1// SPDX-License-Identifier: GPL-2.0 OR MIT
  2/*
  3 * Copyright 2016-2022 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 "v9_structs.h"
 31#include "gc/gc_9_0_offset.h"
 32#include "gc/gc_9_0_sh_mask.h"
 33#include "sdma0/sdma0_4_0_sh_mask.h"
 34#include "amdgpu_amdkfd.h"
 35#include "kfd_device_queue_manager.h"
 36
 37static void update_mqd(struct mqd_manager *mm, void *mqd,
 38		       struct queue_properties *q,
 39		       struct mqd_update_info *minfo);
 40
 41static uint64_t mqd_stride_v9(struct mqd_manager *mm,
 42				struct queue_properties *q)
 43{
 44	if (mm->dev->kfd->cwsr_enabled &&
 45	    q->type == KFD_QUEUE_TYPE_COMPUTE)
 46		return ALIGN(q->ctl_stack_size, PAGE_SIZE) +
 47			ALIGN(sizeof(struct v9_mqd), PAGE_SIZE);
 48
 49	return mm->mqd_size;
 50}
 51
 52static inline struct v9_mqd *get_mqd(void *mqd)
 53{
 54	return (struct v9_mqd *)mqd;
 55}
 56
 57static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
 58{
 59	return (struct v9_sdma_mqd *)mqd;
 60}
 61
 62static void update_cu_mask(struct mqd_manager *mm, void *mqd,
 63			struct mqd_update_info *minfo, uint32_t inst)
 64{
 65	struct v9_mqd *m;
 66	uint32_t se_mask[KFD_MAX_NUM_SE] = {0};
 67
 68	if (!minfo || !minfo->cu_mask.ptr)
 69		return;
 70
 71	mqd_symmetrically_map_cu_mask(mm,
 72		minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask, inst);
 73
 74	m = get_mqd(mqd);
 75
 76	m->compute_static_thread_mgmt_se0 = se_mask[0];
 77	m->compute_static_thread_mgmt_se1 = se_mask[1];
 78	m->compute_static_thread_mgmt_se2 = se_mask[2];
 79	m->compute_static_thread_mgmt_se3 = se_mask[3];
 80	if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3)) {
 81		m->compute_static_thread_mgmt_se4 = se_mask[4];
 82		m->compute_static_thread_mgmt_se5 = se_mask[5];
 83		m->compute_static_thread_mgmt_se6 = se_mask[6];
 84		m->compute_static_thread_mgmt_se7 = se_mask[7];
 85
 86		pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n",
 87			m->compute_static_thread_mgmt_se0,
 88			m->compute_static_thread_mgmt_se1,
 89			m->compute_static_thread_mgmt_se2,
 90			m->compute_static_thread_mgmt_se3,
 91			m->compute_static_thread_mgmt_se4,
 92			m->compute_static_thread_mgmt_se5,
 93			m->compute_static_thread_mgmt_se6,
 94			m->compute_static_thread_mgmt_se7);
 95	} else {
 96		pr_debug("inst: %u, update cu mask to %#x %#x %#x %#x\n",
 97			inst, m->compute_static_thread_mgmt_se0,
 98			m->compute_static_thread_mgmt_se1,
 99			m->compute_static_thread_mgmt_se2,
100			m->compute_static_thread_mgmt_se3);
101	}
102}
103
104static void set_priority(struct v9_mqd *m, struct queue_properties *q)
105{
106	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
107	m->cp_hqd_queue_priority = q->priority;
108}
109
110static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
111		struct queue_properties *q)
112{
113	int retval;
114	struct kfd_mem_obj *mqd_mem_obj = NULL;
115
116	/* For V9 only, due to a HW bug, the control stack of a user mode
117	 * compute queue needs to be allocated just behind the page boundary
118	 * of its regular MQD buffer. So we allocate an enlarged MQD buffer:
119	 * the first page of the buffer serves as the regular MQD buffer
120	 * purpose and the remaining is for control stack. Although the two
121	 * parts are in the same buffer object, they need different memory
122	 * types: MQD part needs UC (uncached) as usual, while control stack
123	 * needs NC (non coherent), which is different from the UC type which
124	 * is used when control stack is allocated in user space.
125	 *
126	 * Because of all those, we use the gtt allocation function instead
127	 * of sub-allocation function for this enlarged MQD buffer. Moreover,
128	 * in order to achieve two memory types in a single buffer object, we
129	 * pass a special bo flag AMDGPU_GEM_CREATE_CP_MQD_GFX9 to instruct
130	 * amdgpu memory functions to do so.
131	 */
132	if (node->kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) {
133		mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
134		if (!mqd_mem_obj)
135			return NULL;
136		retval = amdgpu_amdkfd_alloc_gtt_mem(node->adev,
137			(ALIGN(q->ctl_stack_size, PAGE_SIZE) +
138			ALIGN(sizeof(struct v9_mqd), PAGE_SIZE)) *
139			NUM_XCC(node->xcc_mask),
140			&(mqd_mem_obj->gtt_mem),
141			&(mqd_mem_obj->gpu_addr),
142			(void *)&(mqd_mem_obj->cpu_ptr), true);
143
144		if (retval) {
145			kfree(mqd_mem_obj);
146			return NULL;
147		}
148	} else {
149		retval = kfd_gtt_sa_allocate(node, sizeof(struct v9_mqd),
150				&mqd_mem_obj);
151		if (retval)
152			return NULL;
 
