xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_process.c (revision e0bf6c5ca2d3281f231c5f0c9bf145e9513644de)
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/slab.h>
27 #include <linux/amd-iommu.h>
28 #include <linux/notifier.h>
29 #include <linux/compat.h>
30 
31 struct mm_struct;
32 
33 #include "kfd_priv.h"
34 
35 /*
36  * Initial size for the array of queues.
37  * The allocated size is doubled each time
38  * it is exceeded up to MAX_PROCESS_QUEUES.
39  */
40 #define INITIAL_QUEUE_ARRAY_SIZE 16
41 
42 /*
43  * List of struct kfd_process (field kfd_process).
44  * Unique/indexed by mm_struct*
45  */
46 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
47 static DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
48 static DEFINE_MUTEX(kfd_processes_mutex);
49 
50 DEFINE_STATIC_SRCU(kfd_processes_srcu);
51 
52 static struct workqueue_struct *kfd_process_wq;
53 
54 struct kfd_process_release_work {
55 	struct work_struct kfd_work;
56 	struct kfd_process *p;
57 };
58 
59 static struct kfd_process *find_process(const struct task_struct *thread);
60 static struct kfd_process *create_process(const struct task_struct *thread);
61 
62 void kfd_process_create_wq(void)
63 {
64 	if (!kfd_process_wq)
65 		kfd_process_wq = create_workqueue("kfd_process_wq");
66 }
67 
68 void kfd_process_destroy_wq(void)
69 {
70 	if (kfd_process_wq) {
71 		flush_workqueue(kfd_process_wq);
72 		destroy_workqueue(kfd_process_wq);
73 		kfd_process_wq = NULL;
74 	}
75 }
76 
77 struct kfd_process *kfd_create_process(const struct task_struct *thread)
78 {
79 	struct kfd_process *process;
80 
81 	BUG_ON(!kfd_process_wq);
82 
83 	if (thread->mm == NULL)
84 		return ERR_PTR(-EINVAL);
85 
86 	/* Only the pthreads threading model is supported. */
87 	if (thread->group_leader->mm != thread->mm)
88 		return ERR_PTR(-EINVAL);
89 
90 	/* Take mmap_sem because we call __mmu_notifier_register inside */
91 	down_write(&thread->mm->mmap_sem);
92 
93 	/*
94 	 * take kfd processes mutex before starting of process creation
95 	 * so there won't be a case where two threads of the same process
96 	 * create two kfd_process structures
97 	 */
98 	mutex_lock(&kfd_processes_mutex);
99 
100 	/* A prior open of /dev/kfd could have already created the process. */
101 	process = find_process(thread);
102 	if (process)
103 		pr_debug("kfd: process already found\n");
104 
105 	if (!process)
106 		process = create_process(thread);
107 
108 	mutex_unlock(&kfd_processes_mutex);
109 
110 	up_write(&thread->mm->mmap_sem);
111 
112 	return process;
113 }
114 
115 struct kfd_process *kfd_get_process(const struct task_struct *thread)
116 {
117 	struct kfd_process *process;
118 
119 	if (thread->mm == NULL)
120 		return ERR_PTR(-EINVAL);
121 
122 	/* Only the pthreads threading model is supported. */
123 	if (thread->group_leader->mm != thread->mm)
124 		return ERR_PTR(-EINVAL);
125 
126 	process = find_process(thread);
127 
128 	return process;
129 }
130 
131 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
132 {
133 	struct kfd_process *process;
134 
135 	hash_for_each_possible_rcu(kfd_processes_table, process,
136 					kfd_processes, (uintptr_t)mm)
137 		if (process->mm == mm)
138 			return process;
139 
140 	return NULL;
141 }
142 
143 static struct kfd_process *find_process(const struct task_struct *thread)
144 {
145 	struct kfd_process *p;
146 	int idx;
147 
148 	idx = srcu_read_lock(&kfd_processes_srcu);
149 	p = find_process_by_mm(thread->mm);
150 	srcu_read_unlock(&kfd_processes_srcu, idx);
151 
152 	return p;
153 }
154 
155 static void kfd_process_wq_release(struct work_struct *work)
156 {
