From 72d314042864205c86f935e4bd881955ba7b8cd9 Mon Sep 17 00:00:00 2001 From: Arvind Yadav Date: Tue, 13 Aug 2024 19:37:37 +0530 Subject: [PATCH] amdgpu: UAPI for AMDGPU usermode queues MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This patch adds UAPI interface changes for AMDGPU usermode queues, semaphore and new AMDGPU GEM domain for doorbells. Usermode queues allow a userspace process to create and submit its graphics/compute/sdma work directly to the GPU. v2:(Marek) - Add csa support for SDMA queue. - Rename UAPI objects and struct as per UAPI review. (Shashank) v3:(Yogesh) - Rename UAPI timeline* objects as per UAPI review. (Arvind) v4: (Marek) - Drop AMDGPU_USERQ_BO_WRITE as this should not be a global option of the IOCTL, It should be option per buffer. Hence adding separate array for read and write BO handles. (Arun) - Modify num_fences to __u16, flags changed to __u16 and placed the num_fences next to flags for optimal padding and size. (Arun) - Fix 32-bit holes issue in sturct drm_amdgpu_gem_va as per UAPI review (Arvind). v5:(Marek/Pierre-Eric) - add more detail params description for signal and wait IOCTL calls. - Remove the unused structure fields in signal and wait structs. - Add separate array of read and write for BO handles. (Arun) - Removes the unused flags parameter from the usermode queue UAPI structure - Clarify comments on top of drm_amdgpu_userq_in - Clarify comment for queue_id (in) - Clarify comment for mqd - Clarify comment for compute MQD size - Clarify comment for queue_id (out) - Adds padding variables in userqueue in/out structures. (Shashank) v6:(Pierre-Eric) - Modify the function parameter names and struct field names as per the review comments. (Arun) v7:(Marek) - Modify the structure field name and comments. (Arun) - Rename vm_timeline_syncobj and add comment for vm_timeline_point. - Remove GDS buffer support from MQD. (Arvind) v8:(Pierre-Eric) - Modify the function parameter names. Cc: Koenig, Christian Cc: Alex Deucher Reviewed-by: Marek Olšák Acked-by: Pierre-Eric Pelloux-Prayer Signed-off-by: Shashank Sharma Signed-off-by: Arvind Yadav Signed-off-by: Arunpravin Paneer Selvam --- include/drm/amdgpu_drm.h | 254 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 254 insertions(+) diff --git a/include/drm/amdgpu_drm.h b/include/drm/amdgpu_drm.h index fcfcb804..a81eb7f9 100644 --- a/include/drm/amdgpu_drm.h +++ b/include/drm/amdgpu_drm.h @@ -54,6 +54,9 @@ extern "C" { #define DRM_AMDGPU_VM 0x13 #define DRM_AMDGPU_FENCE_TO_HANDLE 0x14 #define DRM_AMDGPU_SCHED 0x15 +#define DRM_AMDGPU_USERQ 0x16 +#define DRM_AMDGPU_USERQ_SIGNAL 0x17 +#define DRM_AMDGPU_USERQ_WAIT 0x18 #define DRM_IOCTL_AMDGPU_GEM_CREATE DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_CREATE, union drm_amdgpu_gem_create) #define DRM_IOCTL_AMDGPU_GEM_MMAP DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_GEM_MMAP, union drm_amdgpu_gem_mmap) @@ -71,6 +74,9 @@ extern "C" { #define DRM_IOCTL_AMDGPU_VM DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_VM, union drm_amdgpu_vm) #define DRM_IOCTL_AMDGPU_FENCE_TO_HANDLE DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_FENCE_TO_HANDLE, union drm_amdgpu_fence_to_handle) #define DRM_IOCTL_AMDGPU_SCHED DRM_IOW(DRM_COMMAND_BASE + DRM_AMDGPU_SCHED, union drm_amdgpu_sched) +#define DRM_IOCTL_AMDGPU_USERQ DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ, union drm_amdgpu_userq) +#define DRM_IOCTL_AMDGPU_USERQ_SIGNAL DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_SIGNAL, struct drm_amdgpu_userq_signal) +#define DRM_IOCTL_AMDGPU_USERQ_WAIT DRM_IOWR(DRM_COMMAND_BASE + DRM_AMDGPU_USERQ_WAIT, struct drm_amdgpu_userq_wait) /** * DOC: memory domains @@ -317,6 +323,241 @@ union drm_amdgpu_ctx { union drm_amdgpu_ctx_out out; }; +/* user queue IOCTL operations */ +#define AMDGPU_USERQ_OP_CREATE 1 +#define AMDGPU_USERQ_OP_FREE 2 + +/* + * This structure is a container to pass input