 
 
153	}
154
155	return mqd_mem_obj;
 
156}
157
158static void init_mqd(struct mqd_manager *mm, void **mqd,
159			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
160			struct queue_properties *q)
161{
162	uint64_t addr;
163	struct v9_mqd *m;
164
165	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
166	addr = mqd_mem_obj->gpu_addr;
167
168	memset(m, 0, sizeof(struct v9_mqd));
169
170	m->header = 0xC0310800;
171	m->compute_pipelinestat_enable = 1;
172	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
173	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
174	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
175	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
176	m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF;
177	m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF;
178	m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF;
179	m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF;
180
181	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
182			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
183
184	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
185
186	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
187	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
188
189	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
190			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
191			1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
192
193	/* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the
194	 * DISPATCH_PTR.  This is required for the kfd debugger
195	 */
196	m->cp_hqd_hq_status0 = 1 << 14;
197
198	if (q->format == KFD_QUEUE_FORMAT_AQL)
199		m->cp_hqd_aql_control =
200			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
 
201
202	if (q->tba_addr) {
203		m->compute_pgm_rsrc2 |=
204			(1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
205	}
206
207	if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) {
208		m->cp_hqd_persistent_state |=
209			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
210		m->cp_hqd_ctx_save_base_addr_lo =
211			lower_32_bits(q->ctx_save_restore_area_address);
212		m->cp_hqd_ctx_save_base_addr_hi =
213			upper_32_bits(q->ctx_save_restore_area_address);
214		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
215		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
216		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
217		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
218	}
219
220	*mqd = m;
221	if (gart_addr)
222		*gart_addr = addr;
223	update_mqd(mm, m, q, NULL);
224}
225
226static int load_mqd(struct mqd_manager *mm, void *mqd,
227			uint32_t pipe_id, uint32_t queue_id,
228			struct queue_properties *p, struct mm_struct *mms)
229{
230	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
231	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
232
233	return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
234					  (uint32_t __user *)p->write_ptr,
235					  wptr_shift, 0, mms, 0);
 
 
 
 
 
 
 
