xref: /linux/drivers/gpu/drm/xe/xe_exec.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
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 <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_macros.h"
18 #include "xe_ring_ops_types.h"
19 #include "xe_sched_job.h"
20 #include "xe_sync.h"
21 #include "xe_vm.h"
22 
23 /**
24  * DOC: Execbuf (User GPU command submission)
25  *
26  * Execs have historically been rather complicated in DRM drivers (at least in
27  * the i915) because a few things:
28  *
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
32  *
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.
37  *
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).
42  *
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.
47  *
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
53  * dependent execs.
54  *
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).
63  *
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).
66  *
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
70  * ring is available.
71  *
72  * All of this results in a rather simple exec implementation.
73  *
74  * Flow
75  * ~~~~
76  *
77  * .. code-block::
78  *
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                                    |
86  *	Create job                                                             |
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
93  *	Submit job
94  *	Unlock all
95  */
96 
97 static int xe_exec_fn(struct drm_gpuvm_exec *vm_exec)
98 {
99 	return drm_gpuvm_validate(vm_exec->vm, &vm_exec->exec);
100 }
101 
102 int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
103 {
104 	struct xe_device *xe = to_xe_device(dev);
105 	struct xe_file *xef = to_xe_file(file);
106 	struct drm_xe_exec *args = data;
107 	struct drm_xe_sync __user *syncs_user = u64_to_user_ptr(args->syncs);
108 	u64 __user *addresses_user = u64_to_user_ptr(args->address);
109 	struct xe_exec_queue *q;
110 	struct xe_sync_entry *syncs = NULL;
111 	u64 addresses[XE_HW_ENGINE_MAX_INSTANCE];
112 	struct drm_gpuvm_exec vm_exec = {.extra.fn = xe_exec_fn};
113 	struct drm_exec *exec = &vm_exec.exec;
114 	u32 i, num_syncs = 0;
115 	struct xe_sched_job *job;
116 	struct dma_fence *rebind_fence;
117 	struct xe_vm *vm;
118 	bool write_locked, skip_retry = false;
119 	ktime_t end = 0;
120 	int err = 0;
121 
122 	if (XE_IOCTL_DBG(xe, args->extensions) ||
123 	    XE_IOCTL_DBG(xe, args->pad[0] || args->pad[1] || args->pad[2]) ||
124 	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
125 		return -EINVAL;
126 
127 	q = xe_exec_queue_lookup(xef, args->exec_queue_id);
128 	if (XE_IOCTL_DBG(xe, !q))
129 		return -ENOENT;
130 
131 	if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_VM))
132 		return -EINVAL;
133 
134 	if (XE_IOCTL_DBG(xe, args->num_batch_buffer &&
135 			 q->width != args->num_batch_buffer))
136 		return -EINVAL;
137 
138 	if (XE_IOCTL_DBG(xe, q->flags & EXEC_QUEUE_FLAG_BANNED)) {
139 		err = -ECANCELED;
140 		goto err_exec_queue;
141 	}
142 
143 	if (args->num_syncs) {
144 		syncs = kcalloc(args->num_syncs, sizeof(*syncs), GFP_KERNEL);
145 		if (!syncs) {
146 			err = -ENOMEM;
147 			goto err_exec_queue;
148 		}
149 	}
150 
151 	vm = q->vm;
152 
153 	for (i = 0; i < args->num_syncs; i++) {
154 		err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs++],
155 					  &syncs_user[i], SYNC_PARSE_FLAG_EXEC |
156 					  (xe_vm_in_lr_mode(vm) ?
157 					   SYNC_PARSE_FLAG_LR_MODE : 0));
158 		if (err)
159 			goto err_syncs;
160 	}
161 
162 	if (xe_exec_queue_is_parallel(q)) {
163 		err = __copy_from_user(addresses, addresses_user, sizeof(u64) *
164 				       q->width);
165 		if (err) {
166 			err = -EFAULT;
167 			goto err_syncs;
168 		}
169 	}
170 
171 retry:
172 	if (!xe_vm_in_lr_mode(vm) && xe_vm_userptr_check_repin(vm)) {
173 		err = down_write_killable(&vm->lock);
174 		write_locked = true;
175 	} else {
176 		/* We don't allow execs while the VM is in error state */
177 		err = down_read_interruptible(&vm->lock);
178 		write_locked = false;
179 	}
180 	if (err)
181 		goto err_syncs;
182 
183 	if (write_locked) {
184 		err = xe_vm_userptr_pin(vm);
185 		downgrade_write(&vm->lock);
186 		write_locked = false;
187 		if (err)
188 			goto err_unlock_list;
189 	}
190 
191 	vm_exec.vm = &vm->gpuvm;
192 	vm_exec.num_fences = 1 + vm->xe->info.tile_count;
193 	vm_exec.flags = DRM_EXEC_INTERRUPTIBLE_WAIT;
194 	if (xe_vm_in_lr_mode(vm)) {
195 		drm_exec_init(exec, vm_exec.flags, 0);
196 	} else {
197 		err = drm_gpuvm_exec_lock(&vm_exec);
198 		if (err) {
199 			if (xe_vm_validate_should_retry(exec, err, &end))
200 				err = -EAGAIN;
201 			goto err_unlock_list;
202 		}
203 	}
204 
205 	if (xe_vm_is_closed_or_banned(q->vm)) {
206 		drm_warn(&xe->drm, "Trying to schedule after vm is closed or banned\n");
207 		err = -ECANCELED;
208 		goto err_exec;
209 	}
210 
211 	if (!args->num_batch_buffer) {
212 		if (!xe_vm_in_lr_mode(vm)) {
213 			struct dma_fence *fence;
214 
215 			fence = xe_sync_in_fence_get(syncs, num_syncs, q, vm);
216 			if (IS_ERR(fence)) {
217 				err = PTR_ERR(fence);
218 				goto err_exec;
219 			}
220 			for (i = 0; i < num_syncs; i++)
221 				xe_sync_entry_signal(&syncs[i], NULL, fence);
222 			xe_exec_queue_last_fence_set(q, vm, fence);
223 			dma_fence_put(fence);
224 		}
225 
226 		goto err_exec;
227 	}
228 
229 	if (xe_exec_queue_is_lr(q) && xe_exec_queue_ring_full(q)) {
230 		err = -EWOULDBLOCK;	/* Aliased to -EAGAIN */
231 		skip_retry = true;
232 		goto err_exec;
233 	}
234 
235 	job = xe_sched_job_create(q, xe_exec_queue_is_parallel(q) ?
