Linux Audio

Check our new training course

Loading...
v6.8
  1// SPDX-License-Identifier: GPL-2.0 OR MIT
  2/*
  3 * Copyright 2018-2022 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/printk.h>
 26#include <linux/slab.h>
 27#include <linux/uaccess.h>
 28#include "kfd_priv.h"
 29#include "kfd_mqd_manager.h"
 30#include "v10_structs.h"
 31#include "gc/gc_10_1_0_offset.h"
 32#include "gc/gc_10_1_0_sh_mask.h"
 33#include "amdgpu_amdkfd.h"
 34
 35static inline struct v10_compute_mqd *get_mqd(void *mqd)
 36{
 37	return (struct v10_compute_mqd *)mqd;
 38}
 39
 40static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
 41{
 42	return (struct v10_sdma_mqd *)mqd;
 43}
 44
 45static void update_cu_mask(struct mqd_manager *mm, void *mqd,
 46			struct mqd_update_info *minfo)
 47{
 48	struct v10_compute_mqd *m;
 49	uint32_t se_mask[4] = {0}; /* 4 is the max # of SEs */
 50
 51	if (!minfo || !minfo->cu_mask.ptr)
 52		return;
 53
 54	mqd_symmetrically_map_cu_mask(mm,
 55		minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask, 0);
 56
 57	m = get_mqd(mqd);
 58	m->compute_static_thread_mgmt_se0 = se_mask[0];
 59	m->compute_static_thread_mgmt_se1 = se_mask[1];
 60	m->compute_static_thread_mgmt_se2 = se_mask[2];
 61	m->compute_static_thread_mgmt_se3 = se_mask[3];
 62
 63	pr_debug("update cu mask to %#x %#x %#x %#x\n",
 64		m->compute_static_thread_mgmt_se0,
 65		m->compute_static_thread_mgmt_se1,
 66		m->compute_static_thread_mgmt_se2,
 67		m->compute_static_thread_mgmt_se3);
 68}
 69
 70static void set_priority(struct v10_compute_mqd *m, struct queue_properties *q)
 71{
 72	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
 73	m->cp_hqd_queue_priority = q->priority;
 74}
 75
 76static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd,
 77		struct queue_properties *q)
 78{
 79	struct kfd_mem_obj *mqd_mem_obj;
 80
 81	if (kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd),
 82			&mqd_mem_obj))
 83		return NULL;
 84
 85	return mqd_mem_obj;
 86}
 87
 88static void init_mqd(struct mqd_manager *mm, void **mqd,
 89			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
 90			struct queue_properties *q)
 91{
 92	uint64_t addr;
 93	struct v10_compute_mqd *m;
 94
 95	m = (struct v10_compute_mqd *) mqd_mem_obj->cpu_ptr;
 96	addr = mqd_mem_obj->gpu_addr;
 97
 98	memset(m, 0, sizeof(struct v10_compute_mqd));
 99
100	m->header = 0xC0310800;
101	m->compute_pipelinestat_enable = 1;
102	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
103	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
104	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
105	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
106
107	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
108			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
109
110	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
111
112	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
113	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
114
115	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
116			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
117			1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
118
119	/* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the
120	 * DISPATCH_PTR.  This is required for the kfd debugger
121	 */
122	m->cp_hqd_hq_scheduler0 = 1 << 14;
123
124	if (q->format == KFD_QUEUE_FORMAT_AQL) {
125		m->cp_hqd_aql_control =
126			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
127	}
128
129	if (mm->dev->kfd->cwsr_enabled) {
130		m->cp_hqd_persistent_state |=
131			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
132		m->cp_hqd_ctx_save_base_addr_lo =
133			lower_32_bits(q->ctx_save_restore_area_address);
134		m->cp_hqd_ctx_save_base_addr_hi =
135			upper_32_bits(q->ctx_save_restore_area_address);
136		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
137		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
138		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
139		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
140	}
141
142	*mqd = m;
143	if (gart_addr)
144		*gart_addr = addr;
145	mm->update_mqd(mm, m, q, NULL);
146}
147
148static int load_mqd(struct mqd_manager *mm, void *mqd,
149			uint32_t pipe_id, uint32_t queue_id,
150			struct queue_properties *p, struct mm_struct *mms)
151{
152	int r = 0;
153	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
154	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
155
156	r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
157					  (uint32_t __user *)p->write_ptr,
158					  wptr_shift, 0, mms, 0);
159	return r;
160}
161
 