 
236}
237
238static void update_mqd(struct mqd_manager *mm, void *mqd,
239			struct queue_properties *q,
240			struct mqd_update_info *minfo)
241{
242	struct v9_mqd *m;
243
244	m = get_mqd(mqd);
245
246	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
247	m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1;
248	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
249
250	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
251	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
252
253	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
254	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
255	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
256	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
257
258	m->cp_hqd_pq_doorbell_control =
259		q->doorbell_off <<
260			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
261	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
262			m->cp_hqd_pq_doorbell_control);
263
264	m->cp_hqd_ib_control =
265		3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT |
266		1 << CP_HQD_IB_CONTROL__IB_EXE_DISABLE__SHIFT;
267
268	/*
269	 * HW does not clamp this field correctly. Maximum EOP queue size
270	 * is constrained by per-SE EOP done signal count, which is 8-bit.
271	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
272	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
273	 * is safe, giving a maximum field value of 0xA.
274	 *
275	 * Also, do calculation only if EOP is used (size > 0), otherwise
276	 * the order_base_2 calculation provides incorrect result.
277	 *
278	 */
279	m->cp_hqd_eop_control = q->eop_ring_buffer_size ?
280		min(0xA, order_base_2(q->eop_ring_buffer_size / 4) - 1) : 0;
281
282	m->cp_hqd_eop_base_addr_lo =
283			lower_32_bits(q->eop_ring_buffer_address >> 8);
284	m->cp_hqd_eop_base_addr_hi =
285			upper_32_bits(q->eop_ring_buffer_address >> 8);
286
287	m->cp_hqd_iq_timer = 0;
288
289	m->cp_hqd_vmid = q->vmid;
290
291	if (q->format == KFD_QUEUE_FORMAT_AQL) {
292		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
293				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
294				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT |
295				1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT;
296		m->cp_hqd_pq_doorbell_control |= 1 <<
297			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
298	}
299	if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address)
300		m->cp_hqd_ctx_save_control = 0;
301
302	if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3))
303		update_cu_mask(mm, mqd, minfo, 0);
304	set_priority(m, q);
305
306	if (minfo && KFD_GC_VERSION(mm->dev) >= IP_VERSION(9, 4, 2)) {
307		if (minfo->update_flag & UPDATE_FLAG_IS_GWS)
308			m->compute_resource_limits |=
309				COMPUTE_RESOURCE_LIMITS__FORCE_SIMD_DIST_MASK;
310		else
311			m->compute_resource_limits &=
312				~COMPUTE_RESOURCE_LIMITS__FORCE_SIMD_DIST_MASK;
313	}
314
315	q->is_active = QUEUE_IS_ACTIVE(*q);
316}
317
318
319static uint32_t read_doorbell_id(void *mqd)
320{
321	struct v9_mqd *m = (struct v9_mqd *)mqd;
322
323	return m->queue_doorbell_id0;
324}
325
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
326static int get_wave_state(struct mqd_manager *mm, void *mqd,
327			  struct queue_properties *q,
328			  void __user *ctl_stack,
329			  u32 *ctl_stack_used_size,
330			  u32 *save_area_used_size)
331{
332	struct v9_mqd *m;
333	struct kfd_context_save_area_header header;
334
335	/* Control stack is located one page after MQD. */
336	void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
337
338	m = get_mqd(mqd);
339
340	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
341		m->cp_hqd_cntl_stack_offset;
342	*save_area_used_size = m->cp_hqd_wg_state_offset -
343		m->cp_hqd_cntl_stack_size;
344
345	header.wave_state.control_stack_size = *ctl_stack_used_size;
346	header.wave_state.wave_state_size = *save_area_used_size;
347
348	header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset;
349	header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset;
350
351	if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state)))
352		return -EFAULT;
353
354	if (copy_to_user(ctl_stack + m->cp_hqd_cntl_stack_offset,
355				mqd_ctl_stack + m->cp_hqd_cntl_stack_offset,
356				*ctl_stack_used_size))
357		return -EFAULT;
358
359	return 0;
360}
361
362static void get_checkpoint_info(struct mqd_manager *mm, void *mqd, u32 *ctl_stack_size)
363{
364	struct v9_mqd *m = get_mqd(mqd);
365
366	*ctl_stack_size = m->cp_hqd_cntl_stack_size;
367}
368
369static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst)
370{
371	struct v9_mqd *m;
372	/* Control stack is located one page after MQD. */
373	void *ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
374
375	m = get_mqd(mqd);
376
377	memcpy(mqd_dst, m, sizeof(struct v9_mqd));
378	memcpy(ctl_stack_dst, ctl_stack, m->cp_hqd_cntl_stack_size);
379}
380
381static void restore_mqd(struct mqd_manager *mm, void **mqd,
382			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
383			struct queue_properties *qp,
384			const void *mqd_src,
385			const void *ctl_stack_src, u32 ctl_stack_size)
386{
387	uint64_t addr;
388	struct v9_mqd *m;
389	void *ctl_stack;
390
391	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
392	addr = mqd_mem_obj->gpu_addr;
393
394	memcpy(m, mqd_src, sizeof(*m));
395
396	*mqd = m;
397	if (gart_addr)
398		*gart_addr = addr;
399
400	/* Control stack is located one page after MQD. */
401	ctl_stack = (void *)((uintptr_t)*mqd + PAGE_SIZE);
402	memcpy(ctl_stack, ctl_stack_src, ctl_stack_size);
403
404	m->cp_hqd_pq_doorbell_control =
405		qp->doorbell_off <<
406			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
407	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
408				m->cp_hqd_pq_doorbell_control);
409
410	qp->is_active = 0;
411}
412
413static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
414			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
415			struct queue_properties *q)
416{
417	struct v9_mqd *m;
418
419	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
420
421	m = get_mqd(*mqd);
422
423	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
424			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
425}
426
427static int destroy_hiq_mqd(struct mqd_manager *mm, void *mqd,
428			enum kfd_preempt_type type, unsigned int timeout,
429			uint32_t pipe_id, uint32_t queue_id)
430{
431	int err;
432	struct v9_mqd *m;
433	u32 doorbell_off;
434
435	m = get_mqd(mqd);
436
437	doorbell_off = m->cp_hqd_pq_doorbell_control >>
438			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
439	err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, 0);
440	if (err)
441		pr_debug("Destroy HIQ MQD failed: %d\n", err);
442
443	return err;
444}
445
446static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
447		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
448		struct queue_properties *q)
449{
450	struct v9_sdma_mqd *m;
451
452	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
453
454	memset(m, 0, sizeof(struct v9_sdma_mqd));
455
456	*mqd = m;
457	if (gart_addr)
458		*gart_addr = mqd_mem_obj->gpu_addr;
459
460	mm->update_mqd(mm, m, q, NULL);
 
