xref: /linux/drivers/gpu/drm/xe/xe_sync.c (revision ec8c17e5ecb4a5a74069687ccb6d2cfe1851302e)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "xe_sync.h"
7 
8 #include <linux/dma-fence-array.h>
9 #include <linux/kthread.h>
10 #include <linux/sched/mm.h>
11 #include <linux/uaccess.h>
12 
13 #include <drm/drm_print.h>
14 #include <drm/drm_syncobj.h>
15 #include <uapi/drm/xe_drm.h>
16 
17 #include "xe_device_types.h"
18 #include "xe_exec_queue.h"
19 #include "xe_macros.h"
20 #include "xe_sched_job_types.h"
21 
22 struct xe_user_fence {
23 	struct xe_device *xe;
24 	struct kref refcount;
25 	struct dma_fence_cb cb;
26 	struct work_struct worker;
27 	struct mm_struct *mm;
28 	u64 __user *addr;
29 	u64 value;
30 	int signalled;
31 };
32 
user_fence_destroy(struct kref * kref)33 static void user_fence_destroy(struct kref *kref)
34 {
35 	struct xe_user_fence *ufence = container_of(kref, struct xe_user_fence,
36 						 refcount);
37 
38 	mmdrop(ufence->mm);
39 	kfree(ufence);
40 }
41 
user_fence_get(struct xe_user_fence * ufence)42 static void user_fence_get(struct xe_user_fence *ufence)
43 {
44 	kref_get(&ufence->refcount);
45 }
46 
user_fence_put(struct xe_user_fence * ufence)47 static void user_fence_put(struct xe_user_fence *ufence)
48 {
49 	kref_put(&ufence->refcount, user_fence_destroy);
50 }
51 
user_fence_create(struct xe_device * xe,u64 addr,u64 value)52 static struct xe_user_fence *user_fence_create(struct xe_device *xe, u64 addr,
53 					       u64 value)
54 {
55 	struct xe_user_fence *ufence;
56 	u64 __user *ptr = u64_to_user_ptr(addr);
57 	u64 __maybe_unused prefetch_val;
58 
59 	if (get_user(prefetch_val, ptr))
60 		return ERR_PTR(-EFAULT);
61 
62 	ufence = kzalloc(sizeof(*ufence), GFP_KERNEL);
63 	if (!ufence)
64 		return ERR_PTR(-ENOMEM);
65 
66 	ufence->xe = xe;
67 	kref_init(&ufence->refcount);
68 	ufence->addr = ptr;
69 	ufence->value = value;
70 	ufence->mm = current->mm;
71 	mmgrab(ufence->mm);
72 
73 	return ufence;
74 }
75 
user_fence_worker(struct work_struct * w)76 static void user_fence_worker(struct work_struct *w)
77 {
78 	struct xe_user_fence *ufence = container_of(w, struct xe_user_fence, worker);
79 
80 	if (mmget_not_zero(ufence->mm)) {
81 		kthread_use_mm(ufence->mm);
82 		if (copy_to_user(ufence->addr, &ufence->value, sizeof(ufence->value)))
83 			XE_WARN_ON("Copy to user failed");
84 		kthread_unuse_mm(ufence->mm);
85 		mmput(ufence->mm);
86 	} else {
87 		drm_dbg(&ufence->xe->drm, "mmget_not_zero() failed, ufence wasn't signaled\n");
88 	}
89 
90 	/*
91 	 * Wake up waiters only after updating the ufence state, allowing the UMD
92 	 * to safely reuse the same ufence without encountering -EBUSY errors.
