1 /*
2  * Copyright 2018 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/uaccess.h>
27 #include "kfd_priv.h"
28 #include "kfd_mqd_manager.h"
29 #include "v10_structs.h"
30 #include "gc/gc_10_1_0_offset.h"
31 #include "gc/gc_10_1_0_sh_mask.h"
32 #include "amdgpu_amdkfd.h"
33 
34 static inline struct v10_compute_mqd *get_mqd(void *mqd)
35 {
36 	return (struct v10_compute_mqd *)mqd;
37 }
38 
39 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
40 {
41 	return (struct v10_sdma_mqd *)mqd;
42 }
43 
44 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
45 			   struct queue_properties *q)
46 {
47 	struct v10_compute_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 void set_priority(struct v10_compute_mqd *m, struct queue_properties *q)
70 {
71 	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
72 	m->cp_hqd_queue_priority = q->priority;
73 }
74 
75 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd,
76 		struct queue_properties *q)
77 {
78 	struct kfd_mem_obj *mqd_mem_obj;
79 
80 	if (kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd),
81 			&mqd_mem_obj))
82 		return NULL;
83 
84 	return mqd_mem_obj;
85 }
86 
87 static void init_mqd(struct mqd_manager *mm, void **mqd,
88 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
89 			struct queue_properties *q)
90 {
91 	uint64_t addr;
92 	struct v10_compute_mqd *m;
93 
94 	m = (struct v10_compute_mqd *) mqd_mem_obj->cpu_ptr;
95 	addr = mqd_mem_obj->gpu_addr;
96 
97 	memset(m, 0, sizeof(struct v10_compute_mqd));
98 
99 	m->header = 0xC0310800;
100 	m->compute_pipelinestat_enable = 1;
101 	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
102 	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
103 	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
104 	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
105 
106 	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
107 			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
108 
109 	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
110 
111 	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
112 	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
113 
114 	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
115 			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
116 			10 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
117 
118 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
119 		m->cp_hqd_aql_control =
120 			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
121 	}
122 
123 	if (mm->dev->cwsr_enabled) {
124 		m->cp_hqd_persistent_state |=
125 			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
126 		m->cp_hqd_ctx_save_base_addr_lo =
127 			lower_32_bits(q->ctx_save_restore_area_address);
128 		m->cp_hqd_ctx_save_base_addr_hi =
129 			upper_32_bits(q->ctx_save_restore_area_address);
130 		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
131 		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
132 		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
133 		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
134 	}
135 
136 	*mqd = m;
137 	if (gart_addr)
138 		*gart_addr = addr;
139 	mm->update_mqd(mm, m, q);
140 }
141 
142 static int load_mqd(struct mqd_manager *mm, void *mqd,
143 			uint32_t pipe_id, uint32_t queue_id,
144 			struct queue_properties *p, struct mm_struct *mms)
145 {
146 	int r = 0;
147 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
148 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
149 
150 	r = mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id,
151 					  (uint32_t __user *)p->write_ptr,
152 					  wptr_shift, 0, mms);
153 	return r;
154 }
155 
156 static void update_mqd(struct mqd_manager *mm, void *mqd,
157 		      struct queue_properties *q)
158 {
159 	struct v10_compute_mqd *m;
160 
161 	m = get_mqd(mqd);
162 
163 	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
164 	m->cp_hqd_pq_control |=
165 			ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
166 	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
167 
168 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
169 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
170 
171 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
172 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
173 	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
174 	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
175 
176 	m->cp_hqd_pq_doorbell_control =
177 		q->doorbell_off <<
178 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
179 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
180 			m->cp_hqd_pq_doorbell_control);
181 
182 	m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
183 
184 	/*
185 	 * HW does not clamp this field correctly. Maximum EOP queue size
186 	 * is constrained by per-SE EOP done signal count, which is 8-bit.
187 	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
188 	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
189 	 * is safe, giving a maximum field value of 0xA.