 
 
 
 
 
 
 
162static void update_mqd(struct mqd_manager *mm, void *mqd,
163			struct queue_properties *q,
164			struct mqd_update_info *minfo)
165{
166	struct v10_compute_mqd *m;
167
168	m = get_mqd(mqd);
169
170	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
171	m->cp_hqd_pq_control |=
172			ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
173	m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__UNORD_DISPATCH_MASK;
174	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
175
176	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
177	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
178
179	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
180	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
181	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
182	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
183
184	m->cp_hqd_pq_doorbell_control =
185		q->doorbell_off <<
186			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
187	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
188			m->cp_hqd_pq_doorbell_control);
189
190	m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
191
192	/*
193	 * HW does not clamp this field correctly. Maximum EOP queue size
194	 * is constrained by per-SE EOP done signal count, which is 8-bit.
195	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
196	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
197	 * is safe, giving a maximum field value of 0xA.
198	 */
199	m->cp_hqd_eop_control = min(0xA,
200		ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
201	m->cp_hqd_eop_base_addr_lo =
202			lower_32_bits(q->eop_ring_buffer_address >> 8);
203	m->cp_hqd_eop_base_addr_hi =
204			upper_32_bits(q->eop_ring_buffer_address >> 8);
205
206	m->cp_hqd_iq_timer = 0;
207
208	m->cp_hqd_vmid = q->vmid;
209
210	if (q->format == KFD_QUEUE_FORMAT_AQL) {
211		/* GC 10 removed WPP_CLAMP from PQ Control */
212		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
213				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
214				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT;
215		m->cp_hqd_pq_doorbell_control |=
216			1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
217	}
218	if (mm->dev->kfd->cwsr_enabled)
219		m->cp_hqd_ctx_save_control = 0;
220
221	update_cu_mask(mm, mqd, minfo);
222	set_priority(m, q);
223
224	q->is_active = QUEUE_IS_ACTIVE(*q);
225}
226
227static uint32_t read_doorbell_id(void *mqd)
228{
229	struct v10_compute_mqd *m = (struct v10_compute_mqd *)mqd;
230
231	return m->queue_doorbell_id0;
232}
233
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
234static int get_wave_state(struct mqd_manager *mm, void *mqd,
235			  struct queue_properties *q,
236			  void __user *ctl_stack,
237			  u32 *ctl_stack_used_size,
238			  u32 *save_area_used_size)
239{
240	struct v10_compute_mqd *m;
241	struct kfd_context_save_area_header header;
242
243	m = get_mqd(mqd);
244
245	/* Control stack is written backwards, while workgroup context data
246	 * is written forwards. Both starts from m->cp_hqd_cntl_stack_size.
247	 * Current position is at m->cp_hqd_cntl_stack_offset and
248	 * m->cp_hqd_wg_state_offset, respectively.
249	 */
250	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
251		m->cp_hqd_cntl_stack_offset;
252	*save_area_used_size = m->cp_hqd_wg_state_offset -
253		m->cp_hqd_cntl_stack_size;
254
255	/* Control stack is not copied to user mode for GFXv10 because
256	 * it's part of the context save area that is already
257	 * accessible to user mode
258	 */
259
260	header.wave_state.control_stack_size = *ctl_stack_used_size;
261	header.wave_state.wave_state_size = *save_area_used_size;
262
263	header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset;
264	header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset;
265
266	if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state)))
267		return -EFAULT;
268
269	return 0;
270}
271
272static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst)
273{
274	struct v10_compute_mqd *m;
275
276	m = get_mqd(mqd);
277
278	memcpy(mqd_dst, m, sizeof(struct v10_compute_mqd));
279}
280
281static void restore_mqd(struct mqd_manager *mm, void **mqd,
282			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
283			struct queue_properties *qp,
284			const void *mqd_src,
285			const void *ctl_stack_src, const u32 ctl_stack_size)
286{
287	uint64_t addr;
288	struct v10_compute_mqd *m;
289
290	m = (struct v10_compute_mqd *) mqd_mem_obj->cpu_ptr;
291	addr = mqd_mem_obj->gpu_addr;
292
293	memcpy(m, mqd_src, sizeof(*m));
294
295	*mqd = m;
296	if (gart_addr)
297		*gart_addr = addr;
298
299	m->cp_hqd_pq_doorbell_control =
300		qp->doorbell_off <<
301			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
302	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
303			m->cp_hqd_pq_doorbell_control);
304
305	qp->is_active = 0;
306}
307
308static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
309			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
310			struct queue_properties *q)
311{
312	struct v10_compute_mqd *m;
313
314	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
315
316	m = get_mqd(*mqd);
317
318	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
319			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
320}
321
322static int destroy_hiq_mqd(struct mqd_manager *mm, void *mqd,
323			enum kfd_preempt_type type, unsigned int timeout,
324			uint32_t pipe_id, uint32_t queue_id)
325{
326	int err;
327	struct v10_compute_mqd *m;
328	u32 doorbell_off;
329
330	m = get_mqd(mqd);
331
332	doorbell_off = m->cp_hqd_pq_doorbell_control >>
333			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
334
335	err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, 0);
336	if (err)
337		pr_debug("Destroy HIQ MQD failed: %d\n", err);
338
339	return err;
340}
341
342static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
343		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
344		struct queue_properties *q)
345{
346	struct v10_sdma_mqd *m;
347
348	m = (struct v10_sdma_mqd *) mqd_mem_obj->cpu_ptr;
349
350	memset(m, 0, sizeof(struct v10_sdma_mqd));
351
352	*mqd = m;
353	if (gart_addr)
354		*gart_addr = mqd_mem_obj->gpu_addr;
355
356	mm->update_mqd(mm, m, q, NULL);
 
