1 /*
2  * Copyright 2014 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 
24 #include <linux/slab.h>
25 #include <linux/mutex.h>
26 #include "kfd_device_queue_manager.h"
27 #include "kfd_kernel_queue.h"
28 #include "kfd_priv.h"
29 
30 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
31 				unsigned int buffer_size_bytes)
32 {
33 	unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
34 
35 	WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
36 	     "Runlist IB overflow");
37 	*wptr = temp;
38 }
39 
40 static void pm_calc_rlib_size(struct packet_manager *pm,
41 				unsigned int *rlib_size,
42 				bool *over_subscription)
43 {
44 	unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
45 	unsigned int map_queue_size;
46 	unsigned int max_proc_per_quantum = 1;
47 	struct kfd_dev *dev = pm->dqm->dev;
48 
49 	process_count = pm->dqm->processes_count;
50 	queue_count = pm->dqm->active_queue_count;
51 	compute_queue_count = pm->dqm->active_cp_queue_count;
52 	gws_queue_count = pm->dqm->gws_queue_count;
53 
54 	/* check if there is over subscription
55 	 * Note: the arbitration between the number of VMIDs and
56 	 * hws_max_conc_proc has been done in
57 	 * kgd2kfd_device_init().
58 	 */
59 	*over_subscription = false;
60 
61 	if (dev->max_proc_per_quantum > 1)
62 		max_proc_per_quantum = dev->max_proc_per_quantum;
63 
64 	if ((process_count > max_proc_per_quantum) ||
65 	    compute_queue_count > get_cp_queues_num(pm->dqm) ||
66 	    gws_queue_count > 1) {
67 		*over_subscription = true;
68 		pr_debug("Over subscribed runlist\n");
69 	}
70 
71 	map_queue_size = pm->pmf->map_queues_size;
72 	/* calculate run list ib allocation size */
73 	*rlib_size = process_count * pm->pmf->map_process_size +
74 		     queue_count * map_queue_size;
75 
76 	/*
77 	 * Increase the allocation size in case we need a chained run list
78 	 * when over subscription
79 	 */
80 	if (*over_subscription)
81 		*rlib_size += pm->pmf->runlist_size;
82 
83 	pr_debug("runlist ib size %d\n", *rlib_size);
84 }
85 
86 static int pm_allocate_runlist_ib(struct packet_manager *pm,
87 				unsigned int **rl_buffer,
88 				uint64_t *rl_gpu_buffer,
89 				unsigned int *rl_buffer_size,
90 				bool *is_over_subscription)
91 {
92 	int retval;
93 
94 	if (WARN_ON(pm->allocated))
95 		return -EINVAL;
96 
97 	pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
98 
99 	mutex_lock(&pm->lock);
100 
101 	retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
102 					&pm->ib_buffer_obj);
103 
104 	if (retval) {
105 		pr_err("Failed to allocate runlist IB\n");
106 		goto out;
107 	}
108 
109 	*(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
110 	*rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
111 
112 	memset(*rl_buffer, 0, *rl_buffer_size);
113 	pm->allocated = true;
114 
115 out:
116 	mutex_unlock(&pm->lock);
117 	return retval;
118 }
119 
120 static int pm_create_runlist_ib(struct packet_manager *pm,
121 				struct list_head *queues,
122 				uint64_t *rl_gpu_addr,
123 				size_t *rl_size_bytes)
124 {
125 	unsigned int alloc_size_bytes;
126 	unsigned int *rl_buffer, rl_wptr, i;
127 	int retval, processes_mapped;
128 	struct device_process_node *cur;
129 	struct qcm_process_device *qpd;
130 	struct queue *q;
131 	struct kernel_queue *kq;
132 	bool is_over_subscription;
133 
134 	rl_wptr = retval = processes_mapped = 0;
135 
136 	retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
137 				&alloc_size_bytes, &is_over_subscription);
138 	if (retval)
139 		return retval;
140 
141 	*rl_size_bytes = alloc_size_bytes;
142 	pm->ib_size_bytes = alloc_size_bytes;