190 	 */
191 	m->cp_hqd_eop_control = min(0xA,
192 		ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
193 	m->cp_hqd_eop_base_addr_lo =
194 			lower_32_bits(q->eop_ring_buffer_address >> 8);
195 	m->cp_hqd_eop_base_addr_hi =
196 			upper_32_bits(q->eop_ring_buffer_address >> 8);
197 
198 	m->cp_hqd_iq_timer = 0;
199 
200 	m->cp_hqd_vmid = q->vmid;
201 
202 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
203 		/* GC 10 removed WPP_CLAMP from PQ Control */
204 		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
205 				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
206 				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT ;
207 		m->cp_hqd_pq_doorbell_control |=
208 			1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
209 	}
210 	if (mm->dev->cwsr_enabled)
211 		m->cp_hqd_ctx_save_control = 0;
212 
213 	update_cu_mask(mm, mqd, q);
214 	set_priority(m, q);
215 
216 	q->is_active = QUEUE_IS_ACTIVE(*q);
217 }
218 
219 static int destroy_mqd(struct mqd_manager *mm, void *mqd,
220 		       enum kfd_preempt_type type,
221 		       unsigned int timeout, uint32_t pipe_id,
222 		       uint32_t queue_id)
223 {
224 	return mm->dev->kfd2kgd->hqd_destroy
225 		(mm->dev->kgd, mqd, type, timeout,
226 		 pipe_id, queue_id);
227 }
228 
229 static void free_mqd(struct mqd_manager *mm, void *mqd,
230 			struct kfd_mem_obj *mqd_mem_obj)
231 {
232 	kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
233 }
234 
235 static bool is_occupied(struct mqd_manager *mm, void *mqd,
236 			uint64_t queue_address,	uint32_t pipe_id,
237 			uint32_t queue_id)
238 {
239 	return mm->dev->kfd2kgd->hqd_is_occupied(
240 		mm->dev->kgd, queue_address,
241 		pipe_id, queue_id);
242 }
243 
244 static int get_wave_state(struct mqd_manager *mm, void *mqd,
245 			  void __user *ctl_stack,
246 			  u32 *ctl_stack_used_size,
247 			  u32 *save_area_used_size)
248 {
249 	struct v10_compute_mqd *m;
250 
251 	m = get_mqd(mqd);
252 
253 	/* Control stack is written backwards, while workgroup context data
254 	 * is written forwards. Both starts from m->cp_hqd_cntl_stack_size.
255 	 * Current position is at m->cp_hqd_cntl_stack_offset and
256 	 * m->cp_hqd_wg_state_offset, respectively.
257 	 */
258 	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
259 		m->cp_hqd_cntl_stack_offset;
260 	*save_area_used_size = m->cp_hqd_wg_state_offset -
261 		m->cp_hqd_cntl_stack_size;
262 
263 	/* Control stack is not copied to user mode for GFXv10 because
264 	 * it's part of the context save area that is already
265 	 * accessible to user mode
266 	 */
267 
268 	return 0;
269 }
270 
271 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
272 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
273 			struct queue_properties *q)
274 {
275 	struct v10_compute_mqd *m;
276 
277 	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
278 
279 	m = get_mqd(*mqd);
280 
281 	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
282 			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
283 }
284 
285 static void update_mqd_hiq(struct mqd_manager *mm, void *mqd,
286 			struct queue_properties *q)
287 {
288 	struct v10_compute_mqd *m;
289 
290 	update_mqd(mm, mqd, q);
291 
292 	/* TODO: what's the point? update_mqd already does this. */
293 	m = get_mqd(mqd);
294 	m->cp_hqd_vmid = q->vmid;
295 }
296 
297 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
298 		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
299 		struct queue_properties *q)
300 {
301 	struct v10_sdma_mqd *m;
302 
303 	m = (struct v10_sdma_mqd *) mqd_mem_obj->cpu_ptr;
304 
305 	memset(m, 0, sizeof(struct v10_sdma_mqd));
306 
307 	*mqd = m;
308 	if (gart_addr)
309 		*gart_addr = mqd_mem_obj->gpu_addr;
310 
311 	mm->update_mqd(mm, m, q);
312 }
313 
314 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd,
315 		uint32_t pipe_id, uint32_t queue_id,
316 		struct queue_properties *p, struct mm_struct *mms)
317 {
318 	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd,
319 					       (uint32_t __user *)p->write_ptr,
320 					       mms);
321 }
322 
323 #define SDMA_RLC_DUMMY_DEFAULT 0xf
324 
325 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
326 		struct queue_properties *q)
327 {
328 	struct v10_sdma_mqd *m;