 
 
 
 
 
 
 
 
357}
358
359#define SDMA_RLC_DUMMY_DEFAULT 0xf
360
361static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
362			struct queue_properties *q,
363			struct mqd_update_info *minfo)
364{
365	struct v10_sdma_mqd *m;
366
367	m = get_sdma_mqd(mqd);
368	m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
369		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
370		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
371		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
372		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
373
374	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
375	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
376	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
377	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
378	m->sdmax_rlcx_doorbell_offset =
379		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
380
381	m->sdma_engine_id = q->sdma_engine_id;
382	m->sdma_queue_id = q->sdma_queue_id;
383	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
384
385	q->is_active = QUEUE_IS_ACTIVE(*q);
386}
387
388static void checkpoint_mqd_sdma(struct mqd_manager *mm,
389				void *mqd,
390				void *mqd_dst,
391				void *ctl_stack_dst)
 
 
 
 
392{
393	struct v10_sdma_mqd *m;
394
395	m = get_sdma_mqd(mqd);
396
397	memcpy(mqd_dst, m, sizeof(struct v10_sdma_mqd));
398}
399
400static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd,
401			     struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
402			     struct queue_properties *qp,
403			     const void *mqd_src,
404			     const void *ctl_stack_src,
405			     const u32 ctl_stack_size)
406{
407	uint64_t addr;
408	struct v10_sdma_mqd *m;
409
410	m = (struct v10_sdma_mqd *) mqd_mem_obj->cpu_ptr;
411	addr = mqd_mem_obj->gpu_addr;
412
413	memcpy(m, mqd_src, sizeof(*m));
414
415	m->sdmax_rlcx_doorbell_offset =
416		qp->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
417
418	*mqd = m;
419	if (gart_addr)
420		*gart_addr = addr;
421
422	qp->is_active = 0;
423}
424
425#if defined(CONFIG_DEBUG_FS)
426
427static int debugfs_show_mqd(struct seq_file *m, void *data)
428{
429	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
430		     data, sizeof(struct v10_compute_mqd), false);
431	return 0;
432}
433
434static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
435{
436	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
437		     data, sizeof(struct v10_sdma_mqd), false);
438	return 0;
439}
440
441#endif
442
443struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
444		struct kfd_node *dev)
445{
446	struct mqd_manager *mqd;
447
448	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
449		return NULL;
450
