1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2022 Intel Corporation
4 */
5
6 #include "xe_exec.h"
7
8 #include <drm/drm_device.h>
9 #include <drm/drm_exec.h>
10 #include <drm/drm_file.h>
11 #include <uapi/drm/xe_drm.h>
12 #include <linux/delay.h>
13
14 #include "xe_bo.h"
15 #include "xe_device.h"
16 #include "xe_exec_queue.h"
17 #include "xe_hw_engine_group.h"
18 #include "xe_macros.h"
19 #include "xe_ring_ops_types.h"
20 #include "xe_sched_job.h"
21 #include "xe_sync.h"
22 #include "xe_vm.h"
23
24 /**
25 * DOC: Execbuf (User GPU command submission)
26 *
27 * Execs have historically been rather complicated in DRM drivers (at least in
28 * the i915) because a few things:
29 *
30 * - Passing in a list BO which are read / written to creating implicit syncs
31 * - Binding at exec time
32 * - Flow controlling the ring at exec time
33 *
34 * In XE we avoid all of this complication by not allowing a BO list to be
35 * passed into an exec, using the dma-buf implicit sync uAPI, have binds as
36 * seperate operations, and using the DRM scheduler to flow control the ring.
37 * Let's deep dive on each of these.
38 *
39 * We can get away from a BO list by forcing the user to use in / out fences on
40 * every exec rather than the kernel tracking dependencies of BO (e.g. if the
41 * user knows an exec writes to a BO and reads from the BO in the next exec, it
42 * is the user's responsibility to pass in / out fence between the two execs).
43 *
44 * We do not allow a user to trigger a bind at exec time rather we have a VM
45 * bind IOCTL which uses the same in / out fence interface as exec. In that
46 * sense, a VM bind is basically the same operation as an exec from the user
47 * perspective. e.g. If an exec depends on a VM bind use the in / out fence
48 * interface (struct drm_xe_sync) to synchronize like syncing between two
49 * dependent execs.
50 *
51 * Although a user cannot trigger a bind, we still have to rebind userptrs in
52 * the VM that have been invalidated since the last exec, likewise we also have
53 * to rebind BOs that have been evicted by the kernel. We schedule these rebinds
54 * behind any pending kernel operations on any external BOs in VM or any BOs
55 * private to the VM. This is accomplished by the rebinds waiting on BOs
56 * DMA_RESV_USAGE_KERNEL slot (kernel ops) and kernel ops waiting on all BOs
57 * slots (inflight execs are in the DMA_RESV_USAGE_BOOKKEEP for private BOs and
58 * for external BOs).
59 *
60 * Rebinds / dma-resv usage applies to non-compute mode VMs only as for compute
61 * mode VMs we use preempt fences and a rebind worker (TODO: add link).
62 *
63 * There is no need to flow control the ring in the exec as we write the ring at
64 * submission time and set the DRM scheduler max job limit SIZE_OF_RING /
65 * MAX_JOB_SIZE. The DRM scheduler will then hold all jobs until space in the
66 * ring is available.
67 *
68 * All of this results in a rather simple exec implementation.
69 *
70 * Flow
71 * ~~~~
72 *
73 * .. code-block::
74 *
75 * Parse input arguments
76 * Wait for any async VM bind passed as in-fences to start
77 * <----------------------------------------------------------------------|
78 * Lock global VM lock in read mode |
79 * Pin userptrs (also finds userptr invalidated since last exec) |
80 * Lock exec (VM dma-resv lock, external BOs dma-resv locks) |
81 * Validate BOs that have been evicted |
82 * Create job |
83 * Rebind invalidated userptrs + evicted BOs (non-compute-mode) |
84 * Add rebind fence dependency to job |
85 * Add job VM dma-resv bookkeeping slot (non-compute mode) |
86 * Add job to external BOs dma-resv write slots (non-compute mode) |
87 * Check if any userptrs invalidated since pin ------ Drop locks ---------|
88 * Install in / out fences for job
89 * Submit job
90 * Unlock all
91 */
92
93 /*
94 * Add validation and rebinding to the drm_exec locking loop, since both can
95 * trigger eviction which may require sleeping dma_resv locks.