configuration + * info for all supported userqueue related operations. + * For operation AMDGPU_USERQ_OP_CREATE: user is expected + * to set all fields, excep the parameter 'queue_id'. + * For operation AMDGPU_USERQ_OP_FREE: the only input parameter expected + * to be set is 'queue_id', eveything else is ignored. + */ +struct drm_amdgpu_userq_in { + /** AMDGPU_USERQ_OP_* */ + __u32 op; + /** Queue id passed for operation USERQ_OP_FREE */ + __u32 queue_id; + /** the target GPU engine to execute workload (AMDGPU_HW_IP_*) */ + __u32 ip_type; + /** + * @doorbell_handle: the handle of doorbell GEM object + * associated to this userqueue client. + */ + __u32 doorbell_handle; + /** + * @doorbell_offset: 32-bit offset of the doorbell in the doorbell bo. + * Kernel will generate absolute doorbell offset using doorbell_handle + * and doorbell_offset in the doorbell bo. + */ + __u32 doorbell_offset; + __u32 _pad; + /** + * @queue_va: Virtual address of the GPU memory which holds the queue + * object. The queue holds the workload packets. + */ + __u64 queue_va; + /** + * @queue_size: Size of the queue in bytes, this needs to be 256-byte + * aligned. + */ + __u64 queue_size; + /** + * @rptr_va : Virtual address of the GPU memory which holds the ring RPTR. + * This object must be at least 8 byte in size and aligned to 8-byte offset. + */ + __u64 rptr_va; + /** + * @wptr_va : Virtual address of the GPU memory which holds the ring WPTR. + * This object must be at least 8 byte in size and aligned to 8-byte offset. + * + * Queue, RPTR and WPTR can come from the same object, as long as the size + * and alignment related requirements are met. + */ + __u64 wptr_va; + /** + * @mqd: MQD (memory queue descriptor) is a set of parameters which allow + * the GPU to uniquely define and identify a usermode queue. + * + * MQD data can be of different size for different GPU IP/engine and + * their respective versions/revisions, so this points to a __u64 * + * which holds IP specific MQD of this usermode queue. + */ + __u64 mqd; + /** + * @size: size of MQD data in bytes, it must match the MQD structure + * size of the respective engine/revision defined in UAPI for ex, for + * gfx11 workloads, size = sizeof(drm_amdgpu_userq_mqd_gfx11). + */ + __u64 mqd_size; +}; + +/* The structure to carry output of userqueue ops */ +struct drm_amdgpu_userq_out { + /** + * For operation AMDGPU_USERQ_OP_CREATE: This field contains a unique + * queue ID to represent the newly created userqueue in the system, otherwise + * it should be ignored. + */ + __u32 queue_id; + __u32 _pad; +}; + +union drm_amdgpu_userq { + struct drm_amdgpu_userq_in in; + struct drm_amdgpu_userq_out out; +}; + +/* GFX V11 IP specific MQD parameters */ +struct drm_amdgpu_userq_mqd_gfx11 { + /** + * @shadow_va: Virtual address of the GPU memory to hold the shadow buffer. + * Use AMDGPU_INFO_IOCTL to find the exact size of the object. + */ + __u64 shadow_va; + /** + * @csa_va: Virtual address of the GPU memory to hold the CSA buffer. + * Use AMDGPU_INFO_IOCTL to find the exact size of the object. + */ + __u64 csa_va; +}; + +/* GFX V11 SDMA IP specific MQD parameters */ +struct drm_amdgpu_userq_mqd_sdma_gfx11 { + /** + * @csa_va: Virtual address of the GPU memory to hold the CSA buffer. + * This must be a from a separate GPU object, and use AMDGPU_INFO IOCTL + * to get the size. + */ + __u64 csa_va; +}; + +/* GFX V11 Compute IP specific MQD parameters */ +struct drm_amdgpu_userq_mqd_compute_gfx11 { + /** + * @eop_va: Virtual address of the GPU memory to hold the EOP buffer. + * This must be a from a separate GPU object, and use AMDGPU_INFO IOCTL + * to get the size. + */ + __u64 eop_va; +}; + +/* userq signal/wait ioctl */ +struct drm_amdgpu_userq_signal { + /** + * @queue_id: Queue handle used by the userq fence creation function + * to retrieve the WPTR. + */ + __u32 queue_id; + __u32 pad; + /** + * @syncobj_handles: The