451	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
452	if (!mqd)
453		return NULL;
454
455	mqd->dev = dev;
456
457	switch (type) {
458	case KFD_MQD_TYPE_CP:
459		pr_debug("%s@%i\n", __func__, __LINE__);
460		mqd->allocate_mqd = allocate_mqd;
461		mqd->init_mqd = init_mqd;
462		mqd->free_mqd = kfd_free_mqd_cp;
463		mqd->load_mqd = load_mqd;
464		mqd->update_mqd = update_mqd;
465		mqd->destroy_mqd = kfd_destroy_mqd_cp;
466		mqd->is_occupied = kfd_is_occupied_cp;
467		mqd->mqd_size = sizeof(struct v10_compute_mqd);
468		mqd->get_wave_state = get_wave_state;
469		mqd->checkpoint_mqd = checkpoint_mqd;
470		mqd->restore_mqd = restore_mqd;
471		mqd->mqd_stride = kfd_mqd_stride;
472#if defined(CONFIG_DEBUG_FS)
473		mqd->debugfs_show_mqd = debugfs_show_mqd;
474#endif
475		pr_debug("%s@%i\n", __func__, __LINE__);
476		break;
477	case KFD_MQD_TYPE_HIQ:
478		pr_debug("%s@%i\n", __func__, __LINE__);
479		mqd->allocate_mqd = allocate_hiq_mqd;
480		mqd->init_mqd = init_mqd_hiq;
481		mqd->free_mqd = free_mqd_hiq_sdma;
482		mqd->load_mqd = kfd_hiq_load_mqd_kiq;
483		mqd->update_mqd = update_mqd;
484		mqd->destroy_mqd = destroy_hiq_mqd;
485		mqd->is_occupied = kfd_is_occupied_cp;
486		mqd->mqd_size = sizeof(struct v10_compute_mqd);
487		mqd->mqd_stride = kfd_mqd_stride;
488#if defined(CONFIG_DEBUG_FS)
489		mqd->debugfs_show_mqd = debugfs_show_mqd;
490#endif
491		mqd->read_doorbell_id = read_doorbell_id;
492		pr_debug("%s@%i\n", __func__, __LINE__);
493		break;
494	case KFD_MQD_TYPE_DIQ:
495		mqd->allocate_mqd = allocate_mqd;
496		mqd->init_mqd = init_mqd_hiq;
497		mqd->free_mqd = kfd_free_mqd_cp;
498		mqd->load_mqd = load_mqd;
499		mqd->update_mqd = update_mqd;
500		mqd->destroy_mqd = kfd_destroy_mqd_cp;
501		mqd->is_occupied = kfd_is_occupied_cp;
502		mqd->mqd_size = sizeof(struct v10_compute_mqd);
503#if defined(CONFIG_DEBUG_FS)
504		mqd->debugfs_show_mqd = debugfs_show_mqd;
505#endif
506		break;
507	case KFD_MQD_TYPE_SDMA:
508		pr_debug("%s@%i\n", __func__, __LINE__);
509		mqd->allocate_mqd = allocate_sdma_mqd;
510		mqd->init_mqd = init_mqd_sdma;
511		mqd->free_mqd = free_mqd_hiq_sdma;
512		mqd->load_mqd = kfd_load_mqd_sdma;
513		mqd->update_mqd = update_mqd_sdma;
514		mqd->destroy_mqd = kfd_destroy_mqd_sdma;
515		mqd->is_occupied = kfd_is_occupied_sdma;
516		mqd->checkpoint_mqd = checkpoint_mqd_sdma;
517		mqd->restore_mqd = restore_mqd_sdma;
518		mqd->mqd_size = sizeof(struct v10_sdma_mqd);
519		mqd->mqd_stride = kfd_mqd_stride;
520#if defined(CONFIG_DEBUG_FS)
521		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
522#endif
523		pr_debug("%s@%i\n", __func__, __LINE__);
524		break;
525	default:
526		kfree(mqd);
527		return NULL;
528	}
529
530	return mqd;
531}
v5.14.15
 
