xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_process.c (revision 981ab3f1dc3949b45e317e343dfc232a102847aa)
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include <linux/mutex.h>
24 #include <linux/log2.h>
25 #include <linux/sched.h>
26 #include <linux/sched/mm.h>
27 #include <linux/slab.h>
28 #include <linux/amd-iommu.h>
29 #include <linux/notifier.h>
30 #include <linux/compat.h>
31 
32 struct mm_struct;
33 
34 #include "kfd_priv.h"
35 #include "kfd_dbgmgr.h"
36 
37 /*
38  * Initial size for the array of queues.
39  * The allocated size is doubled each time
40  * it is exceeded up to MAX_PROCESS_QUEUES.
41  */
42 #define INITIAL_QUEUE_ARRAY_SIZE 16
43 
44 /*
45  * List of struct kfd_process (field kfd_process).
46  * Unique/indexed by mm_struct*
47  */
48 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
49 static DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
50 static DEFINE_MUTEX(kfd_processes_mutex);
51 
52 DEFINE_STATIC_SRCU(kfd_processes_srcu);
53 
54 static struct workqueue_struct *kfd_process_wq;
55 
56 struct kfd_process_release_work {
57 	struct work_struct kfd_work;
58 	struct kfd_process *p;
59 };
60 
61 static struct kfd_process *find_process(const struct task_struct *thread);
62 static struct kfd_process *create_process(const struct task_struct *thread);
63 
64 void kfd_process_create_wq(void)
65 {
66 	if (!kfd_process_wq)
67 		kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
68 }
69 
70 void kfd_process_destroy_wq(void)
71 {
72 	if (kfd_process_wq) {
73 		destroy_workqueue(kfd_process_wq);
74 		kfd_process_wq = NULL;
75 	}
76 }
77 
78 struct kfd_process *kfd_create_process(const struct task_struct *thread)
79 {
80 	struct kfd_process *process;
81 
82 	if (!thread->mm)
83 		return ERR_PTR(-EINVAL);
84 
85 	/* Only the pthreads threading model is supported. */
86 	if (thread->group_leader->mm != thread->mm)
87 		return ERR_PTR(-EINVAL);
88 
89 	/* Take mmap_sem because we call __mmu_notifier_register inside */
90 	down_write(&thread->mm->mmap_sem);
91 
92 	/*
93 	 * take kfd processes mutex before starting of process creation
94 	 * so there won't be a case where two threads of the same process
95 	 * create two kfd_process structures
96 	 */
97 	mutex_lock(&kfd_processes_mutex);
98 
99 	/* A prior open of /dev/kfd could have already created the process. */
100 	process = find_process(thread);
101 	if (process)
102 		pr_debug("Process already found\n");
103 
104 	if (!process)
105 		process = create_process(thread);
106 
107 	mutex_unlock(&kfd_processes_mutex);
108 
109 	up_write(&thread->mm->mmap_sem);
110 
111 	return process;
112 }
113 
114 struct kfd_process *kfd_get_process(const struct task_struct *thread)
115 {
116 	struct kfd_process *process;
117 
118 	if (!thread->mm)
119 		return ERR_PTR(-EINVAL);
120 
121 	/* Only the pthreads threading model is supported. */
122 	if (thread->group_leader->mm != thread->mm)
123 		return ERR_PTR(-EINVAL);
124 
125 	process = find_process(thread);
126 
127 	return process;
128 }
129 
130 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
131 {
132 	struct kfd_process *process;
133 
134 	hash_for_each_possible_rcu(kfd_processes_table, process,
135 					kfd_processes, (uintptr_t)mm)
136 		if (process->mm == mm)
137 			return process;
138 
139 	return NULL;
140 }
141 
142 static struct kfd_process *find_process(const struct task_struct *thread)
143 {
144 	struct kfd_process *p;
145 	int idx;
146 
147 	idx = srcu_read_lock(&kfd_processes_srcu);
148 	p = find_process_by_mm(thread->mm);
149 	srcu_read_unlock(&kfd_processes_srcu, idx);
150 
151 	return p;
152 }
153 
154 static void kfd_process_wq_release(struct work_struct *work)
155 {
156 	struct kfd_process_release_work *my_work;
157 	struct kfd_process_device *pdd, *temp;
158 	struct kfd_process *p;
159 
160 	my_work = (struct kfd_process_release_work *) work;
161 
162 	p = my_work->p;
163 
164 	pr_debug("Releasing process (pasid %d) in workqueue\n",
165 			p->pasid);
166 
167 	mutex_lock(&p->mutex);
168 
169 	list_for_each_entry_safe(pdd, temp, &p->per_device_data,
170 							per_device_list) {
171 		pr_debug("Releasing pdd (topology id %d) for process (pasid %d) in workqueue\n",
172 				pdd->dev->id, p->pasid);
173 
174 		if (pdd->reset_wavefronts)
175 			dbgdev_wave_reset_wavefronts(pdd->dev, p);
176 