93 	 */
94 	WRITE_ONCE(ufence->signalled, 1);
95 	wake_up_all(&ufence->xe->ufence_wq);
96 	user_fence_put(ufence);
97 }
98 
kick_ufence(struct xe_user_fence * ufence,struct dma_fence * fence)99 static void kick_ufence(struct xe_user_fence *ufence, struct dma_fence *fence)
100 {
101 	INIT_WORK(&ufence->worker, user_fence_worker);
102 	queue_work(ufence->xe->ordered_wq, &ufence->worker);
103 	dma_fence_put(fence);
104 }
105 
user_fence_cb(struct dma_fence * fence,struct dma_fence_cb * cb)106 static void user_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
107 {
108 	struct xe_user_fence *ufence = container_of(cb, struct xe_user_fence, cb);
109 
110 	kick_ufence(ufence, fence);
111 }
112 
xe_sync_entry_parse(struct xe_device * xe,struct xe_file * xef,struct xe_sync_entry * sync,struct drm_xe_sync __user * sync_user,unsigned int flags)113 int xe_sync_entry_parse(struct xe_device *xe, struct xe_file *xef,
114 			struct xe_sync_entry *sync,
115 			struct drm_xe_sync __user *sync_user,
116 			unsigned int flags)
117 {
118 	struct drm_xe_sync sync_in;
119 	int err;
120 	bool exec = flags & SYNC_PARSE_FLAG_EXEC;
121 	bool in_lr_mode = flags & SYNC_PARSE_FLAG_LR_MODE;
122 	bool disallow_user_fence = flags & SYNC_PARSE_FLAG_DISALLOW_USER_FENCE;
123 	bool signal;
124 
125 	if (copy_from_user(&sync_in, sync_user, sizeof(*sync_user)))
126 		return -EFAULT;
127 
128 	if (XE_IOCTL_DBG(xe, sync_in.flags & ~DRM_XE_SYNC_FLAG_SIGNAL) ||
129 	    XE_IOCTL_DBG(xe, sync_in.reserved[0] || sync_in.reserved[1]))
130 		return -EINVAL;
131 
132 	signal = sync_in.flags & DRM_XE_SYNC_FLAG_SIGNAL;
133 	switch (sync_in.type) {
134 	case DRM_XE_SYNC_TYPE_SYNCOBJ:
135 		if (XE_IOCTL_DBG(xe, in_lr_mode && signal))
136 			return -EOPNOTSUPP;
137 
138 		if (XE_IOCTL_DBG(xe, upper_32_bits(sync_in.addr)))
139 			return -EINVAL;
140 
141 		sync->syncobj = drm_syncobj_find(xef->drm, sync_in.handle);
142 		if (XE_IOCTL_DBG(xe, !sync->syncobj))
143 			return -ENOENT;
144 
145 		if (!signal) {
146 			sync->fence = drm_syncobj_fence_get(sync->syncobj);
147 			if (XE_IOCTL_DBG(xe, !sync->fence))
148 				return -EINVAL;
149 		}
150 		break;
151 
152 	case DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ:
153 		if (XE_IOCTL_DBG(xe, in_lr_mode && signal))
154 			return -EOPNOTSUPP;
155 
156 		if (XE_IOCTL_DBG(xe, upper_32_bits(sync_in.addr)))
157 			return -EINVAL;
158 
159 		if (XE_IOCTL_DBG(xe, sync_in.timeline_value == 0))
160 			return -EINVAL;
161 
162 		sync->syncobj = drm_syncobj_find(xef->drm, sync_in.handle);
163 		if (XE_IOCTL_DBG(xe, !sync->syncobj))
164 			return -ENOENT;
165 
166 		if (signal) {
167 			sync->chain_fence = dma_fence_chain_alloc();
168 			if (!sync->chain_fence)
169 				return -ENOMEM;
170 		} else {
171 			sync->fence = drm_syncobj_fence_get(sync->syncobj);
172 			if (XE_IOCTL_DBG(xe, !sync->fence))
173 				return -EINVAL;
174 
175 			err = dma_fence_chain_find_seqno(&sync->fence,