  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
 34static inline struct v10_compute_mqd *get_mqd(void *mqd)
 35{
 36	return (struct v10_compute_mqd *)mqd;
 37}
 38
 39static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
 40{
 41	return (struct v10_sdma_mqd *)mqd;
 42}
 43
 44static 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
 69static 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
 75static 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
 87static 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			1 << 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
142static 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
156static int hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd,
157			    uint32_t pipe_id, uint32_t queue_id,
158			    struct queue_properties *p, struct mm_struct *mms)
159{
160	return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->kgd, mqd, pipe_id,
161					      queue_id, p->doorbell_off);
162}
163
164static void update_mqd(struct mqd_manager *mm, void *mqd,
165		      struct queue_properties *q)
 
166{
167	struct v10_compute_mqd *m;
168
169	m = get_mqd(mqd);
170
171	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
172	m->cp_hqd_pq_control |=
173			ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
 
174	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
175
176	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
177	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
178
179	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
180	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
181	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
182	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
183
184	m->cp_hqd_pq_doorbell_control =
185		q->doorbell_off <<
186			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
187	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
188			m->cp_hqd_pq_doorbell_control);
189
190	m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
191
192	/*
193	 * HW does not clamp this field correctly. Maximum EOP queue size
194	 * is constrained by per-SE EOP done signal count, which is 8-bit.
195	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
196	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
197	 * is safe, giving a maximum field value of 0xA.
198	 */
199	m->cp_hqd_eop_control = min(0xA,
200		ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
201	m->cp_hqd_eop_base_addr_lo =
202			lower_32_bits(q->eop_ring_buffer_address >> 8);
203	m->cp_hqd_eop_base_addr_hi =
204			upper_32_bits(q->eop_ring_buffer_address >> 8);
205
206	m->cp_hqd_iq_timer = 0;
207
208	m->cp_hqd_vmid = q->vmid;
209
210	if (q->format == KFD_QUEUE_FORMAT_AQL) {
211		/* GC 10 removed WPP_CLAMP from PQ Control */
212		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
213				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
214				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT ;
215		m->cp_hqd_pq_doorbell_control |=
216			1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
217	}
218	if (mm->dev->cwsr_enabled)
219		m->cp_hqd_ctx_save_control = 0;
220
221	update_cu_mask(mm, mqd, q);
222	set_priority(m, q);
223
224	q->is_active = QUEUE_IS_ACTIVE(*q);
225}
226
227static uint32_t read_doorbell_id(void *mqd)
228{
229	struct v10_compute_mqd *m = (struct v10_compute_mqd *)mqd;
230
231	return m->queue_doorbell_id0;
232}
233
234static int destroy_mqd(struct mqd_manager *mm, void *mqd,
235		       enum kfd_preempt_type type,
236		       unsigned int timeout, uint32_t pipe_id,
237		       uint32_t queue_id)
238{
239	return mm->dev->kfd2kgd->hqd_destroy
240		(mm->dev->kgd, mqd, type, timeout,
241		 pipe_id, queue_id);
242}
243
244static void free_mqd(struct mqd_manager *mm, void *mqd,
245			struct kfd_mem_obj *mqd_mem_obj)
246{
247	kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
248}
249
250static bool is_occupied(struct mqd_manager *mm, void *mqd,
251			uint64_t queue_address,	uint32_t pipe_id,
252			uint32_t queue_id)
253{
254	return mm->dev->kfd2kgd->hqd_is_occupied(
255		mm->dev->kgd, queue_address,
256		pipe_id, queue_id);
257}
258
259static int get_wave_state(struct mqd_manager *mm, void *mqd,
 
