul 18, 2024 at 05:05:53PM -0400, Philip Yang wrote: Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and check BOs are mapped on the GPU with correct size and take the BO reference. Signed-off-by: Philip Yang <[email protected]> Reviewed-by: Felix Kuehling <[email protected]> This change made it into v6.12-rc1 as 68e599db7a54 ("drm/amdkfd: Validate user queue buffers"). A Debian user (Dieter Faulbaum, on Cc) reported that this change introduced a regression using a gfx803 device resulting in a HSA exception when e.g. darktable is used. I didn't even try to understand the problem, but maybe one of you have an idea about the issue?! Try this patch
https://lore.kernel.org/all/[email protected]/T/
See more details in the Debian bug at https://bugs.debian.org/1093124 Best regards Uwe
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
Quoted 1 line · Philip Yang
[html-content]
Are you aware that your mail is hard for some people (e.g. those like me who read their mail in a terminal, but also consider people who have to rely on braille readers) to read and also isn't properly archived on lore.kernel.org[1]?
Would be great if you could tune the settings of your thunderbird profile to send text replies. The in-kernel documentation[2] might help here.
Best regards
Uwe
[1] https://lore.kernel.org/all/[email protected]/
[2] https://www.kernel.org/doc/html/v4.11/process/email-clients.html#thunderbird-gui
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
In reply to Philip YangHello Philip,
Quoted 25 lines · Philip Yang
On 2025-02-12 17:42, Uwe Kleine-König wrote:
#regzbot introduced:
68e599db7a549f010a329515f3508d8a8c3467a4
#regzbot monitor: https://bugs.debian.org/1093124Hello,
On Thu, Jul 18, 2024 at 05:05:53PM -0400, Philip Yang wrote:
Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and
check BOs are mapped on the GPU with correct size and take the BO reference.Signed-off-by Philip Yang <[email protected]>Reviewed-by Felix Kuehling <[email protected]>This change made it into v6.12-rc1 as
68e599db7a54("drm/amdkfd:
Validate user queue buffers"). A Debian user (Dieter Faulbaum, on Cc) reported that this change introduced a regression using a gfx803 device resulting in a HSA exception when e.g. darktable is used. I didn't even try to understand the problem, but maybe one of you have an idea about the issue?!
Quoted 3 lines
Try this patch
Thank you, this patch helps!
Tested with kernel version 6.12.0+
And (for me) the Debian bug report could be closed.
Thanks to all for your help!
Dieter
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
In reply to Philip YangHello Philip,
Quoted 25 lines · Philip Yang
On 2025-02-12 17:42, Uwe Kleine-König wrote:
#regzbot introduced:
68e599db7a549f010a329515f3508d8a8c3467a4
#regzbot monitor: https://bugs.debian.org/1093124Hello,
On Thu, Jul 18, 2024 at 05:05:53PM -0400, Philip Yang wrote:
Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and
check BOs are mapped on the GPU with correct size and take the BO reference.Signed-off-by Philip Yang <[email protected]>Reviewed-by Felix Kuehling <[email protected]>This change made it into v6.12-rc1 as
68e599db7a54("drm/amdkfd:
Validate user queue buffers"). A Debian user (Dieter Faulbaum, on Cc) reported that this change introduced a regression using a gfx803 device resulting in a HSA exception when e.g. darktable is used. I didn't even try to understand the problem, but maybe one of you have an idea about the issue?!
Quoted 3 lines
Try this patch
It seems for me, that your patch isn't applied in the mainline kernel. What do you think, will it once happen?-)
Is it falling through cracks?