96 */
xe_exec_fn(struct drm_gpuvm_exec * vm_exec)97 static int xe_exec_fn(struct drm_gpuvm_exec *vm_exec)
98 {
99 struct xe_vm *vm = container_of(vm_exec->vm, struct xe_vm, gpuvm);
100
101 /* The fence slot added here is intended for the exec sched job. */
102 return xe_vm_validate_rebind(vm, &vm_exec->exec, 1);
103 }
104
xe_exec_ioctl(struct drm_device * dev,void * data,struct drm_file * file)105 int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
106 {
107 struct xe_device *xe = to_xe_device(dev);
108 struct xe_file *xef = to_xe_file(file);
109 struct drm_xe_exec *args = data;
110 struct drm_xe_sync __user *syncs_user = u64_to_user_ptr(args->syncs);
111 u64 __user *addresses_user = u64_to_user_ptr(args->address);
112 struct xe_exec_queue *q;
113 struct xe_sync_entry *syncs = NULL;
114 u64 addresses[XE_HW_ENGINE_MAX_INSTANCE];
115 struct drm_gpuvm_exec vm_exec = {.extra.fn = xe_exec_fn};
116 struct drm_exec *exec = &vm_exec.exec;
117 u32 i, num_syncs, num_ufence = 0;
118 struct xe_sched_job *job;
119 struct xe_vm *vm;
120 bool write_locked, skip_retry = false;
121 ktime_t end = 0;
122 int err = 0;
123 struct xe_hw_engine_group *group;
124 enum xe_hw_engine_group_execution_mode mode, previous_mode;
125
126 if (XE_IOCTL_DBG(xe, args->extensions) ||
127 XE_IOCTL_DBG(xe, args->pad[0] || args->pad[1] || args->pad[2]) ||
128 XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
129 return -EINVAL;
130
131 q = xe_exec_queue_lookup(xef, args->exec_queue_id);
132 if (XE_IOCTL_DBG(xe, !q))
133 return -ENOENT;
134
135 if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_VM)) {
136 err = -EINVAL;
137 goto err_exec_queue;
138 }
139
140 if (XE_IOCTL_DBG(xe, args->num_batch_buffer &&
141 q->width != args->num_batch_buffer)) {
142 err = -EINVAL;
143 goto err_exec_queue;
144 }
145
146 if (XE_IOCTL_DBG(xe, q->ops->reset_status(q))) {
147 err = -ECANCELED;
148 goto err_exec_queue;
149 }
150
151 if (args->num_syncs) {
152 syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);
153 if (!syncs) {
154 err = -ENOMEM;
155 goto err_exec_queue;
156 }
157 }
158
159 vm = q->vm;
160
161 for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {
162 err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs],
163 &syncs_user[num_syncs], SYNC_PARSE_FLAG_EXEC |
164 (xe_vm_in_lr_mode(vm) ?
165 SYNC_PARSE_FLAG_LR_MODE : 0));
166 if (err)
167 goto err_syncs;
168
169 if (xe_sync_is_ufence(&syncs[num_syncs]))
170 num_ufence++;
171 }
172
173 if (XE_IOCTL_DBG(xe, num_ufence > 1)) {
174 err = -EINVAL;
175 goto err_syncs;
176 }
177
178 if (xe_exec_queue_is_parallel(q)) {
179 err = __copy_from_user(addresses, addresses_user, sizeof(u64) *
180 q->width);
181 if (err) {
182 err = -EFAULT;
183 goto err_syncs;
184 }
185 }
186
187 group = q->hwe->hw_engine_group;
188 mode = xe_hw_engine_group_find_exec_mode(q);
189
190 if (mode == EXEC_MODE_DMA_FENCE) {
191 err = xe_hw_engine_group_get_mode(group, mode, &previous_mode);
192 if (err)
193 goto err_syncs;
194 }
195
196 retry:
197 if (!xe_vm_in_lr_mode(vm) && xe_vm_userptr_check_repin(vm)) {
198 err = down_write_killable(&vm->lock);
199 write_locked = true;
200 } else {
201 /* We don't allow execs while the VM is in error state */
202 err = down_read_interruptible(&vm->lock);
203 write_locked = false;
204 }
205 if (err)
206 goto err_hw_exec_mode;
207
208 if (write_locked) {
209 err = xe_vm_userptr_pin(vm);
210 downgrade_write(&vm->lock);
211 write_locked = false;
212 if (err)
213 goto err_unlock_list;
214 }
215
216 if (!args->num_batch_buffer) {
217 err = xe_vm_lock(vm, true);
218 if (err)
