xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c (revision ebf68996de0ab250c5d520eb2291ab65643e9a1e)
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/printk.h>
25 #include <linux/slab.h>
26 #include <linux/mm_types.h>
27 
28 #include "kfd_priv.h"
29 #include "kfd_mqd_manager.h"
30 #include "cik_regs.h"
31 #include "cik_structs.h"
32 #include "oss/oss_2_4_sh_mask.h"
33 
34 static inline struct cik_mqd *get_mqd(void *mqd)
35 {
36 	return (struct cik_mqd *)mqd;
37 }
38 
39 static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
40 {
41 	return (struct cik_sdma_rlc_registers *)mqd;
42 }
43 
44 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
45 			struct queue_properties *q)
46 {
47 	struct cik_mqd *m;
48 	uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */
49 
50 	if (q->cu_mask_count == 0)
51 		return;
52 
53 	mqd_symmetrically_map_cu_mask(mm,
54 		q->cu_mask, q->cu_mask_count, se_mask);
55 
56 	m = get_mqd(mqd);
57 	m->compute_static_thread_mgmt_se0 = se_mask[0];
58 	m->compute_static_thread_mgmt_se1 = se_mask[1];
59 	m->compute_static_thread_mgmt_se2 = se_mask[2];
60 	m->compute_static_thread_mgmt_se3 = se_mask[3];
61 
62 	pr_debug("Update cu mask to %#x %#x %#x %#x\n",
63 		m->compute_static_thread_mgmt_se0,
64 		m->compute_static_thread_mgmt_se1,
65 		m->compute_static_thread_mgmt_se2,
66 		m->compute_static_thread_mgmt_se3);
67 }
68 
69 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd,
70 					struct queue_properties *q)
71 {
72 	struct kfd_mem_obj *mqd_mem_obj;
73 
74 	if (q->type == KFD_QUEUE_TYPE_HIQ)
75 		return allocate_hiq_mqd(kfd);
76 
77 	if (kfd_gtt_sa_allocate(kfd, sizeof(struct cik_mqd),
78 			&mqd_mem_obj))
79 		return NULL;
80 
81 	return mqd_mem_obj;
82 }
83 
84 
85 static int init_mqd(struct mqd_manager *mm, void **mqd,
86 		struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
87 		struct queue_properties *q)
88 {
89 	uint64_t addr;
90 	struct cik_mqd *m;
91 	int retval;
92 	struct kfd_dev *kfd = mm->dev;
93 
94 	*mqd_mem_obj = allocate_mqd(kfd, q);
95 	if (!*mqd_mem_obj)
96 		return -ENOMEM;
97 
98 	m = (struct cik_mqd *) (*mqd_mem_obj)->cpu_ptr;
99 	addr = (*mqd_mem_obj)->gpu_addr;
100 
101 	memset(m, 0, ALIGN(sizeof(struct cik_mqd), 256));
102 
103 	m->header = 0xC0310800;
104 	m->compute_pipelinestat_enable = 1;
105 	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
106 	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
107 	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
108 	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
109 
110 	/*
111 	 * Make sure to use the last queue state saved on mqd when the cp
112 	 * reassigns the queue, so when queue is switched on/off (e.g over
113 	 * subscription or quantum timeout) the context will be consistent
114 	 */
115 	m->cp_hqd_persistent_state =
116 				DEFAULT_CP_HQD_PERSISTENT_STATE | PRELOAD_REQ;
117 
118 	m->cp_mqd_control             = MQD_CONTROL_PRIV_STATE_EN;
119 	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
120 	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
121 
122 	m->cp_hqd_quantum = QUANTUM_EN | QUANTUM_SCALE_1MS |
123 				QUANTUM_DURATION(10);
124 
125 	/*
126 	 * Pipe Priority
127 	 * Identifies the pipe relative priority when this queue is connected
128 	 * to the pipeline. The pipe priority is against the GFX pipe and HP3D.
129 	 * In KFD we are using a fixed pipe priority set to CS_MEDIUM.