176 							 sync_in.timeline_value);
177 			if (err)
178 				return err;
179 		}
180 		break;
181 
182 	case DRM_XE_SYNC_TYPE_USER_FENCE:
183 		if (XE_IOCTL_DBG(xe, disallow_user_fence))
184 			return -EOPNOTSUPP;
185 
186 		if (XE_IOCTL_DBG(xe, !signal))
187 			return -EOPNOTSUPP;
188 
189 		if (XE_IOCTL_DBG(xe, sync_in.addr & 0x7))
190 			return -EINVAL;
191 
192 		if (exec) {
193 			sync->addr = sync_in.addr;
194 		} else {
195 			sync->ufence = user_fence_create(xe, sync_in.addr,
196 							 sync_in.timeline_value);
197 			if (XE_IOCTL_DBG(xe, IS_ERR(sync->ufence)))
198 				return PTR_ERR(sync->ufence);
199 		}
200 
201 		break;
202 
203 	default:
204 		return -EINVAL;
205 	}
206 
207 	sync->type = sync_in.type;
208 	sync->flags = sync_in.flags;
209 	sync->timeline_value = sync_in.timeline_value;
210 
211 	return 0;
212 }
213 
xe_sync_entry_add_deps(struct xe_sync_entry * sync,struct xe_sched_job * job)214 int xe_sync_entry_add_deps(struct xe_sync_entry *sync, struct xe_sched_job *job)
215 {
216 	if (sync->fence)
217 		return  drm_sched_job_add_dependency(&job->drm,
218 						     dma_fence_get(sync->fence));
219 
220 	return 0;
221 }
222 
xe_sync_entry_signal(struct xe_sync_entry * sync,struct dma_fence * fence)223 void xe_sync_entry_signal(struct xe_sync_entry *sync, struct dma_fence *fence)
224 {
225 	if (!(sync->flags & DRM_XE_SYNC_FLAG_SIGNAL))
226 		return;
227 
228 	if (sync->chain_fence) {
229 		drm_syncobj_add_point(sync->syncobj, sync->chain_fence,
230 				      fence, sync->timeline_value);
231 		/*
232 		 * The chain's ownership is transferred to the
233 		 * timeline.
234 		 */
235 		sync->chain_fence = NULL;
236 	} else if (sync->syncobj) {
237 		drm_syncobj_replace_fence(sync->syncobj, fence);
238 	} else if (sync->ufence) {
239 		int err;
240 
241 		dma_fence_get(fence);
242 		user_fence_get(sync->ufence);
243 		err = dma_fence_add_callback(fence, &sync->ufence->cb,
244 					     user_fence_cb);
245 		if (err == -ENOENT) {
246 			kick_ufence(sync->ufence, fence);
247 		} else if (err) {
248 			XE_WARN_ON("failed to add user fence");
249 			user_fence_put(sync->ufence);
250 			dma_fence_put(fence);
251 		}
252 	}
253 }
254 
xe_sync_entry_cleanup(struct xe_sync_entry * sync)255 void xe_sync_entry_cleanup(struct xe_sync_entry *sync)
256 {
257 	if (sync->syncobj)
258 		drm_syncobj_put(sync->syncobj);
259 	dma_fence_put(sync->fence);
260 	dma_fence_chain_free(sync->chain_fence);
261 	if (sync->ufence)
262 		user_fence_put(sync->ufence);
263 }
264 
265 /**
266  * xe_sync_in_fence_get() - Get a fence from syncs, exec queue, and VM
267  * @sync: input syncs
268  * @num_sync: number of syncs
269  * @q: exec queue
270  * @vm: VM
271  *
272  * Get a fence from syncs, exec queue, and VM. If syncs contain in-fences create
273  * and return a composite fence of all in-fences + last fence. If no in-fences
274  * return last fence on  input exec queue. Caller must drop reference to
275  * returned fence.