list of syncobj handles submitted by the user queue + * job to be signaled. + */ + __u64 syncobj_handles; + /** + * @num_syncobj_handles: A count that represents the number of syncobj handles in + * @syncobj_handles. + */ + __u64 num_syncobj_handles; + /** + * @bo_read_handles: The list of BO handles that the submitted user queue job + * is using for read only. This will update BO fences in the kernel. + */ + __u64 bo_read_handles; + /** + * @bo_write_handles: The list of BO handles that the submitted user queue job + * is using for write only. This will update BO fences in the kernel. + */ + __u64 bo_write_handles; + /** + * @num_bo_read_handles: A count that represents the number of read BO handles in + * @bo_read_handles. + */ + __u32 num_bo_read_handles; + /** + * @num_bo_write_handles: A count that represents the number of write BO handles in + * @bo_write_handles. + */ + __u32 num_bo_write_handles; + +}; + +struct drm_amdgpu_userq_fence_info { + /** + * @va: A gpu address allocated for each queue which stores the + * read pointer (RPTR) value. + */ + __u64 va; + /** + * @value: A 64 bit value represents the write pointer (WPTR) of the + * queue commands which compared with the RPTR value to signal the + * fences. + */ + __u64 value; +}; + +struct drm_amdgpu_userq_wait { + /** + * @syncobj_handles: The list of syncobj handles submitted by the user queue + * job to get the va/value pairs. + */ + __u64 syncobj_handles; + /** + * @syncobj_timeline_handles: The list of timeline syncobj handles submitted by + * the user queue job to get the va/value pairs at given @syncobj_timeline_points. + */ + __u64 syncobj_timeline_handles; + /** + * @syncobj_timeline_points: The list of timeline syncobj points submitted by the + * user queue job for the corresponding @syncobj_timeline_handles. + */ + __u64 syncobj_timeline_points; + /** + * @bo_read_handles: The list of read BO handles submitted by the user queue + * job to get the va/value pairs. + */ + __u64 bo_read_handles; + /** + * @bo_write_handles: The list of write BO handles submitted by the user queue + * job to get the va/value pairs. + */ + __u64 bo_write_handles; + /** + * @num_syncobj_timeline_handles: A count that represents the number of timeline + * syncobj handles in @syncobj_timeline_handles. + */ + __u16 num_syncobj_timeline_handles; + /** + * @num_fences: This field can be used both as input and output. As input it defines + * the maximum number of fences that can be returned and as output it will specify + * how many fences were actually returned from the ioctl. + */ + __u16 num_fences; + /** + * @num_syncobj_handles: A count that represents the number of syncobj handles in + * @syncobj_handles. + */ + __u32 num_syncobj_handles; + /** + * @num_bo_read_handles: A count that represents the number of read BO handles in + * @bo_read_handles. + */ + __u32 num_bo_read_handles; + /** + * @num_bo_write_handles: A count that represents the number of write BO handles in + * @bo_write_handles. + */ + __u32 num_bo_write_handles; + /** + * @out_fences: The field is a return value from the ioctl containing the list of + * address/value pairs to wait for. + */ + __u64 out_fences; +}; + /* vm ioctl */ #define AMDGPU_VM_OP_RESERVE_VMID 1 #define AMDGPU_VM_OP_UNRESERVE_VMID 2 @@ -590,6 +831,19 @@ struct drm_amdgpu_gem_va { __u64 offset_in_bo; /** Specify mapping size. Must be correctly aligned. */ __u64 map_size; + /** + * vm_timeline_point is a sequence number used to add new timeline point. + */ + __u64 vm_timeline_point; + /** + * The vm page table update fence is installed in given vm_timeline_syncobj_out + * at vm_timeline_point. + */ + __u32 vm_timeline_syncobj_out; + /** the number of syncobj handles in @input_fence_syncobj_handles */ + __u32 num_syncobj_handles; + /** Array of sync object handle to wait for given input fences */ + __u64 input_fence_syncobj_handles; }; #define AMDGPU_HW_IP_GFX 0