130 	 * 0 = CS_LOW (typically below GFX)
131 	 * 1 = CS_MEDIUM (typically between HP3D and GFX
132 	 * 2 = CS_HIGH (typically above HP3D)
133 	 */
134 	m->cp_hqd_pipe_priority = 1;
135 	m->cp_hqd_queue_priority = 15;
136 
137 	if (q->format == KFD_QUEUE_FORMAT_AQL)
138 		m->cp_hqd_iq_rptr = AQL_ENABLE;
139 
140 	*mqd = m;
141 	if (gart_addr)
142 		*gart_addr = addr;
143 	retval = mm->update_mqd(mm, m, q);
144 
145 	return retval;
146 }
147 
148 static int init_mqd_sdma(struct mqd_manager *mm, void **mqd,
149 			struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
150 			struct queue_properties *q)
151 {
152 	int retval;
153 	struct cik_sdma_rlc_registers *m;
154 	struct kfd_dev *dev = mm->dev;
155 
156 	*mqd_mem_obj = allocate_sdma_mqd(dev, q);
157 	if (!*mqd_mem_obj)
158 		return -ENOMEM;
159 
160 	m = (struct cik_sdma_rlc_registers *) (*mqd_mem_obj)->cpu_ptr;
161 
162 	memset(m, 0, sizeof(struct cik_sdma_rlc_registers));
163 
164 	*mqd = m;
165 	if (gart_addr)
166 		*gart_addr = (*mqd_mem_obj)->gpu_addr;
167 
168 	retval = mm->update_mqd(mm, m, q);
169 
170 	return retval;
171 }
172 
173 static void uninit_mqd(struct mqd_manager *mm, void *mqd,
174 			struct kfd_mem_obj *mqd_mem_obj)
175 {
176 	kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
177 }
178 
179 
180 static int load_mqd(struct mqd_manager *mm, void *mqd, uint32_t pipe_id,
181 		    uint32_t queue_id, struct queue_properties *p,
182 		    struct mm_struct *mms)
183 {
184 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
185 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
186 	uint32_t wptr_mask = (uint32_t)((p->queue_size / 4) - 1);
187 
188 	return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id,
189 					  (uint32_t __user *)p->write_ptr,
190 					  wptr_shift, wptr_mask, mms);
191 }
192 
193 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd,
194 			 uint32_t pipe_id, uint32_t queue_id,
195 			 struct queue_properties *p, struct mm_struct *mms)
196 {
197 	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd,
198 					       (uint32_t __user *)p->write_ptr,
199 					       mms);
200 }
201 
202 static int __update_mqd(struct mqd_manager *mm, void *mqd,
203 			struct queue_properties *q, unsigned int atc_bit)
204 {
205 	struct cik_mqd *m;
206 
207 	m = get_mqd(mqd);
208 	m->cp_hqd_pq_control = DEFAULT_RPTR_BLOCK_SIZE |
209 				DEFAULT_MIN_AVAIL_SIZE;
210 	m->cp_hqd_ib_control = DEFAULT_MIN_IB_AVAIL_SIZE;
211 	if (atc_bit) {
212 		m->cp_hqd_pq_control |= PQ_ATC_EN;
213 		m->cp_hqd_ib_control |= IB_ATC_EN;
214 	}
215 
216 	/*
217 	 * Calculating queue size which is log base 2 of actual queue size -1
218 	 * dwords and another -1 for ffs
219 	 */
220 	m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1;
221 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
222 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
223 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
224 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
225 	m->cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(q->doorbell_off);
226 
227 	m->cp_hqd_vmid = q->vmid;
228 
229 	if (q->format == KFD_QUEUE_FORMAT_AQL)
