Loading...
1/*
2 * Copyright 2020 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23#include <linux/poll.h>
24#include <linux/wait.h>
25#include <linux/anon_inodes.h>
26#include <uapi/linux/kfd_ioctl.h>
27#include "amdgpu_vm.h"
28#include "kfd_priv.h"
29#include "kfd_smi_events.h"
30
31struct kfd_smi_client {
32 struct list_head list;
33 struct kfifo fifo;
34 wait_queue_head_t wait_queue;
35 /* events enabled */
36 uint64_t events;
37 struct kfd_dev *dev;
38 spinlock_t lock;
39};
40
41#define MAX_KFIFO_SIZE 1024
42
43static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
44static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
45static ssize_t kfd_smi_ev_write(struct file *, const char __user *, size_t,
46 loff_t *);
47static int kfd_smi_ev_release(struct inode *, struct file *);
48
49static const char kfd_smi_name[] = "kfd_smi_ev";
50
51static const struct file_operations kfd_smi_ev_fops = {
52 .owner = THIS_MODULE,
53 .poll = kfd_smi_ev_poll,
54 .read = kfd_smi_ev_read,
55 .write = kfd_smi_ev_write,
56 .release = kfd_smi_ev_release
57};
58
59static __poll_t kfd_smi_ev_poll(struct file *filep,
60 struct poll_table_struct *wait)
61{
62 struct kfd_smi_client *client = filep->private_data;
63 __poll_t mask = 0;
64
65 poll_wait(filep, &client->wait_queue, wait);
66
67 spin_lock(&client->lock);
68 if (!kfifo_is_empty(&client->fifo))
69 mask = EPOLLIN | EPOLLRDNORM;
70 spin_unlock(&client->lock);
71
72 return mask;
73}
74
75static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,
76 size_t size, loff_t *offset)
77{
78 int ret;
79 size_t to_copy;
80 struct kfd_smi_client *client = filep->private_data;
81 unsigned char *buf;
82
83 buf = kmalloc(MAX_KFIFO_SIZE * sizeof(*buf), GFP_KERNEL);
84 if (!buf)
85 return -ENOMEM;
86
87 /* kfifo_to_user can sleep so we can't use spinlock protection around
88 * it. Instead, we kfifo out as spinlocked then copy them to the user.
89 */
90 spin_lock(&client->lock);
91 to_copy = kfifo_len(&client->fifo);
92 if (!to_copy) {
93 spin_unlock(&client->lock);
94 ret = -EAGAIN;
95 goto ret_err;
96 }
97 to_copy = min3(size, sizeof(buf), to_copy);
98 ret = kfifo_out(&client->fifo, buf, to_copy);
99 spin_unlock(&client->lock);
100 if (ret <= 0) {
101 ret = -EAGAIN;
102 goto ret_err;
103 }
104
105 ret = copy_to_user(user, buf, to_copy);
106 if (ret) {
107 ret = -EFAULT;
108 goto ret_err;
109 }
110
111 kfree(buf);
112 return to_copy;
113
114ret_err:
115 kfree(buf);
116 return ret;
117}
118
119static ssize_t kfd_smi_ev_write(struct file *filep, const char __user *user,
120 size_t size, loff_t *offset)
121{
122 struct kfd_smi_client *client = filep->private_data;
123 uint64_t events;
124
125 if (!access_ok(user, size) || size < sizeof(events))
126 return -EFAULT;
127 if (copy_from_user(&events, user, sizeof(events)))
128 return -EFAULT;
129
130 WRITE_ONCE(client->events, events);
131
132 return sizeof(events);
133}
134
135static int kfd_smi_ev_release(struct inode *inode, struct file *filep)
136{
137 struct kfd_smi_client *client = filep->private_data;
138 struct kfd_dev *dev = client->dev;
139
140 spin_lock(&dev->smi_lock);
141 list_del_rcu(&client->list);
142 spin_unlock(&dev->smi_lock);
143
144 synchronize_rcu();
145 kfifo_free(&client->fifo);
146 kfree(client);
147
148 return 0;
149}
150
151void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid)
152{
153 struct amdgpu_device *adev = (struct amdgpu_device *)dev->kgd;
154 struct amdgpu_task_info task_info;
155 /* VmFault msg = (hex)uint32_pid(8) + :(1) + task name(16) = 25 */
156 /* 16 bytes event + 1 byte space + 25 bytes msg + 1 byte \n = 43
157 */
158 char fifo_in[43];
159 struct kfd_smi_client *client;
160 int len;
161
162 if (list_empty(&dev->smi_clients))
163 return;
164