 
 
 
 
 
 
 
 
461}
462
463#define SDMA_RLC_DUMMY_DEFAULT 0xf
464
465static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
466			struct queue_properties *q,
467			struct mqd_update_info *minfo)
468{
469	struct v9_sdma_mqd *m;
470
471	m = get_sdma_mqd(mqd);
472	m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4)
473		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
474		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
475		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
476		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
477
478	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
479	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
480	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
481	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
482	m->sdmax_rlcx_doorbell_offset =
483		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
484
485	m->sdma_engine_id = q->sdma_engine_id;
486	m->sdma_queue_id = q->sdma_queue_id;
487	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
488
489	q->is_active = QUEUE_IS_ACTIVE(*q);
490}
491
492static void checkpoint_mqd_sdma(struct mqd_manager *mm,
493				void *mqd,
494				void *mqd_dst,
495				void *ctl_stack_dst)
496{
497	struct v9_sdma_mqd *m;
498
499	m = get_sdma_mqd(mqd);
500
501	memcpy(mqd_dst, m, sizeof(struct v9_sdma_mqd));
502}
503
504static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd,
505			     struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
506			     struct queue_properties *qp,
507			     const void *mqd_src,
508			     const void *ctl_stack_src, const u32 ctl_stack_size)
509{
510	uint64_t addr;
511	struct v9_sdma_mqd *m;
512
513	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
514	addr = mqd_mem_obj->gpu_addr;
515
516	memcpy(m, mqd_src, sizeof(*m));
517
518	m->sdmax_rlcx_doorbell_offset =
519		qp->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
520
521	*mqd = m;
522	if (gart_addr)
523		*gart_addr = addr;
524
525	qp->is_active = 0;
526}
527
528static void init_mqd_hiq_v9_4_3(struct mqd_manager *mm, void **mqd,
529			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
530			struct queue_properties *q)
531{
532	struct v9_mqd *m;
533	int xcc = 0;
534	struct kfd_mem_obj xcc_mqd_mem_obj;
535	uint64_t xcc_gart_addr = 0;
536
537	memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj));
538
539	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
540		kfd_get_hiq_xcc_mqd(mm->dev, &xcc_mqd_mem_obj, xcc);
541
542		init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q);
543
544		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
545					1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
546					1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
547		m->cp_mqd_stride_size = kfd_hiq_mqd_stride(mm->dev);
548		if (xcc == 0) {
549			/* Set no_update_rptr = 0 in Master XCC */
550			m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
551
552			/* Set the MQD pointer and gart address to XCC0 MQD */
553			*mqd = m;
554			*gart_addr = xcc_gart_addr;
555		}
556	}
557}
558
559static int hiq_load_mqd_kiq_v9_4_3(struct mqd_manager *mm, void *mqd,
560			uint32_t pipe_id, uint32_t queue_id,
561			struct queue_properties *p, struct mm_struct *mms)
562{
563	uint32_t xcc_mask = mm->dev->xcc_mask;
564	int xcc_id, err, inst = 0;
565	void *xcc_mqd;
566	uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
567
568	for_each_inst(xcc_id, xcc_mask) {
569		xcc_mqd = mqd + hiq_mqd_size * inst;
570		err = mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, xcc_mqd,
571						     pipe_id, queue_id,
572						     p->doorbell_off, xcc_id);
573		if (err) {
574			pr_debug("Failed to load HIQ MQD for XCC: %d\n", inst);
575			break;
576		}
577		++inst;
578	}
579
580	return err;
581}
582
583static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
584			enum kfd_preempt_type type, unsigned int timeout,
585			uint32_t pipe_id, uint32_t queue_id)
586{
