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 /* Allocate a new job for @count PTE updates */
51 static int amdgpu_vm_sdma_alloc_job(struct amdgpu_vm_update_params *p,
52 				    unsigned int count)
53 {
54 	enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE
55 		: AMDGPU_IB_POOL_DELAYED;
56 	struct drm_sched_entity *entity = p->immediate ? &p->vm->immediate
57 		: &p->vm->delayed;
58 	unsigned int ndw;
59 	int r;
60 
61 	/* estimate how many dw we need */
62 	ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
63 	if (p->pages_addr)
64 		ndw += count * 2;
65 	ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
66 
67 	r = amdgpu_job_alloc_with_ib(p->adev, entity, AMDGPU_FENCE_OWNER_VM,
68 				     ndw * 4, pool, &p->job);
69 	if (r)
70 		return r;
71 
72 	p->num_dw_left = ndw;
73 	return 0;
74 }
75 
76 /**
77  * amdgpu_vm_sdma_prepare - prepare SDMA command submission
78  *
79  * @p: see amdgpu_vm_update_params definition
80  * @resv: reservation object with embedded fence
81  * @sync_mode: synchronization mode
82  *
83  * Returns:
84  * Negativ errno, 0 for success.
85  */
86 static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
87 				  struct dma_resv *resv,
88 				  enum amdgpu_sync_mode sync_mode)
89 {
90 	struct amdgpu_sync sync;
91 	int r;
92 
93 	r = amdgpu_vm_sdma_alloc_job(p, 0);
94 	if (r)
95 		return r;
96 
97 	if (!resv)
98 		return 0;
99 
100 	amdgpu_sync_create(&sync);
101 	r = amdgpu_sync_resv(p->adev, &sync, resv, sync_mode, p->vm);
102 	if (!r)
103 		r = amdgpu_sync_push_to_job(&sync, p->job);
104 	amdgpu_sync_free(&sync);
105 
106 	if (r) {
107 		p->num_dw_left = 0;
108 		amdgpu_job_free(p->job);
109 	}
110 	return r;
111 }
112 
113 /**
114  * amdgpu_vm_sdma_commit - commit SDMA command submission
115  *
116  * @p: see amdgpu_vm_update_params definition
117  * @fence: resulting fence
118  *
119  * Returns:
120  * Negativ errno, 0 for success.
121  */
122 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
123 				 struct dma_fence **fence)
124 {
125 	struct amdgpu_ib *ib = p->job->ibs;
126 	struct amdgpu_ring *ring;
127 	struct dma_fence *f;
128 
129 	ring = container_of(p->vm->delayed.rq->sched, struct amdgpu_ring,
130 			    sched);
131 
132 	WARN_ON(ib->length_dw == 0);
133 	amdgpu_ring_pad_ib(ring, ib);
134 	WARN_ON(ib->length_dw > p->num_dw_left);
135 	f = amdgpu_job_submit(p->job);
136 
137 	if (p->unlocked) {
138 		struct dma_fence *tmp = dma_fence_get(f);
139 
140 		swap(p->vm->last_unlocked, tmp);
141 		dma_fence_put(tmp);
142 	} else {
143 		dma_resv_add_fence(p->vm->root.bo->tbo.base.resv, f,
144 				   DMA_RESV_USAGE_BOOKKEEP);
145 	}
146 
147 	if (fence && !p->immediate) {
148 		/*
149 		 * Most hw generations now have a separate queue for page table
150 		 * updates, but when the queue is shared with userspace we need
151 		 * the extra CPU round trip to correctly flush the TLB.
152 		 */
153 		set_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &f->flags);
154 		swap(*fence, f);
155 	}
156 	dma_fence_put(f);
157 	return 0;
158 }
159 
160 /**
161  * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
162  *
163  * @p: see amdgpu_vm_update_params definition
164  * @bo: PD/PT to update
165  * @pe: addr of the page entry
166  * @count: number of page entries to copy
167  *
168  * Traces the parameters and calls the DMA function to copy the PTEs.
