Linux Audio

Check our new training course

Loading...
v5.9
  1/*
  2 * Copyright 2019 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#include "amdgpu_vm.h"
 24#include "amdgpu_job.h"
 25#include "amdgpu_object.h"
 26#include "amdgpu_trace.h"
 27
 28#define AMDGPU_VM_SDMA_MIN_NUM_DW	256u
 29#define AMDGPU_VM_SDMA_MAX_NUM_DW	(16u * 1024u)
 30
 31/**
 32 * amdgpu_vm_sdma_map_table - make sure new PDs/PTs are GTT mapped
 33 *
 34 * @table: newly allocated or validated PD/PT
 35 */
 36static int amdgpu_vm_sdma_map_table(struct amdgpu_bo *table)
 37{
 
 
 
 
 
 
 
 
 
 
 
 
 38	int r;
 39
 40	r = amdgpu_ttm_alloc_gart(&table->tbo);
 
 
 
 
 
 
 
 41	if (r)
 42		return r;
 43
 44	if (table->shadow)
 45		r = amdgpu_ttm_alloc_gart(&table->shadow->tbo);
 46
 47	return r;
 48}
 49
 50/**
 51 * amdgpu_vm_sdma_prepare - prepare SDMA command submission
 52 *
 53 * @p: see amdgpu_vm_update_params definition
 54 * @owner: owner we need to sync to
 55 * @exclusive: exclusive move fence we need to sync to
 56 *
 57 * Returns:
 58 * Negativ errno, 0 for success.
 59 */
 60static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
 61				  struct dma_resv *resv,
 62				  enum amdgpu_sync_mode sync_mode)
 63{
 64	enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE
 65		: AMDGPU_IB_POOL_DELAYED;
 66	unsigned int ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
 67	int r;
 68
 69	r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, pool, &p->job);
 70	if (r)
 71		return r;
 72
 73	p->num_dw_left = ndw;
 74
 75	if (!resv)
 76		return 0;
 77
 78	return amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode, p->vm);
 
 
 
 
 
 79}
 80
 81/**
 82 * amdgpu_vm_sdma_commit - commit SDMA command submission
 83 *
 84 * @p: see amdgpu_vm_update_params definition
 85 * @fence: resulting fence
 86 *
 87 * Returns:
 88 * Negativ errno, 0 for success.
 89 */
 90static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
 91				 struct dma_fence **fence)
 92{
 93	struct amdgpu_ib *ib = p->job->ibs;
 94	struct drm_sched_entity *entity;
 95	struct amdgpu_ring *ring;
 96	struct dma_fence *f;
 97	int r;
 98
 99	entity = p->immediate ? &p->vm->immediate : &p->vm->delayed;
100	ring = container_of(entity->rq->sched, struct amdgpu_ring, sched);
101
102	WARN_ON(ib->length_dw == 0);
103	amdgpu_ring_pad_ib(ring, ib);
 
 
 
 
104	WARN_ON(ib->length_dw > p->num_dw_left);
105	r = amdgpu_job_submit(p->job, entity, AMDGPU_FENCE_OWNER_VM, &f);
106	if (r)
107		goto error;
108
109	if (p->unlocked) {
110		struct dma_fence *tmp = dma_fence_get(f);
111
112		swap(p->vm->last_unlocked, f);
113		dma_fence_put(tmp);
114	} else {
115		amdgpu_bo_fence(p->vm->root.base.bo, f, true);
 
116	}
117
118	if (fence && !p->immediate)
 
 
 
 
 
 
119		swap(*fence, f);
 
