1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 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 
25 #include <linux/slab.h>
26 #include <linux/mutex.h>
27 #include "kfd_device_queue_manager.h"
28 #include "kfd_kernel_queue.h"
29 #include "kfd_priv.h"
30 
inc_wptr(unsigned int * wptr,unsigned int increment_bytes,unsigned int buffer_size_bytes)31 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
32 				unsigned int buffer_size_bytes)
33 {
34 	unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
35 
36 	WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
37 	     "Runlist IB overflow");
38 	*wptr = temp;
39 }
40 
pm_calc_rlib_size(struct packet_manager * pm,unsigned int * rlib_size,bool * over_subscription)41 static void pm_calc_rlib_size(struct packet_manager *pm,
42 				unsigned int *rlib_size,
43 				bool *over_subscription)
44 {
45 	unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
46 	unsigned int map_queue_size;
47 	unsigned int max_proc_per_quantum = 1;
48 	struct kfd_node *node = pm->dqm->dev;
49 	struct device *dev = node->adev->dev;
50 
51 	process_count = pm->dqm->processes_count;
52 	queue_count = pm->dqm->active_queue_count;
53 	compute_queue_count = pm->dqm->active_cp_queue_count;
54 	gws_queue_count = pm->dqm->gws_queue_count;
55 
56 	/* check if there is over subscription
57 	 * Note: the arbitration between the number of VMIDs and
58 	 * hws_max_conc_proc has been done in
59 	 * kgd2kfd_device_init().
60 	 */
61 	*over_subscription = false;
62 
63 	if (node->max_proc_per_quantum > 1)
64 		max_proc_per_quantum = node->max_proc_per_quantum;
65 
66 	if ((process_count > max_proc_per_quantum) ||
67 	    compute_queue_count > get_cp_queues_num(pm->dqm) ||
68 	    gws_queue_count > 1) {
69 		*over_subscription = true;
70 		dev_dbg(dev, "Over subscribed runlist\n");
71 	}
72 
73 	map_queue_size = pm->pmf->map_queues_size;
74 	/* calculate run list ib allocation size */
75 	*rlib_size = process_count * pm->pmf->map_process_size +
76 		     queue_count * map_queue_size;
77 
78 	/*
79 	 * Increase the allocation size in case we need a chained run list
80 	 * when over subscription
81 	 */
82 	if (*over_subscription)
83 		*rlib_size += pm->pmf->runlist_size;
84 
85 	dev_dbg(dev, "runlist ib size %d\n", *rlib_size);
86 }
87 
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)88 static int pm_allocate_runlist_ib(struct packet_manager *pm,
89 				unsigned int **rl_buffer,
90 				uint64_t *rl_gpu_buffer,
91 				unsigned int *rl_buffer_size,
92 				bool *is_over_subscription)
93 {
94 	struct kfd_node *node = pm->dqm->dev;
95 	struct device *dev = node->adev->dev;
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(node, *rl_buffer_size, &pm->ib_buffer_obj);
106 
107 	if (retval) {
108 		dev_err(dev, "Failed to allocate runlist IB\n");
109 		goto out;
110 	}
111 
112 	*(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
113 	*rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
114 
115 	memset(*rl_buffer, 0, *rl_buffer_size);
116 	pm->allocated = true;
117 
118 out:
119 	mutex_unlock(&pm->lock);
120 	return retval;
121 }
122 
pm_create_runlist_ib(struct packet_manager * pm,struct list_head * queues,uint64_t * rl_gpu_addr,size_t * rl_size_bytes)123 static int pm_create_runlist_ib(struct packet_manager *pm,
124 				struct list_head *queues,
125 				uint64_t *rl_gpu_addr,
126 				size_t *rl_size_bytes)
127 {
128 	unsigned int alloc_size_bytes;
129 	unsigned int *rl_buffer, rl_wptr, i;
130 	struct kfd_node *node = pm->dqm->dev;
131 	struct device *dev = node->adev->dev;
132 	int retval, processes_mapped;
133 	struct device_process_node *cur;