219 goto err_unlock_list;
220
221 if (!xe_vm_in_lr_mode(vm)) {
222 struct dma_fence *fence;
223
224 fence = xe_sync_in_fence_get(syncs, num_syncs, q, vm);
225 if (IS_ERR(fence)) {
226 err = PTR_ERR(fence);
227 xe_vm_unlock(vm);
228 goto err_unlock_list;
229 }
230 for (i = 0; i < num_syncs; i++)
231 xe_sync_entry_signal(&syncs[i], fence);
232 xe_exec_queue_last_fence_set(q, vm, fence);
233 dma_fence_put(fence);
234 }
235
236 xe_vm_unlock(vm);
237 goto err_unlock_list;
238 }
239
240 vm_exec.vm = &vm->gpuvm;
241 vm_exec.flags = DRM_EXEC_INTERRUPTIBLE_WAIT;
242 if (xe_vm_in_lr_mode(vm)) {
243 drm_exec_init(exec, vm_exec.flags, 0);
244 } else {
245 err = drm_gpuvm_exec_lock(&vm_exec);
246 if (err) {
247 if (xe_vm_validate_should_retry(exec, err, &end))
248 err = -EAGAIN;
249 goto err_unlock_list;
250 }
251 }
252
253 if (xe_vm_is_closed_or_banned(q->vm)) {
254 drm_warn(&xe->drm, "Trying to schedule after vm is closed or banned\n");
255 err = -ECANCELED;
256 goto err_exec;
257 }
258
259 if (xe_exec_queue_is_lr(q) && xe_exec_queue_ring_full(q)) {
260 err = -EWOULDBLOCK; /* Aliased to -EAGAIN */
261 skip_retry = true;
262 goto err_exec;
263 }
264
265 job = xe_sched_job_create(q, xe_exec_queue_is_parallel(q) ?
266 addresses : &args->address);
267 if (IS_ERR(job)) {
268 err = PTR_ERR(job);
269 goto err_exec;
270 }
271
272 /* Wait behind rebinds */
273 if (!xe_vm_in_lr_mode(vm)) {
274 err = xe_sched_job_add_deps(job,
275 xe_vm_resv(vm),
276 DMA_RESV_USAGE_KERNEL);
277 if (err)
278 goto err_put_job;
279 }
280
281 for (i = 0; i < num_syncs && !err; i++)
282 err = xe_sync_entry_add_deps(&syncs[i], job);
283 if (err)
284 goto err_put_job;
285
286 if (!xe_vm_in_lr_mode(vm)) {
287 err = xe_sched_job_last_fence_add_dep(job, vm);
288 if (err)
289 goto err_put_job;
290
291 err = down_read_interruptible(&vm->userptr.notifier_lock);
292 if (err)
293 goto err_put_job;
294
295 err = __xe_vm_userptr_needs_repin(vm);
296 if (err)
297 goto err_repin;
298 }
299
300 /*
301 * Point of no return, if we error after this point just set an error on
302 * the job and let the DRM scheduler / backend clean up the job.
303 */
304 xe_sched_job_arm(job);
305 if (!xe_vm_in_lr_mode(vm))
306 drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, &job->drm.s_fence->finished,
307 DMA_RESV_USAGE_BOOKKEEP,
308 DMA_RESV_USAGE_BOOKKEEP);
309
310 for (i = 0; i < num_syncs; i++) {
311 xe_sync_entry_signal(&syncs[i], &job->drm.s_fence->finished);
312 xe_sched_job_init_user_fence(job, &syncs[i]);
313 }
314
315 if (xe_exec_queue_is_lr(q))
316 q->ring_ops->emit_job(job);
317 if (!xe_vm_in_lr_mode(vm))
318 xe_exec_queue_last_fence_set(q, vm, &job->drm.s_fence->finished);
319 xe_sched_job_push(job);
320 xe_vm_reactivate_rebind(vm);
321
322 if (!err && !xe_vm_in_lr_mode(vm)) {
323 spin_lock(&xe->ttm.lru_lock);
324 ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
325 spin_unlock(&xe->ttm.lru_lock);
326 }
327
328 if (mode == EXEC_MODE_LR)
329 xe_hw_engine_group_resume_faulting_lr_jobs(group);
330
331 err_repin:
332 if (!xe_vm_in_lr_mode(vm))
333 up_read(&vm->userptr.notifier_lock);
334 err_put_job:
335 if (err)
336 xe_sched_job_put(job);
337 err_exec:
338 drm_exec_fini(exec);
339 err_unlock_list:
340 up_read(&vm->lock);
341 if (err == -EAGAIN && !skip_retry)
342 goto retry;
343 err_hw_exec_mode:
344 if (mode == EXEC_MODE_DMA_FENCE)
345 xe_hw_engine_group_put(group);
346 err_syncs:
347 while (num_syncs--)
348 xe_sync_entry_cleanup(&syncs[num_syncs]);
349 kfree(syncs);
350 err_exec_queue:
351 xe_exec_queue_put(q);
352
353 return err;
354 }
355