2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
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>
34 #include "kfd_dbgmgr.h"
37 * Initial size for the array of queues.
38 * The allocated size is doubled each time
39 * it is exceeded up to MAX_PROCESS_QUEUES.
41 #define INITIAL_QUEUE_ARRAY_SIZE 16
44 * List of struct kfd_process (field kfd_process).
45 * Unique/indexed by mm_struct*
47 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
48 static DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
49 static DEFINE_MUTEX(kfd_processes_mutex);
51 DEFINE_STATIC_SRCU(kfd_processes_srcu);
53 static struct workqueue_struct *kfd_process_wq;
55 struct kfd_process_release_work {
56 struct work_struct kfd_work;
57 struct kfd_process *p;
60 static struct kfd_process *find_process(const struct task_struct *thread);
61 static struct kfd_process *create_process(const struct task_struct *thread);
63 void kfd_process_create_wq(void)
66 kfd_process_wq = create_workqueue("kfd_process_wq");
69 void kfd_process_destroy_wq(void)
72 flush_workqueue(kfd_process_wq);
73 destroy_workqueue(kfd_process_wq);
74 kfd_process_wq = NULL;
78 struct kfd_process *kfd_create_process(const struct task_struct *thread)
80 struct kfd_process *process;
82 BUG_ON(!kfd_process_wq);
84 if (thread->mm == NULL)
85 return ERR_PTR(-EINVAL);
87 /* Only the pthreads threading model is supported. */
88 if (thread->group_leader->mm != thread->mm)
89 return ERR_PTR(-EINVAL);
91 /* Take mmap_sem because we call __mmu_notifier_register inside */
92 down_write(&thread->mm->mmap_sem);
95 * take kfd processes mutex before starting of process creation
96 * so there won't be a case where two threads of the same process
97 * create two kfd_process structures
99 mutex_lock(&kfd_processes_mutex);
101 /* A prior open of /dev/kfd could have already created the process. */
102 process = find_process(thread);
104 pr_debug("kfd: process already found\n");
107 process = create_process(thread);
109 mutex_unlock(&kfd_processes_mutex);
111 up_write(&thread->mm->mmap_sem);
116 struct kfd_process *kfd_get_process(const struct task_struct *thread)
118 struct kfd_process *process;
120 if (thread->mm == NULL)
121 return ERR_PTR(-EINVAL);
123 /* Only the pthreads threading model is supported. */
124 if (thread->group_leader->mm != thread->mm)
125 return ERR_PTR(-EINVAL);
127 process = find_process(thread);
132 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
134 struct kfd_process *process;
136 hash_for_each_possible_rcu(kfd_processes_table, process,
137 kfd_processes, (uintptr_t)mm)
138 if (process->mm == mm)
144 static struct kfd_process *find_process(const struct task_struct *thread)
146 struct kfd_process *p;
149 idx = srcu_read_lock(&kfd_processes_srcu);
150 p = find_process_by_mm(thread->mm);
151 srcu_read_unlock(&kfd_processes_srcu, idx);
156 static void kfd_process_wq_release(struct work_struct *work)
158 struct kfd_process_release_work *my_work;
159 struct kfd_process_device *pdd, *temp;
160 struct kfd_process *p;
162 my_work = (struct kfd_process_release_work *) work;
166 pr_debug("Releasing process (pasid %d) in workqueue\n",
169 mutex_lock(&p->mutex);
171 list_for_each_entry_safe(pdd, temp, &p->per_device_data,
173 pr_debug("Releasing pdd (topology id %d) for process (pasid %d) in workqueue\n",
174 pdd->dev->id, p->pasid);
176 if (pdd->reset_wavefronts)
177 dbgdev_wave_reset_wavefronts(pdd->dev, p);
179 amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
180 list_del(&pdd->per_device_list);
185 kfd_event_free_process(p);
187 kfd_pasid_free(p->pasid);
189 mutex_unlock(&p->mutex);
191 mutex_destroy(&p->mutex);
200 static void kfd_process_destroy_delayed(struct rcu_head *rcu)
202 struct kfd_process_release_work *work;
203 struct kfd_process *p;
205 BUG_ON(!kfd_process_wq);
207 p = container_of(rcu, struct kfd_process, rcu);
208 BUG_ON(atomic_read(&p->mm->mm_count) <= 0);
212 work = kmalloc(sizeof(struct kfd_process_release_work), GFP_ATOMIC);
215 INIT_WORK((struct work_struct *) work, kfd_process_wq_release);
217 queue_work(kfd_process_wq, (struct work_struct *) work);
221 static void kfd_process_notifier_release(struct mmu_notifier *mn,
222 struct mm_struct *mm)
224 struct kfd_process *p;
225 struct kfd_process_device *pdd = NULL;
228 * The kfd_process structure can not be free because the
229 * mmu_notifier srcu is read locked
231 p = container_of(mn, struct kfd_process, mmu_notifier);
234 mutex_lock(&kfd_processes_mutex);
235 hash_del_rcu(&p->kfd_processes);
236 mutex_unlock(&kfd_processes_mutex);
237 synchronize_srcu(&kfd_processes_srcu);
239 mutex_lock(&p->mutex);
241 /* In case our notifier is called before IOMMU notifier */
244 /* Iterate over all process device data structure and check
245 * if we should delete debug managers and reset all wavefronts
247 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
248 if ((pdd->dev->dbgmgr) &&
249 (pdd->dev->dbgmgr->pasid == p->pasid))
250 kfd_dbgmgr_destroy(pdd->dev->dbgmgr);
252 if (pdd->reset_wavefronts) {
253 pr_warn("amdkfd: Resetting all wave fronts\n");
254 dbgdev_wave_reset_wavefronts(pdd->dev, p);
255 pdd->reset_wavefronts = false;
259 mutex_unlock(&p->mutex);
262 * Because we drop mm_count inside kfd_process_destroy_delayed
263 * and because the mmu_notifier_unregister function also drop
264 * mm_count we need to take an extra count here.