260			  void __user *ctl_stack,
261			  u32 *ctl_stack_used_size,
262			  u32 *save_area_used_size)
263{
264	struct v10_compute_mqd *m;
 
265
266	m = get_mqd(mqd);
267
268	/* Control stack is written backwards, while workgroup context data
269	 * is written forwards. Both starts from m->cp_hqd_cntl_stack_size.
270	 * Current position is at m->cp_hqd_cntl_stack_offset and
271	 * m->cp_hqd_wg_state_offset, respectively.
272	 */
273	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
274		m->cp_hqd_cntl_stack_offset;
275	*save_area_used_size = m->cp_hqd_wg_state_offset -
276		m->cp_hqd_cntl_stack_size;
277
278	/* Control stack is not copied to user mode for GFXv10 because
279	 * it's part of the context save area that is already
280	 * accessible to user mode
281	 */
282
 
 
 
 
 
 
 
 
 
283	return 0;
284}
285
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
286static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
287			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
288			struct queue_properties *q)
289{
290	struct v10_compute_mqd *m;
291
292	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
293
294	m = get_mqd(*mqd);
295
296	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
297			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
298}
299
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
300static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
301		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
302		struct queue_properties *q)
303{
304	struct v10_sdma_mqd *m;
305
306	m = (struct v10_sdma_mqd *) mqd_mem_obj->cpu_ptr;
307
308	memset(m, 0, sizeof(struct v10_sdma_mqd));
309
310	*mqd = m;
311	if (gart_addr)
312		*gart_addr = mqd_mem_obj->gpu_addr;
313
314	mm->update_mqd(mm, m, q);
315}
316
317static int load_mqd_sdma(struct mqd_manager *mm, void *mqd,
318		uint32_t pipe_id, uint32_t queue_id,
319		struct queue_properties *p, struct mm_struct *mms)
320{
321	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd,
322					       (uint32_t __user *)p->write_ptr,
323					       mms);
324}
325
326#define SDMA_RLC_DUMMY_DEFAULT 0xf
327
328static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
329		struct queue_properties *q)
 
330{
331	struct v10_sdma_mqd *m;
332
333	m = get_sdma_mqd(mqd);
334	m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
335		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
336		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
337		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
338		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
339
340	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
341	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
342	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
343	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
344	m->sdmax_rlcx_doorbell_offset =
345		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
346
347	m->sdma_engine_id = q->sdma_engine_id;
348	m->sdma_queue_id = q->sdma_queue_id;
349	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
350
351	q->is_active = QUEUE_IS_ACTIVE(*q);
352}
353
354/*
355 *  * preempt type here is ignored because there is only one way
356 *  * to preempt sdma queue
357 */
358static int destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
359		enum kfd_preempt_type type,
360		unsigned int timeout, uint32_t pipe_id,
361		uint32_t queue_id)
362{
363	return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->kgd, mqd, timeout);
 
 
 
 
364}
365
366static bool is_occupied_sdma(struct mqd_manager *mm, void *mqd,
367		uint64_t queue_address, uint32_t pipe_id,
368		uint32_t queue_id)
 
 
 