165 memset(&task_info, 0, sizeof(struct amdgpu_task_info));
166 amdgpu_vm_get_task_info(adev, pasid, &task_info);
167 /* Report VM faults from user applications, not retry from kernel */
168 if (!task_info.pid)
169 return;
170
171 len = snprintf(fifo_in, 43, "%x %x:%s\n", KFD_SMI_EVENT_VMFAULT,
172 task_info.pid, task_info.task_name);
173
174 rcu_read_lock();
175
176 list_for_each_entry_rcu(client, &dev->smi_clients, list) {
177 if (!(READ_ONCE(client->events) & KFD_SMI_EVENT_VMFAULT))
178 continue;
179 spin_lock(&client->lock);
180 if (kfifo_avail(&client->fifo) >= len) {
181 kfifo_in(&client->fifo, fifo_in, len);
182 wake_up_all(&client->wait_queue);
183 }
184 else
185 pr_debug("smi_event(vmfault): no space left\n");
186 spin_unlock(&client->lock);
187 }
188
189 rcu_read_unlock();
190}
191
192int kfd_smi_event_open(struct kfd_dev *dev, uint32_t *fd)
193{
194 struct kfd_smi_client *client;
195 int ret;
196
197 client = kzalloc(sizeof(struct kfd_smi_client), GFP_KERNEL);
198 if (!client)
199 return -ENOMEM;
200 INIT_LIST_HEAD(&client->list);
201
202 ret = kfifo_alloc(&client->fifo, MAX_KFIFO_SIZE, GFP_KERNEL);
203 if (ret) {
204 kfree(client);
205 return ret;
206 }
207
208 ret = anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops, (void *)client,
209 O_RDWR);
210 if (ret < 0) {
211 kfifo_free(&client->fifo);
212 kfree(client);
213 return ret;
214 }
215 *fd = ret;
216
217 init_waitqueue_head(&client->wait_queue);
218 spin_lock_init(&client->lock);
219 client->events = 0;
220 client->dev = dev;
221
222 spin_lock(&dev->smi_lock);
223 list_add_rcu(&client->list, &dev->smi_clients);
224 spin_unlock(&dev->smi_lock);
225
226 return 0;
227}
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
34struct 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
49static __poll_t kfd_smi_ev_poll(struct file *, struct poll_table_struct *);
50static ssize_t kfd_smi_ev_read(struct file *, char __user *, size_t, loff_t *);
51static ssize_t kfd_smi_ev_write(struct file *, const char __user *, size_t,
52 loff_t *);
53static int kfd_smi_ev_release(struct inode *, struct file *);
54
55static const char kfd_smi_name[] = "kfd_smi_ev";
56
57static 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
65static __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
81static 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
121ret_err:
122 kfree(buf);
123 return ret;
124}
125
126static 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
142static 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
150static 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
163static bool kfd_smi_ev_enabled(pid_t pid, struct kfd_smi_client *client,
164 unsigned int event)
165{
166 uint64_t all = KFD_SMI_EVENT_MASK_FROM_INDEX(KFD_SMI_EVENT_ALL_PROCESS);
167 uint64_t events = READ_ONCE(client->events);
168
169 if (pid && client->pid != pid && !(client->suser && (events & all)))
170 return false;
171
172 return events & KFD_SMI_EVENT_MASK_FROM_INDEX(event);
173}
174
175static void add_event_to_kfifo(pid_t pid, struct kfd_node *dev,
176 unsigned int smi_event, char *event_msg, int len)
177{
178 struct kfd_smi_client *client;
179
180 rcu_read_lock();
181
182 list_for_each_entry_rcu(client, &dev->smi_clients, list) {
183 if (!kfd_smi_ev_enabled(pid, client, smi_event))
184 continue;
185 spin_lock(&client->lock);
186 if (kfifo_avail(&client->fifo) >= len) {
187 kfifo_in(&client->fifo, event_msg, len);
188 wake_up_all(&client->wait_queue);
189 } else {
190 pr_debug("smi_event(EventID: %u): no space left\n",
191 smi_event);
192 }
193 spin_unlock(&client->lock);
194 }
195
196 rcu_read_unlock();
197}
198
199__printf(4, 5)
200static void kfd_smi_event_add(pid_t pid, struct kfd_node *dev,
201 unsigned int event, char *fmt, ...)