With regards
Dieter
Re: [PATCH v2 4/9] drm/amdkfd: Validate user queue buffers
Quoted 36 lines · Dieter Faulbaum
Hello Philip,
Quoted 30 lines · Philip Yang
On 2025-02-12 17:42, Uwe Kleine-König wrote:
#regzbot introduced:
68e599db7a549f010a329515f3508d8a8c3467a4
#regzbot monitor: https://bugs.debian.org/1093124Hello,
On Thu, Jul 18, 2024 at 05:05:53PM -0400, Philip Yang wrote:
Find user queue rptr, ring buf, eop buffer and cwsr area BOs, and
check BOs are mapped on the GPU with correct size and take the BO
reference.Signed-off-by Philip Yang <[email protected]>Reviewed-by Felix Kuehling <[email protected]>This change made it into v6.12-rc1 as
68e599db7a54("drm/amdkfd:
Validate user queue buffers"). A Debian user (Dieter Faulbaum, on Cc)
reported that this change introduced a regression using a gfx803 device
resulting in a HSA exception when e.g. darktable is used. I didn't even
try to understand the problem, but maybe one of you have an idea about
the issue?!
Quoted 8 lines
Quoted 3 lines
Try this patch
It seems for me, that your patch isn't applied in the mainline kernel.
What do you think, will it once happen?-)
Is it falling through cracks?
It's in drm-next:
https://gitlab.freedesktop.org/drm/kernel/-/commit/e7a477735f1771b9a9346a5fbd09d7ff0641723a
I'll cherry-pick it for stable next week.
Alex
Quoted 5 lines
With regards
Dieter
PATCH v2 5/9 drm/amdkfd: Ensure user queue buffers residency
In reply to Philip YangAdd atomic queue_refcount to struct bo_va, return -EBUSY to fail unmap BO from the GPU if the bo_va queue_refcount is not zero.
Create queue to increase the bo_va queue_refcount, destroy queue to decrease the bo_va queue_refcount, to ensure the queue buffers mapped on the GPU when queue is active.
--- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 14 ++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 6 ++++ drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 3 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 34 ++++++++++++++++--- 5 files changed, 49 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +12 -2
@@ -1252,7 +1252,7 @@ static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,return ret; } -static void unmap_bo_from_gpuvm (struct kgd_mem *mem, +static int unmap_bo_from_gpuvm (struct kgd_mem *mem,struct kfd_mem_attachment *entry,struct amdgpu_sync *sync) {
@@ -1260,11 +1260,18 @@ static void unmap_bo_from_gpuvm(struct kgd_mem *mem,struct amdgpu_device *adev = entry->adev;struct amdgpu_vm *vm = bo_va->base.vm; +if (bo_va->queue_refcount) { +pr_debug ("bo_va->queue_refcount %d\n" , bo_va->queue_refcount); +return -EBUSY ; + } +amdgpu_vm_bo_unmap (adev, bo_va, entry->va);amdgpu_vm_clear_freed (adev, vm, &bo_va->last_pt_update);amdgpu_sync_fence (sync, bo_va->last_pt_update); + +return 0 ; } static int update_gpuvm_pte(struct kgd_mem *mem,
@@ -2191,7 +2198,10 @@ int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(pr_debug ("\t unmap VA 0x%llx - 0x%llx from entry %p\n" , entry->va, entry->va + bo_size, entry); -unmap_bo_from_gpuvm (mem, entry, ctx.sync); + ret =unmap_bo_from_gpuvm (mem, entry, ctx.sync); +if (ret) + goto unreserve_out; + entry->is_mapped = false; mem->mapped_to_gpu_memory--;
drivers/gpu/drm/amd/amdgpu/amdgpu_object.h +6 -0
@@ -90,6 +90,12 @@ struct amdgpu_bo_va {bool cleared;bool is_xgmi; + +/* +* protected by vm reservation lock +* if non-zero, cannot unmap from GPU because user queues may still access it +*/ +unsigned int queue_refcount; };struct amdgpu_bo {
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1 -2
@@ -1384,8 +1384,7 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, err =amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu ( peer_pdd->dev->adev, (struct kgd_mem *)mem, peer_pdd->drm_priv);if (err) { -pr_err ("Failed to unmap from gpu %d/%d\n" , - i, args->n_devices); +pr_debug ("Failed to unmap from gpu %d/%d\n" , i, args->n_devices); goto unmap_memory_from_gpu_failed; } args->n_success = i+1 ;
