xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c (revision e78f70bad29c5ae1e1076698b690b15794e9b81e)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2020-2022 Advanced Micro Devices, Inc.
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/poll.h>
25 #include <linux/wait.h>
26 #include <linux/anon_inodes.h>
27 #include <uapi/linux/kfd_ioctl.h>
28 #include "amdgpu.h"
29 #include "amdgpu_vm.h"
30 #include "kfd_priv.h"
31 #include "kfd_smi_events.h"
32 #include "amdgpu_reset.h"
33 
34 struct kfd_smi_client {
35 	struct list_head list;
36 	struct kfifo fifo;
37 	wait_queue_head_t wait_queue;
38 	/* events enabled */
39 	uint64_t events;
40 	struct kfd_node *dev;
41 	spinlock_t lock;
42 	struct rcu_head rcu;
43 	pid_t pid;
44 	bool suser;
45 };
46 
47 #define KFD_MAX_KFIFO_SIZE	8192
48 
49 static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
50 static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
51 static ssize_t kfd_smi_ev_write(struct file *, const char __user *, size_t,
52 				loff_t *);
53 static int kfd_smi_ev_release(struct inode *, struct file *);
54 
55 static const char kfd_smi_name[] = "kfd_smi_ev";
56 
57 static const struct file_operations kfd_smi_ev_fops = {
58 	.owner = THIS_MODULE,
59 	.poll = kfd_smi_ev_poll,
60 	.read = kfd_smi_ev_read,
61 	.write = kfd_smi_ev_write,
62 	.release = kfd_smi_ev_release
63 };
64 
65 static __poll_t kfd_smi_ev_poll(struct file *filep,
66 				struct poll_table_struct *wait)
67 {
68 	struct kfd_smi_client *client = filep->private_data;
69 	__poll_t mask = 0;
70 
71 	poll_wait(filep, &client->wait_queue, wait);
72 
73 	spin_lock(&client->lock);
74 	if (!kfifo_is_empty(&client->fifo))
75 		mask = EPOLLIN | EPOLLRDNORM;
76 	spin_unlock(&client->lock);
77 
78 	return mask;
79 }
80 
81 static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
82 			       size_t size, loff_t *offset)
83 {
84 	int ret;
85 	size_t to_copy;
86 	struct kfd_smi_client *client = filep->private_data;
87 	unsigned char *buf;
88 
89 	size = min_t(size_t, size, KFD_MAX_KFIFO_SIZE);
90 	buf = kmalloc(size, GFP_KERNEL);
91 	if (!buf)
92 		return -ENOMEM;
93 
94 	/* kfifo_to_user can sleep so we can't use spinlock protection around
95 	 * it. Instead, we kfifo out as spinlocked then copy them to the user.
96 	 */
97 	spin_lock(&client->lock);
98 	to_copy = kfifo_len(&client->fifo);
99 	if (!to_copy) {
100 		spin_unlock(&client->lock);
101 		ret = -EAGAIN;
102 		goto ret_err;
103 	}
104 	to_copy = min(size, to_copy);
105 	ret = kfifo_out(&client->fifo, buf, to_copy);
106 	spin_unlock(&client->lock);
107 	if (ret <= 0) {
108 		ret = -EAGAIN;
109 		goto ret_err;
110 	}
111 
112 	ret = copy_to_user(user, buf, to_copy);
113 	if (ret) {
114 		ret = -EFAULT;
115 		goto ret_err;
116 	}
117 
118 	kfree(buf);
119 	return to_copy;
120 
121 ret_err:
122 	kfree(buf);
123 	return ret;
124 }
125 
126 static ssize_t kfd_smi_ev_write(struct file *filep, const char __user *user,
127 				size_t size, loff_t *offset)
128 {
129 	struct kfd_smi_client *client = filep->private_data;
130 	uint64_t events;
131 
132 	if (!access_ok(user, size) || size < sizeof(events))
133 		return -EFAULT;
134 	if (copy_from_user(&events, user, sizeof(events)))
135 		return -EFAULT;
136 
137 	WRITE_ONCE(client->events, events);
138 
139 	return sizeof(events);
140 }
141 
142 static void kfd_smi_ev_client_free(struct rcu_head *p)
143 {
144 	struct kfd_smi_client *ev = container_of(p, struct kfd_smi_client, rcu);
145 
146 	kfifo_free(&ev->fifo);
147 	kfree(ev);
148 }
149 
150 static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
151 {
152 	struct kfd_smi_client *client = filep->private_data;
153 	struct kfd_node *dev = client->dev;
154 
155 	spin_lock(&dev->smi_lock);
156 	list_del_rcu(&client->list);
157 	spin_unlock(&dev->smi_lock);
158 
159 	call_rcu(&client->rcu, kfd_smi_ev_client_free);
160 	return 0;
