1 /* $NetBSD: kfd_packet_manager.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $ */
2
3 /*
4 * Copyright 2014 Advanced Micro Devices, Inc.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26 #include <sys/cdefs.h>
27 __KERNEL_RCSID(0, "$NetBSD: kfd_packet_manager.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $");
28
29 #include <linux/slab.h>
30 #include <linux/mutex.h>
31 #include "kfd_device_queue_manager.h"
32 #include "kfd_kernel_queue.h"
33 #include "kfd_priv.h"
34
inc_wptr(unsigned int * wptr,unsigned int increment_bytes,unsigned int buffer_size_bytes)35 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
36 unsigned int buffer_size_bytes)
37 {
38 unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
39
40 WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
41 "Runlist IB overflow");
42 *wptr = temp;
43 }
44
pm_calc_rlib_size(struct packet_manager * pm,unsigned int * rlib_size,bool * over_subscription)45 static void pm_calc_rlib_size(struct packet_manager *pm,
46 unsigned int *rlib_size,
47 bool *over_subscription)
48 {
49 unsigned int process_count, queue_count, compute_queue_count;
50 unsigned int map_queue_size;
51 unsigned int max_proc_per_quantum = 1;
52 struct kfd_dev *dev = pm->dqm->dev;
53
54 process_count = pm->dqm->processes_count;
55 queue_count = pm->dqm->queue_count;
56 compute_queue_count = queue_count - pm->dqm->sdma_queue_count -
57 pm->dqm->xgmi_sdma_queue_count;
58
59 /* check if there is over subscription
60 * Note: the arbitration between the number of VMIDs and
61 * hws_max_conc_proc has been done in
62 * kgd2kfd_device_init().
63 */
64 *over_subscription = false;
65
66 if (dev->max_proc_per_quantum > 1)
67 max_proc_per_quantum = dev->max_proc_per_quantum;
68
69 if ((process_count > max_proc_per_quantum) ||
70 compute_queue_count > get_queues_num(pm->dqm)) {
71 *over_subscription = true;
72 pr_debug("Over subscribed runlist\n");
73 }
74
75 map_queue_size = pm->pmf->map_queues_size;
76 /* calculate run list ib allocation size */
77 *rlib_size = process_count * pm->pmf->map_process_size +
78 queue_count * map_queue_size;
79
80 /*
81 * Increase the allocation size in case we need a chained run list
82 * when over subscription
83 */
84 if (*over_subscription)
85 *rlib_size += pm->pmf->runlist_size;
86
87 pr_debug("runlist ib size %d\n", *rlib_size);
88 }
89
pm_allocate_runlist_ib(struct packet_manager * pm,unsigned int ** rl_buffer,uint64_t * rl_gpu_buffer,unsigned int * rl_buffer_size,bool * is_over_subscription)90 static int pm_allocate_runlist_ib(struct packet_manager *pm,
91 unsigned int **rl_buffer,
92 uint64_t *rl_gpu_buffer,
93 unsigned int *rl_buffer_size,
94 bool *is_over_subscription)
95 {
96 int retval;
97
98 if (WARN_ON(pm->allocated))
99 return -EINVAL;
100
101 pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
102
103 mutex_lock(&pm->lock);
104
105 retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
106 &pm->ib_buffer_obj);
107
108 if (retval) {
109 pr_err("Failed to allocate runlist IB\n");
110 goto out;
111 }
112
113 *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
114 *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
115
116 memset(*rl_buffer, 0, *rl_buffer_size);
117 pm->allocated = true;
118
119 out:
120 mutex_unlock(&pm->lock);
121 return retval;
122 }
123
pm_create_runlist_ib(struct packet_manager * pm,struct list_head * queues,uint64_t * rl_gpu_addr,size_t * rl_size_bytes)124 static int pm_create_runlist_ib(struct packet_manager *pm,
125 struct list_head *queues,
126 uint64_t *rl_gpu_addr,
127 size_t *rl_size_bytes)
128 {
129 unsigned int alloc_size_bytes;
130 unsigned int *rl_buffer, rl_wptr, i;
131 int retval, proccesses_mapped;
132 struct device_process_node *cur;
133 struct qcm_process_device *qpd;
134 struct queue *q;
135 struct kernel_queue *kq;
136 bool is_over_subscription;
137
138 rl_wptr = retval = proccesses_mapped = 0;
139
140 retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
141 &alloc_size_bytes, &is_over_subscription);
142 if (retval)
143 return retval;
144
145 *rl_size_bytes = alloc_size_bytes;
146 pm->ib_size_bytes = alloc_size_bytes;
147
148 pr_debug("Building runlist ib process count: %d queues count %d\n",