169  */
170 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
171 				     struct amdgpu_bo *bo, uint64_t pe,
172 				     unsigned count)
173 {
174 	struct amdgpu_ib *ib = p->job->ibs;
175 	uint64_t src = ib->gpu_addr;
176 
177 	src += p->num_dw_left * 4;
178 
179 	pe += amdgpu_bo_gpu_offset_no_check(bo);
180 	trace_amdgpu_vm_copy_ptes(pe, src, count, p->immediate);
181 
182 	amdgpu_vm_copy_pte(p->adev, ib, pe, src, count);
183 }
184 
185 /**
186  * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
187  *
188  * @p: see amdgpu_vm_update_params definition
189  * @bo: PD/PT to update
190  * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
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  * Traces the parameters and calls the right asic functions
197  * to setup the page table using the DMA.
198  */
199 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
200 				    struct amdgpu_bo *bo, uint64_t pe,
201 				    uint64_t addr, unsigned count,
202 				    uint32_t incr, uint64_t flags)
203 {
204 	struct amdgpu_ib *ib = p->job->ibs;
205 
206 	pe += amdgpu_bo_gpu_offset_no_check(bo);
207 	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->immediate);
208 	if (count < 3) {
209 		amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags,
210 				    count, incr);
211 	} else {
212 		amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr,
213 				      count, incr, flags);
214 	}
215 }
216 
217 /**
218  * amdgpu_vm_sdma_update - execute VM update
219  *
220  * @p: see amdgpu_vm_update_params definition
221  * @vmbo: PD/PT to update
222  * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB
223  * @addr: dst addr to write into pe
224  * @count: number of page entries to update
225  * @incr: increase next addr by incr bytes
226  * @flags: hw access flags
227  *
228  * Reserve space in the IB, setup mapping buffer on demand and write commands to
229  * the IB.
230  */
231 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
232 				 struct amdgpu_bo_vm *vmbo, uint64_t pe,
233 				 uint64_t addr, unsigned count, uint32_t incr,
234 				 uint64_t flags)
235 {
236 	struct amdgpu_bo *bo = &vmbo->bo;
237 	struct dma_resv_iter cursor;
238 	unsigned int i, ndw, nptes;
239 	struct dma_fence *fence;
240 	uint64_t *pte;
241 	int r;
242 
243 	/* Wait for PD/PT moves to be completed */
244 	dma_resv_iter_begin(&cursor, bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL);
245 	dma_resv_for_each_fence_unlocked(&cursor, fence) {
246 		dma_fence_get(fence);
247 		r = drm_sched_job_add_dependency(&p->job->base, fence);
248 		if (r) {
249 			dma_fence_put(fence);
250 			dma_resv_iter_end(&cursor);
251 			return r;
252 		}
253 	}
254 	dma_resv_iter_end(&cursor);
255 
256 	do {
257 		ndw = p->num_dw_left;
258 		ndw -= p->job->ibs->length_dw;
259 
260 		if (ndw < 32) {
261 			r = amdgpu_vm_sdma_commit(p, NULL);
262 			if (r)
263 				return r;
264 
265 			r = amdgpu_vm_sdma_alloc_job(p, count);
266 			if (r)
267 				return r;
268 		}
269 
270 		if (!p->pages_addr) {
271 			/* set page commands needed */
272 			if (vmbo->shadow)
273 				amdgpu_vm_sdma_set_ptes(p, vmbo->shadow, pe, addr,
274 							count, incr, flags);
275 			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
276 						incr, flags);
277 			return 0;
278 		}
279 
280 		/* copy commands needed */
281 		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
282 			(vmbo->shadow ? 2 : 1);
283 
284 		/* for padding */
285 		ndw -= 7;
286 
287 		nptes = min(count, ndw / 2);
288 
289 		/* Put the PTEs at the end of the IB. */
290 		p->num_dw_left -= nptes * 2;
291 		pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]);
292 		for (i = 0; i < nptes; ++i, addr += incr) {
293 			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
294 			pte[i] |= flags;
295 		}
296 
297 		if (vmbo->shadow)
298 			amdgpu_vm_sdma_copy_ptes(p, vmbo->shadow, pe, nptes);
299 		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
300 
301 		pe += nptes * 8;
302 		count -= nptes;
303 	} while (count);
304 
305 	return 0;
306 }
307 
308 const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
309 	.map_table = amdgpu_vm_sdma_map_table,
310 	.prepare = amdgpu_vm_sdma_prepare,
311 	.update = amdgpu_vm_sdma_update,
312 	.commit = amdgpu_vm_sdma_commit
313 };
314