587	uint32_t xcc_mask = mm->dev->xcc_mask;
588	int xcc_id, err, inst = 0;
589	uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
590	struct v9_mqd *m;
591	u32 doorbell_off;
592
593	for_each_inst(xcc_id, xcc_mask) {
594		m = get_mqd(mqd + hiq_mqd_size * inst);
595
596		doorbell_off = m->cp_hqd_pq_doorbell_control >>
597				CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
598
599		err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, xcc_id);
600		if (err) {
601			pr_debug("Destroy HIQ MQD failed for xcc: %d\n", inst);
602			break;
603		}
604		++inst;
605	}
606
607	return err;
608}
609
610static void get_xcc_mqd(struct kfd_mem_obj *mqd_mem_obj,
611			       struct kfd_mem_obj *xcc_mqd_mem_obj,
612			       uint64_t offset)
613{
614	xcc_mqd_mem_obj->gtt_mem = (offset == 0) ?
615					mqd_mem_obj->gtt_mem : NULL;
616	xcc_mqd_mem_obj->gpu_addr = mqd_mem_obj->gpu_addr + offset;
617	xcc_mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)mqd_mem_obj->cpu_ptr
618						+ offset);
619}
620
621static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd,
622			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
623			struct queue_properties *q)
624{
625	struct v9_mqd *m;
626	int xcc = 0;
627	struct kfd_mem_obj xcc_mqd_mem_obj;
628	uint64_t xcc_gart_addr = 0;
629	uint64_t xcc_ctx_save_restore_area_address;
630	uint64_t offset = mm->mqd_stride(mm, q);
631	uint32_t local_xcc_start = mm->dev->dqm->current_logical_xcc_start++;
632
633	memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj));
634	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
635		get_xcc_mqd(mqd_mem_obj, &xcc_mqd_mem_obj, offset*xcc);
636
637		init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q);
638
639		m->cp_mqd_stride_size = offset;
640
641		/*
642		 * Update the CWSR address for each XCC if CWSR is enabled
643		 * and CWSR area is allocated in thunk
644		 */
645		if (mm->dev->kfd->cwsr_enabled &&
646		    q->ctx_save_restore_area_address) {
647			xcc_ctx_save_restore_area_address =
648				q->ctx_save_restore_area_address +
649				(xcc * q->ctx_save_restore_area_size);
650
651			m->cp_hqd_ctx_save_base_addr_lo =
652				lower_32_bits(xcc_ctx_save_restore_area_address);
653			m->cp_hqd_ctx_save_base_addr_hi =
654				upper_32_bits(xcc_ctx_save_restore_area_address);
655		}
656
657		if (q->format == KFD_QUEUE_FORMAT_AQL) {
658			m->compute_tg_chunk_size = 1;
659			m->compute_current_logic_xcc_id =
660					(local_xcc_start + xcc) %
661					NUM_XCC(mm->dev->xcc_mask);
662
663			switch (xcc) {
664			case 0:
665				/* Master XCC */
666				m->cp_hqd_pq_control &=
667					~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
668				break;
669			default:
670				break;
671			}
672		} else {
673			/* PM4 Queue */
674			m->compute_current_logic_xcc_id = 0;
675			m->compute_tg_chunk_size = 0;
676			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
677		}
678
679		if (xcc == 0) {
680			/* Set the MQD pointer and gart address to XCC0 MQD */
681			*mqd = m;
682			*gart_addr = xcc_gart_addr;
683		}
684	}
685}
686
687static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
688		      struct queue_properties *q, struct mqd_update_info *minfo)
689{
690	struct v9_mqd *m;
691	int xcc = 0;
692	uint64_t size = mm->mqd_stride(mm, q);
693
694	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
695		m = get_mqd(mqd + size * xcc);
696		update_mqd(mm, m, q, minfo);
697
698		update_cu_mask(mm, mqd, minfo, xcc);
699
700		if (q->format == KFD_QUEUE_FORMAT_AQL) {
701			switch (xcc) {
702			case 0:
703				/* Master XCC */
704				m->cp_hqd_pq_control &=
705					~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
706				break;
707			default:
708				break;
709			}
710			m->compute_tg_chunk_size = 1;
711		} else {
712			/* PM4 Queue */
713			m->compute_current_logic_xcc_id = 0;
714			m->compute_tg_chunk_size = 0;
715			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
716		}
717	}
718}
719