177 		amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
178 		list_del(&pdd->per_device_list);
179 
180 		kfree(pdd);
181 	}
182 
183 	kfd_event_free_process(p);
184 
185 	kfd_pasid_free(p->pasid);
186 	kfd_free_process_doorbells(p);
187 
188 	mutex_unlock(&p->mutex);
189 
190 	mutex_destroy(&p->mutex);
191 
192 	kfree(p->queues);
193 
194 	kfree(p);
195 
196 	kfree(work);
197 }
198 
199 static void kfd_process_destroy_delayed(struct rcu_head *rcu)
200 {
201 	struct kfd_process_release_work *work;
202 	struct kfd_process *p;
203 
204 	p = container_of(rcu, struct kfd_process, rcu);
205 	WARN_ON(atomic_read(&p->mm->mm_count) <= 0);
206 
207 	mmdrop(p->mm);
208 
209 	work = kmalloc(sizeof(struct kfd_process_release_work), GFP_ATOMIC);
210 
211 	if (work) {
212 		INIT_WORK((struct work_struct *) work, kfd_process_wq_release);
213 		work->p = p;
214 		queue_work(kfd_process_wq, (struct work_struct *) work);
215 	}
216 }
217 
218 static void kfd_process_notifier_release(struct mmu_notifier *mn,
219 					struct mm_struct *mm)
220 {
221 	struct kfd_process *p;
222 	struct kfd_process_device *pdd = NULL;
223 
224 	/*
225 	 * The kfd_process structure can not be free because the
226 	 * mmu_notifier srcu is read locked
227 	 */
228 	p = container_of(mn, struct kfd_process, mmu_notifier);
229 	if (WARN_ON(p->mm != mm))
230 		return;
231 
232 	mutex_lock(&kfd_processes_mutex);
233 	hash_del_rcu(&p->kfd_processes);
234 	mutex_unlock(&kfd_processes_mutex);
235 	synchronize_srcu(&kfd_processes_srcu);
236 
237 	mutex_lock(&p->mutex);
238 
239 	/* In case our notifier is called before IOMMU notifier */
240 	pqm_uninit(&p->pqm);
241 
242 	/* Iterate over all process device data structure and check
243 	 * if we should delete debug managers and reset all wavefronts
244 	 */
245 	list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
246 		if ((pdd->dev->dbgmgr) &&
247 				(pdd->dev->dbgmgr->pasid == p->pasid))
248 			kfd_dbgmgr_destroy(pdd->dev->dbgmgr);
249 
250 		if (pdd->reset_wavefronts) {
251 			pr_warn("Resetting all wave fronts\n");
252 			dbgdev_wave_reset_wavefronts(pdd->dev, p);
253 			pdd->reset_wavefronts = false;
254 		}
255 	}
256 
257 	mutex_unlock(&p->mutex);
258 
259 	/*
260 	 * Because we drop mm_count inside kfd_process_destroy_delayed
261 	 * and because the mmu_notifier_unregister function also drop
262 	 * mm_count we need to take an extra count here.
263 	 */
264 	mmgrab(p->mm);
265 	mmu_notifier_unregister_no_release(&p->mmu_notifier, p->mm);
266 	mmu_notifier_call_srcu(&p->rcu, &kfd_process_destroy_delayed);
267 }
268 
269 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
270 	.release = kfd_process_notifier_release,
271 };
272 
273 static struct kfd_process *create_process(const struct task_struct *thread)
274 {
275 	struct kfd_process *process;
276 	int err = -ENOMEM;
277 
278 	process = kzalloc(sizeof(*process), GFP_KERNEL);
279 
280 	if (!process)
281 		goto err_alloc_process;
282 
283 	process->queues = kmalloc_array(INITIAL_QUEUE_ARRAY_SIZE,
284 					sizeof(process->queues[0]), GFP_KERNEL);
285 	if (!process->queues)
286 		goto err_alloc_queues;
287 
288 	process->pasid = kfd_pasid_alloc();
289 	if (process->pasid == 0)
290 		goto err_alloc_pasid;
291 
292 	if (kfd_alloc_process_doorbells(process) < 0)
293 		goto err_alloc_doorbells;
294 
295 	mutex_init(&process->mutex);
296 
297 	process->mm = thread->mm;
298 
299 	/* register notifier */
300 	process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops;
301 	err = __mmu_notifier_register(&process->mmu_notifier, process->mm);
302 	if (err)
303 		goto err_mmu_notifier;
304 
305 	hash_add_rcu(kfd_processes_table, &process->kfd_processes,
306 			(uintptr_t)process->mm);
307 
308 	process->lead_thread = thread->group_leader;
309 
310 	process->queue_array_size = INITIAL_QUEUE_ARRAY_SIZE;
311 
312 	INIT_LIST_HEAD(&process->per_device_data);
313 
314 	kfd_event_init_process(process);
315 
316 	err = pqm_init(&process->pqm, process);
317 	if (err != 0)
318 		goto err_process_pqm_init;
319 
320 	/* init process apertures*/
321 	process->is_32bit_user_mode = in_compat_syscall();
322 	err = kfd_init_apertures(process);
323 	if (err != 0)
324 		goto err_init_apertures;
325 
326 	return process;
327 
328 err_init_apertures:
329 	pqm_uninit(&process->pqm);
330 err_process_pqm_init:
331 	hash_del_rcu(&process->kfd_processes);
332 	synchronize_rcu();
333 	mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm);
334 err_mmu_notifier:
335 	mutex_destroy(&process->mutex);
336 	kfd_free_process_doorbells(process);
337 err_alloc_doorbells:
338 	kfd_pasid_free(process->pasid);
339 err_alloc_pasid:
340 	kfree(process->queues);
341 err_alloc_queues:
342 	kfree(process);
343 err_alloc_process:
344 	return ERR_PTR(err);
345 }
346 
347 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
348 							struct kfd_process *p)
349 {
350 	struct kfd_process_device *pdd = NULL;
351 
352 	list_for_each_entry(pdd, &p->per_device_data, per_device_list)
353 		if (pdd->dev == dev)
354 			break;
355 
356 	return pdd;
357 }
358 
359 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
360 							struct kfd_process *p)
361 {
362 	struct kfd_process_device *pdd = NULL;
363 
364 	pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
365 	if (pdd != NULL) {
366 		pdd->dev = dev;
367 		INIT_LIST_HEAD(&pdd->qpd.queues_list);
368 		INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
369 		pdd->qpd.dqm = dev->dqm;
370 		pdd->reset_wavefronts = false;
371 		list_add(&pdd->per_device_list, &p->per_device_data);
372 	}
373 
374 	return pdd;
375 }
376 
377 /*
378  * Direct the IOMMU to bind the process (specifically the pasid->mm)
379  * to the device.
380  * Unbinding occurs when the process dies or the device is removed.
381  *
382  * Assumes that the process lock is held.
383  */
384 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
385 							struct kfd_process *p)
386 {
387 	struct kfd_process_device *pdd;
388 	int err;
389 
390 	pdd = kfd_get_process_device_data(dev, p);
391 	if (!pdd) {
392 		pr_err("Process device data doesn't exist\n");
393 		return ERR_PTR(-ENOMEM);
394 	}
395 
396 	if (pdd->bound)
397 		return pdd;
398 
399 	err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
400 	if (err < 0)
401 		return ERR_PTR(err);
402 
403 	pdd->bound = true;
404 
405 	return pdd;
406 }
407 
408 void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
409 {
410 	struct kfd_process *p;
411 	struct kfd_process_device *pdd;
412 
413 	/*
414 	 * Look for the process that matches the pasid. If there is no such
415 	 * process, we either released it in amdkfd's own notifier, or there
416 	 * is a bug. Unfortunately, there is no way to tell...
417 	 */
418 	p = kfd_lookup_process_by_pasid(pasid);
419 	if (!p)
420 		return;
421 
422 	pr_debug("Unbinding process %d from IOMMU\n", pasid);
423 
424 	if ((dev->dbgmgr) && (dev->dbgmgr->pasid == p->pasid))
425 		kfd_dbgmgr_destroy(dev->dbgmgr);
426 
427 	pqm_uninit(&p->pqm);
428 
429 	pdd = kfd_get_process_device_data(dev, p);
430 
431 	if (!pdd) {
432 		mutex_unlock(&p->mutex);
433 		return;
434 	}
435 
436 	if (pdd->reset_wavefronts) {
437 		dbgdev_wave_reset_wavefronts(pdd->dev, p);
438 		pdd->reset_wavefronts = false;
439 	}
440 
441 	/*
442 	 * Just mark pdd as unbound, because we still need it
443 	 * to call amd_iommu_unbind_pasid() in when the
444 	 * process exits.
445 	 * We don't call amd_iommu_unbind_pasid() here
446 	 * because the IOMMU called us.
447 	 */
448 	pdd->bound = false;
449 
450 	mutex_unlock(&p->mutex);
451 }
452 
453 struct kfd_process_device *kfd_get_first_process_device_data(
454 						struct kfd_process *p)
455 {
456 	return list_first_entry(&p->per_device_data,
457 				struct kfd_process_device,
458 				per_device_list);
459 }
460 
461 struct kfd_process_device *kfd_get_next_process_device_data(
462 						struct kfd_process *p,
463 						struct kfd_process_device *pdd)
464 {
465 	if (list_is_last(&pdd->per_device_list, &p->per_device_data))
466 		return NULL;
467 	return list_next_entry(pdd, per_device_list);
468 }
469 
470 bool kfd_has_process_device_data(struct kfd_process *p)
471 {
472 	return !(list_empty(&p->per_device_data));
473 }
474 
475 /* This returns with process->mutex locked. */
476 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
477 {
478 	struct kfd_process *p;
479 	unsigned int temp;
480 
481 	int idx = srcu_read_lock(&kfd_processes_srcu);
482 
483 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
484 		if (p->pasid == pasid) {
485 			mutex_lock(&p->mutex);
486 			break;
487 		}
488 	}
489 
490 	srcu_read_unlock(&kfd_processes_srcu, idx);
491 
492 	return p;
493 }
494