161 }
162 
163 static bool kfd_smi_ev_enabled(pid_t pid, struct kfd_smi_client *client,
164 			       unsigned int event)
165 {
166 	uint64_t events = READ_ONCE(client->events);
167 
168 	if (pid && client->pid != pid && !client->suser)
169 		return false;
170 
171 	return events & KFD_SMI_EVENT_MASK_FROM_INDEX(event);
172 }
173 
174 static void add_event_to_kfifo(pid_t pid, struct kfd_node *dev,
175 			       unsigned int smi_event, char *event_msg, int len)
176 {
177 	struct kfd_smi_client *client;
178 
179 	rcu_read_lock();
180 
181 	list_for_each_entry_rcu(client, &dev->smi_clients, list) {
182 		if (!kfd_smi_ev_enabled(pid, client, smi_event))
183 			continue;
184 		spin_lock(&client->lock);
185 		if (kfifo_avail(&client->fifo) >= len) {
186 			kfifo_in(&client->fifo, event_msg, len);
187 			wake_up_all(&client->wait_queue);
188 		} else {
189 			pr_debug("smi_event(EventID: %u): no space left\n",
190 					smi_event);
191 		}
192 		spin_unlock(&client->lock);
193 	}
194 
195 	rcu_read_unlock();
196 }
197 
198 __printf(4, 5)
199 static void kfd_smi_event_add(pid_t pid, struct kfd_node *dev,
200 			      unsigned int event, char *fmt, ...)
201 {
202 	char fifo_in[KFD_SMI_EVENT_MSG_SIZE];
203 	int len;
204 	va_list args;
205 
206 	if (list_empty(&dev->smi_clients))
207 		return;
208 
209 	len = snprintf(fifo_in, sizeof(fifo_in), "%x ", event);
210 
211 	va_start(args, fmt);
212 	len += vsnprintf(fifo_in + len, sizeof(fifo_in) - len, fmt, args);
213 	va_end(args);
214 
215 	add_event_to_kfifo(pid, dev, event, fifo_in, len);
216 }
217 
218 void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset,
219 				    struct amdgpu_reset_context *reset_context)
220 {
221 	unsigned int event;
222 	char reset_cause[64];
223 
224 	if (post_reset) {
225 		event = KFD_SMI_EVENT_GPU_POST_RESET;
226 	} else {
227 		event = KFD_SMI_EVENT_GPU_PRE_RESET;
228 		++(dev->reset_seq_num);
229 	}
230 
231 	memset(reset_cause, 0, sizeof(reset_cause));
232 
233 	if (reset_context)
234 		amdgpu_reset_get_desc(reset_context, reset_cause,
235 				      sizeof(reset_cause));
236 
237 	kfd_smi_event_add(0, dev, event, KFD_EVENT_FMT_UPDATE_GPU_RESET(
238 			  dev->reset_seq_num, reset_cause));
239 }
240 
241 void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev,
242 					     uint64_t throttle_bitmask)
243 {
244 	kfd_smi_event_add(0, dev, KFD_SMI_EVENT_THERMAL_THROTTLE, KFD_EVENT_FMT_THERMAL_THROTTLING(
245 			  throttle_bitmask,
246 			  amdgpu_dpm_get_thermal_throttling_counter(dev->adev)));
247 }
248 
249 void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid)
250 {
251 	struct amdgpu_task_info *task_info;
252 
253 	task_info = amdgpu_vm_get_task_info_pasid(dev->adev, pasid);
254 	if (task_info) {
255 		/* Report VM faults from user applications, not retry from kernel */
256 		if (task_info->pid)
257 			kfd_smi_event_add(0, dev, KFD_SMI_EVENT_VMFAULT, KFD_EVENT_FMT_VMFAULT(
258 					  task_info->pid, task_info->task_name));
259 		amdgpu_vm_put_task_info(task_info);
260 	}
261 }
262 
263 void kfd_smi_event_page_fault_start(struct kfd_node *node, pid_t pid,
264 				    unsigned long address, bool write_fault,
265 				    ktime_t ts)
266 {
267 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_START,
268 			  KFD_EVENT_FMT_PAGEFAULT_START(ktime_to_ns(ts), pid,
269 			  address, node->id, write_fault ? 'W' : 'R'));
270 }
271 
272 void kfd_smi_event_page_fault_end(struct kfd_node *node, pid_t pid,
273 				  unsigned long address, bool migration)
274 {
275 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_END,
276 			  KFD_EVENT_FMT_PAGEFAULT_END(ktime_get_boottime_ns(),
277 			  pid, address, node->id, migration ? 'M' : 'U'));
278 }
279 
280 void kfd_smi_event_migration_start(struct kfd_node *node, pid_t pid,
281 				   unsigned long start, unsigned long end,
282 				   uint32_t from, uint32_t to,
