Commit | Line | Data |
---|---|---|
d87f36a0 | 1 | // SPDX-License-Identifier: GPL-2.0 OR MIT |
19f6d2a6 | 2 | /* |
d87f36a0 | 3 | * Copyright 2014-2022 Advanced Micro Devices, Inc. |
19f6d2a6 OG |
4 | * |
5 | * Permission is hereby granted, free of charge, to any person obtaining a | |
6 | * copy of this software and associated documentation files (the "Software"), | |
7 | * to deal in the Software without restriction, including without limitation | |
8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
9 | * and/or sell copies of the Software, and to permit persons to whom the | |
10 | * Software is furnished to do so, subject to the following conditions: | |
11 | * | |
12 | * The above copyright notice and this permission notice shall be included in | |
13 | * all copies or substantial portions of the Software. | |
14 | * | |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
18 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |
19 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |
20 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |
21 | * OTHER DEALINGS IN THE SOFTWARE. | |
22 | */ | |
23 | ||
24 | #include <linux/mutex.h> | |
25 | #include <linux/log2.h> | |
26 | #include <linux/sched.h> | |
6e84f315 | 27 | #include <linux/sched/mm.h> |
c7b1243e | 28 | #include <linux/sched/task.h> |
32cb59f3 | 29 | #include <linux/mmu_context.h> |
19f6d2a6 OG |
30 | #include <linux/slab.h> |
31 | #include <linux/notifier.h> | |
dd59239a | 32 | #include <linux/compat.h> |
373d7080 | 33 | #include <linux/mman.h> |
b84394e2 | 34 | #include <linux/file.h> |
9593f4d6 | 35 | #include <linux/pm_runtime.h> |
5b87245f | 36 | #include "amdgpu_amdkfd.h" |
ffa02269 | 37 | #include "amdgpu.h" |
dd59239a | 38 | |
19f6d2a6 OG |
39 | struct mm_struct; |
40 | ||
41 | #include "kfd_priv.h" | |
403575c4 | 42 | #include "kfd_device_queue_manager.h" |
42de677f | 43 | #include "kfd_svm.h" |
c7f21978 | 44 | #include "kfd_smi_events.h" |
0ab2d753 | 45 | #include "kfd_debug.h" |
19f6d2a6 | 46 | |
19f6d2a6 OG |
47 | /* |
48 | * List of struct kfd_process (field kfd_process). | |
49 | * Unique/indexed by mm_struct* | |
50 | */ | |
64d1c3a4 | 51 | DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); |
fe1f05df | 52 | DEFINE_MUTEX(kfd_processes_mutex); |
19f6d2a6 | 53 | |
64d1c3a4 | 54 | DEFINE_SRCU(kfd_processes_srcu); |
19f6d2a6 | 55 | |
1679ae8f | 56 | /* For process termination handling */ |
19f6d2a6 OG |
57 | static struct workqueue_struct *kfd_process_wq; |
58 | ||
1679ae8f FK |
59 | /* Ordered, single-threaded workqueue for restoring evicted |
60 | * processes. Restoring multiple processes concurrently under memory | |
61 | * pressure can lead to processes blocking each other from validating | |
62 | * their BOs and result in a live-lock situation where processes | |
63 | * remain evicted indefinitely. | |
64 | */ | |
65 | static struct workqueue_struct *kfd_restore_wq; | |
66 | ||
011bbb03 RB |
67 | static struct kfd_process *find_process(const struct task_struct *thread, |
68 | bool ref); | |
abb208a8 | 69 | static void kfd_process_ref_release(struct kref *ref); |
0029cab3 | 70 | static struct kfd_process *create_process(const struct task_struct *thread); |
373d7080 | 71 | |
26103436 FK |
72 | static void evict_process_worker(struct work_struct *work); |
73 | static void restore_process_worker(struct work_struct *work); | |
74 | ||
68df0f19 LY |
75 | static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd); |
76 | ||
de9f26bb KR |
77 | struct kfd_procfs_tree { |
78 | struct kobject *kobj; | |
79 | }; | |
80 | ||
81 | static struct kfd_procfs_tree procfs; | |
82 | ||
32cb59f3 MJ |
83 | /* |
84 | * Structure for SDMA activity tracking | |
85 | */ | |
86 | struct kfd_sdma_activity_handler_workarea { | |
87 | struct work_struct sdma_activity_work; | |
88 | struct kfd_process_device *pdd; | |
89 | uint64_t sdma_activity_counter; | |
90 | }; | |
91 | ||
d69fd951 | 92 | struct temp_sdma_queue_list { |
818b0324 | 93 | uint64_t __user *rptr; |
d69fd951 MJ |
94 | uint64_t sdma_val; |
95 | unsigned int queue_id; | |
96 | struct list_head list; | |
97 | }; | |
98 | ||
32cb59f3 MJ |
99 | static void kfd_sdma_activity_worker(struct work_struct *work) |
100 | { | |
101 | struct kfd_sdma_activity_handler_workarea *workarea; | |
102 | struct kfd_process_device *pdd; | |
103 | uint64_t val; | |
104 | struct mm_struct *mm; | |
105 | struct queue *q; | |
106 | struct qcm_process_device *qpd; | |
107 | struct device_queue_manager *dqm; | |
108 | int ret = 0; | |
d69fd951 MJ |
109 | struct temp_sdma_queue_list sdma_q_list; |
110 | struct temp_sdma_queue_list *sdma_q, *next; | |
32cb59f3 MJ |
111 | |
112 | workarea = container_of(work, struct kfd_sdma_activity_handler_workarea, | |
113 | sdma_activity_work); | |
32cb59f3 MJ |
114 | |
115 | pdd = workarea->pdd; | |
2652bda7 CIK |
116 | if (!pdd) |
117 | return; | |
32cb59f3 MJ |
118 | dqm = pdd->dev->dqm; |
119 | qpd = &pdd->qpd; | |
2652bda7 | 120 | if (!dqm || !qpd) |
32cb59f3 | 121 | return; |
d69fd951 MJ |
122 | /* |
123 | * Total SDMA activity is current SDMA activity + past SDMA activity | |
124 | * Past SDMA count is stored in pdd. | |
125 | * To get the current activity counters for all active SDMA queues, | |
126 | * we loop over all SDMA queues and get their counts from user-space. | |
127 | * | |
128 | * We cannot call get_user() with dqm_lock held as it can cause | |
129 | * a circular lock dependency situation. To read the SDMA stats, | |
130 | * we need to do the following: | |
131 | * | |
132 | * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list, | |
133 | * with dqm_lock/dqm_unlock(). | |
134 | * 2. Call get_user() for each node in temporary list without dqm_lock. | |
135 | * Save the SDMA count for each node and also add the count to the total | |
136 | * SDMA count counter. | |
137 | * Its possible, during this step, a few SDMA queue nodes got deleted | |
138 | * from the qpd->queues_list. | |
139 | * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted. | |
140 | * If any node got deleted, its SDMA count would be captured in the sdma | |
141 | * past activity counter. So subtract the SDMA counter stored in step 2 | |
142 | * for this node from the total SDMA count. | |
143 | */ | |
144 | INIT_LIST_HEAD(&sdma_q_list.list); | |
32cb59f3 | 145 | |
d69fd951 MJ |
146 | /* |
147 | * Create the temp list of all SDMA queues | |
148 | */ | |
149 | dqm_lock(dqm); | |
150 | ||
151 | list_for_each_entry(q, &qpd->queues_list, list) { | |
152 | if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) && | |
153 | (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI)) | |
154 | continue; | |
155 | ||
156 | sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL); | |
157 | if (!sdma_q) { | |
158 | dqm_unlock(dqm); | |
159 | goto cleanup; | |
160 | } | |
161 | ||
162 | INIT_LIST_HEAD(&sdma_q->list); | |
818b0324 | 163 | sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr; |
d69fd951 MJ |
164 | sdma_q->queue_id = q->properties.queue_id; |
165 | list_add_tail(&sdma_q->list, &sdma_q_list.list); | |
32cb59f3 MJ |
166 | } |
167 | ||
d69fd951 MJ |
168 | /* |
169 | * If the temp list is empty, then no SDMA queues nodes were found in | |
170 | * qpd->queues_list. Return the past activity count as the total sdma | |
171 | * count | |
172 | */ | |
173 | if (list_empty(&sdma_q_list.list)) { | |
174 | workarea->sdma_activity_counter = pdd->sdma_past_activity_counter; | |
175 | dqm_unlock(dqm); | |
176 | return; | |
177 | } | |
32cb59f3 | 178 | |
d69fd951 | 179 | dqm_unlock(dqm); |
32cb59f3 MJ |
180 | |
181 | /* | |
d69fd951 | 182 | * Get the usage count for each SDMA queue in temp_list. |
32cb59f3 | 183 | */ |
d69fd951 MJ |
184 | mm = get_task_mm(pdd->process->lead_thread); |
185 | if (!mm) | |
186 | goto cleanup; | |
187 | ||
9555152b | 188 | kthread_use_mm(mm); |
d69fd951 MJ |
189 | |
190 | list_for_each_entry(sdma_q, &sdma_q_list.list, list) { | |
191 | val = 0; | |
192 | ret = read_sdma_queue_counter(sdma_q->rptr, &val); | |
193 | if (ret) { | |
194 | pr_debug("Failed to read SDMA queue active counter for queue id: %d", | |
195 | sdma_q->queue_id); | |
196 | } else { | |
197 | sdma_q->sdma_val = val; | |
198 | workarea->sdma_activity_counter += val; | |
199 | } | |
200 | } | |
201 | ||
9555152b | 202 | kthread_unuse_mm(mm); |
d69fd951 | 203 | mmput(mm); |
32cb59f3 MJ |
204 | |
205 | /* | |
d69fd951 MJ |
206 | * Do a second iteration over qpd_queues_list to check if any SDMA |
207 | * nodes got deleted while fetching SDMA counter. | |
32cb59f3 | 208 | */ |
d69fd951 MJ |
209 | dqm_lock(dqm); |
210 | ||
211 | workarea->sdma_activity_counter += pdd->sdma_past_activity_counter; | |
212 | ||
32cb59f3 | 213 | list_for_each_entry(q, &qpd->queues_list, list) { |
d69fd951 MJ |
214 | if (list_empty(&sdma_q_list.list)) |
215 | break; | |
216 | ||
217 | if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) && | |
218 | (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI)) | |
219 | continue; | |
220 | ||
221 | list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) { | |
818b0324 | 222 | if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) && |
d69fd951 MJ |
223 | (sdma_q->queue_id == q->properties.queue_id)) { |
224 | list_del(&sdma_q->list); | |
225 | kfree(sdma_q); | |
226 | break; | |
227 | } | |
32cb59f3 MJ |
228 | } |
229 | } | |
230 | ||
231 | dqm_unlock(dqm); | |
d69fd951 MJ |
232 | |
233 | /* | |
234 | * If temp list is not empty, it implies some queues got deleted | |
235 | * from qpd->queues_list during SDMA usage read. Subtract the SDMA | |
236 | * count for each node from the total SDMA count. | |
237 | */ | |
238 | list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) { | |
239 | workarea->sdma_activity_counter -= sdma_q->sdma_val; | |
240 | list_del(&sdma_q->list); | |
241 | kfree(sdma_q); | |
242 | } | |
243 | ||
244 | return; | |
245 | ||
246 | cleanup: | |
247 | list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) { | |
248 | list_del(&sdma_q->list); | |
249 | kfree(sdma_q); | |
250 | } | |
32cb59f3 MJ |
251 | } |
252 | ||
f2fa07b3 | 253 | /** |
bbe04dec | 254 | * kfd_get_cu_occupancy - Collect number of waves in-flight on this device |
f2fa07b3 RE |
255 | * by current process. Translates acquired wave count into number of compute units |
256 | * that are occupied. | |
257 | * | |
bbe04dec | 258 | * @attr: Handle of attribute that allows reporting of wave count. The attribute |
f2fa07b3 RE |
259 | * handle encapsulates GPU device it is associated with, thereby allowing collection |
260 | * of waves in flight, etc | |
f2fa07b3 RE |
