1 // SPDX-License-Identifier: MIT
3 * Copyright © 2022 Intel Corporation
8 #include <drm/drm_device.h>
9 #include <drm/drm_exec.h>
10 #include <drm/drm_file.h>
11 #include <drm/xe_drm.h>
12 #include <linux/delay.h>
15 #include "xe_device.h"
16 #include "xe_exec_queue.h"
17 #include "xe_macros.h"
18 #include "xe_ring_ops_types.h"
19 #include "xe_sched_job.h"
24 * DOC: Execbuf (User GPU command submission)
26 * Execs have historically been rather complicated in DRM drivers (at least in
27 * the i915) because a few things:
29 * - Passing in a list BO which are read / written to creating implicit syncs
30 * - Binding at exec time
31 * - Flow controlling the ring at exec time
33 * In XE we avoid all of this complication by not allowing a BO list to be
34 * passed into an exec, using the dma-buf implicit sync uAPI, have binds as
35 * seperate operations, and using the DRM scheduler to flow control the ring.
36 * Let's deep dive on each of these.
38 * We can get away from a BO list by forcing the user to use in / out fences on
39 * every exec rather than the kernel tracking dependencies of BO (e.g. if the
40 * user knows an exec writes to a BO and reads from the BO in the next exec, it
41 * is the user's responsibility to pass in / out fence between the two execs).
43 * Implicit dependencies for external BOs are handled by using the dma-buf
44 * implicit dependency uAPI (TODO: add link). To make this works each exec must
45 * install the job's fence into the DMA_RESV_USAGE_WRITE slot of every external
46 * BO mapped in the VM.
48 * We do not allow a user to trigger a bind at exec time rather we have a VM
49 * bind IOCTL which uses the same in / out fence interface as exec. In that
50 * sense, a VM bind is basically the same operation as an exec from the user
51 * perspective. e.g. If an exec depends on a VM bind use the in / out fence
52 * interface (struct drm_xe_sync) to synchronize like syncing between two
55 * Although a user cannot trigger a bind, we still have to rebind userptrs in
56 * the VM that have been invalidated since the last exec, likewise we also have
57 * to rebind BOs that have been evicted by the kernel. We schedule these rebinds
58 * behind any pending kernel operations on any external BOs in VM or any BOs
59 * private to the VM. This is accomplished by the rebinds waiting on BOs
60 * DMA_RESV_USAGE_KERNEL slot (kernel ops) and kernel ops waiting on all BOs
61 * slots (inflight execs are in the DMA_RESV_USAGE_BOOKING for private BOs and
62 * in DMA_RESV_USAGE_WRITE for external BOs).
64 * Rebinds / dma-resv usage applies to non-compute mode VMs only as for compute
65 * mode VMs we use preempt fences and a rebind worker (TODO: add link).
67 * There is no need to flow control the ring in the exec as we write the ring at
68 * submission time and set the DRM scheduler max job limit SIZE_OF_RING /
69 * MAX_JOB_SIZE. The DRM scheduler will then hold all jobs until space in the
72 * All of this results in a rather simple exec implementation.
79 * Parse input arguments
80 * Wait for any async VM bind passed as in-fences to start
81 * <----------------------------------------------------------------------|
82 * Lock global VM lock in read mode |
83 * Pin userptrs (also finds userptr invalidated since last exec) |
84 * Lock exec (VM dma-resv lock, external BOs dma-resv locks) |
85 * Validate BOs that have been evicted |
87 * Rebind invalidated userptrs + evicted BOs (non-compute-mode) |
88 * Add rebind fence dependency to job |
89 * Add job VM dma-resv bookkeeping slot (non-compute mode) |
90 * Add job to external BOs dma-resv write slots (non-compute mode) |
91 * Check if any userptrs invalidated since pin ------ Drop locks ---------|
92 * Install in / out fences for job
97 static int xe_exec_fn(struct drm_gpuvm_exec *vm_exec)
99 struct xe_vm *vm = container_of(vm_exec->vm, struct xe_vm, gpuvm);
100 struct drm_gem_object *obj;
105 ret = drm_gpuvm_validate(vm_exec->vm, &vm_exec->exec);
110 * 1 fence slot for the final submit, and 1 more for every per-tile for
111 * GPU bind and 1 extra for CPU bind. Note that there are potentially
112 * many vma per object/dma-resv, however the fence slot will just be
113 * re-used, since they are largely the same timeline and the seqno
114 * should be in order. In the case of CPU bind there is dummy fence used
115 * for all CPU binds, so no need to have a per-tile slot for that.