720static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
721		   enum kfd_preempt_type type, unsigned int timeout,
722		   uint32_t pipe_id, uint32_t queue_id)
723{
724	uint32_t xcc_mask = mm->dev->xcc_mask;
725	int xcc_id, err, inst = 0;
726	void *xcc_mqd;
727	struct v9_mqd *m;
728	uint64_t mqd_offset;
729
730	m = get_mqd(mqd);
731	mqd_offset = m->cp_mqd_stride_size;
732
733	for_each_inst(xcc_id, xcc_mask) {
734		xcc_mqd = mqd + mqd_offset * inst;
735		err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd,
736						    type, timeout, pipe_id,
737						    queue_id, xcc_id);
738		if (err) {
739			pr_debug("Destroy MQD failed for xcc: %d\n", inst);
740			break;
741		}
742		++inst;
743	}
744
745	return err;
746}
747
748static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
749			uint32_t pipe_id, uint32_t queue_id,
750			struct queue_properties *p, struct mm_struct *mms)
751{
752	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
753	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
754	uint32_t xcc_mask = mm->dev->xcc_mask;
755	int xcc_id, err, inst = 0;
756	void *xcc_mqd;
757	uint64_t mqd_stride_size = mm->mqd_stride(mm, p);
758
759	for_each_inst(xcc_id, xcc_mask) {
760		xcc_mqd = mqd + mqd_stride_size * inst;
761		err = mm->dev->kfd2kgd->hqd_load(
762			mm->dev->adev, xcc_mqd, pipe_id, queue_id,
763			(uint32_t __user *)p->write_ptr, wptr_shift, 0, mms,
764			xcc_id);
765		if (err) {
766			pr_debug("Load MQD failed for xcc: %d\n", inst);
767			break;
768		}
769		++inst;
770	}
771
772	return err;
773}
774
775static int get_wave_state_v9_4_3(struct mqd_manager *mm, void *mqd,
776				 struct queue_properties *q,
777				 void __user *ctl_stack,
778				 u32 *ctl_stack_used_size,
779				 u32 *save_area_used_size)
780{
781	int xcc, err = 0;
782	void *xcc_mqd;
783	void __user *xcc_ctl_stack;
784	uint64_t mqd_stride_size = mm->mqd_stride(mm, q);
785	u32 tmp_ctl_stack_used_size = 0, tmp_save_area_used_size = 0;
786
787	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
788		xcc_mqd = mqd + mqd_stride_size * xcc;
789		xcc_ctl_stack = (void __user *)((uintptr_t)ctl_stack +
790					q->ctx_save_restore_area_size * xcc);
791
792		err = get_wave_state(mm, xcc_mqd, q, xcc_ctl_stack,
793				     &tmp_ctl_stack_used_size,
794				     &tmp_save_area_used_size);
795		if (err)
796			break;
797
798		/*
799		 * Set the ctl_stack_used_size and save_area_used_size to
800		 * ctl_stack_used_size and save_area_used_size of XCC 0 when
801		 * passing the info the user-space.
802		 * For multi XCC, user-space would have to look at the header
803		 * info of each Control stack area to determine the control
804		 * stack size and save area used.
805		 */
806		if (xcc == 0) {
807			*ctl_stack_used_size = tmp_ctl_stack_used_size;
808			*save_area_used_size = tmp_save_area_used_size;
809		}
810	}
811
812	return err;
813}
814
815#if defined(CONFIG_DEBUG_FS)
816
817static int debugfs_show_mqd(struct seq_file *m, void *data)
818{
819	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
820		     data, sizeof(struct v9_mqd), false);
821	return 0;
822}
823
824static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
825{
826	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
827		     data, sizeof(struct v9_sdma_mqd), false);
828	return 0;
829}
830
831#endif
832
833struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
834		struct kfd_node *dev)
835{
836	struct mqd_manager *mqd;
837
838	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
839		return NULL;
840
841	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
842	if (!mqd)
843		return NULL;
844
845	mqd->dev = dev;
846
847	switch (type) {
848	case KFD_MQD_TYPE_CP:
849		mqd->allocate_mqd = allocate_mqd;
850		mqd->free_mqd = kfd_free_mqd_cp;
851		mqd->is_occupied = kfd_is_occupied_cp;
852		mqd->get_checkpoint_info = get_checkpoint_info;
853		mqd->checkpoint_mqd = checkpoint_mqd;
854		mqd->restore_mqd = restore_mqd;
 