261 | * @buffer: Handle of user provided buffer updated with wave count |
262 | * | |
263 | * Return: Number of bytes written to user buffer or an error value | |
264 | */ | |
265 | static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer) | |
266 | { | |
267 | int cu_cnt; | |
268 | int wave_cnt; | |
269 | int max_waves_per_cu; | |
8dc1db31 | 270 | struct kfd_node *dev = NULL; |
f2fa07b3 RE |
271 | struct kfd_process *proc = NULL; |
272 | struct kfd_process_device *pdd = NULL; | |
273 | ||
274 | pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy); | |
275 | dev = pdd->dev; | |
276 | if (dev->kfd2kgd->get_cu_occupancy == NULL) | |
277 | return -EINVAL; | |
278 | ||
279 | cu_cnt = 0; | |
280 | proc = pdd->process; | |
281 | if (pdd->qpd.queue_count == 0) { | |
282 | pr_debug("Gpu-Id: %d has no active queues for process %d\n", | |
283 | dev->id, proc->pasid); | |
284 | return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt); | |
285 | } | |
286 | ||
287 | /* Collect wave count from device if it supports */ | |
288 | wave_cnt = 0; | |
289 | max_waves_per_cu = 0; | |
3356c38d | 290 | dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt, |
e2069a7b | 291 | &max_waves_per_cu, 0); |
f2fa07b3 RE |
292 | |
293 | /* Translate wave count to number of compute units */ | |
294 | cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu; | |
295 | return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt); | |
296 | } | |
297 | ||
de9f26bb KR |
298 | static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr, |
299 | char *buffer) | |
300 | { | |
de9f26bb KR |
301 | if (strcmp(attr->name, "pasid") == 0) { |
302 | struct kfd_process *p = container_of(attr, struct kfd_process, | |
303 | attr_pasid); | |
d4566dee MJ |
304 | |
305 | return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid); | |
306 | } else if (strncmp(attr->name, "vram_", 5) == 0) { | |
307 | struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device, | |
308 | attr_vram); | |
32cb59f3 MJ |
309 | return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage)); |
310 | } else if (strncmp(attr->name, "sdma_", 5) == 0) { | |
311 | struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device, | |
312 | attr_sdma); | |
313 | struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler; | |
314 | ||
315 | INIT_WORK(&sdma_activity_work_handler.sdma_activity_work, | |
316 | kfd_sdma_activity_worker); | |
317 | ||
318 | sdma_activity_work_handler.pdd = pdd; | |
5960e022 | 319 | sdma_activity_work_handler.sdma_activity_counter = 0; |
32cb59f3 MJ |
320 | |
321 | schedule_work(&sdma_activity_work_handler.sdma_activity_work); | |
322 | ||
323 | flush_work(&sdma_activity_work_handler.sdma_activity_work); | |
324 | ||
325 | return snprintf(buffer, PAGE_SIZE, "%llu\n", | |
326 | (sdma_activity_work_handler.sdma_activity_counter)/ | |
327 | SDMA_ACTIVITY_DIVISOR); | |
de9f26bb KR |
328 | } else { |
329 | pr_err("Invalid attribute"); | |
330 | return -EINVAL; | |
331 | } | |
332 | ||
d4566dee | 333 | return 0; |
de9f26bb KR |
334 | } |
335 | ||
336 | static void kfd_procfs_kobj_release(struct kobject *kobj) | |
337 | { | |
338 | kfree(kobj); | |
339 | } | |
340 | ||
341 | static const struct sysfs_ops kfd_procfs_ops = { | |
342 | .show = kfd_procfs_show, | |
343 | }; | |
344 | ||
4fa01c63 | 345 | static const struct kobj_type procfs_type = { |
de9f26bb KR |
346 | .release = kfd_procfs_kobj_release, |
347 | .sysfs_ops = &kfd_procfs_ops, | |
348 | }; | |
349 | ||
350 | void kfd_procfs_init(void) | |
351 | { | |
352 | int ret = 0; | |
353 | ||
354 | procfs.kobj = kfd_alloc_struct(procfs.kobj); | |
355 | if (!procfs.kobj) | |
356 | return; | |
357 | ||
358 | ret = kobject_init_and_add(procfs.kobj, &procfs_type, | |
359 | &kfd_device->kobj, "proc"); | |
360 | if (ret) { | |
361 | pr_warn("Could not create procfs proc folder"); | |
362 | /* If we fail to create the procfs, clean up */ | |
363 | kfd_procfs_shutdown(); | |
364 | } | |
365 | } | |
366 | ||
367 | void kfd_procfs_shutdown(void) | |
368 | { | |
369 | if (procfs.kobj) { | |
370 | kobject_del(procfs.kobj); | |
371 | kobject_put(procfs.kobj); | |
372 | procfs.kobj = NULL; | |
373 | } | |
374 | } | |
19f6d2a6 | 375 | |
6d220a7e AL |
376 | static ssize_t kfd_procfs_queue_show(struct kobject *kobj, |
377 | struct attribute *attr, char *buffer) | |
378 | { | |
379 | struct queue *q = container_of(kobj, struct queue, kobj); | |
380 | ||
381 | if (!strcmp(attr->name, "size")) | |
382 | return snprintf(buffer, PAGE_SIZE, "%llu", | |
383 | q->properties.queue_size); | |
384 | else if (!strcmp(attr->name, "type")) | |
385 | return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type); | |
386 | else if (!strcmp(attr->name, "gpuid")) | |
387 | return snprintf(buffer, PAGE_SIZE, "%u", q->device->id); | |
388 | else | |
389 | pr_err("Invalid attribute"); | |
390 | ||
391 | return 0; | |
392 | } | |
393 | ||
4327bed2 PC |
394 | static ssize_t kfd_procfs_stats_show(struct kobject *kobj, |
395 | struct attribute *attr, char *buffer) | |
396 | { | |
397 | if (strcmp(attr->name, "evicted_ms") == 0) { | |
398 | struct kfd_process_device *pdd = container_of(attr, | |
399 | struct kfd_process_device, | |
400 | attr_evict); | |
401 | uint64_t evict_jiffies; | |
402 | ||
403 | evict_jiffies = atomic64_read(&pdd->evict_duration_counter); | |
404 | ||
405 | return snprintf(buffer, | |
406 | PAGE_SIZE, | |
407 | "%llu\n", | |
408 | jiffies64_to_msecs(evict_jiffies)); | |
f2fa07b3 RE |
409 | |
410 | /* Sysfs handle that gets CU occupancy is per device */ | |
411 | } else if (strcmp(attr->name, "cu_occupancy") == 0) { | |
412 | return kfd_get_cu_occupancy(attr, buffer); | |
413 | } else { | |
4327bed2 | 414 | pr_err("Invalid attribute"); |
f2fa07b3 | 415 | } |
4327bed2 PC |
416 | |
417 | return 0; | |
418 | } | |
6d220a7e | 419 | |
751580b3 PY |
420 | static ssize_t kfd_sysfs_counters_show(struct kobject *kobj, |
421 | struct attribute *attr, char *buf) | |
422 | { | |
423 | struct kfd_process_device *pdd; | |
424 | ||
425 | if (!strcmp(attr->name, "faults")) { | |
426 | pdd = container_of(attr, struct kfd_process_device, | |
427 | attr_faults); | |
428 | return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults)); | |
429 | } | |
430 | if (!strcmp(attr->name, "page_in")) { | |
431 | pdd = container_of(attr, struct kfd_process_device, | |
432 | attr_page_in); | |
433 | return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in)); | |
434 | } | |
435 | if (!strcmp(attr->name, "page_out")) { | |
436 | pdd = container_of(attr, struct kfd_process_device, | |
437 | attr_page_out); | |
438 | return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out)); | |
439 | } | |
440 | return 0; | |
441 | } | |
442 | ||
6d220a7e AL |
443 | static struct attribute attr_queue_size = { |
444 | .name = "size", | |
445 | .mode = KFD_SYSFS_FILE_MODE | |
446 | }; | |
447 | ||
448 | static struct attribute attr_queue_type = { | |
449 | .name = "type", | |
450 | .mode = KFD_SYSFS_FILE_MODE | |
451 | }; | |
452 | ||
453 | static struct attribute attr_queue_gpuid = { | |
454 | .name = "gpuid", | |
455 | .mode = KFD_SYSFS_FILE_MODE | |
456 | }; | |
457 | ||
458 | static struct attribute *procfs_queue_attrs[] = { | |
459 | &attr_queue_size, | |
460 | &attr_queue_type, | |
461 | &attr_queue_gpuid, | |
462 | NULL | |
463 | }; | |
5fea167e | 464 | ATTRIBUTE_GROUPS(procfs_queue); |
6d220a7e AL |
465 | |
466 | static const struct sysfs_ops procfs_queue_ops = { | |
467 | .show = kfd_procfs_queue_show, | |
468 | }; | |
469 | ||
4fa01c63 | 470 | static const struct kobj_type procfs_queue_type = { |
6d220a7e | 471 | .sysfs_ops = &procfs_queue_ops, |
5fea167e | 472 | .default_groups = procfs_queue_groups, |
6d220a7e AL |
473 | }; |
474 | ||
4327bed2 PC |
475 | static const struct sysfs_ops procfs_stats_ops = { |
476 | .show = kfd_procfs_stats_show, | |
477 | }; | |
478 | ||
4fa01c63 | 479 | static const struct kobj_type procfs_stats_type = { |
4327bed2 | 480 | .sysfs_ops = &procfs_stats_ops, |
dcdb4d90 | 481 | .release = kfd_procfs_kobj_release, |
4327bed2 PC |
482 | }; |
483 | ||
751580b3 PY |
484 | static const struct sysfs_ops sysfs_counters_ops = { |
485 | .show = kfd_sysfs_counters_show, | |
486 | }; | |
487 | ||
4fa01c63 | 488 | static const struct kobj_type sysfs_counters_type = { |
751580b3 PY |
489 | .sysfs_ops = &sysfs_counters_ops, |
490 | .release = kfd_procfs_kobj_release, | |
491 | }; | |
492 | ||
6d220a7e AL |
493 | int kfd_procfs_add_queue(struct queue *q) |
494 | { | |
495 | struct kfd_process *proc; | |
496 | int ret; | |
497 | ||
498 | if (!q || !q->process) | |
499 | return -EINVAL; | |
500 | proc = q->process; | |
501 | ||
502 | /* Create proc/<pid>/queues/<queue id> folder */ | |
503 | if (!proc->kobj_queues) | |
504 | return -EFAULT; | |
505 | ret = kobject_init_and_add(&q->kobj, &procfs_queue_type, | |
506 | proc->kobj_queues, "%u", q->properties.queue_id); | |
507 | if (ret < 0) { | |
508 | pr_warn("Creating proc/<pid>/queues/%u failed", | |
509 | q->properties.queue_id); | |
510 | kobject_put(&q->kobj); | |
511 | return ret; | |
512 | } | |
513 | ||
514 | return 0; | |
515 | } | |
516 | ||
75ae84c8 | 517 | static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr, |
32cb59f3 MJ |
518 | char *name) |
519 | { | |
75ae84c8 | 520 | int ret; |
32cb59f3 | 521 | |
75ae84c8 PY |
522 | if (!kobj || !attr || !name) |
523 | return; | |
32cb59f3 MJ |
524 | |
525 | attr->name = name; | |
526 | attr->mode = KFD_SYSFS_FILE_MODE; | |
527 | sysfs_attr_init(attr); | |
528 | ||
75ae84c8 PY |
529 | ret = sysfs_create_file(kobj, attr); |
530 | if (ret) | |
531 | pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret); | |
32cb59f3 MJ |
532 | } |
533 | ||
75ae84c8 | 534 | static void kfd_procfs_add_sysfs_stats(struct kfd_process *p) |
4327bed2 | 535 | { |
75ae84c8 | 536 | int ret; |
6ae27841 | 537 | int i; |
4327bed2 PC |
538 | char stats_dir_filename[MAX_SYSFS_FILENAME_LEN]; |
539 | ||
75ae84c8 PY |
540 | if (!p || !p->kobj) |
541 | return; | |
4327bed2 PC |
542 | |
543 | /* | |
544 | * Create sysfs files for each GPU: | |
545 | * - proc/<pid>/stats_<gpuid>/ | |
546 | * - proc/<pid>/stats_<gpuid>/evicted_ms | |
f2fa07b3 | 547 | * - proc/<pid>/stats_<gpuid>/cu_occupancy |
4327bed2 | 548 | */ |
6ae27841 AS |
549 | for (i = 0; i < p->n_pdds; i++) { |
550 | struct kfd_process_device *pdd = p->pdds[i]; | |
4327bed2 PC |
551 | |
552 | snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN, | |
553 | "stats_%u", pdd->dev->id); | |
75ae84c8 PY |
554 | pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats); |
555 | if (!pdd->kobj_stats) | |
556 | return; | |