202{
203 char fifo_in[KFD_SMI_EVENT_MSG_SIZE];
204 int len;
205 va_list args;
206
207 if (list_empty(&dev->smi_clients))
208 return;
209
210 len = snprintf(fifo_in, sizeof(fifo_in), "%x ", event);
211
212 va_start(args, fmt);
213 len += vsnprintf(fifo_in + len, sizeof(fifo_in) - len, fmt, args);
214 va_end(args);
215
216 add_event_to_kfifo(pid, dev, event, fifo_in, len);
217}
218
219void kfd_smi_event_update_gpu_reset(struct kfd_node *dev, bool post_reset,
220 struct amdgpu_reset_context *reset_context)
221{
222 unsigned int event;
223 char reset_cause[64];
224
225 if (post_reset) {
226 event = KFD_SMI_EVENT_GPU_POST_RESET;
227 } else {
228 event = KFD_SMI_EVENT_GPU_PRE_RESET;
229 ++(dev->reset_seq_num);
230 }
231
232 memset(reset_cause, 0, sizeof(reset_cause));
233
234 if (reset_context)
235 amdgpu_reset_get_desc(reset_context, reset_cause,
236 sizeof(reset_cause));
237
238 kfd_smi_event_add(0, dev, event, KFD_EVENT_FMT_UPDATE_GPU_RESET(
239 dev->reset_seq_num, reset_cause));
240}
241
242void kfd_smi_event_update_thermal_throttling(struct kfd_node *dev,
243 uint64_t throttle_bitmask)
244{
245 kfd_smi_event_add(0, dev, KFD_SMI_EVENT_THERMAL_THROTTLE, KFD_EVENT_FMT_THERMAL_THROTTLING(
246 throttle_bitmask,
247 amdgpu_dpm_get_thermal_throttling_counter(dev->adev)));
248}
249
250void kfd_smi_event_update_vmfault(struct kfd_node *dev, uint16_t pasid)
251{
252 struct amdgpu_task_info *task_info;
253
254 task_info = amdgpu_vm_get_task_info_pasid(dev->adev, pasid);
255 if (task_info) {
256 /* Report VM faults from user applications, not retry from kernel */
257 if (task_info->pid)
258 kfd_smi_event_add(0, dev, KFD_SMI_EVENT_VMFAULT, KFD_EVENT_FMT_VMFAULT(
259 task_info->pid, task_info->task_name));
260 amdgpu_vm_put_task_info(task_info);
261 }
262}
263
264void kfd_smi_event_page_fault_start(struct kfd_node *node, pid_t pid,
265 unsigned long address, bool write_fault,
266 ktime_t ts)
267{
268 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_START,
269 KFD_EVENT_FMT_PAGEFAULT_START(ktime_to_ns(ts), pid,
270 address, node->id, write_fault ? 'W' : 'R'));
271}
272
273void kfd_smi_event_page_fault_end(struct kfd_node *node, pid_t pid,
274 unsigned long address, bool migration)
275{
276 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_PAGE_FAULT_END,
277 KFD_EVENT_FMT_PAGEFAULT_END(ktime_get_boottime_ns(),
278 pid, address, node->id, migration ? 'M' : 'U'));
279}
280
281void kfd_smi_event_migration_start(struct kfd_node *node, pid_t pid,
282 unsigned long start, unsigned long end,
283 uint32_t from, uint32_t to,
284 uint32_t prefetch_loc, uint32_t preferred_loc,
285 uint32_t trigger)
286{
287 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_START,
288 KFD_EVENT_FMT_MIGRATE_START(
289 ktime_get_boottime_ns(), pid, start, end - start,
290 from, to, prefetch_loc, preferred_loc, trigger));