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
@@ -1292,6 +1292,7 @@ void print_queue_properties(struct queue_properties *q);void print_queue (struct queue *q);int kfd_queue_buffer_get (struct amdgpu_vm *vm,void __user *addr,struct amdgpu_bo **pbo,u64 expected_size); +void kfd_queue_buffer_put (struct amdgpu_vm *vm,struct amdgpu_bo **bo);int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties);int kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties);
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +29 -5
@@ -106,6 +106,7 @@ int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_ } *pbo =amdgpu_bo_ref (mapping->bo_va->base.bo); + mapping->bo_va->queue_refcount++;return 0 ; out_err:
@@ -113,6 +114,19 @@ int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_return -EINVAL ; } +void kfd_queue_buffer_put (struct amdgpu_vm *vm,struct amdgpu_bo **bo) +{ +if (*bo) { +struct amdgpu_bo_va *bo_va; + + bo_va =amdgpu_vm_bo_find (vm, *bo); +if (bo_va) + bo_va->queue_refcount--; + } + +amdgpu_bo_unref (bo); +} +int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) {struct amdgpu_vm *vm;
@@ -166,10 +180,20 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_propeint kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) { -amdgpu_bo_unref (&properties->wptr_bo); -amdgpu_bo_unref (&properties->rptr_bo); -amdgpu_bo_unref (&properties->ring_bo); -amdgpu_bo_unref (&properties->eop_buf_bo); -amdgpu_bo_unref (&properties->cwsr_bo); +struct amdgpu_vm *vm; +int err; + + vm =drm_priv_to_vm (pdd->drm_priv); + err =amdgpu_bo_reserve (vm->root.bo, false); +if (err) +return err; + +kfd_queue_buffer_put (vm, &properties->wptr_bo); +kfd_queue_buffer_put (vm, &properties->rptr_bo); +kfd_queue_buffer_put (vm, &properties->ring_bo); +kfd_queue_buffer_put (vm, &properties->eop_buf_bo); +kfd_queue_buffer_put (vm, &properties->cwsr_bo); + +amdgpu_bo_unreserve (vm->root.bo);return 0 ; }
Signature · 1 line
2.43.2
PATCH v2 6/9 drm/amdkfd: Validate user queue svm memory residency
In reply to Philip YangQueue CWSR area maybe registered to GPU as svm memory, create queue to ensure svm mapped to GPU with KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED flag.
Add queue_refcount to struct svm_range, to track queue CWSR area usage.
Because unmap mmu notifier callback return value is ignored, if application unmap the CWSR area while queue is active, pr_warn message in dmesg log. To be safe, evict user queue.
--- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 110 ++++++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 12 +++ drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 + 3 files changed, 122 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +109 -1
@@ -24,6 +24,7 @@#include <linux/slab.h> #include "kfd_priv.h" +#include "kfd_svm.h" void print_queue_properties (struct queue_properties *q) {
@@ -83,6 +84,100 @@ void uninit_queue(struct queue *q)kfree (q); } +static int kfd_queue_buffer_svm_get (struct kfd_process_device *pdd,u64 addr,u64 size) +{ +struct kfd_process *p = pdd->process; +struct list_head update_list; +struct svm_range *prange; +int ret = -EINVAL ; + +INIT_LIST_HEAD (&update_list); + addr >>=PAGE_SHIFT ; + size >>=PAGE_SHIFT ; + +mutex_lock (&p->svms.lock); + +/* +* range may split to multiple svm pranges aligned to granularity boundaery. +*/ +while (size) { +uint32_t gpuid, gpuidx; +int r; + + prange =svm_range_from_addr (&p->svms, addr,NULL ); +if (!prange) +break ; + +if (!prange->mapped_to_gpu) +break ; + + r =kfd_process_gpuid_from_node (p, pdd->dev, &gpuid, &gpuidx); +if (r <0 ) +break ; +if (!test_bit (gpuidx, prange->bitmap_access) && + !test_bit (gpuidx, prange->bitmap_aip)) +break ; + +if (!