134 	struct qcm_process_device *qpd;
135 	struct queue *q;
136 	struct kernel_queue *kq;
137 	bool is_over_subscription;
138 
139 	rl_wptr = retval = processes_mapped = 0;
140 
141 	retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
142 				&alloc_size_bytes, &is_over_subscription);
143 	if (retval)
144 		return retval;
145 
146 	*rl_size_bytes = alloc_size_bytes;
147 	pm->ib_size_bytes = alloc_size_bytes;
148 
149 	dev_dbg(dev, "Building runlist ib process count: %d queues count %d\n",
150 		pm->dqm->processes_count, pm->dqm->active_queue_count);
151 
152 	/* build the run list ib packet */
153 	list_for_each_entry(cur, queues, list) {
154 		qpd = cur->qpd;
155 		/* build map process packet */
156 		if (processes_mapped >= pm->dqm->processes_count) {
157 			dev_dbg(dev, "Not enough space left in runlist IB\n");
158 			pm_release_ib(pm);
159 			return -ENOMEM;
160 		}
161 
162 		retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
163 		if (retval)
164 			return retval;
165 
166 		processes_mapped++;
167 		inc_wptr(&rl_wptr, pm->pmf->map_process_size,
168 				alloc_size_bytes);
169 
170 		list_for_each_entry(kq, &qpd->priv_queue_list, list) {
171 			if (!kq->queue->properties.is_active)
172 				continue;
173 
174 			dev_dbg(dev,
175 				"static_queue, mapping kernel q %d, is debug status %d\n",
176 				kq->queue->queue, qpd->is_debug);
177 
178 			retval = pm->pmf->map_queues(pm,
179 						&rl_buffer[rl_wptr],
180 						kq->queue,
181 						qpd->is_debug);
182 			if (retval)
183 				return retval;
184 
185 			inc_wptr(&rl_wptr,
186 				pm->pmf->map_queues_size,
187 				alloc_size_bytes);
188 		}
189 
190 		list_for_each_entry(q, &qpd->queues_list, list) {
191 			if (!q->properties.is_active)
192 				continue;
193 
194 			dev_dbg(dev,
195 				"static_queue, mapping user queue %d, is debug status %d\n",
196 				q->queue, qpd->is_debug);
197 
198 			retval = pm->pmf->map_queues(pm,
199 						&rl_buffer[rl_wptr],
200 						q,
201 						qpd->is_debug);
202 
203 			if (retval)
204 				return retval;
205 
206 			inc_wptr(&rl_wptr,
207 				pm->pmf->map_queues_size,
208 				alloc_size_bytes);
209 		}
210 	}
211 
212 	dev_dbg(dev, "Finished map process and queues to runlist\n");
213 
214 	if (is_over_subscription) {
215 		if (!pm->is_over_subscription)
216 			dev_warn(
217 				dev,
218 				"Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
219 		retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
220 					*rl_gpu_addr,
221 					alloc_size_bytes / sizeof(uint32_t),
222 					true);
223 	}
224 	pm->is_over_subscription = is_over_subscription;
225 
226 	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
227 		pr_debug("0x%2X ", rl_buffer[i]);
228 	pr_debug("\n");
229 
230 	return retval;
231 }
232 
pm_init(struct packet_manager * pm,struct device_queue_manager * dqm)233 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
234 {
235 	switch (dqm->dev->adev->asic_type) {
236 	case CHIP_KAVERI:
237 	case CHIP_HAWAII:
238 		/* PM4 packet structures on CIK are the same as on VI */
239 	case CHIP_CARRIZO:
240 	case CHIP_TONGA:
241 	case CHIP_FIJI:
242 	case CHIP_POLARIS10:
243 	case CHIP_POLARIS11:
244 	case CHIP_POLARIS12:
245 	case CHIP_VEGAM:
246 		pm->pmf = &kfd_vi_pm_funcs;
247 		break;
248 	default:
249 		if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 2) ||
250 		    KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3) ||
251 		    KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 4))
252 			pm->pmf = &kfd_aldebaran_pm_funcs;
253 		else if (KFD_GC_VERSION(dqm->dev) >= IP_VERSION(9, 0, 1))
254 			pm->pmf = &kfd_v9_pm_funcs;
255 		else {
