2 * Copyright 2019 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include "amdgpu_vm.h"
24 #include "amdgpu_job.h"
25 #include "amdgpu_object.h"
26 #include "amdgpu_trace.h"
28 #define AMDGPU_VM_SDMA_MIN_NUM_DW 256u
29 #define AMDGPU_VM_SDMA_MAX_NUM_DW (16u * 1024u)
32 * amdgpu_vm_sdma_map_table - make sure new PDs/PTs are GTT mapped
34 * @table: newly allocated or validated PD/PT
36 static int amdgpu_vm_sdma_map_table(struct amdgpu_bo *table)
40 r = amdgpu_ttm_alloc_gart(&table->tbo);
45 r = amdgpu_ttm_alloc_gart(&table->shadow->tbo);
51 * amdgpu_vm_sdma_prepare - prepare SDMA command submission
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
58 * Negativ errno, 0 for success.
60 static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
61 void *owner, struct dma_fence *exclusive)
63 struct amdgpu_bo *root = p->vm->root.base.bo;
64 unsigned int ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
67 r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
71 r = amdgpu_sync_fence(p->adev, &p->job->sync, exclusive, false);
75 r = amdgpu_sync_resv(p->adev, &p->job->sync, root->tbo.resv,
81 p->ib = &p->job->ibs[0];
86 * amdgpu_vm_sdma_commit - commit SDMA command submission
88 * @p: see amdgpu_vm_update_params definition
89 * @fence: resulting fence
92 * Negativ errno, 0 for success.
94 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
95 struct dma_fence **fence)
97 struct amdgpu_bo *root = p->vm->root.base.bo;
98 struct amdgpu_ring *ring;
102 ring = container_of(p->vm->entity.rq->sched, struct amdgpu_ring, sched);
104 WARN_ON(p->ib->length_dw == 0);
105 amdgpu_ring_pad_ib(ring, p->ib);
106 WARN_ON(p->ib->length_dw > p->num_dw_left);
107 r = amdgpu_job_submit(p->job, &p->vm->entity,
108 AMDGPU_FENCE_OWNER_VM, &f);
112 amdgpu_bo_fence(root, f, true);
119 amdgpu_job_free(p->job);
125 * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
127 * @p: see amdgpu_vm_update_params definition
128 * @bo: PD/PT to update
129 * @pe: addr of the page entry
130 * @count: number of page entries to copy
132 * Traces the parameters and calls the DMA function to copy the PTEs.
134 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
135 struct amdgpu_bo *bo, uint64_t pe,
138 uint64_t src = p->ib->gpu_addr;
140 src += p->num_dw_left * 4;
142 pe += amdgpu_bo_gpu_offset(bo);
143 trace_amdgpu_vm_copy_ptes(pe, src, count);
145 amdgpu_vm_copy_pte(p->adev, p->ib, pe, src, count);
149 * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
151 * @p: see amdgpu_vm_update_params definition
152 * @bo: PD/PT to update
153 * @pe: addr of the page entry
154 * @addr: dst addr to write into pe
155 * @count: number of page entries to update
156 * @incr: increase next addr by incr bytes
157 * @flags: hw access flags
159 * Traces the parameters and calls the right asic functions
160 * to setup the page table using the DMA.
162 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
163 struct amdgpu_bo *bo, uint64_t pe,
164 uint64_t addr, unsigned count,
165 uint32_t incr, uint64_t flags)
167 pe += amdgpu_bo_gpu_offset(bo);
168 trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
170 amdgpu_vm_write_pte(p->adev, p->ib, pe, addr | flags,
173 amdgpu_vm_set_pte_pde(p->adev, p->ib, pe, addr,
179 * amdgpu_vm_sdma_update - execute VM update
181 * @p: see amdgpu_vm_update_params definition
182 * @bo: PD/PT to update
183 * @pe: addr of the page entry
184 * @addr: dst addr to write into pe
185 * @count: number of page entries to update
186 * @incr: increase next addr by incr bytes
187 * @flags: hw access flags
189 * Reserve space in the IB, setup mapping buffer on demand and write commands to
192 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
193 struct amdgpu_bo *bo, uint64_t pe,
194 uint64_t addr, unsigned count, uint32_t incr,
197 unsigned int i, ndw, nptes;
202 ndw = p->num_dw_left;
203 ndw -= p->ib->length_dw;
206 r = amdgpu_vm_sdma_commit(p, NULL);
210 /* estimate how many dw we need */
214 ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW);
215 ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
217 r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
221 p->num_dw_left = ndw;
222 p->ib = &p->job->ibs[0];
225 if (!p->pages_addr) {
226 /* set page commands needed */
228 amdgpu_vm_sdma_set_ptes(p, bo->shadow, pe, addr,
230 amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
235 /* copy commands needed */
236 ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
237 (bo->shadow ? 2 : 1);
242 nptes = min(count, ndw / 2);
244 /* Put the PTEs at the end of the IB. */
245 p->num_dw_left -= nptes * 2;
246 pte = (uint64_t *)&(p->ib->ptr[p->num_dw_left]);
247 for (i = 0; i < nptes; ++i, addr += incr) {
248 pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
253 amdgpu_vm_sdma_copy_ptes(p, bo->shadow, pe, nptes);
254 amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
263 const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
264 .map_table = amdgpu_vm_sdma_map_table,
265 .prepare = amdgpu_vm_sdma_prepare,
266 .update = amdgpu_vm_sdma_update,
267 .commit = amdgpu_vm_sdma_commit