283 				   uint32_t prefetch_loc, uint32_t preferred_loc,
284 				   uint32_t trigger)
285 {
286 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_START,
287 			  KFD_EVENT_FMT_MIGRATE_START(
288 			  ktime_get_boottime_ns(), pid, start, end - start,
289 			  from, to, prefetch_loc, preferred_loc, trigger));
290 }
291 
292 void kfd_smi_event_migration_end(struct kfd_node *node, pid_t pid,
293 				 unsigned long start, unsigned long end,
294 				 uint32_t from, uint32_t to, uint32_t trigger,
295 				 int error_code)
296 {
297 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_END,
298 			  KFD_EVENT_FMT_MIGRATE_END(
299 			  ktime_get_boottime_ns(), pid, start, end - start,
300 			  from, to, trigger, error_code));
301 }
302 
303 void kfd_smi_event_queue_eviction(struct kfd_node *node, pid_t pid,
304 				  uint32_t trigger)
305 {
306 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_EVICTION,
307 			  KFD_EVENT_FMT_QUEUE_EVICTION(ktime_get_boottime_ns(), pid,
308 			  node->id, trigger));
309 }
310 
311 void kfd_smi_event_queue_restore(struct kfd_node *node, pid_t pid)
312 {
313 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_RESTORE,
314 			  KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(), pid,
315 			  node->id, 0));
316 }
317 
318 void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm)
319 {
320 	struct kfd_process *p;
321 	int i;
322 
323 	p = kfd_lookup_process_by_mm(mm);
324 	if (!p)
325 		return;
326 
327 	for (i = 0; i < p->n_pdds; i++) {
328 		struct kfd_process_device *pdd = p->pdds[i];
329 
330 		kfd_smi_event_add(p->lead_thread->pid, pdd->dev,
331 				  KFD_SMI_EVENT_QUEUE_RESTORE,
332 				  KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(),
333 				  p->lead_thread->pid, pdd->dev->id, 'R'));
334 	}
335 	kfd_unref_process(p);
336 }
337 
338 void kfd_smi_event_unmap_from_gpu(struct kfd_node *node, pid_t pid,
339 				  unsigned long address, unsigned long last,
340 				  uint32_t trigger)
341 {
342 	kfd_smi_event_add(pid, node, KFD_SMI_EVENT_UNMAP_FROM_GPU,
343 			  KFD_EVENT_FMT_UNMAP_FROM_GPU(ktime_get_boottime_ns(),
344 			  pid, address, last - address + 1, node->id, trigger));
345 }
346 
347 void kfd_smi_event_process(struct kfd_process_device *pdd, bool start)
348 {
349 	struct amdgpu_task_info *task_info;
350 	struct amdgpu_vm *avm;
351 
352 	if (!pdd->drm_priv)
353 		return;
354 
355 	avm = drm_priv_to_vm(pdd->drm_priv);
356 	task_info = amdgpu_vm_get_task_info_vm(avm);
357 
358 	if (task_info) {
359 		kfd_smi_event_add(0, pdd->dev,
360 				  start ? KFD_SMI_EVENT_PROCESS_START :
361 				  KFD_SMI_EVENT_PROCESS_END,
362 				  KFD_EVENT_FMT_PROCESS(task_info->pid,
363 				  task_info->task_name));
364 		amdgpu_vm_put_task_info(task_info);
365 	}
366 }
367 
368 int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd)
369 {
370 	struct kfd_smi_client *client;
371 	int ret;
372 
373 	client = kzalloc(sizeof(struct kfd_smi_client), GFP_KERNEL);
374 	if (!client)
375 		return -ENOMEM;
376 	INIT_LIST_HEAD(&client->list);
377 
378 	ret = kfifo_alloc(&client->fifo, KFD_MAX_KFIFO_SIZE, GFP_KERNEL);
379 	if (ret) {
380 		kfree(client);
381 		return ret;
382 	}
383 
384 	init_waitqueue_head(&client->wait_queue);
385 	spin_lock_init(&client->lock);
386 	client->events = 0;
387 	client->dev = dev;
388 	client->pid = current->tgid;
389 	client->suser = capable(CAP_SYS_ADMIN);
390 
391 	spin_lock(&dev->smi_lock);
392 	list_add_rcu(&client->list, &dev->smi_clients);
393 	spin_unlock(&dev->smi_lock);
394 
395 	ret = anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops, (void *)client,
396 			       O_RDWR);
397 	if (ret < 0) {
398 		spin_lock(&dev->smi_lock);
399 		list_del_rcu(&client->list);
400 		spin_unlock(&dev->smi_lock);
401 
402 		synchronize_rcu();
403 
404 		kfifo_free(&client->fifo);
405 		kfree(client);
406 		return ret;
407 	}
408 	*fd = ret;
409 
410 	return 0;
411 }
412