4327bed2 | 557 | |
75ae84c8 PY |
558 | ret = kobject_init_and_add(pdd->kobj_stats, |
559 | &procfs_stats_type, | |
560 | p->kobj, | |
561 | stats_dir_filename); | |
4327bed2 PC |
562 | |
563 | if (ret) { | |
564 | pr_warn("Creating KFD proc/stats_%s folder failed", | |
75ae84c8 PY |
565 | stats_dir_filename); |
566 | kobject_put(pdd->kobj_stats); | |
567 | pdd->kobj_stats = NULL; | |
568 | return; | |
4327bed2 PC |
569 | } |
570 | ||
75ae84c8 PY |
571 | kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict, |
572 | "evicted_ms"); | |
f2fa07b3 | 573 | /* Add sysfs file to report compute unit occupancy */ |
75ae84c8 PY |
574 | if (pdd->dev->kfd2kgd->get_cu_occupancy) |
575 | kfd_sysfs_create_file(pdd->kobj_stats, | |
576 | &pdd->attr_cu_occupancy, | |
577 | "cu_occupancy"); | |
4327bed2 | 578 | } |
4327bed2 PC |
579 | } |
580 | ||
751580b3 PY |
581 | static void kfd_procfs_add_sysfs_counters(struct kfd_process *p) |
582 | { | |
583 | int ret = 0; | |
584 | int i; | |
585 | char counters_dir_filename[MAX_SYSFS_FILENAME_LEN]; | |
586 | ||
587 | if (!p || !p->kobj) | |
588 | return; | |
589 | ||
590 | /* | |
591 | * Create sysfs files for each GPU which supports SVM | |
592 | * - proc/<pid>/counters_<gpuid>/ | |
593 | * - proc/<pid>/counters_<gpuid>/faults | |
594 | * - proc/<pid>/counters_<gpuid>/page_in | |
595 | * - proc/<pid>/counters_<gpuid>/page_out | |
596 | */ | |
597 | for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) { | |
598 | struct kfd_process_device *pdd = p->pdds[i]; | |
599 | struct kobject *kobj_counters; | |
600 | ||
601 | snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN, | |
602 | "counters_%u", pdd->dev->id); | |
603 | kobj_counters = kfd_alloc_struct(kobj_counters); | |
604 | if (!kobj_counters) | |
605 | return; | |
606 | ||
607 | ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type, | |
608 | p->kobj, counters_dir_filename); | |
609 | if (ret) { | |
610 | pr_warn("Creating KFD proc/%s folder failed", | |
611 | counters_dir_filename); | |
612 | kobject_put(kobj_counters); | |
613 | return; | |
614 | } | |
615 | ||
616 | pdd->kobj_counters = kobj_counters; | |
617 | kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults, | |
618 | "faults"); | |
619 | kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in, | |
620 | "page_in"); | |
621 | kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out, | |
622 | "page_out"); | |
623 | } | |
624 | } | |
4327bed2 | 625 | |
75ae84c8 | 626 | static void kfd_procfs_add_sysfs_files(struct kfd_process *p) |
d4566dee | 627 | { |
6ae27841 | 628 | int i; |
d4566dee | 629 | |
75ae84c8 PY |
630 | if (!p || !p->kobj) |
631 | return; | |
d4566dee | 632 | |
32cb59f3 MJ |
633 | /* |
634 | * Create sysfs files for each GPU: | |
635 | * - proc/<pid>/vram_<gpuid> | |
636 | * - proc/<pid>/sdma_<gpuid> | |
637 | */ | |
6ae27841 AS |
638 | for (i = 0; i < p->n_pdds; i++) { |
639 | struct kfd_process_device *pdd = p->pdds[i]; | |
640 | ||
32cb59f3 | 641 | snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u", |
d4566dee | 642 | pdd->dev->id); |
75ae84c8 PY |
643 | kfd_sysfs_create_file(p->kobj, &pdd->attr_vram, |
644 | pdd->vram_filename); | |
32cb59f3 MJ |
645 | |
646 | snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u", | |
647 | pdd->dev->id); | |
75ae84c8 PY |
648 | kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma, |
649 | pdd->sdma_filename); | |
d4566dee | 650 | } |
d4566dee MJ |
651 | } |
652 | ||
6d220a7e AL |
653 | void kfd_procfs_del_queue(struct queue *q) |
654 | { | |
655 | if (!q) | |
656 | return; | |
657 | ||
658 | kobject_del(&q->kobj); | |
659 | kobject_put(&q->kobj); | |
660 | } | |
661 | ||
1679ae8f | 662 | int kfd_process_create_wq(void) |
19f6d2a6 OG |
663 | { |
664 | if (!kfd_process_wq) | |
fd320bf6 | 665 | kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0); |
1679ae8f | 666 | if (!kfd_restore_wq) |
9a1c1339 FK |
667 | kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", |
668 | WQ_FREEZABLE); | |
1679ae8f FK |
669 | |
670 | if (!kfd_process_wq || !kfd_restore_wq) { | |
671 | kfd_process_destroy_wq(); | |
672 | return -ENOMEM; | |
673 | } | |
674 | ||
675 | return 0; | |
19f6d2a6 OG |
676 | } |
677 | ||
678 | void kfd_process_destroy_wq(void) | |
679 | { | |
680 | if (kfd_process_wq) { | |
19f6d2a6 OG |
681 | destroy_workqueue(kfd_process_wq); |
682 | kfd_process_wq = NULL; | |
683 | } | |
1679ae8f FK |
684 | if (kfd_restore_wq) { |
685 | destroy_workqueue(kfd_restore_wq); | |
686 | kfd_restore_wq = NULL; | |
687 | } | |
19f6d2a6 OG |
688 | } |
689 | ||
f35751b8 | 690 | static void kfd_process_free_gpuvm(struct kgd_mem *mem, |
cb8dc232 | 691 | struct kfd_process_device *pdd, void **kptr) |
f35751b8 | 692 | { |
8dc1db31 | 693 | struct kfd_node *dev = pdd->dev; |
f35751b8 | 694 | |
cb8dc232 | 695 | if (kptr && *kptr) { |
4e2d1044 | 696 | amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem); |
cb8dc232 | 697 | *kptr = NULL; |
68df0f19 LY |
698 | } |
699 | ||
dff63da9 GS |
700 | amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv); |
701 | amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv, | |
d4ec4bdc | 702 | NULL); |
f35751b8 FK |
703 | } |
704 | ||
705 | /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process | |
706 | * This function should be only called right after the process | |
707 | * is created and when kfd_processes_mutex is still being held | |
708 | * to avoid concurrency. Because of that exclusiveness, we do | |
709 | * not need to take p->mutex. | |
710 | */ | |
711 | static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd, | |
712 | uint64_t gpu_va, uint32_t size, | |
68df0f19 | 713 | uint32_t flags, struct kgd_mem **mem, void **kptr) |
f35751b8 | 714 | { |
8dc1db31 | 715 | struct kfd_node *kdev = pdd->dev; |
f35751b8 FK |
716 | int err; |
717 | ||
dff63da9 | 718 | err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size, |
011bbb03 RB |
719 | pdd->drm_priv, mem, NULL, |
720 | flags, false); | |
f35751b8 FK |
721 | if (err) |
722 | goto err_alloc_mem; | |
723 | ||
dff63da9 | 724 | err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem, |
4d30a83c | 725 | pdd->drm_priv); |
f35751b8 FK |
726 | if (err) |
727 | goto err_map_mem; | |
728 | ||
dff63da9 | 729 | err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true); |
f35751b8 FK |
730 | if (err) { |
731 | pr_debug("Sync memory failed, wait interrupted by user signal\n"); | |
732 | goto sync_memory_failed; | |
733 | } | |
734 | ||
f35751b8 | 735 | if (kptr) { |
4e2d1044 | 736 | err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel( |
68df0f19 | 737 | (struct kgd_mem *)*mem, kptr, NULL); |
f35751b8 FK |
738 | if (err) { |
739 | pr_debug("Map GTT BO to kernel failed\n"); | |
68df0f19 | 740 | goto sync_memory_failed; |
f35751b8 FK |
741 | } |
742 | } | |
743 | ||
744 | return err; | |
745 | ||
f35751b8 | 746 | sync_memory_failed: |
dff63da9 | 747 | amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv); |
f35751b8 FK |
748 | |
749 | err_map_mem: | |
dff63da9 | 750 | amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv, |
d4ec4bdc | 751 | NULL); |
f35751b8 | 752 | err_alloc_mem: |
68df0f19 | 753 | *mem = NULL; |
f35751b8 FK |
754 | *kptr = NULL; |
755 | return err; | |
756 | } | |
757 | ||
552764b6 FK |
758 | /* kfd_process_device_reserve_ib_mem - Reserve memory inside the |
759 | * process for IB usage The memory reserved is for KFD to submit | |
760 | * IB to AMDGPU from kernel. If the memory is reserved | |
761 | * successfully, ib_kaddr will have the CPU/kernel | |
762 | * address. Check ib_kaddr before accessing the memory. | |
763 | */ | |
764 | static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd) | |
765 | { | |
766 | struct qcm_process_device *qpd = &pdd->qpd; | |
1d251d90 YZ |
767 | uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT | |
768 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE | | |
769 | KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE | | |
770 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE; | |
68df0f19 | 771 | struct kgd_mem *mem; |
552764b6 FK |
772 | void *kaddr; |
773 | int ret; | |
774 | ||
775 | if (qpd->ib_kaddr || !qpd->ib_base) | |
776 | return 0; | |
777 | ||
778 | /* ib_base is only set for dGPU */ | |
779 | ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags, | |
68df0f19 | 780 | &mem, &kaddr); |
552764b6 FK |
781 | if (ret) |
782 | return ret; | |
783 | ||
68df0f19 | 784 | qpd->ib_mem = mem; |
552764b6 FK |
785 | qpd->ib_kaddr = kaddr; |
786 | ||
787 | return 0; | |
788 | } | |
789 | ||
68df0f19 LY |
790 | static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd) |
791 | { | |
792 | struct qcm_process_device *qpd = &pdd->qpd; | |
793 | ||
794 | if (!qpd->ib_kaddr || !qpd->ib_base) | |
795 | return; | |
796 | ||
cb8dc232 | 797 | kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr); |
68df0f19 LY |
798 | } |
799 | ||
0ab2d753 | 800 | struct kfd_process *kfd_create_process(struct task_struct *thread) |
19f6d2a6 OG |
801 | { |
802 | struct kfd_process *process; | |
de9f26bb | 803 | int ret; |
19f6d2a6 | 804 | |
0ab2d753 | 805 | if (!(thread->mm && mmget_not_zero(thread->mm))) |
19f6d2a6 OG |
806 | return ERR_PTR(-EINVAL); |
807 | ||
808 | /* Only the pthreads threading model is supported. */ | |
0ab2d753 JK |
809 | if (thread->group_leader->mm != thread->mm) { |
810 | mmput(thread->mm); | |
19f6d2a6 | 811 | return ERR_PTR(-EINVAL); |
0ab2d753 | 812 | } |
19f6d2a6 | 813 | |
19f6d2a6 OG |
814 | /* |
815 | * take kfd processes mutex before starting of process creation | |
816 | * so there won't be a case where two threads of the same process | |
817 | * create two kfd_process structures | |
818 | */ | |
819 | mutex_lock(&kfd_processes_mutex); | |
820 | ||
fe1f05df | 821 | if (kfd_is_locked()) { |
fe1f05df | 822 | pr_debug("KFD is locked! Cannot create process"); |
18921b20 FK |
823 | process = ERR_PTR(-EINVAL); |
824 | goto out; | |
fe1f05df MJ |
825 | } |
826 | ||
19f6d2a6 | 827 | /* A prior open of /dev/kfd could have already created the process. */ |
011bbb03 | 828 | process = find_process(thread, false); |
de9f26bb | 829 | if (process) { |
79775b62 | 830 | pr_debug("Process already found\n"); |
de9f26bb | 831 | } else { |
0029cab3 JG |
832 | process = create_process(thread); |
833 | if (IS_ERR(process)) | |
834 | goto out; | |
835 | ||
de9f26bb KR |
836 | if (!procfs.kobj) |
837 | goto out; | |
838 | ||
839 | process->kobj = kfd_alloc_struct(process->kobj); | |
840 | if (!process->kobj) { | |
841 | pr_warn("Creating procfs kobject failed"); | |
842 | goto out; | |
843 | } | |
844 | ret = kobject_init_and_add(process->kobj, &procfs_type, | |