369{
370	return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->kgd, mqd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
371}
372
373#if defined(CONFIG_DEBUG_FS)
374
375static int debugfs_show_mqd(struct seq_file *m, void *data)
376{
377	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
378		     data, sizeof(struct v10_compute_mqd), false);
379	return 0;
380}
381
382static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
383{
384	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
385		     data, sizeof(struct v10_sdma_mqd), false);
386	return 0;
387}
388
389#endif
390
391struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
392		struct kfd_dev *dev)
393{
394	struct mqd_manager *mqd;
395
396	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
397		return NULL;
398
399	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
400	if (!mqd)
401		return NULL;
402
403	mqd->dev = dev;
404
405	switch (type) {
406	case KFD_MQD_TYPE_CP:
407		pr_debug("%s@%i\n", __func__, __LINE__);
408		mqd->allocate_mqd = allocate_mqd;
409		mqd->init_mqd = init_mqd;
410		mqd->free_mqd = free_mqd;
411		mqd->load_mqd = load_mqd;
412		mqd->update_mqd = update_mqd;
413		mqd->destroy_mqd = destroy_mqd;
414		mqd->is_occupied = is_occupied;
415		mqd->mqd_size = sizeof(struct v10_compute_mqd);
416		mqd->get_wave_state = get_wave_state;
 
 
 
417#if defined(CONFIG_DEBUG_FS)
418		mqd->debugfs_show_mqd = debugfs_show_mqd;
419#endif
420		pr_debug("%s@%i\n", __func__, __LINE__);
421		break;
422	case KFD_MQD_TYPE_HIQ:
423		pr_debug("%s@%i\n", __func__, __LINE__);
424		mqd->allocate_mqd = allocate_hiq_mqd;
425		mqd->init_mqd = init_mqd_hiq;
426		mqd->free_mqd = free_mqd_hiq_sdma;
427		mqd->load_mqd = hiq_load_mqd_kiq;
428		mqd->update_mqd = update_mqd;
429		mqd->destroy_mqd = destroy_mqd;
430		mqd->is_occupied = is_occupied;
431		mqd->mqd_size = sizeof(struct v10_compute_mqd);
 
432#if defined(CONFIG_DEBUG_FS)
433		mqd->debugfs_show_mqd = debugfs_show_mqd;
434#endif
435		mqd->read_doorbell_id = read_doorbell_id;
436		pr_debug("%s@%i\n", __func__, __LINE__);
437		break;
438	case KFD_MQD_TYPE_DIQ:
439		mqd->allocate_mqd = allocate_mqd;
440		mqd->init_mqd = init_mqd_hiq;
441		mqd->free_mqd = free_mqd;
442		mqd->load_mqd = load_mqd;
443		mqd->update_mqd = update_mqd;
444		mqd->destroy_mqd = destroy_mqd;
445		mqd->is_occupied = is_occupied;
446		mqd->mqd_size = sizeof(struct v10_compute_mqd);
447#if defined(CONFIG_DEBUG_FS)
448		mqd->debugfs_show_mqd = debugfs_show_mqd;
449#endif
450		break;
451	case KFD_MQD_TYPE_SDMA:
452		pr_debug("%s@%i\n", __func__, __LINE__);
453		mqd->allocate_mqd = allocate_sdma_mqd;
454		mqd->init_mqd = init_mqd_sdma;
455		mqd->free_mqd = free_mqd_hiq_sdma;
456		mqd->load_mqd = load_mqd_sdma;
457		mqd->update_mqd = update_mqd_sdma;
458		mqd->destroy_mqd = destroy_mqd_sdma;
459		mqd->is_occupied = is_occupied_sdma;
 
 
460		mqd->mqd_size = sizeof(struct v10_sdma_mqd);
 
461#if defined(CONFIG_DEBUG_FS)
462		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
463#endif
464		pr_debug("%s@%i\n", __func__, __LINE__);
465		break;
466	default:
467		kfree(mqd);
468		return NULL;
469	}
470
471	return mqd;
472}