 
855		mqd->mqd_size = sizeof(struct v9_mqd);
856		mqd->mqd_stride = mqd_stride_v9;
857#if defined(CONFIG_DEBUG_FS)
858		mqd->debugfs_show_mqd = debugfs_show_mqd;
859#endif
860		if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) {
861			mqd->init_mqd = init_mqd_v9_4_3;
862			mqd->load_mqd = load_mqd_v9_4_3;
863			mqd->update_mqd = update_mqd_v9_4_3;
864			mqd->destroy_mqd = destroy_mqd_v9_4_3;
865			mqd->get_wave_state = get_wave_state_v9_4_3;
866		} else {
867			mqd->init_mqd = init_mqd;
868			mqd->load_mqd = load_mqd;
869			mqd->update_mqd = update_mqd;
870			mqd->destroy_mqd = kfd_destroy_mqd_cp;
871			mqd->get_wave_state = get_wave_state;
872		}
873		break;
874	case KFD_MQD_TYPE_HIQ:
875		mqd->allocate_mqd = allocate_hiq_mqd;
 
876		mqd->free_mqd = free_mqd_hiq_sdma;
 
877		mqd->update_mqd = update_mqd;
878		mqd->is_occupied = kfd_is_occupied_cp;
 
879		mqd->mqd_size = sizeof(struct v9_mqd);
880		mqd->mqd_stride = kfd_mqd_stride;
881#if defined(CONFIG_DEBUG_FS)
882		mqd->debugfs_show_mqd = debugfs_show_mqd;
883#endif
884		mqd->read_doorbell_id = read_doorbell_id;
885		if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) {
886			mqd->init_mqd = init_mqd_hiq_v9_4_3;
887			mqd->load_mqd = hiq_load_mqd_kiq_v9_4_3;
888			mqd->destroy_mqd = destroy_hiq_mqd_v9_4_3;
889		} else {
890			mqd->init_mqd = init_mqd_hiq;
891			mqd->load_mqd = kfd_hiq_load_mqd_kiq;
892			mqd->destroy_mqd = destroy_hiq_mqd;
893		}
894		break;
895	case KFD_MQD_TYPE_DIQ:
896		mqd->allocate_mqd = allocate_mqd;
897		mqd->init_mqd = init_mqd_hiq;
898		mqd->free_mqd = kfd_free_mqd_cp;
899		mqd->load_mqd = load_mqd;
900		mqd->update_mqd = update_mqd;
901		mqd->destroy_mqd = kfd_destroy_mqd_cp;
902		mqd->is_occupied = kfd_is_occupied_cp;
903		mqd->mqd_size = sizeof(struct v9_mqd);
904#if defined(CONFIG_DEBUG_FS)
905		mqd->debugfs_show_mqd = debugfs_show_mqd;
906#endif
907		break;
908	case KFD_MQD_TYPE_SDMA:
909		mqd->allocate_mqd = allocate_sdma_mqd;
910		mqd->init_mqd = init_mqd_sdma;
911		mqd->free_mqd = free_mqd_hiq_sdma;
912		mqd->load_mqd = kfd_load_mqd_sdma;
913		mqd->update_mqd = update_mqd_sdma;
914		mqd->destroy_mqd = kfd_destroy_mqd_sdma;
915		mqd->is_occupied = kfd_is_occupied_sdma;
916		mqd->checkpoint_mqd = checkpoint_mqd_sdma;
917		mqd->restore_mqd = restore_mqd_sdma;
918		mqd->mqd_size = sizeof(struct v9_sdma_mqd);
919		mqd->mqd_stride = kfd_mqd_stride;
920#if defined(CONFIG_DEBUG_FS)
921		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
922#endif
923		break;
924	default:
925		kfree(mqd);
926		return NULL;
927	}
928
929	return mqd;
930}