266 atomic_inc(&p->mm->mm_count);
267 mmu_notifier_unregister_no_release(&p->mmu_notifier, p->mm);
268 mmu_notifier_call_srcu(&p->rcu, &kfd_process_destroy_delayed);
271 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
272 .release = kfd_process_notifier_release,
275 static struct kfd_process *create_process(const struct task_struct *thread)
277 struct kfd_process *process;
280 process = kzalloc(sizeof(*process), GFP_KERNEL);
283 goto err_alloc_process;
285 process->queues = kmalloc_array(INITIAL_QUEUE_ARRAY_SIZE,
286 sizeof(process->queues[0]), GFP_KERNEL);
287 if (!process->queues)
288 goto err_alloc_queues;
290 process->pasid = kfd_pasid_alloc();
291 if (process->pasid == 0)
292 goto err_alloc_pasid;
294 mutex_init(&process->mutex);
296 process->mm = thread->mm;
298 /* register notifier */
299 process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops;
300 err = __mmu_notifier_register(&process->mmu_notifier, process->mm);
302 goto err_mmu_notifier;
304 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
305 (uintptr_t)process->mm);
307 process->lead_thread = thread->group_leader;
309 process->queue_array_size = INITIAL_QUEUE_ARRAY_SIZE;
311 INIT_LIST_HEAD(&process->per_device_data);
313 kfd_event_init_process(process);
315 err = pqm_init(&process->pqm, process);
317 goto err_process_pqm_init;
319 /* init process apertures*/
320 process->is_32bit_user_mode = in_compat_syscall();
321 if (kfd_init_apertures(process) != 0)
322 goto err_init_apretures;
327 pqm_uninit(&process->pqm);
328 err_process_pqm_init:
329 hash_del_rcu(&process->kfd_processes);
331 mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm);
333 kfd_pasid_free(process->pasid);
335 kfree(process->queues);
342 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
343 struct kfd_process *p)
345 struct kfd_process_device *pdd = NULL;
347 list_for_each_entry(pdd, &p->per_device_data, per_device_list)
354 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
355 struct kfd_process *p)
357 struct kfd_process_device *pdd = NULL;
359 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
362 INIT_LIST_HEAD(&pdd->qpd.queues_list);
363 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
364 pdd->qpd.dqm = dev->dqm;
365 pdd->reset_wavefronts = false;
366 list_add(&pdd->per_device_list, &p->per_device_data);
373 * Direct the IOMMU to bind the process (specifically the pasid->mm)
375 * Unbinding occurs when the process dies or the device is removed.
377 * Assumes that the process lock is held.
379 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
380 struct kfd_process *p)
382 struct kfd_process_device *pdd;
385 pdd = kfd_get_process_device_data(dev, p);
387 pr_err("Process device data doesn't exist\n");
388 return ERR_PTR(-ENOMEM);
394 err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
403 void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
405 struct kfd_process *p;
406 struct kfd_process_device *pdd;
411 idx = srcu_read_lock(&kfd_processes_srcu);
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...
418 hash_for_each_rcu(kfd_processes_table, i, p, kfd_processes)
419 if (p->pasid == pasid) {
421 srcu_read_unlock(&kfd_processes_srcu, idx);
423 pr_debug("Unbinding process %d from IOMMU\n", pasid);
425 mutex_lock(&p->mutex);
427 if ((dev->dbgmgr) && (dev->dbgmgr->pasid == p->pasid))
428 kfd_dbgmgr_destroy(dev->dbgmgr);
432 pdd = kfd_get_process_device_data(dev, p);
435 mutex_unlock(&p->mutex);
439 if (pdd->reset_wavefronts) {
440 dbgdev_wave_reset_wavefronts(pdd->dev, p);
441 pdd->reset_wavefronts = false;
445 * Just mark pdd as unbound, because we still need it
446 * to call amd_iommu_unbind_pasid() in when the
448 * We don't call amd_iommu_unbind_pasid() here
449 * because the IOMMU called us.
453 mutex_unlock(&p->mutex);
458 srcu_read_unlock(&kfd_processes_srcu, idx);
461 struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p)
463 return list_first_entry(&p->per_device_data,
464 struct kfd_process_device,
468 struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p,
469 struct kfd_process_device *pdd)
471 if (list_is_last(&pdd->per_device_list, &p->per_device_data))
473 return list_next_entry(pdd, per_device_list);
476 bool kfd_has_process_device_data(struct kfd_process *p)
478 return !(list_empty(&p->per_device_data));
481 /* This returns with process->mutex locked. */
482 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
484 struct kfd_process *p;
487 int idx = srcu_read_lock(&kfd_processes_srcu);
489 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
490 if (p->pasid == pasid) {
491 mutex_lock(&p->mutex);
496 srcu_read_unlock(&kfd_processes_srcu, idx);