1 /* $NetBSD: amdgpu_vm_sdma.c,v 1.2 2021/12/18 23:44:58 riastradh Exp $ */ 2 3 /* 4 * Copyright 2019 Advanced Micro Devices, Inc. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 */ 24 25 #include <sys/cdefs.h> 26 __KERNEL_RCSID(0, "$NetBSD: amdgpu_vm_sdma.c,v 1.2 2021/12/18 23:44:58 riastradh Exp $"); 27 28 #include "amdgpu_vm.h" 29 #include "amdgpu_job.h" 30 #include "amdgpu_object.h" 31 #include "amdgpu_trace.h" 32 33 #define AMDGPU_VM_SDMA_MIN_NUM_DW 256u 34 #define AMDGPU_VM_SDMA_MAX_NUM_DW (16u * 1024u) 35 36 /** 37 * amdgpu_vm_sdma_map_table - make sure new PDs/PTs are GTT mapped 38 * 39 * @table: newly allocated or validated PD/PT 40 */ 41 static int amdgpu_vm_sdma_map_table(struct amdgpu_bo *table) 42 { 43 int r; 44 45 r = amdgpu_ttm_alloc_gart(&table->tbo); 46 if (r) 47 return r; 48 49 if (table->shadow) 50 r = amdgpu_ttm_alloc_gart(&table->shadow->tbo); 51 52 return r; 53 } 54 55 /** 56 * amdgpu_vm_sdma_prepare - prepare SDMA command submission 57 * 58 * @p: see amdgpu_vm_update_params definition 59 * @owner: owner we need to sync to 60 * @exclusive: exclusive move fence we need to sync to 61 * 62 * Returns: 63 * Negativ errno, 0 for success. 64 */ 65 static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p, 66 void *owner, struct dma_fence *exclusive) 67 { 68 struct amdgpu_bo *root = p->vm->root.base.bo; 69 unsigned int ndw = AMDGPU_VM_SDMA_MIN_NUM_DW; 70 int r; 71 72 r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job); 73 if (r) 74 return r; 75 76 p->num_dw_left = ndw; 77 78 /* Wait for moves to be completed */ 79 r = amdgpu_sync_fence(&p->job->sync, exclusive, false); 80 if (r) 81 return r; 82 83 /* Don't wait for any submissions during page fault handling */ 84 if (p->direct) 85 return 0; 86 87 return amdgpu_sync_resv(p->adev, &p->job->sync, root->tbo.base.resv, 88 owner, false); 89 } 90 91 /** 92 * amdgpu_vm_sdma_commit - commit SDMA command submission 93 * 94 * @p: see amdgpu_vm_update_params definition 95 * @fence: resulting fence 96 * 97 * Returns: 98 * Negativ errno, 0 for success. 99 */ 100 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p, 101 struct dma_fence **fence) 102 { 103 struct amdgpu_ib *ib = p->job->ibs; 104 struct drm_sched_entity *entity; 105 struct dma_fence *f, *tmp; 106 struct amdgpu_ring *ring; 107 int r; 108 109 entity = p->direct ? &p->vm->direct : &p->vm->delayed; 110 ring = container_of(entity->rq->sched, struct amdgpu_ring, sched); 111 112 WARN_ON(ib->length_dw == 0); 113 amdgpu_ring_pad_ib(ring, ib); 114 WARN_ON(ib->length_dw > p->num_dw_left); 115 r = amdgpu_job_submit(p->job, entity, AMDGPU_FENCE_OWNER_VM, &f); 116 if (r) 117 goto error; 118 119 tmp = dma_fence_get(f); 120 if (p->direct) 121 swap(p->vm->last_direct, tmp); 122 else 123 swap(p->vm->last_delayed, tmp); 124 dma_fence_put(tmp); 125 126 if (fence && !p->direct) 127 swap(*fence, f); 128 dma_fence_put(f); 129 return 0; 130 131 error: 132 amdgpu_job_free(p->job); 133 return r; 134 } 135 136 /** 137 * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping 138 * 139 * @p: see amdgpu_vm_update_params definition 140 * @bo: PD/PT to update 141 * @pe: addr of the page entry 142 * @count: number of page entries to copy 143 * 144 * Traces the parameters and calls the DMA function to copy the PTEs. 145 */ 146 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p, 147 struct amdgpu_bo *bo, uint64_t pe, 148 unsigned count) 149 { 150 struct amdgpu_ib *ib = p->job->ibs; 151 uint64_t src = ib->gpu_addr; 152 153 src += p->num_dw_left * 4; 154 155 pe += amdgpu_bo_gpu_offset(bo); 156 trace_amdgpu_vm_copy_ptes(pe, src, count, p->direct); 157 158 amdgpu_vm_copy_pte(p->adev, ib, pe, src, count); 159 } 160 161 /** 162 * amdgpu_vm_sdma_set_ptes - helper to call the right asic function 163 * 164 * @p: see amdgpu_vm_update_params definition 165 * @bo: PD/PT to update 166 * @pe: addr of the page entry 167 * @addr: dst addr to write into pe 168 * @count: number of page entries to update 169 * @incr: increase next addr by incr bytes 170 * @flags: hw access flags 171 * 172 * Traces the parameters and calls the right asic functions 173 * to setup the page table using the DMA. 174 */ 175 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p, 176 struct amdgpu_bo *bo, uint64_t pe, 177 uint64_t addr, unsigned count, 178 uint32_t incr, uint64_t flags) 179 { 180 struct amdgpu_ib *ib = p->job->ibs; 181 182 pe += amdgpu_bo_gpu_offset(bo); 183 trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->direct); 184 if (count < 3) { 185 amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags, 186 count, incr); 187 } else { 188 amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr, 189 count, incr, flags); 190 } 191 } 192 193 /** 194 * amdgpu_vm_sdma_update - execute VM update 195 * 196 * @p: see amdgpu_vm_update_params definition 197 * @bo: PD/PT to update 198 * @pe: addr of the page entry 199 * @addr: dst addr to write into pe 200 * @count: number of page entries to update 201 * @incr: increase next addr by incr bytes 202 * @flags: hw access flags 203 * 204 * Reserve space in the IB, setup mapping buffer on demand and write commands to 205 * the IB. 206 */ 207 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p, 208 struct amdgpu_bo *bo, uint64_t pe, 209 uint64_t addr, unsigned count, uint32_t incr, 210 uint64_t flags) 211 { 212 unsigned int i, ndw, nptes; 213 uint64_t *pte; 214 int r; 215 216 do { 217 ndw = p->num_dw_left; 218 ndw -= p->job->ibs->length_dw; 219 220 if (ndw < 32) { 221 r = amdgpu_vm_sdma_commit(p, NULL); 222 if (r) 223 return r; 224 225 /* estimate how many dw we need */ 226 ndw = 32; 227 if (p->pages_addr) 228 ndw += count * 2; 229 ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW); 230 ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW); 231 232 r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &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 } 276 277 const 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 }; 283
Indexes created Sun Sep 21 20:09:37 GMT 2025