(prange->flags &KFD_IOCTL_SVM_FLAG_GPU_ALWAYS_MAPPED )) +break ; + +list_add (&prange->update_list, &update_list); + +if (prange->last - prange->start +1 >= size) { + size =0 ; +break ; + } + + size -= prange->last - prange->start +1 ; + addr += prange->last - prange->start +1 ; + } +if (size) { +pr_debug ("[0x%llx 0x%llx] not registered\n" , addr, addr + size -1 ); + goto out_unlock; + } + +list_for_each_entry (prange, &update_list, update_list) +atomic_inc (&prange->queue_refcount); + ret =0 ; + +out_unlock: +mutex_unlock (&p->svms.lock); +return ret; +} + +static void kfd_queue_buffer_svm_put (struct kfd_process_device *pdd,u64 addr,u64 size) +{ +struct kfd_process *p = pdd->process; +struct svm_range *prange, *pchild; +struct interval_tree_node *node; +unsigned long last; + + addr >>=PAGE_SHIFT ; + last = addr + (size >>PAGE_SHIFT ) -1 ; + +mutex_lock (&p->svms.lock); + + node =interval_tree_iter_first (&p->svms.objects, addr, last); +while (node) { +struct interval_tree_node *next_node; +unsigned long next_start; + + prange = container_of(node, struct svm_range, it_node); + next_node =interval_tree_iter_next (node, addr, last); + next_start =min (node->last, last) +1 ; + +if (atomic_add_unless (&prange->queue_refcount,-1 ,0 )) { +list_for_each_entry (pchild, &prange->child_list, child_list) +atomic_add_unless (&pchild->queue_refcount,-1 ,0 ); + } + + node = next_node; + addr = next_start; + } + +mutex_unlock (&p->svms.lock); +} +int kfd_queue_buffer_get (struct amdgpu_vm *vm,void __user *addr,struct amdgpu_bo **pbo,u64 expected_size) {
@@ -165,8 +260,17 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope err =kfd_queue_buffer_get (vm, (void *)properties->ctx_save_restore_area_address, &properties->cwsr_bo,0 ); +if (!err) + goto out_unreserve; + +amdgpu_bo_unreserve (vm->root.bo); + + err =kfd_queue_buffer_svm_get (pdd, properties->ctx_save_restore_area_address, + properties->ctx_save_restore_area_size);if (err) - goto out_err_unreserve; + goto out_err_release; + +return 0 ; out_unreserve:amdgpu_bo_unreserve (vm->root.bo);
@@ -174,6 +278,7 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope out_err_unreserve:amdgpu_bo_unreserve (vm->root.bo); +out_err_release:kfd_queue_release_buffers (pdd, properties);return err; }
@@ -195,5 +300,8 @@ int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_propekfd_queue_buffer_put (vm, &properties->cwsr_bo);amdgpu_bo_unreserve (vm->root.bo); + +kfd_queue_buffer_svm_put (pdd, properties->ctx_save_restore_area_address, + properties->ctx_save_restore_area_size);return 0 ; }
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +12 -0
@@ -1051,6 +1051,7 @@ svm_range_split_adjust(struct svm_range *new, struct svm_range *old, new->mapped_to_gpu = old->mapped_to_gpu;bitmap_copy (new->bitmap_access, old->bitmap_access,MAX_GPU_INSTANCE );bitmap_copy (new->bitmap_aip, old->bitmap_aip,MAX_GPU_INSTANCE ); +atomic_set (&new->queue_refcount,atomic_read (&old->queue_refcount));return 0 ; }
@@ -1986,6 +1987,7 @@ static struct svm_range *svm_range_clone(struct svm_range *old) new->vram_pages = old->vram_pages;bitmap_copy (new->bitmap_access, old->bitmap_access,MAX_GPU_INSTANCE );bitmap_copy (new->bitmap_aip, old->bitmap_aip,MAX_GPU_INSTANCE ); +atomic_set (&new->queue_refcount,atomic_read (&old->queue_refcount));return new; }
@@ -2438,6 +2440,16 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,unsigned long s, l;bool unmap_parent; +if (atomic_read (&prange->queue_refcount)) { +int r; + +pr_warn ("Freeing queue vital buffer 0x%lx, queue evicted\n" , + prange->start <<PAGE_SHIFT ); + r =kgd2kfd_quiesce_mm (mm,KFD_QUEUE_EVICTION_TRIGGER_SVM ); +if (r) +pr_debug ("failed %d to quiesce KFD queues\n" , r); + } + p =kfd_lookup_process_by_mm (mm);if (!p)return ;
drivers/gpu/drm/amd/amdkfd/kfd_svm.h +1 -0
@@ -137,6 +137,7 @@ struct svm_range {DECLARE_BITMAP (bitmap_access,MAX_GPU_INSTANCE );DECLARE_BITMAP (bitmap_aip,MAX_GPU_INSTANCE );bool mapped_to_gpu; +atomic_t queue_refcount; };static inline void svm_range_lock (struct svm_range *prange)
Signature · 1 line
2.43.2
PATCH v2 7/9 drm/amdkfd: Validate user queue update
In reply to Philip YangEnsure update queue new ring buffer is mapped on GPU with correct size.