157 	struct kfd_process_release_work *my_work;
158 	struct kfd_process_device *pdd, *temp;
159 	struct kfd_process *p;
160 
161 	my_work = (struct kfd_process_release_work *) work;
162 
163 	p = my_work->p;
164 
165 	mutex_lock(&p->mutex);
166 
167 	list_for_each_entry_safe(pdd, temp, &p->per_device_data,
168 							per_device_list) {
169 		amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
170 		list_del(&pdd->per_device_list);
171 
172 		kfree(pdd);
173 	}
174 
175 	kfd_pasid_free(p->pasid);
176 
177 	mutex_unlock(&p->mutex);
178 
179 	mutex_destroy(&p->mutex);
180 
181 	kfree(p->queues);
182 
183 	kfree(p);
184 
185 	kfree((void *)work);
186 }
187 
188 static void kfd_process_destroy_delayed(struct rcu_head *rcu)
189 {
190 	struct kfd_process_release_work *work;
191 	struct kfd_process *p;
192 
193 	BUG_ON(!kfd_process_wq);
194 
195 	p = container_of(rcu, struct kfd_process, rcu);
196 	BUG_ON(atomic_read(&p->mm->mm_count) <= 0);
197 
198 	mmdrop(p->mm);
199 
200 	work = (struct kfd_process_release_work *)
201 		kmalloc(sizeof(struct kfd_process_release_work), GFP_ATOMIC);
202 
203 	if (work) {
204 		INIT_WORK((struct work_struct *) work, kfd_process_wq_release);
205 		work->p = p;
206 		queue_work(kfd_process_wq, (struct work_struct *) work);
207 	}
208 }
209 
210 static void kfd_process_notifier_release(struct mmu_notifier *mn,
211 					struct mm_struct *mm)
212 {
213 	struct kfd_process *p;
214 
215 	/*
216 	 * The kfd_process structure can not be free because the
217 	 * mmu_notifier srcu is read locked
218 	 */
219 	p = container_of(mn, struct kfd_process, mmu_notifier);
220 	BUG_ON(p->mm != mm);
221 
222 	mutex_lock(&kfd_processes_mutex);
223 	hash_del_rcu(&p->kfd_processes);
224 	mutex_unlock(&kfd_processes_mutex);
225 	synchronize_srcu(&kfd_processes_srcu);
226 
227 	mutex_lock(&p->mutex);
228 
229 	/* In case our notifier is called before IOMMU notifier */
230 	pqm_uninit(&p->pqm);
231 
232 	mutex_unlock(&p->mutex);
233 
234 	/*
235 	 * Because we drop mm_count inside kfd_process_destroy_delayed
236 	 * and because the mmu_notifier_unregister function also drop
237 	 * mm_count we need to take an extra count here.
238 	 */
239 	atomic_inc(&p->mm->mm_count);
240 	mmu_notifier_unregister_no_release(&p->mmu_notifier, p->mm);
241 	mmu_notifier_call_srcu(&p->rcu, &kfd_process_destroy_delayed);
242 }
243 
244 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
245 	.release = kfd_process_notifier_release,
246 };
247 
248 static struct kfd_process *create_process(const struct task_struct *thread)
249 {
250 	struct kfd_process *process;
251 	int err = -ENOMEM;
252 
253 	process = kzalloc(sizeof(*process), GFP_KERNEL);
254 
255 	if (!process)
256 		goto err_alloc_process;
257 
258 	process->queues = kmalloc_array(INITIAL_QUEUE_ARRAY_SIZE,
259 					sizeof(process->queues[0]), GFP_KERNEL);
260 	if (!process->queues)
261 		goto err_alloc_queues;
262 
263 	process->pasid = kfd_pasid_alloc();
264 	if (process->pasid == 0)
265 		goto err_alloc_pasid;
266 
267 	mutex_init(&process->mutex);
268 
269 	process->mm = thread->mm;
270 
271 	/* register notifier */
272 	process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops;
273 	err = __mmu_notifier_register(&process->mmu_notifier, process->mm);
274 	if (err)
275 		goto err_mmu_notifier;
276 
277 	hash_add_rcu(kfd_processes_table, &process->kfd_processes,
278 			(uintptr_t)process->mm);
279 
280 	process->lead_thread = thread->group_leader;
281 
282 	process->queue_array_size = INITIAL_QUEUE_ARRAY_SIZE;
283 
284 	INIT_LIST_HEAD(&process->per_device_data);
285 
286 	err = pqm_init(&process->pqm, process);
287 	if (err != 0)