329 
330 	m = get_sdma_mqd(mqd);
331 	m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
332 		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
333 		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
334 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
335 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
336 
337 	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
338 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
339 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
340 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
341 	m->sdmax_rlcx_doorbell_offset =
342 		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
343 
344 	m->sdma_engine_id = q->sdma_engine_id;
345 	m->sdma_queue_id = q->sdma_queue_id;
346 	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
347 
348 	q->is_active = QUEUE_IS_ACTIVE(*q);
349 }
350 
351 /*
352  *  * preempt type here is ignored because there is only one way
353  *  * to preempt sdma queue
354  */
355 static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
356 		enum kfd_preempt_type type,
357 		unsigned int timeout, uint32_t pipe_id,
358 		uint32_t queue_id)
359 {
360 	return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout);
361 }
362 
363 static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd,
364 		uint64_t queue_address, uint32_t pipe_id,
365 		uint32_t queue_id)
366 {
367 	return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd);
368 }
369 
370 #if defined(CONFIG_DEBUG_FS)
371 
372 static int debugfs_show_mqd(struct seq_file *m, void *data)
373 {
374 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
375 		     data, sizeof(struct v10_compute_mqd), false);
376 	return 0;
377 }
378 
379 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
380 {
381 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
382 		     data, sizeof(struct v10_sdma_mqd), false);
383 	return 0;
384 }
385 
386 #endif
387 
388 struct mqd_manager *mqd_manager_init_v10(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 		pr_debug("%s@%i\n", __func__, __LINE__);
406 		mqd->allocate_mqd = allocate_mqd;
407 		mqd->init_mqd = init_mqd;
408 		mqd->free_mqd = free_mqd;
409 		mqd->load_mqd = load_mqd;
410 		mqd->update_mqd = update_mqd;
411 		mqd->destroy_mqd = destroy_mqd;
412 		mqd->is_occupied = is_occupied;
413 		mqd->mqd_size = sizeof(struct v10_compute_mqd);
414 		mqd->get_wave_state = get_wave_state;
415 #if defined(CONFIG_DEBUG_FS)
416 		mqd->debugfs_show_mqd = debugfs_show_mqd;
417 #endif
418 		pr_debug("%s@%i\n", __func__, __LINE__);
419 		break;
420 	case KFD_MQD_TYPE_HIQ:
421 		pr_debug("%s@%i\n", __func__, __LINE__);
422 		mqd->allocate_mqd = allocate_hiq_mqd;
423 		mqd->init_mqd = init_mqd_hiq;
424 		mqd->free_mqd = free_mqd_hiq_sdma;
425 		mqd->load_mqd = load_mqd;
426 		mqd->update_mqd = update_mqd_hiq;
427 		mqd->destroy_mqd = destroy_mqd;
428 		mqd->is_occupied = is_occupied;
429 		mqd->mqd_size = sizeof(struct v10_compute_mqd);
430 #if defined(CONFIG_DEBUG_FS)
431 		mqd->debugfs_show_mqd = debugfs_show_mqd;
432 #endif
433 		pr_debug("%s@%i\n", __func__, __LINE__);
434 		break;
435 	case KFD_MQD_TYPE_DIQ:
436 		mqd->allocate_mqd = allocate_hiq_mqd;
437 		mqd->init_mqd = init_mqd_hiq;
438 		mqd->free_mqd = free_mqd;
439 		mqd->load_mqd = load_mqd;
440 		mqd->update_mqd = update_mqd_hiq;
441 		mqd->destroy_mqd = destroy_mqd;
442 		mqd->is_occupied = is_occupied;
443 		mqd->mqd_size = sizeof(struct v10_compute_mqd);
444 #if defined(CONFIG_DEBUG_FS)
445 		mqd->debugfs_show_mqd = debugfs_show_mqd;
446 #endif
447 		break;
448 	case KFD_MQD_TYPE_SDMA:
449 		pr_debug("%s@%i\n", __func__, __LINE__);
450 		mqd->allocate_mqd = allocate_sdma_mqd;
451 		mqd->init_mqd = init_mqd_sdma;
452 		mqd->free_mqd = free_mqd_hiq_sdma;
453 		mqd->load_mqd = load_mqd_sdma;
454 		mqd->update_mqd = update_mqd_sdma;
455 		mqd->destroy_mqd = destroy_mqd_sdma;
456 		mqd->is_occupied = is_occupied_sdma;
457 		mqd->mqd_size = sizeof(struct v10_sdma_mqd);
458 #if defined(CONFIG_DEBUG_FS)
459 		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
460 #endif
461 		pr_debug("%s@%i\n", __func__, __LINE__);
462 		break;
463 	default:
464 		kfree(mqd);
465 		return NULL;
466 	}
467 
468 	return mqd;
469 }
470