230 		m->cp_hqd_pq_control |= NO_UPDATE_RPTR;
231 
232 	update_cu_mask(mm, mqd, q);
233 
234 	q->is_active = (q->queue_size > 0 &&
235 			q->queue_address != 0 &&
236 			q->queue_percent > 0 &&
237 			!q->is_evicted);
238 
239 	return 0;
240 }
241 
242 static int update_mqd(struct mqd_manager *mm, void *mqd,
243 			struct queue_properties *q)
244 {
245 	return __update_mqd(mm, mqd, q, 1);
246 }
247 
248 static int update_mqd_hawaii(struct mqd_manager *mm, void *mqd,
249 			struct queue_properties *q)
250 {
251 	return __update_mqd(mm, mqd, q, 0);
252 }
253 
254 static int update_mqd_sdma(struct mqd_manager *mm, void *mqd,
255 				struct queue_properties *q)
256 {
257 	struct cik_sdma_rlc_registers *m;
258 
259 	m = get_sdma_mqd(mqd);
260 	m->sdma_rlc_rb_cntl = order_base_2(q->queue_size / 4)
261 			<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
262 			q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
263 			1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
264 			6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
265 
266 	m->sdma_rlc_rb_base = lower_32_bits(q->queue_address >> 8);
267 	m->sdma_rlc_rb_base_hi = upper_32_bits(q->queue_address >> 8);
268 	m->sdma_rlc_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
269 	m->sdma_rlc_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
270 	m->sdma_rlc_doorbell =
271 		q->doorbell_off << SDMA0_RLC0_DOORBELL__OFFSET__SHIFT;
272 
273 	m->sdma_rlc_virtual_addr = q->sdma_vm_addr;
274 
275 	m->sdma_engine_id = q->sdma_engine_id;
276 	m->sdma_queue_id = q->sdma_queue_id;
277 
278 	q->is_active = (q->queue_size > 0 &&
279 			q->queue_address != 0 &&
280 			q->queue_percent > 0 &&
281 			!q->is_evicted);
282 
283 	return 0;
284 }
285 
286 static int destroy_mqd(struct mqd_manager *mm, void *mqd,
287 			enum kfd_preempt_type type,
288 			unsigned int timeout, uint32_t pipe_id,
289 			uint32_t queue_id)
290 {
291 	return mm->dev->kfd2kgd->hqd_destroy(mm->dev->kgd, mqd, type, timeout,
292 					pipe_id, queue_id);
293 }
294 
295 /*
296  * preempt type here is ignored because there is only one way
297  * to preempt sdma queue
298  */
299 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
300 				enum kfd_preempt_type type,
301 				unsigned int timeout, uint32_t pipe_id,
302 				uint32_t queue_id)
303 {
304 	return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout);
305 }
306 
307 static bool is_occupied(struct mqd_manager *mm, void *mqd,
308 			uint64_t queue_address,	uint32_t pipe_id,
309 			uint32_t queue_id)
310 {
311 
312 	return mm->dev->kfd2kgd->hqd_is_occupied(mm->dev->kgd, queue_address,
313 					pipe_id, queue_id);
314 
315 }
316 
317 static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd,
318 			uint64_t queue_address,	uint32_t pipe_id,
319 			uint32_t queue_id)
320 {
321 	return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd);
322 }
323 
324 /*
325  * HIQ MQD Implementation, concrete implementation for HIQ MQD implementation.
326  * The HIQ queue in Kaveri is using the same MQD structure as all the user mode
327  * queues but with different initial values.