143 
144 	pr_debug("Building runlist ib process count: %d queues count %d\n",
145 		pm->dqm->processes_count, pm->dqm->active_queue_count);
146 
147 	/* build the run list ib packet */
148 	list_for_each_entry(cur, queues, list) {
149 		qpd = cur->qpd;
150 		/* build map process packet */
151 		if (processes_mapped >= pm->dqm->processes_count) {
152 			pr_debug("Not enough space left in runlist IB\n");
153 			pm_release_ib(pm);
154 			return -ENOMEM;
155 		}
156 
157 		retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
158 		if (retval)
159 			return retval;
160 
161 		processes_mapped++;
162 		inc_wptr(&rl_wptr, pm->pmf->map_process_size,
163 				alloc_size_bytes);
164 
165 		list_for_each_entry(kq, &qpd->priv_queue_list, list) {
166 			if (!kq->queue->properties.is_active)
167 				continue;
168 
169 			pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
170 				kq->queue->queue, qpd->is_debug);
171 
172 			retval = pm->pmf->map_queues(pm,
173 						&rl_buffer[rl_wptr],
174 						kq->queue,
175 						qpd->is_debug);
176 			if (retval)
177 				return retval;
178 
179 			inc_wptr(&rl_wptr,
180 				pm->pmf->map_queues_size,
181 				alloc_size_bytes);
182 		}
183 
184 		list_for_each_entry(q, &qpd->queues_list, list) {
185 			if (!q->properties.is_active)
186 				continue;
187 
188 			pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
189 				q->queue, qpd->is_debug);
190 
191 			retval = pm->pmf->map_queues(pm,
192 						&rl_buffer[rl_wptr],
193 						q,
194 						qpd->is_debug);
195 
196 			if (retval)
197 				return retval;
198 
199 			inc_wptr(&rl_wptr,
200 				pm->pmf->map_queues_size,
201 				alloc_size_bytes);
202 		}
203 	}
204 
205 	pr_debug("Finished map process and queues to runlist\n");
206 
207 	if (is_over_subscription) {
208 		if (!pm->is_over_subscription)
209 			pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
210 		retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
211 					*rl_gpu_addr,
212 					alloc_size_bytes / sizeof(uint32_t),
213 					true);
214 	}
215 	pm->is_over_subscription = is_over_subscription;
216 
217 	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
218 		pr_debug("0x%2X ", rl_buffer[i]);
219 	pr_debug("\n");
220 
221 	return retval;
222 }
223 
224 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
225 {
226 	switch (dqm->dev->device_info->asic_family) {
227 	case CHIP_KAVERI:
228 	case CHIP_HAWAII:
229 		/* PM4 packet structures on CIK are the same as on VI */
230 	case CHIP_CARRIZO:
231 	case CHIP_TONGA:
232 	case CHIP_FIJI:
233 	case CHIP_POLARIS10:
234 	case CHIP_POLARIS11:
235 	case CHIP_POLARIS12:
236 	case CHIP_VEGAM:
237 		pm->pmf = &kfd_vi_pm_funcs;
238 		break;
239 	case CHIP_VEGA10:
240 	case CHIP_VEGA12:
241 	case CHIP_VEGA20:
242 	case CHIP_RAVEN:
243 	case CHIP_RENOIR:
244 	case CHIP_ARCTURUS:
245 	case CHIP_NAVI10:
246 	case CHIP_NAVI12:
247 	case CHIP_NAVI14:
248 	case CHIP_SIENNA_CICHLID:
249 	case CHIP_NAVY_FLOUNDER:
250 	case CHIP_VANGOGH:
251 	case CHIP_DIMGREY_CAVEFISH:
252 	case CHIP_BEIGE_GOBY:
253 	case CHIP_YELLOW_CARP:
254 		pm->pmf = &kfd_v9_pm_funcs;
255 		break;
256 	case CHIP_ALDEBARAN:
257 		pm->pmf = &kfd_aldebaran_pm_funcs;
258 		break;
259 	default:
260 		WARN(1, "Unexpected ASIC family %u",
261 		     dqm->dev->device_info->asic_family);
262 		return -EINVAL;
263 	}
264 
265 	pm->dqm = dqm;
266 	mutex_init(&pm->lock);
267 	pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
268 	if (!pm->priv_queue) {
269 		mutex_destroy(&pm->lock);
270 		return -ENOMEM;
271 	}
272 	pm->allocated = false;
273 
274 	return 0;
275 }
