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  */
36 static 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  */
60 static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
61 				  void *owner, struct dma_fence *exclusive)
62 {
63 	struct amdgpu_bo *root = p->vm->root.base.bo;
64 	unsigned int ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
65 	int r;
66 
67 	r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
68 	if (r)
69 		return r;
70 
71 	p->num_dw_left = ndw;
72 
73 	/* Wait for moves to be completed */
74 	r = amdgpu_sync_fence(&p->job->sync, exclusive, false);
75 	if (r)
76 		return r;
77 
78 	/* Don't wait for any submissions during page fault handling */
79 	if (p->direct)
80 		return 0;
81 
82 	return amdgpu_sync_resv(p->adev, &p->job->sync, root->tbo.base.resv,
83 				owner, false);
84 }
85 
86 /**
87  * amdgpu_vm_sdma_commit - commit SDMA command submission
88  *
89  * @p: see amdgpu_vm_update_params definition
90  * @fence: resulting fence
91  *
92  * Returns:
93  * Negativ errno, 0 for success.
94  */
95 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
96 				 struct dma_fence **fence)
97 {
98 	struct amdgpu_ib *ib = p->job->ibs;
99 	struct drm_sched_entity *entity;
100 	struct dma_fence *f, *tmp;
101 	struct amdgpu_ring *ring;
102 	int r;
103 
104 	entity = p->direct ? &p->vm->direct : &p->vm->delayed;
105 	ring = container_of(entity->rq->sched, struct amdgpu_ring, sched);
106 
107 	WARN_ON(ib->length_dw == 0);
108 	amdgpu_ring_pad_ib(ring, ib);
109 	WARN_ON(ib->length_dw > p->num_dw_left);
110 	r = amdgpu_job_submit(p->job, entity, AMDGPU_FENCE_OWNER_VM, &f);
111 	if (r)
112 		goto error;
113 
114 	tmp = dma_fence_get(f);
115 	if (p->direct)
116 		swap(p->vm->last_direct, tmp);
117 	else
118 		swap(p->vm->last_delayed, tmp);
119 	dma_fence_put(tmp);
120 
121 	if (fence && !p->direct)
122 		swap(*fence, f);
123 	dma_fence_put(f);
124 	return 0;
125 
126 error:
127 	amdgpu_job_free(p->job);
128 	return r;
129 }
130 
131 /**
132  * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
133  *
134  * @p: see amdgpu_vm_update_params definition
135  * @bo: PD/PT to update
136  * @pe: addr of the page entry
137  * @count: number of page entries to copy
138  *
139  * Traces the parameters and calls the DMA function to copy the PTEs.
140  */
141 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
142 				     struct amdgpu_bo *bo, uint64_t pe,
143 				     unsigned count)
144 {
145 	struct amdgpu_ib *ib = p->job->ibs;
146 	uint64_t src = ib->gpu_addr;
147 
148 	src += p->num_dw_left * 4;
149 
150 	pe += amdgpu_bo_gpu_offset(bo);
151 	trace_amdgpu_vm_copy_ptes(pe, src, count, p->direct);
152 
153 	amdgpu_vm_copy_pte(p->adev, ib, pe, src, count);
154 }
155 
156 /**
157  * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
158  *
159  * @p: see amdgpu_vm_update_params definition
160  * @bo: PD/PT to update
161  * @pe: addr of the page entry
162  * @addr: dst addr to write into pe
163  * @count: number of page entries to update
164  * @incr: increase next addr by incr bytes
165  * @flags: hw access flags
166  *
167  * Traces the parameters and calls the right asic functions
168  * to setup the page table using the DMA.
169  */
170 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
171 				    struct amdgpu_bo *bo, uint64_t pe,
172 				    uint64_t addr, unsigned count,
173 				    uint32_t incr, uint64_t flags)
174 {
175 	struct amdgpu_ib *ib = p->job->ibs;
176 
177 	pe += amdgpu_bo_gpu_offset(bo);
178 	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->direct);
179 	if (count < 3) {
180 		amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags,
181 				    count, incr);
182 	} else {
183 		amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr,
184 				      count, incr, flags);
185 	}
186 }
187 
188 /**
189  * amdgpu_vm_sdma_update - execute VM update
190  *
191  * @p: see amdgpu_vm_update_params definition
192  * @bo: PD/PT to update
193  * @pe: addr of the page entry
194  * @addr: dst addr to write into pe
195  * @count: number of page entries to update
196  * @incr: increase next addr by incr bytes
197  * @flags: hw access flags
198  *
199  * Reserve space in the IB, setup mapping buffer on demand and write commands to
200  * the IB.
201  */
202 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
203 				 struct amdgpu_bo *bo, uint64_t pe,
204 				 uint64_t addr, unsigned count, uint32_t incr,
205 				 uint64_t flags)
206 {
207 	unsigned int i, ndw, nptes;
208 	uint64_t *pte;
209 	int r;
210 
211 	do {
212 		ndw = p->num_dw_left;
213 		ndw -= p->job->ibs->length_dw;
214 
215 		if (ndw < 32) {
216 			r = amdgpu_vm_sdma_commit(p, NULL);
217 			if (r)
218 				return r;
219 
220 			/* estimate how many dw we need */
221 			ndw = 32;
222 			if (p->pages_addr)
223 				ndw += count * 2;
224 			ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW);
225 			ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
226 
227 			r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
228 			if (r)
229 				return r;
230 
231 			p->num_dw_left = ndw;
232 		}
233 
234 		if (!p->pages_addr) {
235 			/* set page commands needed */
236 			if (bo->shadow)
237 				amdgpu_vm_sdma_set_ptes(p, bo->shadow, pe, addr,
238 							count, incr, flags);
239 			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
240 						incr, flags);
241 			return 0;
242 		}
243 
244 		/* copy commands needed */
245 		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
246 			(bo->shadow ? 2 : 1);
247 
248 		/* for padding */
249 		ndw -= 7;
250 
251 		nptes = min(count, ndw / 2);
252 
253 		/* Put the PTEs at the end of the IB. */
254 		p->num_dw_left -= nptes * 2;
255 		pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]);
256 		for (i = 0; i < nptes; ++i, addr += incr) {
257 			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
258 			pte[i] |= flags;
259 		}
260 
261 		if (bo->shadow)
262 			amdgpu_vm_sdma_copy_ptes(p, bo->shadow, pe, nptes);
263 		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
264 
265 		pe += nptes * 8;
266 		count -= nptes;
267 	} while (count);
268 
269 	return 0;
270 }
271 
272 const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
273 	.map_table = amdgpu_vm_sdma_map_table,
274 	.prepare = amdgpu_vm_sdma_prepare,
275 	.update = amdgpu_vm_sdma_update,
276 	.commit = amdgpu_vm_sdma_commit
277 };
278