845 | procfs.kobj, "%d", | |
846 | (int)process->lead_thread->pid); | |
847 | if (ret) { | |
848 | pr_warn("Creating procfs pid directory failed"); | |
dc2f832e | 849 | kobject_put(process->kobj); |
de9f26bb KR |
850 | goto out; |
851 | } | |
852 | ||
75ae84c8 PY |
853 | kfd_sysfs_create_file(process->kobj, &process->attr_pasid, |
854 | "pasid"); | |
6d220a7e AL |
855 | |
856 | process->kobj_queues = kobject_create_and_add("queues", | |
857 | process->kobj); | |
858 | if (!process->kobj_queues) | |
859 | pr_warn("Creating KFD proc/queues folder failed"); | |
d4566dee | 860 | |
75ae84c8 PY |
861 | kfd_procfs_add_sysfs_stats(process); |
862 | kfd_procfs_add_sysfs_files(process); | |
751580b3 | 863 | kfd_procfs_add_sysfs_counters(process); |
12fb1ad7 JK |
864 | |
865 | init_waitqueue_head(&process->wait_irq_drain); | |
de9f26bb KR |
866 | } |
867 | out: | |
0f899fd4 FK |
868 | if (!IS_ERR(process)) |
869 | kref_get(&process->ref); | |
19f6d2a6 | 870 | mutex_unlock(&kfd_processes_mutex); |
0ab2d753 | 871 | mmput(thread->mm); |
19f6d2a6 | 872 | |
19f6d2a6 OG |
873 | return process; |
874 | } | |
875 | ||
876 | struct kfd_process *kfd_get_process(const struct task_struct *thread) | |
877 | { | |
878 | struct kfd_process *process; | |
879 | ||
4eacc26b | 880 | if (!thread->mm) |
19f6d2a6 OG |
881 | return ERR_PTR(-EINVAL); |
882 | ||
883 | /* Only the pthreads threading model is supported. */ | |
884 | if (thread->group_leader->mm != thread->mm) | |
885 | return ERR_PTR(-EINVAL); | |
886 | ||
011bbb03 | 887 | process = find_process(thread, false); |
e47cb828 WL |
888 | if (!process) |
889 | return ERR_PTR(-EINVAL); | |
19f6d2a6 OG |
890 | |
891 | return process; | |
892 | } | |
893 | ||
894 | static struct kfd_process *find_process_by_mm(const struct mm_struct *mm) | |
895 | { | |
896 | struct kfd_process *process; | |
897 | ||
898 | hash_for_each_possible_rcu(kfd_processes_table, process, | |
899 | kfd_processes, (uintptr_t)mm) | |
900 | if (process->mm == mm) | |
901 | return process; | |
902 | ||
903 | return NULL; | |
904 | } | |
905 | ||
011bbb03 RB |
906 | static struct kfd_process *find_process(const struct task_struct *thread, |
907 | bool ref) | |
19f6d2a6 OG |
908 | { |
909 | struct kfd_process *p; | |
910 | int idx; | |
911 | ||
912 | idx = srcu_read_lock(&kfd_processes_srcu); | |
913 | p = find_process_by_mm(thread->mm); | |
011bbb03 RB |
914 | if (p && ref) |
915 | kref_get(&p->ref); | |
19f6d2a6 OG |
916 | srcu_read_unlock(&kfd_processes_srcu, idx); |
917 | ||
918 | return p; | |
919 | } | |
920 | ||
abb208a8 FK |
921 | void kfd_unref_process(struct kfd_process *p) |
922 | { | |
923 | kref_put(&p->ref, kfd_process_ref_release); | |
924 | } | |
925 | ||
011bbb03 RB |
926 | /* This increments the process->ref counter. */ |
927 | struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid) | |
928 | { | |
929 | struct task_struct *task = NULL; | |
930 | struct kfd_process *p = NULL; | |
931 | ||
932 | if (!pid) { | |
933 | task = current; | |
934 | get_task_struct(task); | |
935 | } else { | |
936 | task = get_pid_task(pid, PIDTYPE_PID); | |
937 | } | |
938 | ||
939 | if (task) { | |
940 | p = find_process(task, true); | |
941 | put_task_struct(task); | |
942 | } | |
943 | ||
944 | return p; | |
945 | } | |
6ae27841 | 946 | |
52b29d73 FK |
947 | static void kfd_process_device_free_bos(struct kfd_process_device *pdd) |
948 | { | |
949 | struct kfd_process *p = pdd->process; | |
950 | void *mem; | |
951 | int id; | |
6ae27841 | 952 | int i; |
52b29d73 FK |
953 | |
954 | /* | |
955 | * Remove all handles from idr and release appropriate | |
956 | * local memory object | |
957 | */ | |
958 | idr_for_each_entry(&pdd->alloc_idr, mem, id) { | |
52b29d73 | 959 | |
6ae27841 AS |
960 | for (i = 0; i < p->n_pdds; i++) { |
961 | struct kfd_process_device *peer_pdd = p->pdds[i]; | |
962 | ||
b40a6ab2 | 963 | if (!peer_pdd->drm_priv) |
52b29d73 | 964 | continue; |
5b87245f | 965 | amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( |
dff63da9 | 966 | peer_pdd->dev->adev, mem, peer_pdd->drm_priv); |
52b29d73 FK |
967 | } |
968 | ||
dff63da9 | 969 | amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem, |
d4ec4bdc | 970 | pdd->drm_priv, NULL); |
52b29d73 FK |
971 | kfd_process_device_remove_obj_handle(pdd, id); |
972 | } | |
973 | } | |
974 | ||
68df0f19 LY |
975 | /* |
976 | * Just kunmap and unpin signal BO here. It will be freed in | |
977 | * kfd_process_free_outstanding_kfd_bos() | |
978 | */ | |
979 | static void kfd_process_kunmap_signal_bo(struct kfd_process *p) | |
980 | { | |
981 | struct kfd_process_device *pdd; | |
8dc1db31 | 982 | struct kfd_node *kdev; |
68df0f19 LY |
983 | void *mem; |
984 | ||
985 | kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle)); | |
986 | if (!kdev) | |
987 | return; | |
988 | ||
989 | mutex_lock(&p->mutex); | |
990 | ||
991 | pdd = kfd_get_process_device_data(kdev, p); | |
992 | if (!pdd) | |
993 | goto out; | |
994 | ||
995 | mem = kfd_process_device_translate_handle( | |
996 | pdd, GET_IDR_HANDLE(p->signal_handle)); | |
997 | if (!mem) | |
998 | goto out; | |
999 | ||
4e2d1044 | 1000 | amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem); |
68df0f19 LY |
1001 | |
1002 | out: | |
1003 | mutex_unlock(&p->mutex); | |
1004 | } | |
1005 | ||
52b29d73 FK |
1006 | static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p) |
1007 | { | |
6ae27841 | 1008 | int i; |
52b29d73 | 1009 | |
6ae27841 AS |
1010 | for (i = 0; i < p->n_pdds; i++) |
1011 | kfd_process_device_free_bos(p->pdds[i]); | |
52b29d73 FK |
1012 | } |
1013 | ||
de1450a5 | 1014 | static void kfd_process_destroy_pdds(struct kfd_process *p) |
19f6d2a6 | 1015 | { |
6ae27841 AS |
1016 | int i; |
1017 | ||
1018 | for (i = 0; i < p->n_pdds; i++) { | |
1019 | struct kfd_process_device *pdd = p->pdds[i]; | |
19f6d2a6 | 1020 | |
6027b1bf | 1021 | pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n", |
94a1ee09 OG |
1022 | pdd->dev->id, p->pasid); |
1023 | ||
68df0f19 LY |
1024 | kfd_process_device_destroy_cwsr_dgpu(pdd); |
1025 | kfd_process_device_destroy_ib_mem(pdd); | |
1026 | ||
bf47afba | 1027 | if (pdd->drm_file) { |
5b87245f | 1028 | amdgpu_amdkfd_gpuvm_release_process_vm( |
dff63da9 | 1029 | pdd->dev->adev, pdd->drm_priv); |
b84394e2 | 1030 | fput(pdd->drm_file); |
bf47afba | 1031 | } |
403575c4 | 1032 | |
f35751b8 | 1033 | if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base) |
373d7080 FK |
1034 | free_pages((unsigned long)pdd->qpd.cwsr_kaddr, |
1035 | get_order(KFD_CWSR_TBA_TMA_SIZE)); | |
1036 | ||
52b29d73 FK |
1037 | idr_destroy(&pdd->alloc_idr); |
1038 | ||
2105a15a | 1039 | kfd_free_process_doorbells(pdd->dev->kfd, pdd); |
59d7115d | 1040 | |
8dc1db31 | 1041 | if (pdd->dev->kfd->shared_resources.enable_mes) |
cc009e61 MJ |
1042 | amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev, |
1043 | pdd->proc_ctx_bo); | |
9593f4d6 RB |
1044 | /* |
1045 | * before destroying pdd, make sure to report availability | |
1046 | * for auto suspend | |
1047 | */ | |
1048 | if (pdd->runtime_inuse) { | |
d69a3b76 MJ |
1049 | pm_runtime_mark_last_busy(adev_to_drm(pdd->dev->adev)->dev); |
1050 | pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev); | |
9593f4d6 RB |
1051 | pdd->runtime_inuse = false; |
1052 | } | |
1053 | ||
19f6d2a6 | 1054 | kfree(pdd); |
6ae27841 | 1055 | p->pdds[i] = NULL; |
19f6d2a6 | 1056 | } |
6ae27841 | 1057 | p->n_pdds = 0; |
de1450a5 FK |
1058 | } |
1059 | ||
751580b3 | 1060 | static void kfd_process_remove_sysfs(struct kfd_process *p) |
de1450a5 | 1061 | { |
751580b3 | 1062 | struct kfd_process_device *pdd; |
6ae27841 | 1063 | int i; |
de1450a5 | 1064 | |
751580b3 PY |
1065 | if (!p->kobj) |
1066 | return; | |
d4566dee | 1067 | |
751580b3 PY |
1068 | sysfs_remove_file(p->kobj, &p->attr_pasid); |
1069 | kobject_del(p->kobj_queues); | |
1070 | kobject_put(p->kobj_queues); | |
1071 | p->kobj_queues = NULL; | |
6ae27841 | 1072 | |
751580b3 PY |
1073 | for (i = 0; i < p->n_pdds; i++) { |
1074 | pdd = p->pdds[i]; | |
dcdb4d90 | 1075 | |
751580b3 PY |
1076 | sysfs_remove_file(p->kobj, &pdd->attr_vram); |
1077 | sysfs_remove_file(p->kobj, &pdd->attr_sdma); | |
d4566dee | 1078 | |
751580b3 PY |
1079 | sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict); |
1080 | if (pdd->dev->kfd2kgd->get_cu_occupancy) | |
1081 | sysfs_remove_file(pdd->kobj_stats, | |
1082 | &pdd->attr_cu_occupancy); | |
1083 | kobject_del(pdd->kobj_stats); | |
1084 | kobject_put(pdd->kobj_stats); | |
1085 | pdd->kobj_stats = NULL; | |
1086 | } | |
1087 | ||
1088 | for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) { | |
1089 | pdd = p->pdds[i]; | |
1090 | ||
1091 | sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults); | |
1092 | sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in); | |
1093 | sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out); | |
1094 | kobject_del(pdd->kobj_counters); | |
1095 | kobject_put(pdd->kobj_counters); | |
1096 | pdd->kobj_counters = NULL; | |
de9f26bb KR |
1097 | } |
1098 | ||
751580b3 PY |
1099 | kobject_del(p->kobj); |
1100 | kobject_put(p->kobj); | |
1101 | p->kobj = NULL; | |
1102 | } | |
1103 | ||
1104 | /* No process locking is needed in this function, because the process | |
1105 | * is not findable any more. We must assume that no other thread is | |
1106 | * using it any more, otherwise we couldn't safely free the process | |
1107 | * structure in the end. | |
1108 | */ | |
1109 | static void kfd_process_wq_release(struct work_struct *work) | |
1110 | { | |
1111 | struct kfd_process *p = container_of(work, struct kfd_process, | |
1112 | release_work); | |
c147ddc6 | 1113 | struct dma_fence *ef; |
68df0f19 | 1114 | |
74097f9f PY |
1115 | kfd_process_dequeue_from_all_devices(p); |
1116 | pqm_uninit(&p->pqm); | |
1117 | ||
1118 | /* Signal the eviction fence after user mode queues are | |
1119 | * destroyed. This allows any BOs to be freed without | |
1120 | * triggering pointless evictions or waiting for fences. | |
1121 | */ | |
c147ddc6 FK |
1122 | synchronize_rcu(); |
1123 | ef = rcu_access_pointer(p->ef); | |
1124 | dma_fence_signal(ef); | |
74097f9f | 1125 | |
751580b3 | 1126 | kfd_process_remove_sysfs(p); |
de1450a5 | 1127 | |
68df0f19 | 1128 | kfd_process_kunmap_signal_bo(p); |
52b29d73 | 1129 | kfd_process_free_outstanding_kfd_bos(p); |
42de677f | 1130 | svm_range_list_fini(p); |
52b29d73 | 1131 | |
de1450a5 | 1132 | kfd_process_destroy_pdds(p); |
c147ddc6 | 1133 | dma_fence_put(ef); |
19f6d2a6 | 1134 | |
f3a39818 AL |
1135 | kfd_event_free_process(p); |
1136 | ||
19f6d2a6 | 1137 | kfd_pasid_free(p->pasid); |