120	dma_fence_put(f);
121	return 0;
122
123error:
124	amdgpu_job_free(p->job);
125	return r;
126}
127
128/**
129 * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
130 *
131 * @p: see amdgpu_vm_update_params definition
132 * @bo: PD/PT to update
133 * @pe: addr of the page entry
134 * @count: number of page entries to copy
135 *
136 * Traces the parameters and calls the DMA function to copy the PTEs.
137 */
138static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
139				     struct amdgpu_bo *bo, uint64_t pe,
140				     unsigned count)
141{
142	struct amdgpu_ib *ib = p->job->ibs;
143	uint64_t src = ib->gpu_addr;
144
145	src += p->num_dw_left * 4;
146
147	pe += amdgpu_gmc_sign_extend(amdgpu_bo_gpu_offset_no_check(bo));
148	trace_amdgpu_vm_copy_ptes(pe, src, count, p->immediate);
149
150	amdgpu_vm_copy_pte(p->adev, ib, pe, src, count);
151}
152
153/**
154 * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
155 *
156 * @p: see amdgpu_vm_update_params definition
157 * @bo: PD/PT to update
158 * @pe: addr of the page entry
159 * @addr: dst addr to write into pe
160 * @count: number of page entries to update
161 * @incr: increase next addr by incr bytes
162 * @flags: hw access flags
163 *
164 * Traces the parameters and calls the right asic functions
165 * to setup the page table using the DMA.
166 */
167static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
168				    struct amdgpu_bo *bo, uint64_t pe,
169				    uint64_t addr, unsigned count,
170				    uint32_t incr, uint64_t flags)
171{
172	struct amdgpu_ib *ib = p->job->ibs;
173
174	pe += amdgpu_gmc_sign_extend(amdgpu_bo_gpu_offset_no_check(bo));
175	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->immediate);
176	if (count < 3) {
177		amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags,
178				    count, incr);
179	} else {
180		amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr,
181				      count, incr, flags);
182	}
183}
184
185/**
186 * amdgpu_vm_sdma_update - execute VM update
187 *
188 * @p: see amdgpu_vm_update_params definition
189 * @bo: PD/PT to update
190 * @pe: addr of the page entry
191 * @addr: dst addr to write into pe
192 * @count: number of page entries to update
193 * @incr: increase next addr by incr bytes
194 * @flags: hw access flags
195 *
196 * Reserve space in the IB, setup mapping buffer on demand and write commands to
197 * the IB.
198 */
199static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
200				 struct amdgpu_bo *bo, uint64_t pe,
201				 uint64_t addr, unsigned count, uint32_t incr,
202				 uint64_t flags)
203{
204	enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE
205		: AMDGPU_IB_POOL_DELAYED;
206	unsigned int i, ndw, nptes;
 
207	uint64_t *pte;
208	int r;
209
210	/* Wait for PD/PT moves to be completed */
211	r = amdgpu_sync_fence(&p->job->sync, bo->tbo.moving);
212	if (r)
213		return r;
 
 
 
 
 
 
 
 
214
215	do {
216		ndw = p->num_dw_left;
217		ndw -= p->job->ibs->length_dw;
218
219		if (ndw < 32) {
220			r = amdgpu_vm_sdma_commit(p, NULL);
221			if (r)
222				return r;
223
224			/* estimate how many dw we need */
225			ndw = 32;
226			if (p->pages_addr)
227				ndw += count * 2;
228			ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW);
229			ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
230
231			r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, pool,
232						     &p->job);
233			if (r)
234				return r;
235
236			p->num_dw_left = ndw;
237		}
238
239		if (!p->pages_addr) {
240			/* set page commands needed */
241			if (bo->shadow)
242				amdgpu_vm_sdma_set_ptes(p, bo->shadow, pe, addr,
243							count, incr, flags);
244			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
245						incr, flags);
246			return 0;
247		}
248
249		/* copy commands needed */
250		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
251			(bo->shadow ? 2 : 1);
252
253		/* for padding */
254		ndw -= 7;
255
256		nptes = min(count, ndw / 2);
257
258		/* Put the PTEs at the end of the IB. */
259		p->num_dw_left -= nptes * 2;
260		pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]);
261		for (i = 0; i < nptes; ++i, addr += incr) {
262			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
263			pte[i] |= flags;
264		}
265
266		if (bo->shadow)
267			amdgpu_vm_sdma_copy_ptes(p, bo->shadow, pe, nptes);
268		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
269
270		pe += nptes * 8;
271		count -= nptes;
272	} while (count);
273
274	return 0;
275}
276
277const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
278	.map_table = amdgpu_vm_sdma_map_table,
279	.prepare = amdgpu_vm_sdma_prepare,
280	.update = amdgpu_vm_sdma_update,
281	.commit = amdgpu_vm_sdma_commit
282};
v6.13.7
  1/*
  2 * Copyright 2019 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#include "amdgpu_vm.h"
 24#include "amdgpu_job.h"
 25#include "amdgpu_object.h"
 26#include "amdgpu_trace.h"
 27
 28#define AMDGPU_VM_SDMA_MIN_NUM_DW	256u
 29#define AMDGPU_VM_SDMA_MAX_NUM_DW	(16u * 1024u)
 30
 31/**
 32 * amdgpu_vm_sdma_map_table - make sure new PDs/PTs are GTT mapped
 33 *
 34 * @table: newly allocated or validated PD/PT
 35 */
 36static int amdgpu_vm_sdma_map_table(struct amdgpu_bo_vm *table)
 37{
 38	return amdgpu_ttm_alloc_gart(&table->bo.tbo);
 39}
 40
 41/* Allocate a new job for @count PTE updates */
 42static int amdgpu_vm_sdma_alloc_job(struct amdgpu_vm_update_params *p,
 43				    unsigned int count)
 44{
 45	enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE
 46		: AMDGPU_IB_POOL_DELAYED;
 47	struct drm_sched_entity *entity = p->immediate ? &p->vm->immediate
 48		: &p->vm->delayed;
 49	unsigned int ndw;
 50	int r;
 51
 52	/* estimate how many dw we need */
 53	ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
 54	if (p->pages_addr)
 55		ndw += count * 2;
 56	ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
 57
 58	r = amdgpu_job_alloc_with_ib(p->adev, entity, AMDGPU_FENCE_OWNER_VM,
 59				     ndw * 4, pool, &p->job);
 60	if (r)
 61		return r;
 62
 63	p->num_dw_left = ndw;
 64	return 0;
 
