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_vm *table)
37 {
38 	int r;
39 
40 	r = amdgpu_ttm_alloc_gart(&table->bo.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  * @resv: reservation object with embedded fence
55  * @sync_mode: synchronization mode
56  *
57  * Returns:
58  * Negativ errno, 0 for success.
59  */
60 static 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  */
90 static 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, tmp);
113 		dma_fence_put(tmp);
114 	} else {
115 		dma_resv_add_fence(p->vm->root.bo->tbo.base.resv, f,
116 				   DMA_RESV_USAGE_BOOKKEEP);
117 	}
118 
119 	if (fence && !p->immediate)
120 		swap(*fence, f);
121 	dma_fence_put(f);
122 	return 0;
123 
124 error:
125 	amdgpu_job_free(p->job);
126 	return r;
127 }
128 
129 /**
130  * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
131  *
132  * @p: see amdgpu_vm_update_params definition
133  * @bo: PD/PT to update
134  * @pe: addr of the page entry
135  * @count: number of page entries to copy
136  *
137  * Traces the parameters and calls the DMA function to copy the PTEs.
138  */
139 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
140 				     struct amdgpu_bo *bo, uint64_t pe,
141 				     unsigned count)
142 {
143 	struct amdgpu_ib *ib = p->job->ibs;
144 	uint64_t src = ib->gpu_addr;
145 
146 	src += p->num_dw_left * 4;
147 
148 	pe += amdgpu_gmc_sign_extend(amdgpu_bo_gpu_offset_no_check(bo));
149 	trace_amdgpu_vm_copy_ptes(pe, src, count, p->immediate);
150 
151 	amdgpu_vm_copy_pte(p->adev, ib, pe, src, count);
152 }
153 
154 /**
155  * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
156  *
157  * @p: see amdgpu_vm_update_params definition
158  * @bo: PD/PT to update
159  * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
160  * @addr: dst addr to write into pe
161  * @count: number of page entries to update
162  * @incr: increase next addr by incr bytes
163  * @flags: hw access flags
164  *
165  * Traces the parameters and calls the right asic functions
166  * to setup the page table using the DMA.
167  */
168 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
169 				    struct amdgpu_bo *bo, uint64_t pe,
170 				    uint64_t addr, unsigned count,
171 				    uint32_t incr, uint64_t flags)
172 {
173 	struct amdgpu_ib *ib = p->job->ibs;
174 
175 	pe += amdgpu_gmc_sign_extend(amdgpu_bo_gpu_offset_no_check(bo));
176 	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->immediate);
177 	if (count < 3) {
178 		amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags,
179 				    count, incr);
180 	} else {
181 		amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr,
182 				      count, incr, flags);
183 	}
184 }
185 
186 /**
187  * amdgpu_vm_sdma_update - execute VM update
188  *
189  * @p: see amdgpu_vm_update_params definition
190  * @vmbo: PD/PT to update
191  * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
192  * @addr: dst addr to write into pe
193  * @count: number of page entries to update
194  * @incr: increase next addr by incr bytes
195  * @flags: hw access flags
196  *
197  * Reserve space in the IB, setup mapping buffer on demand and write commands to
198  * the IB.
199  */
200 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
201 				 struct amdgpu_bo_vm *vmbo, uint64_t pe,
202 				 uint64_t addr, unsigned count, uint32_t incr,
203 				 uint64_t flags)
204 {
205 	struct amdgpu_bo *bo = &vmbo->bo;
206 	enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE
207 		: AMDGPU_IB_POOL_DELAYED;
208 	struct dma_resv_iter cursor;
209 	unsigned int i, ndw, nptes;
210 	struct dma_fence *fence;
211 	uint64_t *pte;
212 	int r;
213 
214 	/* Wait for PD/PT moves to be completed */
215 	dma_resv_iter_begin(&cursor, bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL);
216 	dma_resv_for_each_fence_unlocked(&cursor, fence) {
217 		r = amdgpu_sync_fence(&p->job->sync, fence);
218 		if (r) {
219 			dma_resv_iter_end(&cursor);
220 			return r;
221 		}
222 	}
223 	dma_resv_iter_end(&cursor);
224 
225 	do {
226 		ndw = p->num_dw_left;
227 		ndw -= p->job->ibs->length_dw;
228 
229 		if (ndw < 32) {
230 			r = amdgpu_vm_sdma_commit(p, NULL);
231 			if (r)
232 				return r;
233 
234 			/* estimate how many dw we need */
235 			ndw = 32;
236 			if (p->pages_addr)
237 				ndw += count * 2;
238 			ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW);
239 			ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
240 
241 			r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, pool,
242 						     &p->job);
243 			if (r)
244 				return r;
245 
246 			p->num_dw_left = ndw;
247 		}
248 
249 		if (!p->pages_addr) {
250 			/* set page commands needed */
251 			if (vmbo->shadow)
252 				amdgpu_vm_sdma_set_ptes(p, vmbo->shadow, pe, addr,
253 							count, incr, flags);
254 			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
255 						incr, flags);
256 			return 0;
257 		}
258 
259 		/* copy commands needed */
260 		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
261 			(vmbo->shadow ? 2 : 1);
262 
263 		/* for padding */
264 		ndw -= 7;
265 
266 		nptes = min(count, ndw / 2);
267 
268 		/* Put the PTEs at the end of the IB. */
269 		p->num_dw_left -= nptes * 2;
270 		pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]);
271 		for (i = 0; i < nptes; ++i, addr += incr) {
272 			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
273 			pte[i] |= flags;
274 		}
275 
276 		if (vmbo->shadow)
277 			amdgpu_vm_sdma_copy_ptes(p, vmbo->shadow, pe, nptes);
278 		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
279 
280 		pe += nptes * 8;
281 		count -= nptes;
282 	} while (count);
283 
284 	return 0;
285 }
286 
287 const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
288 	.map_table = amdgpu_vm_sdma_map_table,
289 	.prepare = amdgpu_vm_sdma_prepare,
290 	.update = amdgpu_vm_sdma_update,
291 	.commit = amdgpu_vm_sdma_commit
292 };
293