117 num_fences = 1 + 1 + vm->xe->info.tile_count;
120 * We don't know upfront exactly how many fence slots we will need at
121 * the start of the exec, since the TTM bo_validate above can consume
122 * numerous fence slots. Also due to how the dma_resv_reserve_fences()
123 * works it only ensures that at least that many fence slots are
124 * available i.e if there are already 10 slots available and we reserve
125 * two more, it can just noop without reserving anything. With this it
126 * is quite possible that TTM steals some of the fence slots and then
127 * when it comes time to do the vma binding and final exec stage we are
128 * lacking enough fence slots, leading to some nasty BUG_ON() when
129 * adding the fences. Hence just add our own fences here, after the
132 drm_exec_for_each_locked_object(&vm_exec->exec, index, obj) {
133 ret = dma_resv_reserve_fences(obj->resv, num_fences);
141 int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
143 struct xe_device *xe = to_xe_device(dev);
144 struct xe_file *xef = to_xe_file(file);
145 struct drm_xe_exec *args = data;
146 struct drm_xe_sync __user *syncs_user = u64_to_user_ptr(args->syncs);
147 u64 __user *addresses_user = u64_to_user_ptr(args->address);
148 struct xe_exec_queue *q;
149 struct xe_sync_entry *syncs = NULL;
150 u64 addresses[XE_HW_ENGINE_MAX_INSTANCE];
151 struct drm_gpuvm_exec vm_exec = {.extra.fn = xe_exec_fn};
152 struct drm_exec *exec = &vm_exec.exec;
153 u32 i, num_syncs = 0, num_ufence = 0;
154 struct xe_sched_job *job;
155 struct dma_fence *rebind_fence;
157 bool write_locked, skip_retry = false;
161 if (XE_IOCTL_DBG(xe, args->extensions) ||
162 XE_IOCTL_DBG(xe, args->pad[0] || args->pad[1] || args->pad[2]) ||
163 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
166 q = xe_exec_queue_lookup(xef, args->exec_queue_id);
167 if (XE_IOCTL_DBG(xe, !q))
170 if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_VM))
173 if (XE_IOCTL_DBG(xe, args->num_batch_buffer &&
174 q->width != args->num_batch_buffer))
177 if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_BANNED)) {
182 if (args->num_syncs) {
183 syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);
192 for (i = 0; i < args->num_syncs; i++) {
193 err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs++],
194 &syncs_user[i], SYNC_PARSE_FLAG_EXEC |
195 (xe_vm_in_lr_mode(vm) ?