19f6d2a6 OG |
1138 | mutex_destroy(&p->mutex); |
1139 | ||
c7b1243e FK |
1140 | put_task_struct(p->lead_thread); |
1141 | ||
19f6d2a6 | 1142 | kfree(p); |
19f6d2a6 OG |
1143 | } |
1144 | ||
5ce10687 | 1145 | static void kfd_process_ref_release(struct kref *ref) |
19f6d2a6 | 1146 | { |
5ce10687 | 1147 | struct kfd_process *p = container_of(ref, struct kfd_process, ref); |
19f6d2a6 | 1148 | |
5ce10687 FK |
1149 | INIT_WORK(&p->release_work, kfd_process_wq_release); |
1150 | queue_work(kfd_process_wq, &p->release_work); | |
1151 | } | |
19f6d2a6 | 1152 | |
3248b6d3 FK |
1153 | static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm) |
1154 | { | |
1155 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
1156 | struct kfd_process *p = find_process_by_mm(mm); | |
1157 | ||
1158 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
1159 | ||
1160 | return p ? &p->mmu_notifier : ERR_PTR(-ESRCH); | |
1161 | } | |
1162 | ||
471f3902 | 1163 | static void kfd_process_free_notifier(struct mmu_notifier *mn) |
5ce10687 | 1164 | { |
471f3902 | 1165 | kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier)); |
19f6d2a6 OG |
1166 | } |
1167 | ||
22e3d934 DB |
1168 | static void kfd_process_notifier_release_internal(struct kfd_process *p) |
1169 | { | |
21889582 JK |
1170 | int i; |
1171 | ||
22e3d934 DB |
1172 | cancel_delayed_work_sync(&p->eviction_work); |
1173 | cancel_delayed_work_sync(&p->restore_work); | |
1174 | ||
21889582 JK |
1175 | for (i = 0; i < p->n_pdds; i++) { |
1176 | struct kfd_process_device *pdd = p->pdds[i]; | |
1177 | ||
1178 | /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */ | |
1179 | if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup) | |
1180 | amdgpu_gfx_off_ctrl(pdd->dev->adev, true); | |
1181 | } | |
1182 | ||
22e3d934 DB |
1183 | /* Indicate to other users that MM is no longer valid */ |
1184 | p->mm = NULL; | |
0ab2d753 JK |
1185 | kfd_dbg_trap_disable(p); |
1186 | ||
1187 | if (atomic_read(&p->debugged_process_count) > 0) { | |
1188 | struct kfd_process *target; | |
1189 | unsigned int temp; | |
1190 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
1191 | ||
1192 | hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) { | |
1193 | if (target->debugger_process && target->debugger_process == p) { | |
1194 | mutex_lock_nested(&target->mutex, 1); | |
1195 | kfd_dbg_trap_disable(target); | |
1196 | mutex_unlock(&target->mutex); | |
1197 | if (atomic_read(&p->debugged_process_count) == 0) | |
1198 | break; | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
1203 | } | |
22e3d934 DB |
1204 | |
1205 | mmu_notifier_put(&p->mmu_notifier); | |
1206 | } | |
1207 | ||
19f6d2a6 OG |
1208 | static void kfd_process_notifier_release(struct mmu_notifier *mn, |
1209 | struct mm_struct *mm) | |
1210 | { | |
1211 | struct kfd_process *p; | |
1212 | ||
1213 | /* | |
1214 | * The kfd_process structure can not be free because the | |
1215 | * mmu_notifier srcu is read locked | |
1216 | */ | |
1217 | p = container_of(mn, struct kfd_process, mmu_notifier); | |
32fa8219 FK |
1218 | if (WARN_ON(p->mm != mm)) |
1219 | return; | |
19f6d2a6 OG |
1220 | |
1221 | mutex_lock(&kfd_processes_mutex); | |
22e3d934 DB |
1222 | /* |
1223 | * Do early return if table is empty. | |
1224 | * | |
1225 | * This could potentially happen if this function is called concurrently | |
1226 | * by mmu_notifier and by kfd_cleanup_pocesses. | |
1227 | * | |
1228 | */ | |
1229 | if (hash_empty(kfd_processes_table)) { | |
1230 | mutex_unlock(&kfd_processes_mutex); | |
1231 | return; | |
1232 | } | |
19f6d2a6 OG |
1233 | hash_del_rcu(&p->kfd_processes); |
1234 | mutex_unlock(&kfd_processes_mutex); | |
1235 | synchronize_srcu(&kfd_processes_srcu); | |
1236 | ||
22e3d934 | 1237 | kfd_process_notifier_release_internal(p); |
19f6d2a6 OG |
1238 | } |
1239 | ||
1240 | static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = { | |
1241 | .release = kfd_process_notifier_release, | |
3248b6d3 | 1242 | .alloc_notifier = kfd_process_alloc_notifier, |
471f3902 | 1243 | .free_notifier = kfd_process_free_notifier, |
19f6d2a6 OG |
1244 | }; |
1245 | ||
22e3d934 DB |
1246 | /* |
1247 | * This code handles the case when driver is being unloaded before all | |
1248 | * mm_struct are released. We need to safely free the kfd_process and | |
1249 | * avoid race conditions with mmu_notifier that might try to free them. | |
1250 | * | |
1251 | */ | |
1252 | void kfd_cleanup_processes(void) | |
1253 | { | |
1254 | struct kfd_process *p; | |
1255 | struct hlist_node *p_temp; | |
1256 | unsigned int temp; | |
1257 | HLIST_HEAD(cleanup_list); | |
1258 | ||
1259 | /* | |
1260 | * Move all remaining kfd_process from the process table to a | |
1261 | * temp list for processing. Once done, callback from mmu_notifier | |
1262 | * release will not see the kfd_process in the table and do early return, | |
1263 | * avoiding double free issues. | |
1264 | */ | |
1265 | mutex_lock(&kfd_processes_mutex); | |
1266 | hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) { | |
1267 | hash_del_rcu(&p->kfd_processes); | |
1268 | synchronize_srcu(&kfd_processes_srcu); | |
1269 | hlist_add_head(&p->kfd_processes, &cleanup_list); | |
1270 | } | |
1271 | mutex_unlock(&kfd_processes_mutex); | |
1272 | ||
1273 | hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes) | |
1274 | kfd_process_notifier_release_internal(p); | |
1275 | ||
1276 | /* | |
1277 | * Ensures that all outstanding free_notifier get called, triggering | |
1278 | * the release of the kfd_process struct. | |
1279 | */ | |
1280 | mmu_notifier_synchronize(); | |
1281 | } | |
1282 | ||
0ab2d753 | 1283 | int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep) |
373d7080 | 1284 | { |
373d7080 | 1285 | unsigned long offset; |
6ae27841 | 1286 | int i; |
373d7080 | 1287 | |
0ab2d753 JK |
1288 | if (p->has_cwsr) |
1289 | return 0; | |
1290 | ||
6ae27841 | 1291 | for (i = 0; i < p->n_pdds; i++) { |
8dc1db31 | 1292 | struct kfd_node *dev = p->pdds[i]->dev; |
6ae27841 | 1293 | struct qcm_process_device *qpd = &p->pdds[i]->qpd; |
f35751b8 | 1294 | |
8dc1db31 | 1295 | if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base) |
373d7080 | 1296 | continue; |
f35751b8 | 1297 | |
29453755 | 1298 | offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id); |
373d7080 FK |
1299 | qpd->tba_addr = (int64_t)vm_mmap(filep, 0, |
1300 | KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC, | |
1301 | MAP_SHARED, offset); | |
1302 | ||
1303 | if (IS_ERR_VALUE(qpd->tba_addr)) { | |
c0ede1f8 YZ |
1304 | int err = qpd->tba_addr; |
1305 | ||
1306 | pr_err("Failure to set tba address. error %d.\n", err); | |
373d7080 FK |
1307 | qpd->tba_addr = 0; |
1308 | qpd->cwsr_kaddr = NULL; | |
c0ede1f8 | 1309 | return err; |
373d7080 FK |
1310 | } |
1311 | ||
8dc1db31 | 1312 | memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); |
373d7080 | 1313 | |
50cff45e JC |
1314 | kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled); |
1315 | ||
373d7080 FK |
1316 | qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; |
1317 | pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", | |
1318 | qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr); | |
1319 | } | |
c0ede1f8 | 1320 | |
0ab2d753 JK |
1321 | p->has_cwsr = true; |
1322 | ||
c0ede1f8 | 1323 | return 0; |
373d7080 FK |
1324 | } |
1325 | ||
f35751b8 FK |
1326 | static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd) |
1327 | { | |
8dc1db31 | 1328 | struct kfd_node *dev = pdd->dev; |
f35751b8 | 1329 | struct qcm_process_device *qpd = &pdd->qpd; |
1d251d90 YZ |
1330 | uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
1331 | | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE | |
1332 | | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE; | |
68df0f19 | 1333 | struct kgd_mem *mem; |
f35751b8 FK |
1334 | void *kaddr; |
1335 | int ret; | |
1336 | ||
8dc1db31 | 1337 | if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base) |
f35751b8 FK |
1338 | return 0; |
1339 | ||
1340 | /* cwsr_base is only set for dGPU */ | |
1341 | ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base, | |
68df0f19 | 1342 | KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr); |
f35751b8 FK |
1343 | if (ret) |
1344 | return ret; | |
1345 | ||
68df0f19 | 1346 | qpd->cwsr_mem = mem; |
f35751b8 FK |
1347 | qpd->cwsr_kaddr = kaddr; |
1348 | qpd->tba_addr = qpd->cwsr_base; | |
1349 | ||
8dc1db31 | 1350 | memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size); |
f35751b8 | 1351 | |
50cff45e JC |
1352 | kfd_process_set_trap_debug_flag(&pdd->qpd, |
1353 | pdd->process->debug_trap_enabled); | |
1354 | ||
f35751b8 FK |
1355 | qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET; |
1356 | pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n", | |
1357 | qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr); | |
1358 | ||
1359 | return 0; | |
1360 | } | |
1361 | ||
68df0f19 LY |
1362 | static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd) |
1363 | { | |
8dc1db31 | 1364 | struct kfd_node *dev = pdd->dev; |
68df0f19 LY |
1365 | struct qcm_process_device *qpd = &pdd->qpd; |
1366 | ||
8dc1db31 | 1367 | if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base) |
68df0f19 LY |
1368 | return; |
1369 | ||
cb8dc232 | 1370 | kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr); |
68df0f19 LY |
1371 | } |
1372 | ||
7c9631af JC |
1373 | void kfd_process_set_trap_handler(struct qcm_process_device *qpd, |
1374 | uint64_t tba_addr, | |
1375 | uint64_t tma_addr) | |
1376 | { | |
1377 | if (qpd->cwsr_kaddr) { | |
1378 | /* KFD trap handler is bound, record as second-level TBA/TMA | |
1379 | * in first-level TMA. First-level trap will jump to second. | |
1380 | */ | |
1381 | uint64_t *tma = | |
1382 | (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET); | |
1383 | tma[0] = tba_addr; | |
1384 | tma[1] = tma_addr; | |
1385 | } else { | |
1386 | /* No trap handler bound, bind as first-level TBA/TMA. */ | |
1387 | qpd->tba_addr = tba_addr; | |
1388 | qpd->tma_addr = tma_addr; | |
1389 | } | |
1390 | } | |
1391 | ||
063e33c5 AS |
1392 | bool kfd_process_xnack_mode(struct kfd_process *p, bool supported) |
1393 | { | |
1394 | int i; | |
1395 | ||
1396 | /* On most GFXv9 GPUs, the retry mode in the SQ must match the | |
1397 | * boot time retry setting. Mixing processes with different | |
1398 | * XNACK/retry settings can hang the GPU. | |
1399 | * | |
1400 | * Different GPUs can have different noretry settings depending | |
1401 | * on HW bugs or limitations. We need to find at least one | |
1402 | * XNACK mode for this process that's compatible with all GPUs. | |