291}
292
293void kfd_smi_event_migration_end(struct kfd_node *node, pid_t pid,
294 unsigned long start, unsigned long end,
295 uint32_t from, uint32_t to, uint32_t trigger,
296 int error_code)
297{
298 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_MIGRATE_END,
299 KFD_EVENT_FMT_MIGRATE_END(
300 ktime_get_boottime_ns(), pid, start, end - start,
301 from, to, trigger, error_code));
302}
303
304void kfd_smi_event_queue_eviction(struct kfd_node *node, pid_t pid,
305 uint32_t trigger)
306{
307 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_EVICTION,
308 KFD_EVENT_FMT_QUEUE_EVICTION(ktime_get_boottime_ns(), pid,
309 node->id, trigger));
310}
311
312void kfd_smi_event_queue_restore(struct kfd_node *node, pid_t pid)
313{
314 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_QUEUE_RESTORE,
315 KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(), pid,
316 node->id, 0));
317}
318
319void kfd_smi_event_queue_restore_rescheduled(struct mm_struct *mm)
320{
321 struct kfd_process *p;
322 int i;
323
324 p = kfd_lookup_process_by_mm(mm);
325 if (!p)
326 return;
327
328 for (i = 0; i < p->n_pdds; i++) {
329 struct kfd_process_device *pdd = p->pdds[i];
330
331 kfd_smi_event_add(p->lead_thread->pid, pdd->dev,
332 KFD_SMI_EVENT_QUEUE_RESTORE,
333 KFD_EVENT_FMT_QUEUE_RESTORE(ktime_get_boottime_ns(),
334 p->lead_thread->pid, pdd->dev->id, 'R'));
335 }
336 kfd_unref_process(p);
337}
338
339void kfd_smi_event_unmap_from_gpu(struct kfd_node *node, pid_t pid,
340 unsigned long address, unsigned long last,
341 uint32_t trigger)
342{
343 kfd_smi_event_add(pid, node, KFD_SMI_EVENT_UNMAP_FROM_GPU,
344 KFD_EVENT_FMT_UNMAP_FROM_GPU(ktime_get_boottime_ns(),
345 pid, address, last - address + 1, node->id, trigger));
346}
347
348int kfd_smi_event_open(struct kfd_node *dev, uint32_t *fd)
349{
350 struct kfd_smi_client *client;
351 int ret;
352
353 client = kzalloc(sizeof(struct kfd_smi_client), GFP_KERNEL);
354 if (!client)
355 return -ENOMEM;
356 INIT_LIST_HEAD(&client->list);
357
358 ret = kfifo_alloc(&client->fifo, KFD_MAX_KFIFO_SIZE, GFP_KERNEL);
359 if (ret) {
360 kfree(client);
361 return ret;
362 }
363
364 init_waitqueue_head(&client->wait_queue);
365 spin_lock_init(&client->lock);
366 client->events = 0;
367 client->dev = dev;
368 client->pid = current->tgid;
369 client->suser = capable(CAP_SYS_ADMIN);
370
371 spin_lock(&dev->smi_lock);
372 list_add_rcu(&client->list, &dev->smi_clients);
373 spin_unlock(&dev->smi_lock);
374
375 ret = anon_inode_getfd(kfd_smi_name, &kfd_smi_ev_fops, (void *)client,
376 O_RDWR);
377 if (ret < 0) {
378 spin_lock(&dev->smi_lock);
379 list_del_rcu(&client->list);
380 spin_unlock(&dev->smi_lock);
381
382 synchronize_rcu();
383
384 kfifo_free(&client->fifo);
385 kfree(client);
386 return ret;
387 }
388 *fd = ret;
389
390 return 0;
391}