 
 65}
 66
 67/**
 68 * amdgpu_vm_sdma_prepare - prepare SDMA command submission
 69 *
 70 * @p: see amdgpu_vm_update_params definition
 71 * @sync: amdgpu_sync object with fences to wait for
 
 72 *
 73 * Returns:
 74 * Negativ errno, 0 for success.
 75 */
 76static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
 77				  struct amdgpu_sync *sync)
 
 78{
 
 
 
 79	int r;
 80
 81	r = amdgpu_vm_sdma_alloc_job(p, 0);
 82	if (r)
 83		return r;
 84
 85	if (!sync)
 
 
 86		return 0;
 87
 88	r = amdgpu_sync_push_to_job(sync, p->job);
 89	if (r) {
 90		p->num_dw_left = 0;
 91		amdgpu_job_free(p->job);
 92	}
 93	return r;
 94}
 95
 96/**
 97 * amdgpu_vm_sdma_commit - commit SDMA command submission
 98 *
 99 * @p: see amdgpu_vm_update_params definition
100 * @fence: resulting fence
101 *
102 * Returns:
103 * Negativ errno, 0 for success.
104 */
105static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
106				 struct dma_fence **fence)
107{
108	struct amdgpu_ib *ib = p->job->ibs;
 
109	struct amdgpu_ring *ring;
110	struct dma_fence *f;
 
111
112	ring = container_of(p->vm->delayed.rq->sched, struct amdgpu_ring,
113			    sched);
114
115	WARN_ON(ib->length_dw == 0);
116	amdgpu_ring_pad_ib(ring, ib);
117
118	if (p->needs_flush)
119		atomic64_inc(&p->vm->tlb_seq);
120
121	WARN_ON(ib->length_dw > p->num_dw_left);
122	f = amdgpu_job_submit(p->job);
 
 
123
124	if (p->unlocked) {
125		struct dma_fence *tmp = dma_fence_get(f);
126
127		swap(p->vm->last_unlocked, tmp);
128		dma_fence_put(tmp);
129	} else {
130		dma_resv_add_fence(p->vm->root.bo->tbo.base.resv, f,
131				   DMA_RESV_USAGE_BOOKKEEP);
132	}
133
134	if (fence && !p->immediate) {
135		/*
136		 * Most hw generations now have a separate queue for page table
137		 * updates, but when the queue is shared with userspace we need
138		 * the extra CPU round trip to correctly flush the TLB.
139		 */
140		set_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &f->flags);
141		swap(*fence, f);
142	}
143	dma_fence_put(f);
144	return 0;
 
 
 