149 pm->dqm->processes_count, pm->dqm->queue_count);
150
151 /* build the run list ib packet */
152 list_for_each_entry(cur, queues, list) {
153 qpd = cur->qpd;
154 /* build map process packet */
155 if (proccesses_mapped >= pm->dqm->processes_count) {
156 pr_debug("Not enough space left in runlist IB\n");
157 pm_release_ib(pm);
158 return -ENOMEM;
159 }
160
161 retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
162 if (retval)
163 return retval;
164
165 proccesses_mapped++;
166 inc_wptr(&rl_wptr, pm->pmf->map_process_size,
167 alloc_size_bytes);
168
169 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
170 if (!kq->queue->properties.is_active)
171 continue;
172
173 pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
174 kq->queue->queue, qpd->is_debug);
175
176 retval = pm->pmf->map_queues(pm,
177 &rl_buffer[rl_wptr],
178 kq->queue,
179 qpd->is_debug);
180 if (retval)
181 return retval;
182
183 inc_wptr(&rl_wptr,
184 pm->pmf->map_queues_size,
185 alloc_size_bytes);
186 }
187
188 list_for_each_entry(q, &qpd->queues_list, list) {
189 if (!q->properties.is_active)
190 continue;
191
192 pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
193 q->queue, qpd->is_debug);
194
195 retval = pm->pmf->map_queues(pm,
196 &rl_buffer[rl_wptr],
197 q,
198 qpd->is_debug);
199
200 if (retval)
201 return retval;
202
203 inc_wptr(&rl_wptr,
204 pm->pmf->map_queues_size,
205 alloc_size_bytes);
206 }
207 }
208
209 pr_debug("Finished map process and queues to runlist\n");
210
211 if (is_over_subscription) {
212 if (!pm->is_over_subscription)
213 pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
214 retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
215 *rl_gpu_addr,
216 alloc_size_bytes / sizeof(uint32_t),
217 true);
218 }
219 pm->is_over_subscription = is_over_subscription;
220
221 for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
222 pr_debug("0x%2X ", rl_buffer[i]);
223 pr_debug("\n");
224
225 return retval;
226 }
227
pm_init(struct packet_manager * pm,struct device_queue_manager * dqm)228 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
229 {
230 switch (dqm->dev->device_info->asic_family) {
231 case CHIP_KAVERI:
232 case CHIP_HAWAII:
233 /* PM4 packet structures on CIK are the same as on VI */
234 case CHIP_CARRIZO:
235 case CHIP_TONGA:
236 case CHIP_FIJI:
237 case CHIP_POLARIS10:
238 case CHIP_POLARIS11:
239 case CHIP_POLARIS12:
240 case CHIP_VEGAM:
241 pm->pmf = &kfd_vi_pm_funcs;
242 break;
243 case CHIP_VEGA10:
244 case CHIP_VEGA12:
245 case CHIP_VEGA20:
246 case CHIP_RAVEN:
247 case CHIP_RENOIR:
248 case CHIP_ARCTURUS:
249 case CHIP_NAVI10:
250 case CHIP_NAVI12:
251 case CHIP_NAVI14:
252 pm->pmf = &kfd_v9_pm_funcs;
253 break;
254 default:
255 WARN(1, "Unexpected ASIC family %u",
256 dqm->dev->device_info->asic_family);
257 return -EINVAL;
258 }
259
260 pm->dqm = dqm;
261 mutex_init(&pm->lock);
262 pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
263 if (!pm->priv_queue) {
264 mutex_destroy(&pm->lock);
265 return -ENOMEM;
266 }
267 pm->allocated = false;
268
269 return 0;
270 }
271
pm_uninit(struct packet_manager * pm,bool hanging)272 void pm_uninit(struct packet_manager *pm, bool hanging)
273 {
274 mutex_destroy(&pm->lock);
275 kernel_queue_uninit(pm->priv_queue, hanging);
276 }
277
pm_send_set_resources(struct packet_manager * pm,struct scheduling_resources * res)278 int pm_send_set_resources(struct packet_manager *pm,
279 struct scheduling_resources *res)
280 {
281 uint32_t *buffer, size;
282 int retval = 0;
283
284 size = pm->pmf->set_resources_size;
285 mutex_lock(&pm->lock);
286 kq_acquire_packet_buffer(pm->priv_queue,
287 size / sizeof(uint32_t),
288 (unsigned int **)&buffer);
289 if (!buffer) {
290 pr_err("Failed to allocate buffer on kernel queue\n");
291 retval = -ENOMEM;
292 goto out;
293 }
294
295 retval = pm->pmf->set_resources(pm, buffer, res);
296 if (!retval)
297 kq_submit_packet(pm->priv_queue);
298 else
299 kq_rollback_packet(pm->priv_queue);
300
301 out:
302 mutex_unlock(&pm->lock);
303
304 return retval;
305 }
306
pm_send_runlist(struct packet_manager * pm,struct list_head * dqm_queues)307 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
308 {
309 uint64_t rl_gpu_ib_addr;
310 uint32_t *rl_buffer;
