1 // SPDX-License-Identifier: MIT
31  * - Passing in a list BO which are read / written to creating implicit syncs
32  * - Binding at exec time
33  * - Flow controlling the ring at exec time
36  * passed into an exec, using the dma-buf implicit sync uAPI, have binds as
37  * separate operations, and using the DRM scheduler to flow control the ring.
61  * Rebinds / dma-resv usage applies to non-compute mode VMs only as for compute
64  * There is no need to flow control the ring in the exec as we write the ring at
65  * submission time and set the DRM scheduler max job limit SIZE_OF_RING /
67  * ring is available.
74  * .. code-block::
77  *	Wait for any async VM bind passed as in-fences to start
78  *	<----------------------------------------------------------------------|
81  *	Lock exec (VM dma-resv lock, external BOs dma-resv locks)              |
83  *	Create job                                                             |
84  *	Rebind invalidated userptrs + evicted BOs (non-compute-mode)           |
85  *	Add rebind fence dependency to job                                     |
86  *	Add job VM dma-resv bookkeeping slot (non-compute mode)                |
87  *	Add job to external BOs dma-resv write slots (non-compute mode)        |
88  *	Check if any userptrs invalidated since pin ------ Drop locks ---------|
89  *	Install in / out fences for job
90  *	Submit job
100 	struct xe_vm *vm = container_of(vm_exec->vm, struct xe_vm, gpuvm);  in xe_exec_fn()
103 	/* The fence slot added here is intended for the exec sched job. */  in xe_exec_fn()
104 	xe_vm_set_validation_exec(vm, &vm_exec->exec);  in xe_exec_fn()
105 	ret = xe_vm_validate_rebind(vm, &vm_exec->exec, 1);  in xe_exec_fn()
115 	struct drm_xe_sync __user *syncs_user = u64_to_user_ptr(args->syncs);  in xe_exec_ioctl()
116 	u64 __user *addresses_user = u64_to_user_ptr(args->address);  in xe_exec_ioctl()
122 	u32 i, num_syncs, num_ufence = 0;  in xe_exec_ioctl()
124 	struct xe_sched_job *job;  in xe_exec_ioctl()  local
127 	int err = 0;  in xe_exec_ioctl()
131 	if (XE_IOCTL_DBG(xe, args->extensions) ||  in xe_exec_ioctl()
132 	    XE_IOCTL_DBG(xe, args->pad[0] || args->pad[1] || args->pad[2]) ||  in xe_exec_ioctl()
133 	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))  in xe_exec_ioctl()
134 		return -EINVAL;  in xe_exec_ioctl()
136 	q = xe_exec_queue_lookup(xef, args->exec_queue_id);  in xe_exec_ioctl()
138 		return -ENOENT;  in xe_exec_ioctl()
140 	if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_VM)) {  in xe_exec_ioctl()
141 		err = -EINVAL;  in xe_exec_ioctl()
145 	if (XE_IOCTL_DBG(xe, args->num_batch_buffer &&  in xe_exec_ioctl()
146 			 q->width != args->num_batch_buffer)) {  in xe_exec_ioctl()
147 		err = -EINVAL;  in xe_exec_ioctl()
151 	if (XE_IOCTL_DBG(xe, q->ops->reset_status(q))) {  in xe_exec_ioctl()
152 		err = -ECANCELED;  in xe_exec_ioctl()
156 	if (args->num_syncs) {  in xe_exec_ioctl()
157 		syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);  in xe_exec_ioctl()
159 			err = -ENOMEM;  in xe_exec_ioctl()
164 	vm = q->vm;  in xe_exec_ioctl()
166 	for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {  in xe_exec_ioctl()
170 					   SYNC_PARSE_FLAG_LR_MODE : 0));  in xe_exec_ioctl()
179 		err = -EINVAL;  in xe_exec_ioctl()
185 				     q->width);  in xe_exec_ioctl()
187 			err = -EFAULT;  in xe_exec_ioctl()
192 	group = q->hwe->hw_engine_group;  in xe_exec_ioctl()
203 		err = down_write_killable(&vm->lock);  in xe_exec_ioctl()
207 		err = down_read_interruptible(&vm->lock);  in xe_exec_ioctl()
215 		downgrade_write(&vm->lock);  in xe_exec_ioctl()
221 	if (!args->num_batch_buffer) {  in xe_exec_ioctl()
235 			for (i = 0; i < num_syncs; i++)  in xe_exec_ioctl()
248 	 * return -ERESTARTSYS and the IOCTL will be rerun.  in xe_exec_ioctl()
250 	err = wait_for_completion_interruptible(&xe->pm_block);  in xe_exec_ioctl()
255 		vm_exec.vm = &vm->gpuvm;  in xe_exec_ioctl()
257 		err = xe_validation_exec_lock(&ctx, &vm_exec, &xe->val);  in xe_exec_ioctl()
262 	if (xe_vm_is_closed_or_banned(q->vm)) {  in xe_exec_ioctl()
263 		drm_warn(&xe->drm, "Trying to schedule after vm is closed or banned\n");  in xe_exec_ioctl()
264 		err = -ECANCELED;  in xe_exec_ioctl()
269 		err = -EWOULDBLOCK;	/* Aliased to -EAGAIN */  in xe_exec_ioctl()
275 		err = xe_vm_validate_protected(q->vm);  in xe_exec_ioctl()
280 	job = xe_sched_job_create(q, xe_exec_queue_is_parallel(q) ?  in xe_exec_ioctl()
281 				  addresses : &args->address);  in xe_exec_ioctl()
282 	if (IS_ERR(job)) {  in xe_exec_ioctl()
283 		err = PTR_ERR(job);  in xe_exec_ioctl()
289 		err = xe_sched_job_add_deps(job,  in xe_exec_ioctl()
296 	for (i = 0; i < num_syncs && !err; i++)  in xe_exec_ioctl()
297 		err = xe_sync_entry_add_deps(&syncs[i], job);  in xe_exec_ioctl()
302 		err = xe_sched_job_last_fence_add_dep(job, vm);  in xe_exec_ioctl()
317 	 * the job and let the DRM scheduler / backend clean up the job.  in xe_exec_ioctl()
319 	xe_sched_job_arm(job);  in xe_exec_ioctl()
321 		drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, &job->drm.s_fence->finished,  in xe_exec_ioctl()
325 	for (i = 0; i < num_syncs; i++) {  in xe_exec_ioctl()
326 		xe_sync_entry_signal(&syncs[i], &job->drm.s_fence->finished);  in xe_exec_ioctl()
327 		xe_sched_job_init_user_fence(job, &syncs[i]);  in xe_exec_ioctl()
331 		q->ring_ops->emit_job(job);  in xe_exec_ioctl()
333 		xe_exec_queue_last_fence_set(q, vm, &job->drm.s_fence->finished);  in xe_exec_ioctl()
334 	xe_sched_job_push(job);  in xe_exec_ioctl()
338 		spin_lock(&xe->ttm.lru_lock);  in xe_exec_ioctl()
339 		ttm_lru_bulk_move_tail(&vm->lru_bulk_move);  in xe_exec_ioctl()
340 		spin_unlock(&xe->ttm.lru_lock);  in xe_exec_ioctl()
351 		xe_sched_job_put(job);  in xe_exec_ioctl()
356 	up_read(&vm->lock);  in xe_exec_ioctl()
357 	if (err == -EAGAIN && !skip_retry)  in xe_exec_ioctl()
363 	while (num_syncs--)  in xe_exec_ioctl()