276  *
277  * Return: fence on success, ERR_PTR(-ENOMEM) on failure
278  */
279 struct dma_fence *
xe_sync_in_fence_get(struct xe_sync_entry * sync,int num_sync,struct xe_exec_queue * q,struct xe_vm * vm)280 xe_sync_in_fence_get(struct xe_sync_entry *sync, int num_sync,
281 		     struct xe_exec_queue *q, struct xe_vm *vm)
282 {
283 	struct dma_fence **fences = NULL;
284 	struct dma_fence_array *cf = NULL;
285 	struct dma_fence *fence;
286 	int i, num_in_fence = 0, current_fence = 0;
287 
288 	lockdep_assert_held(&vm->lock);
289 
290 	/* Count in-fences */
291 	for (i = 0; i < num_sync; ++i) {
292 		if (sync[i].fence) {
293 			++num_in_fence;
294 			fence = sync[i].fence;
295 		}
296 	}
297 
298 	/* Easy case... */
299 	if (!num_in_fence) {
300 		fence = xe_exec_queue_last_fence_get(q, vm);
301 		return fence;
302 	}
303 
304 	/* Create composite fence */
305 	fences = kmalloc_array(num_in_fence + 1, sizeof(*fences), GFP_KERNEL);
306 	if (!fences)
307 		return ERR_PTR(-ENOMEM);
308 	for (i = 0; i < num_sync; ++i) {
309 		if (sync[i].fence) {
310 			dma_fence_get(sync[i].fence);
311 			fences[current_fence++] = sync[i].fence;
312 		}
313 	}
314 	fences[current_fence++] = xe_exec_queue_last_fence_get(q, vm);
315 	cf = dma_fence_array_create(num_in_fence, fences,
316 				    vm->composite_fence_ctx,
317 				    vm->composite_fence_seqno++,
318 				    false);
319 	if (!cf) {
320 		--vm->composite_fence_seqno;
321 		goto err_out;
322 	}
323 
324 	return &cf->base;
325 
326 err_out:
327 	while (current_fence)
328 		dma_fence_put(fences[--current_fence]);
329 	kfree(fences);
330 	kfree(cf);
331 
332 	return ERR_PTR(-ENOMEM);
333 }
334 
335 /**
336  * __xe_sync_ufence_get() - Get user fence from user fence
337  * @ufence: input user fence
338  *
339  * Get a user fence reference from user fence
340  *
341  * Return: xe_user_fence pointer with reference
342  */
__xe_sync_ufence_get(struct xe_user_fence * ufence)343 struct xe_user_fence *__xe_sync_ufence_get(struct xe_user_fence *ufence)
344 {
345 	user_fence_get(ufence);
346 
347 	return ufence;
348 }
349 
350 /**
351  * xe_sync_ufence_get() - Get user fence from sync
352  * @sync: input sync
353  *
354  * Get a user fence reference from sync.
355  *
356  * Return: xe_user_fence pointer with reference
357  */
xe_sync_ufence_get(struct xe_sync_entry * sync)358 struct xe_user_fence *xe_sync_ufence_get(struct xe_sync_entry *sync)
359 {
360 	user_fence_get(sync->ufence);
361 
362 	return sync->ufence;
363 }
364 
365 /**
366  * xe_sync_ufence_put() - Put user fence reference
367  * @ufence: user fence reference
368  *
369  */
xe_sync_ufence_put(struct xe_user_fence * ufence)370 void xe_sync_ufence_put(struct xe_user_fence *ufence)
371 {
372 	user_fence_put(ufence);
373 }
374 
375 /**
376  * xe_sync_ufence_get_status() - Get user fence status
377  * @ufence: user fence
378  *
379  * Return: 1 if signalled, 0 not signalled, <0 on error
380  */
xe_sync_ufence_get_status(struct xe_user_fence * ufence)381 int xe_sync_ufence_get_status(struct xe_user_fence *ufence)
382 {
383 	return READ_ONCE(ufence->signalled);
384 }
385