328  */
329 
330 static int init_mqd_hiq(struct mqd_manager *mm, void **mqd,
331 		struct kfd_mem_obj **mqd_mem_obj, uint64_t *gart_addr,
332 		struct queue_properties *q)
333 {
334 	return init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
335 }
336 
337 static int update_mqd_hiq(struct mqd_manager *mm, void *mqd,
338 				struct queue_properties *q)
339 {
340 	struct cik_mqd *m;
341 
342 	m = get_mqd(mqd);
343 	m->cp_hqd_pq_control = DEFAULT_RPTR_BLOCK_SIZE |
344 				DEFAULT_MIN_AVAIL_SIZE |
345 				PRIV_STATE |
346 				KMD_QUEUE;
347 
348 	/*
349 	 * Calculating queue size which is log base 2 of actual queue
350 	 * size -1 dwords
351 	 */
352 	m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1;
353 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
354 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
355 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
356 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
357 	m->cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(q->doorbell_off);
358 
359 	m->cp_hqd_vmid = q->vmid;
360 
361 	q->is_active = (q->queue_size > 0 &&
362 			q->queue_address != 0 &&
363 			q->queue_percent > 0 &&
364 			!q->is_evicted);
365 
366 	return 0;
367 }
368 
369 #if defined(CONFIG_DEBUG_FS)
370 
371 static int debugfs_show_mqd(struct seq_file *m, void *data)
372 {
373 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
374 		     data, sizeof(struct cik_mqd), false);
375 	return 0;
376 }
377 
378 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
379 {
380 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
381 		     data, sizeof(struct cik_sdma_rlc_registers), false);
382 	return 0;
383 }
384 
385 #endif
386 
387 
388 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
389 		struct kfd_dev *dev)
390 {
391 	struct mqd_manager *mqd;
392 
393 	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
394 		return NULL;
395 
396 	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
397 	if (!mqd)
398 		return NULL;
399 
400 	mqd->dev = dev;
401 
402 	switch (type) {
403 	case KFD_MQD_TYPE_CP:
404 	case KFD_MQD_TYPE_COMPUTE:
405 		mqd->init_mqd = init_mqd;
406 		mqd->uninit_mqd = uninit_mqd;
407 		mqd->load_mqd = load_mqd;
408 		mqd->update_mqd = update_mqd;
409 		mqd->destroy_mqd = destroy_mqd;
410 		mqd->is_occupied = is_occupied;
411 		mqd->mqd_size = sizeof(struct cik_mqd);
412 #if defined(CONFIG_DEBUG_FS)
413 		mqd->debugfs_show_mqd = debugfs_show_mqd;
414 #endif
415 		break;
416 	case KFD_MQD_TYPE_HIQ:
417 		mqd->init_mqd = init_mqd_hiq;
418 		mqd->uninit_mqd = uninit_mqd_hiq_sdma;
419 		mqd->load_mqd = load_mqd;
420 		mqd->update_mqd = update_mqd_hiq;
421 		mqd->destroy_mqd = destroy_mqd;
422 		mqd->is_occupied = is_occupied;
423 		mqd->mqd_size = sizeof(struct cik_mqd);
424 #if defined(CONFIG_DEBUG_FS)
425 		mqd->debugfs_show_mqd = debugfs_show_mqd;
426 #endif
427 		break;
428 	case KFD_MQD_TYPE_DIQ:
429 		mqd->init_mqd = init_mqd_hiq;
430 		mqd->uninit_mqd = uninit_mqd;
431 		mqd->load_mqd = load_mqd;
432 		mqd->update_mqd = update_mqd_hiq;
433 		mqd->destroy_mqd = destroy_mqd;
434 		mqd->is_occupied = is_occupied;
435 		mqd->mqd_size = sizeof(struct cik_mqd);
436 #if defined(CONFIG_DEBUG_FS)
437 		mqd->debugfs_show_mqd = debugfs_show_mqd;
438 #endif
439 		break;
440 	case KFD_MQD_TYPE_SDMA:
441 		mqd->init_mqd = init_mqd_sdma;
442 		mqd->uninit_mqd = uninit_mqd_hiq_sdma;
443 		mqd->load_mqd = load_mqd_sdma;
444 		mqd->update_mqd = update_mqd_sdma;
445 		mqd->destroy_mqd = destroy_mqd_sdma;
446 		mqd->is_occupied = is_occupied_sdma;
447 		mqd->mqd_size = sizeof(struct cik_sdma_rlc_registers);
448 #if defined(CONFIG_DEBUG_FS)
449 		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
450 #endif
451 		break;
452 	default:
453 		kfree(mqd);
454 		return NULL;
455 	}
456 
457 	return mqd;
458 }
459 
460 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
461 			struct kfd_dev *dev)
462 {
463 	struct mqd_manager *mqd;
464 
465 	mqd = mqd_manager_init_cik(type, dev);
466 	if (!mqd)
467 		return NULL;
468 	if ((type == KFD_MQD_TYPE_CP) || (type == KFD_MQD_TYPE_COMPUTE))
469 		mqd->update_mqd = update_mqd_hawaii;
470 	return mqd;
471 }
472