256 			WARN(1, "Unexpected ASIC family %u",
257 			     dqm->dev->adev->asic_type);
258 			return -EINVAL;
259 		}
260 	}
261 
262 	pm->dqm = dqm;
263 	mutex_init(&pm->lock);
264 	pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
265 	if (!pm->priv_queue) {
266 		mutex_destroy(&pm->lock);
267 		return -ENOMEM;
268 	}
269 	pm->allocated = false;
270 
271 	return 0;
272 }
273 
pm_uninit(struct packet_manager * pm)274 void pm_uninit(struct packet_manager *pm)
275 {
276 	mutex_destroy(&pm->lock);
277 	kernel_queue_uninit(pm->priv_queue);
278 	pm->priv_queue = NULL;
279 }
280 
pm_send_set_resources(struct packet_manager * pm,struct scheduling_resources * res)281 int pm_send_set_resources(struct packet_manager *pm,
282 				struct scheduling_resources *res)
283 {
284 	struct kfd_node *node = pm->dqm->dev;
285 	struct device *dev = node->adev->dev;
286 	uint32_t *buffer, size;
287 	int retval = 0;
288 
289 	size = pm->pmf->set_resources_size;
290 	mutex_lock(&pm->lock);
291 	kq_acquire_packet_buffer(pm->priv_queue,
292 					size / sizeof(uint32_t),
293 					(unsigned int **)&buffer);
294 	if (!buffer) {
295 		dev_err(dev, "Failed to allocate buffer on kernel queue\n");
296 		retval = -ENOMEM;
297 		goto out;
298 	}
299 
300 	retval = pm->pmf->set_resources(pm, buffer, res);
301 	if (!retval)
302 		retval = kq_submit_packet(pm->priv_queue);
303 	else
304 		kq_rollback_packet(pm->priv_queue);
305 
306 out:
307 	mutex_unlock(&pm->lock);
308 
309 	return retval;
310 }
311 
pm_send_runlist(struct packet_manager * pm,struct list_head * dqm_queues)312 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
313 {
314 	uint64_t rl_gpu_ib_addr;
315 	uint32_t *rl_buffer;
316 	size_t rl_ib_size, packet_size_dwords;
317 	int retval;
318 
319 	retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
320 					&rl_ib_size);
321 	if (retval)
322 		goto fail_create_runlist_ib;
323 
324 	pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
325 
326 	packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
327 	mutex_lock(&pm->lock);
328 
329 	retval = kq_acquire_packet_buffer(pm->priv_queue,
330 					packet_size_dwords, &rl_buffer);
331 	if (retval)
332 		goto fail_acquire_packet_buffer;
333 
334 	retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
335 					rl_ib_size / sizeof(uint32_t), false);
336 	if (retval)
337 		goto fail_create_runlist;
338 
339 	retval = kq_submit_packet(pm->priv_queue);
340 
341 	mutex_unlock(&pm->lock);
342 
343 	return retval;
344 
345 fail_create_runlist:
346 	kq_rollback_packet(pm->priv_queue);
347 fail_acquire_packet_buffer:
348 	mutex_unlock(&pm->lock);
349 fail_create_runlist_ib:
350 	pm_release_ib(pm);
351 	return retval;
352 }
353 
pm_send_query_status(struct packet_manager * pm,uint64_t fence_address,uint64_t fence_value)354 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
355 			uint64_t fence_value)
356 {
357 	struct kfd_node *node = pm->dqm->dev;
358 	struct device *dev = node->adev->dev;
359 	uint32_t *buffer, size;
360 	int retval = 0;
361 
362 	if (WARN_ON(!fence_address))
363 		return -EFAULT;
364 
365 	size = pm->pmf->query_status_size;
366 	mutex_lock(&pm->lock);
367 	kq_acquire_packet_buffer(pm->priv_queue,
368 			size / sizeof(uint32_t), (unsigned int **)&buffer);
369 	if (!buffer) {
370 		dev_err(dev, "Failed to allocate buffer on kernel queue\n");
371 		retval = -ENOMEM;
372 		goto out;
373 	}
374 
375 	retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
376 	if (!retval)
377 		retval = kq_submit_packet(pm->priv_queue);
378 	else
379 		kq_rollback_packet(pm->priv_queue);
380 
381 out:
382 	mutex_unlock(&pm->lock);
383 	return retval;
384 }
385 