1403 | * Fortunately GPUs with retry enabled (noretry=0) can run code | |
1404 | * built for XNACK-off. On GFXv9 it may perform slower. | |
1405 | * | |
1406 | * Therefore applications built for XNACK-off can always be | |
1407 | * supported and will be our fallback if any GPU does not | |
1408 | * support retry. | |
1409 | */ | |
1410 | for (i = 0; i < p->n_pdds; i++) { | |
8dc1db31 | 1411 | struct kfd_node *dev = p->pdds[i]->dev; |
063e33c5 AS |
1412 | |
1413 | /* Only consider GFXv9 and higher GPUs. Older GPUs don't | |
1414 | * support the SVM APIs and don't need to be considered | |
1415 | * for the XNACK mode selection. | |
1416 | */ | |
046e674b | 1417 | if (!KFD_IS_SOC15(dev)) |
063e33c5 AS |
1418 | continue; |
1419 | /* Aldebaran can always support XNACK because it can support | |
1420 | * per-process XNACK mode selection. But let the dev->noretry | |
1421 | * setting still influence the default XNACK mode. | |
1422 | */ | |
9256e8d4 SK |
1423 | if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) { |
1424 | if (!amdgpu_sriov_xnack_support(dev->kfd->adev)) { | |
1425 | pr_debug("SRIOV platform xnack not supported\n"); | |
1426 | return false; | |
1427 | } | |
063e33c5 | 1428 | continue; |
9256e8d4 | 1429 | } |
063e33c5 AS |
1430 | |
1431 | /* GFXv10 and later GPUs do not support shader preemption | |
1432 | * during page faults. This can lead to poor QoS for queue | |
1433 | * management and memory-manager-related preemptions or | |
1434 | * even deadlocks. | |
1435 | */ | |
046e674b | 1436 | if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) |
063e33c5 AS |
1437 | return false; |
1438 | ||
8dc1db31 | 1439 | if (dev->kfd->noretry) |
063e33c5 AS |
1440 | return false; |
1441 | } | |
1442 | ||
1443 | return true; | |
1444 | } | |
1445 | ||
50cff45e JC |
1446 | void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd, |
1447 | bool enabled) | |
1448 | { | |
1449 | if (qpd->cwsr_kaddr) { | |
1450 | uint64_t *tma = | |
1451 | (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET); | |
1452 | tma[2] = enabled; | |
1453 | } | |
1454 | } | |
1455 | ||
0029cab3 JG |
1456 | /* |
1457 | * On return the kfd_process is fully operational and will be freed when the | |
1458 | * mm is released | |
1459 | */ | |
1460 | static struct kfd_process *create_process(const struct task_struct *thread) | |
19f6d2a6 OG |
1461 | { |
1462 | struct kfd_process *process; | |
3248b6d3 | 1463 | struct mmu_notifier *mn; |
19f6d2a6 OG |
1464 | int err = -ENOMEM; |
1465 | ||
1466 | process = kzalloc(sizeof(*process), GFP_KERNEL); | |
19f6d2a6 OG |
1467 | if (!process) |
1468 | goto err_alloc_process; | |
1469 | ||
5ce10687 | 1470 | kref_init(&process->ref); |
19f6d2a6 | 1471 | mutex_init(&process->mutex); |
19f6d2a6 | 1472 | process->mm = thread->mm; |
19f6d2a6 | 1473 | process->lead_thread = thread->group_leader; |
6ae27841 | 1474 | process->n_pdds = 0; |
cd9f7910 | 1475 | process->queues_paused = false; |
0029cab3 JG |
1476 | INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker); |
1477 | INIT_DELAYED_WORK(&process->restore_work, restore_process_worker); | |
1478 | process->last_restore_timestamp = get_jiffies_64(); | |
c3eb12df FK |
1479 | err = kfd_event_init_process(process); |
1480 | if (err) | |
1481 | goto err_event_init; | |
0029cab3 | 1482 | process->is_32bit_user_mode = in_compat_syscall(); |
0ab2d753 JK |
1483 | process->debug_trap_enabled = false; |
1484 | process->debugger_process = NULL; | |
1485 | process->exception_enable_mask = 0; | |
1486 | atomic_set(&process->debugged_process_count, 0); | |
c2d2588c | 1487 | sema_init(&process->runtime_enable_sema, 0); |
0029cab3 JG |
1488 | |
1489 | process->pasid = kfd_pasid_alloc(); | |
c3eb12df FK |
1490 | if (process->pasid == 0) { |
1491 | err = -ENOSPC; | |
0029cab3 | 1492 | goto err_alloc_pasid; |
c3eb12df | 1493 | } |
0029cab3 | 1494 | |
45102048 BG |
1495 | err = pqm_init(&process->pqm, process); |
1496 | if (err != 0) | |
1497 | goto err_process_pqm_init; | |
1498 | ||
dd59239a | 1499 | /* init process apertures*/ |
b312b2b2 DC |
1500 | err = kfd_init_apertures(process); |
1501 | if (err != 0) | |
7a10d63f | 1502 | goto err_init_apertures; |
dd59239a | 1503 | |
063e33c5 AS |
1504 | /* Check XNACK support after PDDs are created in kfd_init_apertures */ |
1505 | process->xnack_enabled = kfd_process_xnack_mode(process, false); | |
1506 | ||
42de677f PY |
1507 | err = svm_range_list_init(process); |
1508 | if (err) | |
1509 | goto err_init_svm_range_list; | |
1510 | ||
3248b6d3 FK |
1511 | /* alloc_notifier needs to find the process in the hash table */ |
1512 | hash_add_rcu(kfd_processes_table, &process->kfd_processes, | |
1513 | (uintptr_t)process->mm); | |
1514 | ||
0593ad21 PY |
1515 | /* Avoid free_notifier to start kfd_process_wq_release if |
1516 | * mmu_notifier_get failed because of pending signal. | |
1517 | */ | |
1518 | kref_get(&process->ref); | |
1519 | ||
3248b6d3 FK |
1520 | /* MMU notifier registration must be the last call that can fail |
1521 | * because after this point we cannot unwind the process creation. | |
1522 | * After this point, mmu_notifier_put will trigger the cleanup by | |
1523 | * dropping the last process reference in the free_notifier. | |
1524 | */ | |
1525 | mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm); | |
1526 | if (IS_ERR(mn)) { | |
1527 | err = PTR_ERR(mn); | |
0029cab3 | 1528 | goto err_register_notifier; |
3248b6d3 FK |
1529 | } |
1530 | BUG_ON(mn != &process->mmu_notifier); | |
0029cab3 | 1531 | |
0593ad21 | 1532 | kfd_unref_process(process); |
0029cab3 | 1533 | get_task_struct(process->lead_thread); |
c0ede1f8 | 1534 | |
44b87bb0 JK |
1535 | INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler); |
1536 | ||
19f6d2a6 OG |
1537 | return process; |
1538 | ||
0029cab3 | 1539 | err_register_notifier: |
3248b6d3 | 1540 | hash_del_rcu(&process->kfd_processes); |
42de677f PY |
1541 | svm_range_list_fini(process); |
1542 | err_init_svm_range_list: | |
52b29d73 | 1543 | kfd_process_free_outstanding_kfd_bos(process); |
c0ede1f8 | 1544 | kfd_process_destroy_pdds(process); |
7a10d63f | 1545 | err_init_apertures: |
dd59239a | 1546 | pqm_uninit(&process->pqm); |
45102048 | 1547 | err_process_pqm_init: |
19f6d2a6 OG |
1548 | kfd_pasid_free(process->pasid); |
1549 | err_alloc_pasid: | |
c3eb12df FK |
1550 | kfd_event_free_process(process); |
1551 | err_event_init: | |
0029cab3 | 1552 | mutex_destroy(&process->mutex); |
19f6d2a6 OG |
1553 | kfree(process); |
1554 | err_alloc_process: | |
1555 | return ERR_PTR(err); | |
1556 | } | |
1557 | ||
8dc1db31 | 1558 | struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev, |
093c7d8c | 1559 | struct kfd_process *p) |
19f6d2a6 | 1560 | { |
6ae27841 | 1561 | int i; |
19f6d2a6 | 1562 | |
6ae27841 AS |
1563 | for (i = 0; i < p->n_pdds; i++) |
1564 | if (p->pdds[i]->dev == dev) | |
1565 | return p->pdds[i]; | |
093c7d8c | 1566 | |
733fa1f7 | 1567 | return NULL; |
093c7d8c AS |
1568 | } |
1569 | ||
8dc1db31 | 1570 | struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, |
093c7d8c AS |
1571 | struct kfd_process *p) |
1572 | { | |
1573 | struct kfd_process_device *pdd = NULL; | |
cc009e61 | 1574 | int retval = 0; |
093c7d8c | 1575 | |
6ae27841 AS |
1576 | if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE)) |
1577 | return NULL; | |
093c7d8c | 1578 | pdd = kzalloc(sizeof(*pdd), GFP_KERNEL); |
2d9b36f9 FK |
1579 | if (!pdd) |
1580 | return NULL; | |
1581 | ||
1582 | pdd->dev = dev; | |
1583 | INIT_LIST_HEAD(&pdd->qpd.queues_list); | |
1584 | INIT_LIST_HEAD(&pdd->qpd.priv_queue_list); | |
1585 | pdd->qpd.dqm = dev->dqm; | |
1586 | pdd->qpd.pqm = &p->pqm; | |
26103436 | 1587 | pdd->qpd.evicted = 0; |
b8020b03 | 1588 | pdd->qpd.mapped_gws_queue = false; |
2d9b36f9 FK |
1589 | pdd->process = p; |
1590 | pdd->bound = PDD_UNBOUND; | |
1591 | pdd->already_dequeued = false; | |
9593f4d6 | 1592 | pdd->runtime_inuse = false; |
d4566dee | 1593 | pdd->vram_usage = 0; |
32cb59f3 | 1594 | pdd->sdma_past_activity_counter = 0; |
bef153b7 | 1595 | pdd->user_gpu_id = dev->id; |
4327bed2 | 1596 | atomic64_set(&pdd->evict_duration_counter, 0); |
cc009e61 | 1597 | |
8dc1db31 | 1598 | if (dev->kfd->shared_resources.enable_mes) { |
cc009e61 MJ |
1599 | retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, |
1600 | AMDGPU_MES_PROC_CTX_SIZE, | |
1601 | &pdd->proc_ctx_bo, | |
1602 | &pdd->proc_ctx_gpu_addr, | |
1603 | &pdd->proc_ctx_cpu_ptr, | |
1604 | false); | |
1605 | if (retval) { | |
1606 | pr_err("failed to allocate process context bo\n"); | |
1607 | goto err_free_pdd; | |
1608 | } | |
1609 | memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE); | |
1610 | } | |
1611 | ||
6ae27841 | 1612 | p->pdds[p->n_pdds++] = pdd; |
0de4ec9a JK |
1613 | if (kfd_dbg_is_per_vmid_supported(pdd->dev)) |
1614 | pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap( | |
1615 | pdd->dev->adev, | |
1616 | false, | |
1617 | 0); | |
19f6d2a6 | 1618 | |
52b29d73 FK |
1619 | /* Init idr used for memory handle translation */ |
1620 | idr_init(&pdd->alloc_idr); | |
1621 | ||
b84394e2 | 1622 | return pdd; |
59d7115d MJ |
1623 | |
1624 | err_free_pdd: | |
1625 | kfree(pdd); | |
1626 | return NULL; | |
b84394e2 FK |
1627 | } |
1628 | ||
1629 | /** | |
1630 | * kfd_process_device_init_vm - Initialize a VM for a process-device | |
1631 | * | |
1632 | * @pdd: The process-device | |
1633 | * @drm_file: Optional pointer to a DRM file descriptor | |
1634 | * | |
1635 | * If @drm_file is specified, it will be used to acquire the VM from | |
1636 | * that file descriptor. If successful, the @pdd takes ownership of | |
1637 | * the file descriptor. | |
1638 | * | |
1639 | * If @drm_file is NULL, a new VM is created. | |
1640 | * | |
1641 | * Returns 0 on success, -errno on failure. | |
1642 | */ | |
1643 | int kfd_process_device_init_vm(struct kfd_process_device *pdd, | |
1644 | struct file *drm_file) | |
1645 | { | |
23b02b0e PY |
1646 | struct amdgpu_fpriv *drv_priv; |
1647 | struct amdgpu_vm *avm; | |
b84394e2 | 1648 | struct kfd_process *p; |
9a1c1339 | 1649 | struct dma_fence *ef; |
8dc1db31 | 1650 | struct kfd_node *dev; |
b84394e2 FK |
1651 | int ret; |
1652 | ||
f45e6b9d FK |
1653 | if (!drm_file) |
1654 | return -EINVAL; | |
1655 | ||
b40a6ab2 | 1656 | if (pdd->drm_priv) |
f45e6b9d | 1657 | return -EBUSY; |
b84394e2 | 1658 | |
23b02b0e PY |
1659 | ret = amdgpu_file_to_fpriv(drm_file, &drv_priv); |
1660 | if (ret) | |
1661 | return ret; | |
1662 | avm = &drv_priv->vm; | |
1663 | ||
b84394e2 FK |
1664 | p = pdd->process; |
1665 | dev = pdd->dev; | |
1666 | ||
23b02b0e | 1667 | ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm, |