Decrease queue old ring_bo queue_refcount and increase new ring_bo queue_refcount.
--- .../amd/amdkfd/kfd_process_queue_manager.c | 32 ++++++++++++++++++- 1 file changed, 31 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +31 -1
@@ -549,11 +549,41 @@ int pqm_update_queue_properties(struct process_queue_manager *pqm,struct process_queue_node *pqn; pqn =get_queue_by_qid (pqm, qid); -if (!pqn) { +if (!pqn || !pqn->q) {pr_debug ("No queue %d exists for update operation\n" , qid);return -EFAULT ; } +/* +* Update with NULL ring address is used to disable the queue +*/ +if (p->queue_address && p->queue_size) { +struct kfd_process_device *pdd; +struct amdgpu_vm *vm; +struct queue *q = pqn->q; +int err; + + pdd =kfd_get_process_device_data (q->device, q->process); +if (!pdd) +return -ENODEV ; + vm =drm_priv_to_vm (pdd->drm_priv); + err =amdgpu_bo_reserve (vm->root.bo, false); +if (err) +return err; + +if (kfd_queue_buffer_get (vm, (void *)p->queue_address, &p->ring_bo, + p->queue_size)) { +pr_debug ("ring buf 0x%llx size 0x%llx not mapped on GPU\n" , + p->queue_address, p->queue_size); +return -EFAULT ; + } + +kfd_queue_buffer_put (vm, &pqn->q->properties.ring_bo); +amdgpu_bo_unreserve (vm->root.bo); + + pqn->q->properties.ring_bo = p->ring_bo; + } + pqn->q->properties.queue_address = p->queue_address; pqn->q->properties.queue_size = p->queue_size; pqn->q->properties.queue_percent = p->queue_percent;
Signature · 1 line
2.43.2
PATCH v2 8/9 drm/amdkfd: Store queue cwsr area size to node properties
In reply to Philip YangUse the queue eop buffer size, cwsr area size, ctl stack size calculation from Thunk, store the value to KFD node properties.
Those will be used to validate queue eop buffer size, cwsr area size, ctl stack size when creating KFD user compute queue.
Those will be exposed to user space via sysfs KFD node properties, to remove the duplicate calculation code from Thunk.
--- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 75 +++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 4 ++ 4 files changed, 82 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
@@ -1295,6 +1295,7 @@ int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_void kfd_queue_buffer_put (struct amdgpu_vm *vm,struct amdgpu_bo **bo);int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties);int kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties); +void kfd_queue_ctx_save_restore_size (struct kfd_topology_device *dev);struct mqd_manager *mqd_manager_init_cik (enum KFD_MQD_TYPE type,struct kfd_node *dev);
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +75 -0
84 lines not on this page · raw message
drivers/gpu/drm/amd/amdkfd/kfd_topology.c +2 -0
@@ -2120,6 +2120,8 @@ int kfd_topology_add_device(struct kfd_node *gpu) dev->gpu->adev->gmc.xgmi.connected_to_cpu) dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS; +kfd_queue_ctx_save_restore_size (dev); +kfd_debug_print_topology ();kfd_notify_gpu_change (gpu_id,1 );
drivers/gpu/drm/amd/amdkfd/kfd_topology.h +4 -0
@@ -74,6 +74,10 @@ struct kfd_node_properties {uint32_t num_sdma_xgmi_engines;uint32_t num_sdma_queues_per_engine;uint32_t num_cp_queues; +uint32_t cwsr_size; +uint32_t ctl_stack_size; +uint32_t eop_buffer_size; +uint32_t debug_memory_size;char name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE ]; };
Signature · 1 line
2.43.2
PATCH v2 9/9 drm/amdkfd: Validate queue cwsr area and eop buffer size
In reply to Philip YangWhen creating KFD user compute queue, check if queue eop buffer size, cwsr area size, ctl stack size equal to the size of KFD node properities.
Check the entire cwsr area which may split into multiple svm ranges aligned to gramularity boundary.
--- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 46 +++++++++++++++++++++++--- 1 file changed, 42 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_queue.c +42 -4
@@ -225,9 +225,15 @@ void kfd_queue_buffer_put(struct amdgpu_vm *vm, struct amdgpu_bo **bo)int kfd_queue_acquire_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) { +struct kfd_topology_device *topo_dev;struct amdgpu_vm *vm; +u32 total_cwsr_size;int err; + topo_dev =kfd_topology_device_by_id (pdd->dev->id); +if (!topo_dev) +return -EINVAL ; + vm =drm_priv_to_vm (pdd->drm_priv); err =amdgpu_bo_reserve (vm->root.bo, false);if (err)
@@ -252,6 +258,12 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope/* EOP buffer is not required for all ASICs */ if (properties->eop_ring_buffer_address) { +if (properties->eop_ring_buffer_size != topo_dev->node_props.eop_buffer_size) { +pr_debug ("queue eop bo size 0x%lx not equal to node eop buf size 0x%x\n" , + properties->eop_buf_bo->tbo.base.size, + topo_dev->node_props.eop_buffer_size); + goto out_err_unreserve; + } err =kfd_queue_buffer_get (vm, (void *)properties->eop_ring_buffer_address, &properties->eop_buf_bo, properties->eop_ring_buffer_size);
@@ -259,15 +271,33 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope goto out_err_unreserve; } +if (properties->ctl_stack_size != topo_dev->node_props.ctl_stack_size) { +pr_debug ("queue ctl stack size 0x%x not equal to node ctl stack size 0x%x\n" , + properties->ctl_stack_size, + topo_dev->node_props.ctl_stack_size); + goto out_err_unreserve; + } + +if (properties->ctx_save_restore_area_size != topo_dev->node_props.cwsr_size) { +pr_debug ("queue cwsr size 0x%x not equal to node cwsr size 0x%x\n" , + properties->ctx_save_restore_area_size, + topo_dev->node_props.cwsr_size); + goto out_err_unreserve; + } + + total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size) + *NUM_XCC (pdd->dev->xcc_mask); + total_cwsr_size =ALIGN (total_cwsr_size,PAGE_SIZE ); + err =kfd_queue_buffer_get (vm, (void *)properties->ctx_save_restore_area_address, - &properties->cwsr_bo,0 ); + &properties->cwsr_bo, total_cwsr_size);if (!err) goto out_unreserve;amdgpu_bo_unreserve (vm->root.bo); err =kfd_queue_buffer_svm_get (pdd, properties->ctx_save_restore_area_address, - properties->ctx_save_restore_area_size); + total_cwsr_size);if (err) goto out_err_release;
@@ -286,7 +316,9 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_propeint kfd_queue_release_buffers (struct kfd_process_device *pdd,struct queue_properties *properties) { +struct kfd_topology_device *topo_dev;struct amdgpu_vm *vm; +u32 total_cwsr_size;int err; vm =drm_priv_to_vm (pdd->drm_priv);
@@ -302,8 +334,14 @@ int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_propeamdgpu_bo_unreserve (vm->root.bo); -kfd_queue_buffer_svm_put (pdd, properties->ctx_save_restore_area_address, - properties->ctx_save_restore_area_size); + topo_dev =kfd_topology_device_by_id (pdd->dev->id); +if (!topo_dev) +return -EINVAL ; + total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size) + *NUM_XCC (pdd->dev->xcc_mask); + total_cwsr_size =ALIGN (total_cwsr_size,PAGE_SIZE ); + +kfd_queue_buffer_svm_put (pdd, properties->ctx_save_restore_area_address, total_cwsr_size);return 0 ; }
Signature · 1 line
2.43.2
Quoted 9 lines · Philip Yang
This patch series do additional queue buffers validation in the queue creation IOCTLS, fail the queue creation if buffers not mapped on the GPU with the expected size.