311 size_t rl_ib_size, packet_size_dwords;
312 int retval;
313
314 retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
315 &rl_ib_size);
316 if (retval)
317 goto fail_create_runlist_ib;
318
319 pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
320
321 packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
322 mutex_lock(&pm->lock);
323
324 retval = kq_acquire_packet_buffer(pm->priv_queue,
325 packet_size_dwords, &rl_buffer);
326 if (retval)
327 goto fail_acquire_packet_buffer;
328
329 retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
330 rl_ib_size / sizeof(uint32_t), false);
331 if (retval)
332 goto fail_create_runlist;
333
334 kq_submit_packet(pm->priv_queue);
335
336 mutex_unlock(&pm->lock);
337
338 return retval;
339
340 fail_create_runlist:
341 kq_rollback_packet(pm->priv_queue);
342 fail_acquire_packet_buffer:
343 mutex_unlock(&pm->lock);
344 fail_create_runlist_ib:
345 pm_release_ib(pm);
346 return retval;
347 }
348
pm_send_query_status(struct packet_manager * pm,uint64_t fence_address,uint32_t fence_value)349 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
350 uint32_t fence_value)
351 {
352 uint32_t *buffer, size;
353 int retval = 0;
354
355 if (WARN_ON(!fence_address))
356 return -EFAULT;
357
358 size = pm->pmf->query_status_size;
359 mutex_lock(&pm->lock);
360 kq_acquire_packet_buffer(pm->priv_queue,
361 size / sizeof(uint32_t), (unsigned int **)&buffer);
362 if (!buffer) {
363 pr_err("Failed to allocate buffer on kernel queue\n");
364 retval = -ENOMEM;
365 goto out;
366 }
367
368 retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
369 if (!retval)
370 kq_submit_packet(pm->priv_queue);
371 else
372 kq_rollback_packet(pm->priv_queue);
373
374 out:
375 mutex_unlock(&pm->lock);
376 return retval;
377 }
378
pm_send_unmap_queue(struct packet_manager * pm,enum kfd_queue_type type,enum kfd_unmap_queues_filter filter,uint32_t filter_param,bool reset,unsigned int sdma_engine)379 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
380 enum kfd_unmap_queues_filter filter,
381 uint32_t filter_param, bool reset,
382 unsigned int sdma_engine)
383 {
384 uint32_t *buffer, size;
385 int retval = 0;
386
387 size = pm->pmf->unmap_queues_size;
388 mutex_lock(&pm->lock);
389 kq_acquire_packet_buffer(pm->priv_queue,
390 size / sizeof(uint32_t), (unsigned int **)&buffer);
391 if (!buffer) {
392 pr_err("Failed to allocate buffer on kernel queue\n");
393 retval = -ENOMEM;
394 goto out;
395 }
396
397 retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
398 reset, sdma_engine);
399 if (!retval)
400 kq_submit_packet(pm->priv_queue);
401 else
402 kq_rollback_packet(pm->priv_queue);
403
404 out:
405 mutex_unlock(&pm->lock);
406 return retval;
407 }
408
pm_release_ib(struct packet_manager * pm)409 void pm_release_ib(struct packet_manager *pm)
410 {
411 mutex_lock(&pm->lock);
412 if (pm->allocated) {
413 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
414 pm->allocated = false;
415 }
416 mutex_unlock(&pm->lock);
417 }
418
419 #if defined(CONFIG_DEBUG_FS)
420
pm_debugfs_runlist(struct seq_file * m,void * data)421 int pm_debugfs_runlist(struct seq_file *m, void *data)
422 {
423 struct packet_manager *pm = data;
424
425 mutex_lock(&pm->lock);
426
427 if (!pm->allocated) {
428 seq_puts(m, " No active runlist\n");
429 goto out;
430 }
431
432 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
433 pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
434
435 out:
436 mutex_unlock(&pm->lock);
437 return 0;
438 }
439
pm_debugfs_hang_hws(struct packet_manager * pm)440 int pm_debugfs_hang_hws(struct packet_manager *pm)
441 {
442 uint32_t *buffer, size;
443 int r = 0;
444
445 size = pm->pmf->query_status_size;
446 mutex_lock(&pm->lock);
447 kq_acquire_packet_buffer(pm->priv_queue,
448 size / sizeof(uint32_t), (unsigned int **)&buffer);
449 if (!buffer) {
450 pr_err("Failed to allocate buffer on kernel queue\n");
451 r = -ENOMEM;
452 goto out;
453 }
454 memset(buffer, 0x55, size);
455 kq_submit_packet(pm->priv_queue);
456
457 pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
458 buffer[0], buffer[1], buffer[2], buffer[3],
459 buffer[4], buffer[5], buffer[6]);
460 out:
461 mutex_unlock(&pm->lock);
462 return r;
463 }
464
465
466 #endif
467