41d82649 | 1668 | &p->kgd_process_info, |
9a1c1339 | 1669 | &ef); |
b84394e2 | 1670 | if (ret) { |
403575c4 | 1671 | pr_err("Failed to create process VM object\n"); |
b84394e2 | 1672 | return ret; |
403575c4 | 1673 | } |
9a1c1339 | 1674 | RCU_INIT_POINTER(p->ef, ef); |
b40a6ab2 | 1675 | pdd->drm_priv = drm_file->private_data; |
f40c6912 | 1676 | |
552764b6 FK |
1677 | ret = kfd_process_device_reserve_ib_mem(pdd); |
1678 | if (ret) | |
1679 | goto err_reserve_ib_mem; | |
f35751b8 FK |
1680 | ret = kfd_process_device_init_cwsr_dgpu(pdd); |
1681 | if (ret) | |
1682 | goto err_init_cwsr; | |
1683 | ||
23b02b0e | 1684 | ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid); |
41d82649 PY |
1685 | if (ret) |
1686 | goto err_set_pasid; | |
1687 | ||
b84394e2 FK |
1688 | pdd->drm_file = drm_file; |
1689 | ||
1690 | return 0; | |
f35751b8 | 1691 | |
41d82649 PY |
1692 | err_set_pasid: |
1693 | kfd_process_device_destroy_cwsr_dgpu(pdd); | |
f35751b8 | 1694 | err_init_cwsr: |
cb8dc232 | 1695 | kfd_process_device_destroy_ib_mem(pdd); |
552764b6 | 1696 | err_reserve_ib_mem: |
b40a6ab2 | 1697 | pdd->drm_priv = NULL; |
23b02b0e | 1698 | amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm); |
f35751b8 FK |
1699 | |
1700 | return ret; | |
19f6d2a6 OG |
1701 | } |
1702 | ||
1703 | /* | |
1704 | * Direct the IOMMU to bind the process (specifically the pasid->mm) | |
1705 | * to the device. | |
1706 | * Unbinding occurs when the process dies or the device is removed. | |
1707 | * | |
1708 | * Assumes that the process lock is held. | |
1709 | */ | |
8dc1db31 | 1710 | struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev, |
19f6d2a6 OG |
1711 | struct kfd_process *p) |
1712 | { | |
093c7d8c | 1713 | struct kfd_process_device *pdd; |
b17f068a | 1714 | int err; |
19f6d2a6 | 1715 | |
093c7d8c AS |
1716 | pdd = kfd_get_process_device_data(dev, p); |
1717 | if (!pdd) { | |
1718 | pr_err("Process device data doesn't exist\n"); | |
19f6d2a6 | 1719 | return ERR_PTR(-ENOMEM); |
093c7d8c | 1720 | } |
19f6d2a6 | 1721 | |
b40a6ab2 | 1722 | if (!pdd->drm_priv) |
f45e6b9d FK |
1723 | return ERR_PTR(-ENODEV); |
1724 | ||
9593f4d6 RB |
1725 | /* |
1726 | * signal runtime-pm system to auto resume and prevent | |
1727 | * further runtime suspend once device pdd is created until | |
1728 | * pdd is destroyed. | |
1729 | */ | |
1730 | if (!pdd->runtime_inuse) { | |
d69a3b76 | 1731 | err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev); |
1c1ada37 | 1732 | if (err < 0) { |
d69a3b76 | 1733 | pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev); |
9593f4d6 | 1734 | return ERR_PTR(err); |
1c1ada37 | 1735 | } |
9593f4d6 RB |
1736 | } |
1737 | ||
9593f4d6 RB |
1738 | /* |
1739 | * make sure that runtime_usage counter is incremented just once | |
1740 | * per pdd | |
1741 | */ | |
1742 | pdd->runtime_inuse = true; | |
b84394e2 | 1743 | |
19f6d2a6 OG |
1744 | return pdd; |
1745 | } | |
1746 | ||
52b29d73 FK |
1747 | /* Create specific handle mapped to mem from process local memory idr |
1748 | * Assumes that the process lock is held. | |
1749 | */ | |
1750 | int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, | |
1751 | void *mem) | |
1752 | { | |
1753 | return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL); | |
1754 | } | |
1755 | ||
1756 | /* Translate specific handle from process local memory idr | |
1757 | * Assumes that the process lock is held. | |
1758 | */ | |
1759 | void *kfd_process_device_translate_handle(struct kfd_process_device *pdd, | |
1760 | int handle) | |
1761 | { | |
1762 | if (handle < 0) | |
1763 | return NULL; | |
1764 | ||
1765 | return idr_find(&pdd->alloc_idr, handle); | |
1766 | } | |
1767 | ||
1768 | /* Remove specific handle from process local memory idr | |
1769 | * Assumes that the process lock is held. | |
1770 | */ | |
1771 | void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, | |
1772 | int handle) | |
1773 | { | |
1774 | if (handle >= 0) | |
1775 | idr_remove(&pdd->alloc_idr, handle); | |
1776 | } | |
1777 | ||
abb208a8 | 1778 | /* This increments the process->ref counter. */ |
c7b6bac9 | 1779 | struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid) |
f3a39818 | 1780 | { |
82c16b42 | 1781 | struct kfd_process *p, *ret_p = NULL; |
f3a39818 AL |
1782 | unsigned int temp; |
1783 | ||
1784 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
1785 | ||
1786 | hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { | |
1787 | if (p->pasid == pasid) { | |
abb208a8 | 1788 | kref_get(&p->ref); |
82c16b42 | 1789 | ret_p = p; |
f3a39818 AL |
1790 | break; |
1791 | } | |
1792 | } | |
1793 | ||
1794 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
1795 | ||
82c16b42 | 1796 | return ret_p; |
f3a39818 | 1797 | } |
373d7080 | 1798 | |
26103436 FK |
1799 | /* This increments the process->ref counter. */ |
1800 | struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm) | |
1801 | { | |
1802 | struct kfd_process *p; | |
1803 | ||
1804 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
1805 | ||
1806 | p = find_process_by_mm(mm); | |
1807 | if (p) | |
1808 | kref_get(&p->ref); | |
1809 | ||
1810 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
1811 | ||
1812 | return p; | |
1813 | } | |
1814 | ||
0aeaaf64 | 1815 | /* kfd_process_evict_queues - Evict all user queues of a process |
26103436 FK |
1816 | * |
1817 | * Eviction is reference-counted per process-device. This means multiple | |
1818 | * evictions from different sources can be nested safely. | |
1819 | */ | |
c7f21978 | 1820 | int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger) |
26103436 | 1821 | { |
26103436 | 1822 | int r = 0; |
6ae27841 | 1823 | int i; |
26103436 FK |
1824 | unsigned int n_evicted = 0; |
1825 | ||
6ae27841 AS |
1826 | for (i = 0; i < p->n_pdds; i++) { |
1827 | struct kfd_process_device *pdd = p->pdds[i]; | |
1828 | ||
d6e924ad | 1829 | kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid, |
c7f21978 PY |
1830 | trigger); |
1831 | ||
26103436 FK |
1832 | r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm, |
1833 | &pdd->qpd); | |
b8c20c74 | 1834 | /* evict return -EIO if HWS is hang or asic is resetting, in this case |
1835 | * we would like to set all the queues to be in evicted state to prevent | |
1836 | * them been add back since they actually not be saved right now. | |
1837 | */ | |
1838 | if (r && r != -EIO) { | |
26103436 FK |
1839 | pr_err("Failed to evict process queues\n"); |
1840 | goto fail; | |
1841 | } | |
1842 | n_evicted++; | |
1843 | } | |
1844 | ||
1845 | return r; | |
1846 | ||
1847 | fail: | |
1848 | /* To keep state consistent, roll back partial eviction by | |
1849 | * restoring queues | |
1850 | */ | |
6ae27841 AS |
1851 | for (i = 0; i < p->n_pdds; i++) { |
1852 | struct kfd_process_device *pdd = p->pdds[i]; | |
1853 | ||
26103436 FK |
1854 | if (n_evicted == 0) |
1855 | break; | |
c7f21978 | 1856 | |
d6e924ad | 1857 | kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); |
c7f21978 | 1858 | |
26103436 FK |
1859 | if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, |
1860 | &pdd->qpd)) | |
1861 | pr_err("Failed to restore queues\n"); | |
1862 | ||
1863 | n_evicted--; | |
1864 | } | |
1865 | ||
1866 | return r; | |
1867 | } | |
1868 | ||
0aeaaf64 | 1869 | /* kfd_process_restore_queues - Restore all user queues of a process */ |
6b95e797 | 1870 | int kfd_process_restore_queues(struct kfd_process *p) |
26103436 | 1871 | { |
26103436 | 1872 | int r, ret = 0; |
6ae27841 AS |
1873 | int i; |
1874 | ||
1875 | for (i = 0; i < p->n_pdds; i++) { | |
1876 | struct kfd_process_device *pdd = p->pdds[i]; | |
26103436 | 1877 | |
d6e924ad | 1878 | kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid); |
c7f21978 | 1879 | |
26103436 FK |
1880 | r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm, |
1881 | &pdd->qpd); | |
1882 | if (r) { | |
1883 | pr_err("Failed to restore process queues\n"); | |
1884 | if (!ret) | |
1885 | ret = r; | |
1886 | } | |
1887 | } | |
1888 | ||
1889 | return ret; | |
1890 | } | |
1891 | ||
2aeb742b AS |
1892 | int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id) |
1893 | { | |
1894 | int i; | |
1895 | ||
1896 | for (i = 0; i < p->n_pdds; i++) | |
d763d803 | 1897 | if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id) |
2aeb742b AS |
1898 | return i; |
1899 | return -EINVAL; | |
1900 | } | |
1901 | ||
cda0f85b | 1902 | int |
5fb34bd9 AS |
1903 | kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node, |
1904 | uint32_t *gpuid, uint32_t *gpuidx) | |
cda0f85b | 1905 | { |
cda0f85b FK |
1906 | int i; |
1907 | ||
1908 | for (i = 0; i < p->n_pdds; i++) | |
5fb34bd9 | 1909 | if (p->pdds[i] && p->pdds[i]->dev == node) { |
d763d803 | 1910 | *gpuid = p->pdds[i]->user_gpu_id; |
cda0f85b FK |
1911 | *gpuidx = i; |
1912 | return 0; | |
1913 | } | |
1914 | return -EINVAL; | |
1915 | } | |
1916 | ||
9a1c1339 FK |
1917 | static int signal_eviction_fence(struct kfd_process *p) |
1918 | { | |
1919 | struct dma_fence *ef; | |
1920 | int ret; | |
1921 | ||
1922 | rcu_read_lock(); | |
1923 | ef = dma_fence_get_rcu_safe(&p->ef); | |
1924 | rcu_read_unlock(); | |
37865e02 FK |
1925 | if (!ef) |
1926 | return -EINVAL; | |
9a1c1339 FK |
1927 | |
1928 | ret = dma_fence_signal(ef); | |
1929 | dma_fence_put(ef); | |
1930 | ||
1931 | return ret; | |
1932 | } | |
1933 | ||
26103436 FK |
1934 | static void evict_process_worker(struct work_struct *work) |
1935 | { | |
1936 | int ret; | |
1937 | struct kfd_process *p; | |
1938 | struct delayed_work *dwork; | |
1939 | ||
1940 | dwork = to_delayed_work(work); | |
1941 | ||
1942 | /* Process termination destroys this worker thread. So during the | |
1943 | * lifetime of this thread, kfd_process p will be valid | |
1944 | */ | |
1945 | p = container_of(dwork, struct kfd_process, eviction_work); | |
26103436 | 1946 | |
6027b1bf | 1947 | pr_debug("Started evicting pasid 0x%x\n", p->pasid); |
c7f21978 | 1948 | ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM); |
26103436 | 1949 | if (!ret) { |
9a1c1339 FK |
1950 | /* If another thread already signaled the eviction fence, |
1951 | * they are responsible stopping the queues and scheduling | |
1952 | * the restore work. | |
1953 | */ | |
37865e02 FK |
1954 | if (signal_eviction_fence(p) || |
1955 | mod_delayed_work(kfd_restore_wq, &p->restore_work, | |
1956 | msecs_to_jiffies(PROCESS_RESTORE_TIME_MS))) | |
9a1c1339 | 1957 | kfd_process_restore_queues(p); |
26103436 | 1958 | |
6027b1bf | 1959 | pr_debug("Finished evicting pasid 0x%x\n", p->pasid); |
26103436 | 1960 | } else |
6027b1bf | 1961 | pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid); |
