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_dev *dev = pm->dqm->dev;
49 
50 	process_count = pm->dqm->processes_count;
51 	queue_count = pm->dqm->active_queue_count;
52 	compute_queue_count = pm->dqm->active_cp_queue_count;
53 	gws_queue_count = pm->dqm->gws_queue_count;
54 
55 	/* check if there is over subscription
56 	 * Note: the arbitration between the number of VMIDs and
57 	 * hws_max_conc_proc has been done in
58 	 * kgd2kfd_device_init().
59 	 */
60 	*over_subscription = false;
61 
62 	if (dev->max_proc_per_quantum > 1)
63 		max_proc_per_quantum = dev->max_proc_per_quantum;
64 
65 	if ((process_count > max_proc_per_quantum) ||
66 	    compute_queue_count > get_cp_queues_num(pm->dqm) ||
67 	    gws_queue_count > 1) {
68 		*over_subscription = true;
69 		pr_debug("Over subscribed runlist\n");
70 	}
71 
72 	map_queue_size = pm->pmf->map_queues_size;
73 	/* calculate run list ib allocation size */
74 	*rlib_size = process_count * pm->pmf->map_process_size +
75 		     queue_count * map_queue_size;
76 
77 	/*
78 	 * Increase the allocation size in case we need a chained run list
79 	 * when over subscription
80 	 */
81 	if (*over_subscription)
82 		*rlib_size += pm->pmf->runlist_size;
83 
84 	pr_debug("runlist ib size %d\n", *rlib_size);
85 }
86 
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)87 static int pm_allocate_runlist_ib(struct packet_manager *pm,
88 				unsigned int **rl_buffer,
89 				uint64_t *rl_gpu_buffer,
90 				unsigned int *rl_buffer_size,
91 				bool *is_over_subscription)
92 {
93 	int retval;
94 
95 	if (WARN_ON(pm->allocated))
96 		return -EINVAL;
97 
98 	pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
99 
100 	mutex_lock(&pm->lock);
101 
102 	retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
103 					&pm->ib_buffer_obj);
104 
105 	if (retval) {
106 		pr_err("Failed to allocate runlist IB\n");
107 		goto out;
108 	}
109 
110 	*(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
111 	*rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
112 
113 	memset(*rl_buffer, 0, *rl_buffer_size);
114 	pm->allocated = true;
115 
116 out:
117 	mutex_unlock(&pm->lock);
118 	return retval;
119 }
120 
pm_create_runlist_ib(struct packet_manager * pm,struct list_head * queues,uint64_t * rl_gpu_addr,size_t * rl_size_bytes)121 static int pm_create_runlist_ib(struct packet_manager *pm,
122 				struct list_head *queues,
123 				uint64_t *rl_gpu_addr,
124 				size_t *rl_size_bytes)
125 {
126 	unsigned int alloc_size_bytes;
127 	unsigned int *rl_buffer, rl_wptr, i;
128 	int retval, processes_mapped;
129 	struct device_process_node *cur;
130 	struct qcm_process_device *qpd;
131 	struct queue *q;
132 	struct kernel_queue *kq;
133 	bool is_over_subscription;
134 
135 	rl_wptr = retval = processes_mapped = 0;
136 
137 	retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
138 				&alloc_size_bytes, &is_over_subscription);
139 	if (retval)
140 		return retval;
141 
142 	*rl_size_bytes = alloc_size_bytes;
143 	pm->ib_size_bytes = alloc_size_bytes;
144 
145 	pr_debug("Building runlist ib process count: %d queues count %d\n",
146 		pm->dqm->processes_count, pm->dqm->active_queue_count);
147 
148 	/* build the run list ib packet */
149 	list_for_each_entry(cur, queues, list) {
150 		qpd = cur->qpd;
151 		/* build map process packet */
152 		if (processes_mapped >= pm->dqm->processes_count) {
153 			pr_debug("Not enough space left in runlist IB\n");
154 			pm_release_ib(pm);
155 			return -ENOMEM;
156 		}
157 
158 		retval = pm->pmf->map_process(pm, &rl_buffer[rl_wptr], qpd);
159 		if (retval)
160 			return retval;
161 
162 		processes_mapped++;
163 		inc_wptr(&rl_wptr, pm->pmf->map_process_size,
164 				alloc_size_bytes);
165 
166 		list_for_each_entry(kq, &qpd->priv_queue_list, list) {
167 			if (!kq->queue->properties.is_active)
168 				continue;
169 
170 			pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
171 				kq->queue->queue, qpd->is_debug);
172 
173 			retval = pm->pmf->map_queues(pm,
174 						&rl_buffer[rl_wptr],