288 		goto err_process_pqm_init;
289 
290 	/* init process apertures*/
291 	process->is_32bit_user_mode = is_compat_task();
292 	if (kfd_init_apertures(process) != 0)
293 		goto err_init_apretures;
294 
295 	return process;
296 
297 err_init_apretures:
298 	pqm_uninit(&process->pqm);
299 err_process_pqm_init:
300 	hash_del_rcu(&process->kfd_processes);
301 	synchronize_rcu();
302 	mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm);
303 err_mmu_notifier:
304 	kfd_pasid_free(process->pasid);
305 err_alloc_pasid:
306 	kfree(process->queues);
307 err_alloc_queues:
308 	kfree(process);
309 err_alloc_process:
310 	return ERR_PTR(err);
311 }
312 
313 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
314 							struct kfd_process *p)
315 {
316 	struct kfd_process_device *pdd = NULL;
317 
318 	list_for_each_entry(pdd, &p->per_device_data, per_device_list)
319 		if (pdd->dev == dev)
320 			break;
321 
322 	return pdd;
323 }
324 
325 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
326 							struct kfd_process *p)
327 {
328 	struct kfd_process_device *pdd = NULL;
329 
330 	pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
331 	if (pdd != NULL) {
332 		pdd->dev = dev;
333 		INIT_LIST_HEAD(&pdd->qpd.queues_list);
334 		INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
335 		pdd->qpd.dqm = dev->dqm;
336 		list_add(&pdd->per_device_list, &p->per_device_data);
337 	}
338 
339 	return pdd;
340 }
341 
342 /*
343  * Direct the IOMMU to bind the process (specifically the pasid->mm)
344  * to the device.
345  * Unbinding occurs when the process dies or the device is removed.
346  *
347  * Assumes that the process lock is held.
348  */
349 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
350 							struct kfd_process *p)
351 {
352 	struct kfd_process_device *pdd;
353 	int err;
354 
355 	pdd = kfd_get_process_device_data(dev, p);
356 	if (!pdd) {
357 		pr_err("Process device data doesn't exist\n");
358 		return ERR_PTR(-ENOMEM);
359 	}
360 
361 	if (pdd->bound)
362 		return pdd;
363 
364 	err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
365 	if (err < 0)
366 		return ERR_PTR(err);
367 
368 	pdd->bound = true;
369 
370 	return pdd;
371 }
372 
373 void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
374 {
375 	struct kfd_process *p;
376 	struct kfd_process_device *pdd;
377 	int idx, i;
378 
379 	BUG_ON(dev == NULL);
380 
381 	idx = srcu_read_lock(&kfd_processes_srcu);
382 
383 	hash_for_each_rcu(kfd_processes_table, i, p, kfd_processes)
384 		if (p->pasid == pasid)
385 			break;
386 
387 	srcu_read_unlock(&kfd_processes_srcu, idx);
388 
389 	BUG_ON(p->pasid != pasid);
390 
391 	mutex_lock(&p->mutex);
392 
393 	pqm_uninit(&p->pqm);
394 
395 	pdd = kfd_get_process_device_data(dev, p);
396 
397 	/*
398 	 * Just mark pdd as unbound, because we still need it to call
399 	 * amd_iommu_unbind_pasid() in when the process exits.
400 	 * We don't call amd_iommu_unbind_pasid() here
401 	 * because the IOMMU called us.
402 	 */
403 	if (pdd)
404 		pdd->bound = false;
405 
406 	mutex_unlock(&p->mutex);
407 }
408 
409 struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p)
410 {
411 	return list_first_entry(&p->per_device_data,
412 				struct kfd_process_device,
413 				per_device_list);
414 }
415 
416 struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p,
417 						struct kfd_process_device *pdd)
418 {
419 	if (list_is_last(&pdd->per_device_list, &p->per_device_data))
420 		return NULL;
421 	return list_next_entry(pdd, per_device_list);
422 }
423 
424 bool kfd_has_process_device_data(struct kfd_process *p)
425 {
426 	return !(list_empty(&p->per_device_data));
427 }
428