26103436 FK |
1962 | } |
1963 | ||
9a1c1339 FK |
1964 | static int restore_process_helper(struct kfd_process *p) |
1965 | { | |
1966 | int ret = 0; | |
1967 | ||
1968 | /* VMs may not have been acquired yet during debugging. */ | |
1969 | if (p->kgd_process_info) { | |
1970 | ret = amdgpu_amdkfd_gpuvm_restore_process_bos( | |
1971 | p->kgd_process_info, &p->ef); | |
1972 | if (ret) | |
1973 | return ret; | |
1974 | } | |
1975 | ||
1976 | ret = kfd_process_restore_queues(p); | |
1977 | if (!ret) | |
1978 | pr_debug("Finished restoring pasid 0x%x\n", p->pasid); | |
1979 | else | |
1980 | pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid); | |
1981 | ||
1982 | return ret; | |
1983 | } | |
1984 | ||
26103436 FK |
1985 | static void restore_process_worker(struct work_struct *work) |
1986 | { | |
1987 | struct delayed_work *dwork; | |
1988 | struct kfd_process *p; | |
26103436 FK |
1989 | int ret = 0; |
1990 | ||
1991 | dwork = to_delayed_work(work); | |
1992 | ||
1993 | /* Process termination destroys this worker thread. So during the | |
1994 | * lifetime of this thread, kfd_process p will be valid | |
1995 | */ | |
1996 | p = container_of(dwork, struct kfd_process, restore_work); | |
6027b1bf | 1997 | pr_debug("Started restoring pasid 0x%x\n", p->pasid); |
26103436 FK |
1998 | |
1999 | /* Setting last_restore_timestamp before successful restoration. | |
2000 | * Otherwise this would have to be set by KGD (restore_process_bos) | |
2001 | * before KFD BOs are unreserved. If not, the process can be evicted | |
2002 | * again before the timestamp is set. | |
2003 | * If restore fails, the timestamp will be set again in the next | |
2004 | * attempt. This would mean that the minimum GPU quanta would be | |
2005 | * PROCESS_ACTIVE_TIME_MS - (time to execute the following two | |
2006 | * functions) | |
2007 | */ | |
2008 | ||
2009 | p->last_restore_timestamp = get_jiffies_64(); | |
9a1c1339 FK |
2010 | |
2011 | ret = restore_process_helper(p); | |
26103436 | 2012 | if (ret) { |
6027b1bf | 2013 | pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n", |
26103436 | 2014 | p->pasid, PROCESS_BACK_OFF_TIME_MS); |
e26305f3 FK |
2015 | if (mod_delayed_work(kfd_restore_wq, &p->restore_work, |
2016 | msecs_to_jiffies(PROCESS_RESTORE_TIME_MS))) | |
2017 | kfd_process_restore_queues(p); | |
26103436 | 2018 | } |
26103436 FK |
2019 | } |
2020 | ||
2021 | void kfd_suspend_all_processes(void) | |
2022 | { | |
2023 | struct kfd_process *p; | |
2024 | unsigned int temp; | |
2025 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
2026 | ||
8a491bb3 | 2027 | WARN(debug_evictions, "Evicting all processes"); |
26103436 | 2028 | hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { |
c7f21978 | 2029 | if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND)) |
6027b1bf | 2030 | pr_err("Failed to suspend process 0x%x\n", p->pasid); |
9a1c1339 | 2031 | signal_eviction_fence(p); |
26103436 FK |
2032 | } |
2033 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
2034 | } | |
2035 | ||
2036 | int kfd_resume_all_processes(void) | |
2037 | { | |
2038 | struct kfd_process *p; | |
2039 | unsigned int temp; | |
2040 | int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu); | |
2041 | ||
2042 | hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { | |
9a1c1339 | 2043 | if (restore_process_helper(p)) { |
26103436 FK |
2044 | pr_err("Restore process %d failed during resume\n", |
2045 | p->pasid); | |
2046 | ret = -EFAULT; | |
2047 | } | |
2048 | } | |
2049 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
2050 | return ret; | |
2051 | } | |
2052 | ||
8dc1db31 | 2053 | int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process, |
373d7080 FK |
2054 | struct vm_area_struct *vma) |
2055 | { | |
373d7080 FK |
2056 | struct kfd_process_device *pdd; |
2057 | struct qcm_process_device *qpd; | |
2058 | ||
373d7080 FK |
2059 | if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) { |
2060 | pr_err("Incorrect CWSR mapping size.\n"); | |
2061 | return -EINVAL; | |
2062 | } | |
2063 | ||
2064 | pdd = kfd_get_process_device_data(dev, process); | |
2065 | if (!pdd) | |
2066 | return -EINVAL; | |
2067 | qpd = &pdd->qpd; | |
2068 | ||
2069 | qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, | |
2070 | get_order(KFD_CWSR_TBA_TMA_SIZE)); | |
2071 | if (!qpd->cwsr_kaddr) { | |
2072 | pr_err("Error allocating per process CWSR buffer.\n"); | |
2073 | return -ENOMEM; | |
2074 | } | |
2075 | ||
1c71222e SB |
2076 | vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND |
2077 | | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP); | |
373d7080 FK |
2078 | /* Mapping pages to user process */ |
2079 | return remap_pfn_range(vma, vma->vm_start, | |
2080 | PFN_DOWN(__pa(qpd->cwsr_kaddr)), | |
2081 | KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot); | |
2082 | } | |
851a645e | 2083 | |
12fb1ad7 JK |
2084 | /* assumes caller holds process lock. */ |
2085 | int kfd_process_drain_interrupts(struct kfd_process_device *pdd) | |
2086 | { | |
2087 | uint32_t irq_drain_fence[8]; | |
d4300362 | 2088 | uint8_t node_id = 0; |
12fb1ad7 JK |
2089 | int r = 0; |
2090 | ||
2091 | if (!KFD_IS_SOC15(pdd->dev)) | |
2092 | return 0; | |
2093 | ||
2094 | pdd->process->irq_drain_is_open = true; | |
2095 | ||
2096 | memset(irq_drain_fence, 0, sizeof(irq_drain_fence)); | |
2097 | irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) | | |
2098 | KFD_IRQ_FENCE_CLIENTID; | |
2099 | irq_drain_fence[3] = pdd->process->pasid; | |
2100 | ||
d4300362 MJ |
2101 | /* |
2102 | * For GFX 9.4.3, send the NodeId also in IH cookie DW[3] | |
2103 | */ | |
2104 | if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3)) { | |
2105 | node_id = ffs(pdd->dev->interrupt_bitmap) - 1; | |
2106 | irq_drain_fence[3] |= node_id << 16; | |
2107 | } | |
2108 | ||
12fb1ad7 JK |
2109 | /* ensure stale irqs scheduled KFD interrupts and send drain fence. */ |
2110 | if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev, | |
2111 | irq_drain_fence)) { | |
2112 | pdd->process->irq_drain_is_open = false; | |
2113 | return 0; | |
2114 | } | |
2115 | ||
2116 | r = wait_event_interruptible(pdd->process->wait_irq_drain, | |
2117 | !READ_ONCE(pdd->process->irq_drain_is_open)); | |
2118 | if (r) | |
2119 | pdd->process->irq_drain_is_open = false; | |
2120 | ||
2121 | return r; | |
2122 | } | |
2123 | ||
2124 | void kfd_process_close_interrupt_drain(unsigned int pasid) | |
2125 | { | |
2126 | struct kfd_process *p; | |
2127 | ||
2128 | p = kfd_lookup_process_by_pasid(pasid); | |
2129 | ||
2130 | if (!p) | |
2131 | return; | |
2132 | ||
2133 | WRITE_ONCE(p->irq_drain_is_open, false); | |
2134 | wake_up_all(&p->wait_irq_drain); | |
2135 | kfd_unref_process(p); | |
2136 | } | |
2137 | ||
c2d2588c JK |
2138 | struct send_exception_work_handler_workarea { |
2139 | struct work_struct work; | |
2140 | struct kfd_process *p; | |
2141 | unsigned int queue_id; | |
2142 | uint64_t error_reason; | |
2143 | }; | |
2144 | ||
2145 | static void send_exception_work_handler(struct work_struct *work) | |
2146 | { | |
2147 | struct send_exception_work_handler_workarea *workarea; | |
2148 | struct kfd_process *p; | |
2149 | struct queue *q; | |
2150 | struct mm_struct *mm; | |
2151 | struct kfd_context_save_area_header __user *csa_header; | |
2152 | uint64_t __user *err_payload_ptr; | |
2153 | uint64_t cur_err; | |
2154 | uint32_t ev_id; | |
2155 | ||
2156 | workarea = container_of(work, | |
2157 | struct send_exception_work_handler_workarea, | |
2158 | work); | |
2159 | p = workarea->p; | |
2160 | ||
2161 | mm = get_task_mm(p->lead_thread); | |
2162 | ||
2163 | if (!mm) | |
2164 | return; | |
2165 | ||
2166 | kthread_use_mm(mm); | |
2167 | ||
2168 | q = pqm_get_user_queue(&p->pqm, workarea->queue_id); | |
2169 | ||
2170 | if (!q) | |
2171 | goto out; | |
2172 | ||
2173 | csa_header = (void __user *)q->properties.ctx_save_restore_area_address; | |
2174 | ||
2175 | get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr); | |
2176 | get_user(cur_err, err_payload_ptr); | |
2177 | cur_err |= workarea->error_reason; | |
2178 | put_user(cur_err, err_payload_ptr); | |
2179 | get_user(ev_id, &csa_header->err_event_id); | |
2180 | ||
2181 | kfd_set_event(p, ev_id); | |
2182 | ||
2183 | out: | |
2184 | kthread_unuse_mm(mm); | |
2185 | mmput(mm); | |
2186 | } | |
2187 | ||
2188 | int kfd_send_exception_to_runtime(struct kfd_process *p, | |
2189 | unsigned int queue_id, | |
2190 | uint64_t error_reason) | |
2191 | { | |
2192 | struct send_exception_work_handler_workarea worker; | |
2193 | ||
2194 | INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler); | |
2195 | ||
2196 | worker.p = p; | |
2197 | worker.queue_id = queue_id; | |
2198 | worker.error_reason = error_reason; | |
2199 | ||
2200 | schedule_work(&worker.work); | |
2201 | flush_work(&worker.work); | |
2202 | destroy_work_on_stack(&worker.work); | |
2203 | ||
2204 | return 0; | |
2205 | } | |
2206 | ||
bef153b7 DYS |
2207 | struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id) |
2208 | { | |
2209 | int i; | |
2210 | ||
2211 | if (gpu_id) { | |
2212 | for (i = 0; i < p->n_pdds; i++) { | |
2213 | struct kfd_process_device *pdd = p->pdds[i]; | |
2214 | ||
2215 | if (pdd->user_gpu_id == gpu_id) | |
2216 | return pdd; | |
2217 | } | |
2218 | } | |
2219 | return NULL; | |
2220 | } | |
2221 | ||
2222 | int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id) | |
2223 | { | |
2224 | int i; | |
2225 | ||
2226 | if (!actual_gpu_id) | |
2227 | return 0; | |
2228 | ||
2229 | for (i = 0; i < p->n_pdds; i++) { | |
2230 | struct kfd_process_device *pdd = p->pdds[i]; | |
2231 | ||
2232 | if (pdd->dev->id == actual_gpu_id) | |
2233 | return pdd->user_gpu_id; | |
2234 | } | |
2235 | return -EINVAL; | |
2236 | } | |
2237 | ||
851a645e FK |
2238 | #if defined(CONFIG_DEBUG_FS) |
2239 | ||
2240 | int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data) | |
2241 | { | |
2242 | struct kfd_process *p; | |
2243 | unsigned int temp; | |
2244 | int r = 0; | |
2245 | ||
2246 | int idx = srcu_read_lock(&kfd_processes_srcu); | |
2247 | ||
2248 | hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) { | |
6027b1bf | 2249 | seq_printf(m, "Process %d PASID 0x%x:\n", |
851a645e FK |
2250 | p->lead_thread->tgid, p->pasid); |
2251 | ||
2252 | mutex_lock(&p->mutex); | |
2253 | r = pqm_debugfs_mqds(m, &p->pqm); | |
2254 | mutex_unlock(&p->mutex); | |
2255 | ||
2256 | if (r) | |
2257 | break; | |
2258 | } | |
2259 | ||
2260 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
2261 | ||
2262 | return r; | |
2263 | } | |
2264 | ||
2265 | #endif |