175 						kq->queue,
176 						qpd->is_debug);
177 			if (retval)
178 				return retval;
179 
180 			inc_wptr(&rl_wptr,
181 				pm->pmf->map_queues_size,
182 				alloc_size_bytes);
183 		}
184 
185 		list_for_each_entry(q, &qpd->queues_list, list) {
186 			if (!q->properties.is_active)
187 				continue;
188 
189 			pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
190 				q->queue, qpd->is_debug);
191 
192 			retval = pm->pmf->map_queues(pm,
193 						&rl_buffer[rl_wptr],
194 						q,
195 						qpd->is_debug);
196 
197 			if (retval)
198 				return retval;
199 
200 			inc_wptr(&rl_wptr,
201 				pm->pmf->map_queues_size,
202 				alloc_size_bytes);
203 		}
204 	}
205 
206 	pr_debug("Finished map process and queues to runlist\n");
207 
208 	if (is_over_subscription) {
209 		if (!pm->is_over_subscription)
210 			pr_warn("Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
211 		retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
212 					*rl_gpu_addr,
213 					alloc_size_bytes / sizeof(uint32_t),
214 					true);
215 	}
216 	pm->is_over_subscription = is_over_subscription;
217 
218 	for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
219 		pr_debug("0x%2X ", rl_buffer[i]);
220 	pr_debug("\n");
221 
222 	return retval;
223 }
224 
pm_init(struct packet_manager * pm,struct device_queue_manager * dqm)225 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
226 {
227 	switch (dqm->dev->adev->asic_type) {
228 	case CHIP_KAVERI:
229 	case CHIP_HAWAII:
230 		/* PM4 packet structures on CIK are the same as on VI */
231 	case CHIP_CARRIZO:
232 	case CHIP_TONGA:
233 	case CHIP_FIJI:
234 	case CHIP_POLARIS10:
235 	case CHIP_POLARIS11:
236 	case CHIP_POLARIS12:
237 	case CHIP_VEGAM:
238 		pm->pmf = &kfd_vi_pm_funcs;
239 		break;
240 	default:
241 		if (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 2))
242 			pm->pmf = &kfd_aldebaran_pm_funcs;
243 		else if (KFD_GC_VERSION(dqm->dev) >= IP_VERSION(9, 0, 1))
244 			pm->pmf = &kfd_v9_pm_funcs;
245 		else {
246 			WARN(1, "Unexpected ASIC family %u",
247 			     dqm->dev->adev->asic_type);
248 			return -EINVAL;
249 		}
250 	}
251 
252 	pm->dqm = dqm;
253 	mutex_init(&pm->lock);
254 	pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
255 	if (!pm->priv_queue) {
256 		mutex_destroy(&pm->lock);
257 		return -ENOMEM;
258 	}
259 	pm->allocated = false;
260 
261 	return 0;
262 }
263 
pm_uninit(struct packet_manager * pm,bool hanging)264 void pm_uninit(struct packet_manager *pm, bool hanging)
265 {
266 	mutex_destroy(&pm->lock);
267 	kernel_queue_uninit(pm->priv_queue, hanging);
268 	pm->priv_queue = NULL;
269 }
270 
pm_send_set_resources(struct packet_manager * pm,struct scheduling_resources * res)271 int pm_send_set_resources(struct packet_manager *pm,
272 				struct scheduling_resources *res)
273 {
274 	uint32_t *buffer, size;
275 	int retval = 0;
276 
277 	size = pm->pmf->set_resources_size;
278 	mutex_lock(&pm->lock);
279 	kq_acquire_packet_buffer(pm->priv_queue,
280 					size / sizeof(uint32_t),
281 					(unsigned int **)&buffer);
282 	if (!buffer) {
283 		pr_err("Failed to allocate buffer on kernel queue\n");
284 		retval = -ENOMEM;
285 		goto out;
286 	}
287 
288 	retval = pm->pmf->set_resources(pm, buffer, res);
289 	if (!retval)
290 		kq_submit_packet(pm->priv_queue);
291 	else
292 		kq_rollback_packet(pm->priv_queue);
293 
294 out:
295 	mutex_unlock(&pm->lock);
296 
297 	return retval;
298 }
299 
pm_send_runlist(struct packet_manager * pm,struct list_head * dqm_queues)300 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
301 {
302 	uint64_t rl_gpu_ib_addr;
303 	uint32_t *rl_buffer;
304 	size_t rl_ib_size, packet_size_dwords;
305 	int retval;
306 
307 	retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
308 					&rl_ib_size);
309 	if (retval)
310 		goto fail_create_runlist_ib;
311 
312 	pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
313 
314 	packet_size_dwords = pm->pmf->runlist_size / sizeof(uint32_t);
315 	mutex_lock(&pm->lock);
316 