276 
277 void pm_uninit(struct packet_manager *pm, bool hanging)
278 {
279 	mutex_destroy(&pm->lock);
280 	kernel_queue_uninit(pm->priv_queue, hanging);
281 }
282 
283 int pm_send_set_resources(struct packet_manager *pm,
284 				struct scheduling_resources *res)
285 {
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 		pr_err("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 		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 
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 	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 
354 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
355 			uint64_t fence_value)
356 {
357 	uint32_t *buffer, size;
358 	int retval = 0;
359 
360 	if (WARN_ON(!fence_address))
361 		return -EFAULT;
362 
363 	size = pm->pmf->query_status_size;
364 	mutex_lock(&pm->lock);
365 	kq_acquire_packet_buffer(pm->priv_queue,
366 			size / sizeof(uint32_t), (unsigned int **)&buffer);
367 	if (!buffer) {
368 		pr_err("Failed to allocate buffer on kernel queue\n");
369 		retval = -ENOMEM;
370 		goto out;
371 	}
372 
373 	retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
374 	if (!retval)
375 		kq_submit_packet(pm->priv_queue);
376 	else
377 		kq_rollback_packet(pm->priv_queue);
378 
379 out:
380 	mutex_unlock(&pm->lock);
381 	return retval;
382 }
383 
384 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
385 			enum kfd_unmap_queues_filter filter,
386 			uint32_t filter_param, bool reset,
387 			unsigned int sdma_engine)
388 {
389 	uint32_t *buffer, size;
390 	int retval = 0;
391 
392 	size = pm->pmf->unmap_queues_size;
393 	mutex_lock(&pm->lock);
394 	kq_acquire_packet_buffer(pm->priv_queue,
395 			size / sizeof(uint32_t), (unsigned int **)&buffer);
396 	if (!buffer) {
397 		pr_err("Failed to allocate buffer on kernel queue\n");
398 		retval = -ENOMEM;
399 		goto out;
400 	}
401 
402 	retval = pm->pmf->unmap_queues(pm, buffer, type, filter, filter_param,
403 				       reset, sdma_engine);
404 	if (!retval)
405 		kq_submit_packet(pm->priv_queue);
406 	else
407 		kq_rollback_packet(pm->priv_queue);
408 
409 out:
410 	mutex_unlock(&pm->lock);
411 	return retval;
412 }
413 
414 void pm_release_ib(struct packet_manager *pm)
415 {
416 	mutex_lock(&pm->lock);
417 	if (pm->allocated) {
418 		kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
419 		pm->allocated = false;
420 	}
421 	mutex_unlock(&pm->lock);
422 }
423 
424 #if defined(CONFIG_DEBUG_FS)
425 
426 int pm_debugfs_runlist(struct seq_file *m, void *data)
427 {
428 	struct packet_manager *pm = data;
429 
430 	mutex_lock(&pm->lock);
431 
432 	if (!pm->allocated) {
433 		seq_puts(m, "  No active runlist\n");
434 		goto out;
435 	}
436 
437 	seq_hex_dump(m, "  ", DUMP_PREFIX_OFFSET, 32, 4,
438 		     pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
439 
440 out:
441 	mutex_unlock(&pm->lock);
442 	return 0;
443 }
444 
445 int pm_debugfs_hang_hws(struct packet_manager *pm)
446 {
447 	uint32_t *buffer, size;
448 	int r = 0;
449 
450 	size = pm->pmf->query_status_size;
451 	mutex_lock(&pm->lock);
452 	kq_acquire_packet_buffer(pm->priv_queue,
453 			size / sizeof(uint32_t), (unsigned int **)&buffer);
454 	if (!buffer) {
455 		pr_err("Failed to allocate buffer on kernel queue\n");
456 		r = -ENOMEM;
457 		goto out;
458 	}
459 	memset(buffer, 0x55, size);
460 	kq_submit_packet(pm->priv_queue);
461 
462 	pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
463 		buffer[0], buffer[1], buffer[2], buffer[3],
464 		buffer[4], buffer[5], buffer[6]);
465 out:
466 	mutex_unlock(&pm->lock);
467 	return r;
468 }
469 
470 
471 #endif
472