 
145}
146
147/**
148 * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
149 *
150 * @p: see amdgpu_vm_update_params definition
151 * @bo: PD/PT to update
152 * @pe: addr of the page entry
153 * @count: number of page entries to copy
154 *
155 * Traces the parameters and calls the DMA function to copy the PTEs.
156 */
157static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
158				     struct amdgpu_bo *bo, uint64_t pe,
159				     unsigned count)
160{
161	struct amdgpu_ib *ib = p->job->ibs;
162	uint64_t src = ib->gpu_addr;
163
164	src += p->num_dw_left * 4;
165
166	pe += amdgpu_bo_gpu_offset_no_check(bo);
167	trace_amdgpu_vm_copy_ptes(pe, src, count, p->immediate);
168
169	amdgpu_vm_copy_pte(p->adev, ib, pe, src, count);
170}
171
172/**
173 * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
174 *
175 * @p: see amdgpu_vm_update_params definition
176 * @bo: PD/PT to update
177 * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
178 * @addr: dst addr to write into pe
179 * @count: number of page entries to update
180 * @incr: increase next addr by incr bytes
181 * @flags: hw access flags
182 *
183 * Traces the parameters and calls the right asic functions
184 * to setup the page table using the DMA.
185 */
186static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
187				    struct amdgpu_bo *bo, uint64_t pe,
188				    uint64_t addr, unsigned count,
189				    uint32_t incr, uint64_t flags)
190{
191	struct amdgpu_ib *ib = p->job->ibs;
192
193	pe += amdgpu_bo_gpu_offset_no_check(bo);
194	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->immediate);
195	if (count < 3) {
196		amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags,
197				    count, incr);
198	} else {
199		amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr,
200				      count, incr, flags);
201	}
202}
203
204/**
205 * amdgpu_vm_sdma_update - execute VM update
206 *
207 * @p: see amdgpu_vm_update_params definition
208 * @vmbo: PD/PT to update
209 * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
210 * @addr: dst addr to write into pe
211 * @count: number of page entries to update
212 * @incr: increase next addr by incr bytes
213 * @flags: hw access flags
214 *
215 * Reserve space in the IB, setup mapping buffer on demand and write commands to
216 * the IB.
217 */
218static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
219				 struct amdgpu_bo_vm *vmbo, uint64_t pe,
220				 uint64_t addr, unsigned count, uint32_t incr,
221				 uint64_t flags)
222{
223	struct amdgpu_bo *bo = &vmbo->bo;
224	struct dma_resv_iter cursor;
225	unsigned int i, ndw, nptes;
226	struct dma_fence *fence;
227	uint64_t *pte;
228	int r;
229
230	/* Wait for PD/PT moves to be completed */
231	dma_resv_iter_begin(&cursor, bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL);
232	dma_resv_for_each_fence_unlocked(&cursor, fence) {
233		dma_fence_get(fence);
234		r = drm_sched_job_add_dependency(&p->job->base, fence);
235		if (r) {
236			dma_fence_put(fence);
237			dma_resv_iter_end(&cursor);
238			return r;
239		}
240	}
241	dma_resv_iter_end(&cursor);
242
243	do {
244		ndw = p->num_dw_left;
245		ndw -= p->job->ibs->length_dw;
246
247		if (ndw < 32) {
248			r = amdgpu_vm_sdma_commit(p, NULL);
249			if (r)
250				return r;
251
252			r = amdgpu_vm_sdma_alloc_job(p, count);
 
 
 
 
 
 
 
 
253			if (r)
254				return r;
 
 
255		}
256
257		if (!p->pages_addr) {
258			/* set page commands needed */
 
 
 
259			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
260						incr, flags);
261			return 0;
262		}
263
264		/* copy commands needed */
265		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw;
 
266
267		/* for padding */
268		ndw -= 7;
269
270		nptes = min(count, ndw / 2);
271
272		/* Put the PTEs at the end of the IB. */
273		p->num_dw_left -= nptes * 2;
274		pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]);
275		for (i = 0; i < nptes; ++i, addr += incr) {
276			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
277			pte[i] |= flags;
278		}
279
 
 
280		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
281
282		pe += nptes * 8;
283		count -= nptes;
284	} while (count);
285
286	return 0;
287}
288
289const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
290	.map_table = amdgpu_vm_sdma_map_table,
291	.prepare = amdgpu_vm_sdma_prepare,
292	.update = amdgpu_vm_sdma_update,
293	.commit = amdgpu_vm_sdma_commit
294};