317 	retval = kq_acquire_packet_buffer(pm->priv_queue,
318 					packet_size_dwords, &rl_buffer);
319 	if (retval)
320 		goto fail_acquire_packet_buffer;
321 
322 	retval = pm->pmf->runlist(pm, rl_buffer, rl_gpu_ib_addr,
323 					rl_ib_size / sizeof(uint32_t), false);
324 	if (retval)
325 		goto fail_create_runlist;
326 
327 	kq_submit_packet(pm->priv_queue);
328 
329 	mutex_unlock(&pm->lock);
330 
331 	return retval;
332 
333 fail_create_runlist:
334 	kq_rollback_packet(pm->priv_queue);
335 fail_acquire_packet_buffer:
336 	mutex_unlock(&pm->lock);
337 fail_create_runlist_ib:
338 	pm_release_ib(pm);
339 	return retval;
340 }
341 
pm_send_query_status(struct packet_manager * pm,uint64_t fence_address,uint64_t fence_value)342 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
343 			uint64_t fence_value)
344 {
345 	uint32_t *buffer, size;
346 	int retval = 0;
347 
348 	if (WARN_ON(!fence_address))
349 		return -EFAULT;
350 
351 	size = pm->pmf->query_status_size;
352 	mutex_lock(&pm->lock);
353 	kq_acquire_packet_buffer(pm->priv_queue,
354 			size / sizeof(uint32_t), (unsigned int **)&buffer);
355 	if (!buffer) {
356 		pr_err("Failed to allocate buffer on kernel queue\n");
357 		retval = -ENOMEM;
358 		goto out;
359 	}
360 
361 	retval = pm->pmf->query_status(pm, buffer, fence_address, fence_value);
362 	if (!retval)
363 		kq_submit_packet(pm->priv_queue);
364 	else
365 		kq_rollback_packet(pm->priv_queue);
366 
367 out:
368 	mutex_unlock(&pm->lock);
369 	return retval;
370 }
371 
pm_send_unmap_queue(struct packet_manager * pm,enum kfd_unmap_queues_filter filter,uint32_t filter_param,bool reset)372 int pm_send_unmap_queue(struct packet_manager *pm,
373 			enum kfd_unmap_queues_filter filter,
374 			uint32_t filter_param, bool reset)
375 {
376 	uint32_t *buffer, size;
377 	int retval = 0;
378 
379 	size = pm->pmf->unmap_queues_size;
380 	mutex_lock(&pm->lock);
381 	kq_acquire_packet_buffer(pm->priv_queue,
382 			size / sizeof(uint32_t), (unsigned int **)&buffer);
383 	if (!buffer) {
384 		pr_err("Failed to allocate buffer on kernel queue\n");
385 		retval = -ENOMEM;
386 		goto out;
387 	}
388 
389 	retval = pm->pmf->unmap_queues(pm, buffer, filter, filter_param, reset);
390 	if (!retval)
391 		kq_submit_packet(pm->priv_queue);
392 	else
393 		kq_rollback_packet(pm->priv_queue);
394 
395 out:
396 	mutex_unlock(&pm->lock);
397 	return retval;
398 }
399 
pm_release_ib(struct packet_manager * pm)400 void pm_release_ib(struct packet_manager *pm)
401 {
402 	mutex_lock(&pm->lock);
403 	if (pm->allocated) {
404 		kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
405 		pm->allocated = false;
406 	}
407 	mutex_unlock(&pm->lock);
408 }
409 
410 #if defined(CONFIG_DEBUG_FS)
411 
pm_debugfs_runlist(struct seq_file * m,void * data)412 int pm_debugfs_runlist(struct seq_file *m, void *data)
413 {
414 	struct packet_manager *pm = data;
415 
416 	mutex_lock(&pm->lock);
417 
418 	if (!pm->allocated) {
419 		seq_puts(m, "  No active runlist\n");
420 		goto out;
421 	}
422 
423 	seq_hex_dump(m, "  ", DUMP_PREFIX_OFFSET, 32, 4,
424 		     pm->ib_buffer_obj->cpu_ptr, pm->ib_size_bytes, false);
425 
426 out:
427 	mutex_unlock(&pm->lock);
428 	return 0;
429 }
430 
pm_debugfs_hang_hws(struct packet_manager * pm)431 int pm_debugfs_hang_hws(struct packet_manager *pm)
432 {
433 	uint32_t *buffer, size;
434 	int r = 0;
435 
436 	if (!pm->priv_queue)
437 		return -EAGAIN;
438 
439 	size = pm->pmf->query_status_size;
440 	mutex_lock(&pm->lock);
441 	kq_acquire_packet_buffer(pm->priv_queue,
442 			size / sizeof(uint32_t), (unsigned int **)&buffer);
443 	if (!buffer) {
444 		pr_err("Failed to allocate buffer on kernel queue\n");
445 		r = -ENOMEM;
446 		goto out;
447 	}
448 	memset(buffer, 0x55, size);
449 	kq_submit_packet(pm->priv_queue);
450 
451 	pr_info("Submitting %x %x %x %x %x %x %x to HIQ to hang the HWS.",
452 		buffer[0], buffer[1], buffer[2], buffer[3],
453 		buffer[4], buffer[5], buffer[6]);
454 out:
455 	mutex_unlock(&pm->lock);
456 	return r;
457 }
458 
459 
460 #endif
461