pm_update_grace_period(struct packet_manager * pm,uint32_t grace_period)386 int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period)
387 {
388 	struct kfd_node *node = pm->dqm->dev;
389 	struct device *dev = node->adev->dev;
390 	int retval = 0;
391 	uint32_t *buffer, size;
392 
393 	size = pm->pmf->set_grace_period_size;
394 
395 	mutex_lock(&pm->lock);
396 
397 	if (size) {
398 		kq_acquire_packet_buffer(pm->priv_queue,
399 			size / sizeof(uint32_t),
400 			(unsigned int **)&buffer);
401 
402 		if (!buffer) {
403 			dev_err(dev,
404 				"Failed to allocate buffer on kernel queue\n");
405 			retval = -ENOMEM;
406 			goto out;
407 		}
408 
409 		retval = pm->pmf->set_grace_period(pm, buffer, grace_period);
410 		if (!retval)
411 			retval = kq_submit_packet(pm->priv_queue);
412 		else
413 			kq_rollback_packet(pm->priv_queue);
414 	}
415 
416 out:
417 	mutex_unlock(&pm->lock);
418 	return retval;
419 }
420 
pm_send_unmap_queue(struct packet_manager * pm,enum kfd_unmap_queues_filter filter,uint32_t filter_param,bool reset)421 int pm_send_unmap_queue(struct packet_manager *pm,
422 			enum kfd_unmap_queues_filter filter,
423 			uint32_t filter_param, bool reset)
424 {
425 	struct kfd_node *node = pm->dqm->dev;
426 	struct device *dev = node->adev->dev;
427 	uint32_t *buffer, size;
428 	int retval = 0;
429 
430 	size = pm->pmf->unmap_queues_size;
431 	mutex_lock(&pm->lock);
432 	kq_acquire_packet_buffer(pm->priv_queue,
433 			size / sizeof(uint32_t), (unsigned int **)&buffer);
434 	if (!buffer) {
435 		dev_err(dev, "Failed to allocate buffer on kernel queue\n");
436 		retval = -ENOMEM;
437 		goto out;
438 	}
439 
440 	retval = pm->pmf->unmap_queues(pm, buffer, filter, filter_param, reset);
441 	if (!retval)
442 		retval = kq_submit_packet(pm->priv_queue);
443 	else
444 		kq_rollback_packet(pm->priv_queue);
445 
446 out:
447 	mutex_unlock(&pm->lock);
448 	return retval;
449 }
450 
pm_release_ib(struct packet_manager * pm)451 void pm_release_ib(struct packet_manager *pm)
452 {
453 	mutex_lock(&pm->lock);
454 	if (pm->allocated) {
455 		kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
456 		pm->allocated = false;
457 	}
458 	mutex_unlock(&pm->lock);
459 }
460 
461 #if defined(CONFIG_DEBUG_FS)
462 
pm_debugfs_runlist(struct seq_file * m,void * data)463 int pm_debugfs_runlist(struct seq_file *m, void *data)
464 {
465 	struct packet_manager *pm = data;
466 
467 	mutex_lock(&pm->lock);
468 
469 	if (!pm->allocated) {
470 		seq_puts(m, "  No active runlist\n");
471 		goto out;
472 	}
473 
474 	seq_hex_dump(m, "  ", DUMP_PREFIX_OFFSET, 32, 4,
475 		     pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
476 
477 out:
478 	mutex_unlock(&pm->lock);
479 	return 0;
480 }
481 
pm_debugfs_hang_hws(struct packet_manager * pm)482 int pm_debugfs_hang_hws(struct packet_manager *pm)
483 {
484 	struct kfd_node *node = pm->dqm->dev;
485 	struct device *dev = node->adev->dev;
486 	uint32_t *buffer, size;
487 	int r = 0;
488 
489 	if (!pm->priv_queue)
490 		return -EAGAIN;
491 
492 	size = pm->pmf->query_status_size;
493 	mutex_lock(&pm->lock);
494 	kq_acquire_packet_buffer(pm->priv_queue,
495 			size / sizeof(uint32_t), (unsigned int **)&buffer);
496 	if (!buffer) {
497 		dev_err(dev, "Failed to allocate buffer on kernel queue\n");
498 		r = -ENOMEM;
499 		goto out;
500 	}
501 	memset(buffer, 0x55, size);
502 	kq_submit_packet(pm->priv_queue);
503 
504 	dev_info(dev, "Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
505 		 buffer[0], buffer[1], buffer[2], buffer[3], buffer[4],
506 		 buffer[5], buffer[6]);
507 out:
508 	mutex_unlock(&pm->lock);
509 	return r;
510 }
511 
512 
513 #endif
514