236 				  addresses : &args->address);
237 	if (IS_ERR(job)) {
238 		err = PTR_ERR(job);
239 		goto err_exec;
240 	}
241 
242 	/*
243 	 * Rebind any invalidated userptr or evicted BOs in the VM, non-compute
244 	 * VM mode only.
245 	 */
246 	rebind_fence = xe_vm_rebind(vm, false);
247 	if (IS_ERR(rebind_fence)) {
248 		err = PTR_ERR(rebind_fence);
249 		goto err_put_job;
250 	}
251 
252 	/*
253 	 * We store the rebind_fence in the VM so subsequent execs don't get
254 	 * scheduled before the rebinds of userptrs / evicted BOs is complete.
255 	 */
256 	if (rebind_fence) {
257 		dma_fence_put(vm->rebind_fence);
258 		vm->rebind_fence = rebind_fence;
259 	}
260 	if (vm->rebind_fence) {
261 		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
262 			     &vm->rebind_fence->flags)) {
263 			dma_fence_put(vm->rebind_fence);
264 			vm->rebind_fence = NULL;
265 		} else {
266 			dma_fence_get(vm->rebind_fence);
267 			err = drm_sched_job_add_dependency(&job->drm,
268 							   vm->rebind_fence);
269 			if (err)
270 				goto err_put_job;
271 		}
272 	}
273 
274 	/* Wait behind munmap style rebinds */
275 	if (!xe_vm_in_lr_mode(vm)) {
276 		err = drm_sched_job_add_resv_dependencies(&job->drm,
277 							  xe_vm_resv(vm),
278 							  DMA_RESV_USAGE_KERNEL);
279 		if (err)
280 			goto err_put_job;
281 	}
282 
283 	for (i = 0; i < num_syncs && !err; i++)
284 		err = xe_sync_entry_add_deps(&syncs[i], job);
285 	if (err)
286 		goto err_put_job;
287 
288 	if (!xe_vm_in_lr_mode(vm)) {
289 		err = xe_sched_job_last_fence_add_dep(job, vm);
290 		if (err)
291 			goto err_put_job;
292 
293 		err = down_read_interruptible(&vm->userptr.notifier_lock);
294 		if (err)
295 			goto err_put_job;
296 
297 		err = __xe_vm_userptr_needs_repin(vm);
298 		if (err)
299 			goto err_repin;
300 	}
301 
302 	/*
303 	 * Point of no return, if we error after this point just set an error on
304 	 * the job and let the DRM scheduler / backend clean up the job.
305 	 */
306 	xe_sched_job_arm(job);
307 	if (!xe_vm_in_lr_mode(vm))
308 		drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, &job->drm.s_fence->finished,
309 					 DMA_RESV_USAGE_BOOKKEEP, DMA_RESV_USAGE_WRITE);
310 
311 	for (i = 0; i < num_syncs; i++)
312 		xe_sync_entry_signal(&syncs[i], job,
313 				     &job->drm.s_fence->finished);
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 err_repin:
329 	if (!xe_vm_in_lr_mode(vm))
330 		up_read(&vm->userptr.notifier_lock);
331 err_put_job:
332 	if (err)
333 		xe_sched_job_put(job);
334 err_exec:
335 	drm_exec_fini(exec);
336 err_unlock_list:
337 	if (write_locked)
338 		up_write(&vm->lock);
339 	else
340 		up_read(&vm->lock);
341 	if (err == -EAGAIN && !skip_retry)
342 		goto retry;
343 err_syncs:
344 	for (i = 0; i < num_syncs; i++)
345 		xe_sync_entry_cleanup(&syncs[i]);
346 	kfree(syncs);
347 err_exec_queue:
348 	xe_exec_queue_put(q);
349 
350 	return err;
351 }
352