Ensure queue buffers residency by tracking the GPUVM virtual addresses for queue buffers to return error if the user tries to free and unmap them when the qeueu is active, or evict the queue if SVM memory is unmapped and freed from CPU.
Quoted 4 lines
Patch 1-2 is prepration work and general fix.
v2:
- patch 3/9, keepwptr_bo_gartin struct queue
The series is
Quoted 30 lines
Philip Yang (9):
drm/amdkfd:kfd_bo_mapped_devsupport partition drm/amdkfd:amdkfd_free_gtt_memclear the correct pointer drm/amdkfd: Refactor queuewptr_boGART mapping drm/amdkfd: Validate user queue buffers drm/amdkfd: Ensure user queue buffers residency drm/amdkfd: Validate user queue svm memory residency drm/amdkfd: Validate user queue update drm/amdkfd: Store queue cwsr area size to node properties drm/amdkfd: Validate queue cwsr area and eop buffer sizedrivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 6 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 6 + drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 61 +--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 19 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- .../amd/amdkfd/kfd_process_queue_manager.c | 79 +++- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 336 ++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 12 + drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 4 + 16 files changed, 489 insertions(+), 91 deletions(-)
Christian König
maintainerAm 18.07.24 um 23:12 schrieb Felix Kuehling:
Quoted 10 lines · Philip Yang
Quoted 9 lines
This patch series do additional queue buffers validation in the queue creation IOCTLS, fail the queue creation if buffers not mapped on the GPU with the expected size.
Ensure queue buffers residency by tracking the GPUVM virtual addresses for queue buffers to return error if the user tries to free and unmap them when the qeueu is active, or evict the queue if SVM memory is unmapped and freed from CPU.
Quoted 6 lines
Quoted 4 lines
Patch 1-2 is prepration work and general fix.
v2:
- patch 3/9, keepwptr_bo_gartin struct queueThe series is
Reviewed-by Felix Kuehling <[email protected]>
I only skimmed over it and will probably find something to complain on later.
But we need to get this out of the door, so feel free to add Acked-by: Christian König <[email protected]> to the series for now.
Thanks,
Christian.
Quoted 30 lines
Quoted 30 lines
Philip Yang (9):
drm/amdkfd:kfd_bo_mapped_devsupport partition drm/amdkfd:amdkfd_free_gtt_memclear the correct pointer drm/amdkfd: Refactor queuewptr_boGART mapping drm/amdkfd: Validate user queue buffers drm/amdkfd: Ensure user queue buffers residency drm/amdkfd: Validate user queue svm memory residency drm/amdkfd: Validate user queue update drm/amdkfd: Store queue cwsr area size to node properties drm/amdkfd: Validate queue cwsr area and eop buffer sizedrivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 14 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 6 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 6 + drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 61 +--- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 8 +- drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 19 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 2 +- .../amd/amdkfd/kfd_process_queue_manager.c | 79 +++- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 336 ++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 12 + drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 + drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_topology.h | 4 + 16 files changed, 489 insertions(+), 91 deletions(-)