196 SYNC_PARSE_FLAG_LR_MODE : 0));
200 if (xe_sync_is_ufence(&syncs[i]))
204 if (XE_IOCTL_DBG(xe, num_ufence > 1)) {
209 if (xe_exec_queue_is_parallel(q)) {
210 err = __copy_from_user(addresses, addresses_user, sizeof(u64) *
219 if (!xe_vm_in_lr_mode(vm) && xe_vm_userptr_check_repin(vm)) {
220 err = down_write_killable(&vm->lock);
223 /* We don't allow execs while the VM is in error state */
224 err = down_read_interruptible(&vm->lock);
225 write_locked = false;
231 err = xe_vm_userptr_pin(vm);
232 downgrade_write(&vm->lock);
233 write_locked = false;
235 goto err_unlock_list;
238 if (!args->num_batch_buffer) {
239 err = xe_vm_lock(vm, true);
241 goto err_unlock_list;
243 if (!xe_vm_in_lr_mode(vm)) {
244 struct dma_fence *fence;
246 fence = xe_sync_in_fence_get(syncs, num_syncs, q, vm);
248 err = PTR_ERR(fence);
249 goto err_unlock_list;
251 for (i = 0; i < num_syncs; i++)
252 xe_sync_entry_signal(&syncs[i], NULL, fence);
253 xe_exec_queue_last_fence_set(q, vm, fence);
254 dma_fence_put(fence);
258 goto err_unlock_list;
261 vm_exec.vm = &vm->gpuvm;
262 vm_exec.flags = DRM_EXEC_INTERRUPTIBLE_WAIT;
263 if (xe_vm_in_lr_mode(vm)) {
264 drm_exec_init(exec, vm_exec.flags, 0);
266 err = drm_gpuvm_exec_lock(&vm_exec);
268 if (xe_vm_validate_should_retry(exec, err, &end))
270 goto err_unlock_list;
274 if (xe_vm_is_closed_or_banned(q->vm)) {
275 drm_warn(&xe->drm, "Trying to schedule after vm is closed or banned\n");
280 if (xe_exec_queue_is_lr(q) && xe_exec_queue_ring_full(q)) {
281 err = -EWOULDBLOCK; /* Aliased to -EAGAIN */
286 job = xe_sched_job_create(q, xe_exec_queue_is_parallel(q) ?
287 addresses : &args->address);
294 * Rebind any invalidated userptr or evicted BOs in the VM, non-compute
297 rebind_fence = xe_vm_rebind(vm, false);
298 if (IS_ERR(rebind_fence)) {
299 err = PTR_ERR(rebind_fence);
304 * We store the rebind_fence in the VM so subsequent execs don't get
305 * scheduled before the rebinds of userptrs / evicted BOs is complete.
308 dma_fence_put(vm->rebind_fence);
309 vm->rebind_fence = rebind_fence;
311 if (vm->rebind_fence) {
312 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
313 &vm->rebind_fence->flags)) {
314 dma_fence_put(vm->rebind_fence);
315 vm->rebind_fence = NULL;
317 dma_fence_get(vm->rebind_fence);
318 err = drm_sched_job_add_dependency(&job->drm,
325 /* Wait behind munmap style rebinds */
326 if (!xe_vm_in_lr_mode(vm)) {
327 err = drm_sched_job_add_resv_dependencies(&job->drm,
329 DMA_RESV_USAGE_KERNEL);
334 for (i = 0; i < num_syncs && !err; i++)
335 err = xe_sync_entry_add_deps(&syncs[i], job);
339 if (!xe_vm_in_lr_mode(vm)) {
340 err = xe_sched_job_last_fence_add_dep(job, vm);
344 err = down_read_interruptible(&vm->userptr.notifier_lock);
348 err = __xe_vm_userptr_needs_repin(vm);
354 * Point of no return, if we error after this point just set an error on
355 * the job and let the DRM scheduler / backend clean up the job.
357 xe_sched_job_arm(job);
358 if (!xe_vm_in_lr_mode(vm))
359 drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, &job->drm.s_fence->finished,
360 DMA_RESV_USAGE_BOOKKEEP, DMA_RESV_USAGE_WRITE);
362 for (i = 0; i < num_syncs; i++)
363 xe_sync_entry_signal(&syncs[i], job,
364 &job->drm.s_fence->finished);
366 if (xe_exec_queue_is_lr(q))
367 q->ring_ops->emit_job(job);
368 if (!xe_vm_in_lr_mode(vm))
369 xe_exec_queue_last_fence_set(q, vm, &job->drm.s_fence->finished);
370 xe_sched_job_push(job);
371 xe_vm_reactivate_rebind(vm);
373 if (!err && !xe_vm_in_lr_mode(vm)) {
374 spin_lock(&xe->ttm.lru_lock);
375 ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
376 spin_unlock(&xe->ttm.lru_lock);
380 if (!xe_vm_in_lr_mode(vm))
381 up_read(&vm->userptr.notifier_lock);
384 xe_sched_job_put(job);
392 if (err == -EAGAIN && !skip_retry)
395 for (i = 0; i < num_syncs; i++)
396 xe_sync_entry_cleanup(&syncs[i]);
399 xe_exec_queue_put(q);