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 return amdgpu_ttm_alloc_gart(&table->bo.tbo); 39 } 40 41 /* Allocate a new job for @count PTE updates */ 42 static int amdgpu_vm_sdma_alloc_job(struct amdgpu_vm_update_params *p, 43 unsigned int count) 44 { 45 enum amdgpu_ib_pool_type pool = p->immediate ? AMDGPU_IB_POOL_IMMEDIATE 46 : AMDGPU_IB_POOL_DELAYED; 47 struct drm_sched_entity *entity = p->immediate ? &p->vm->immediate 48 : &p->vm->delayed; 49 unsigned int ndw; 50 int r; 51 52 /* estimate how many dw we need */ 53 ndw = AMDGPU_VM_SDMA_MIN_NUM_DW; 54 if (p->pages_addr) 55 ndw += count * 2; 56 ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW); 57 58 r = amdgpu_job_alloc_with_ib(p->adev, entity, AMDGPU_FENCE_OWNER_VM, 59 ndw * 4, pool, &p->job); 60 if (r) 61 return r; 62 63 p->num_dw_left = ndw; 64 return 0; 65 } 66 67 /** 68 * amdgpu_vm_sdma_prepare - prepare SDMA command submission 69 * 70 * @p: see amdgpu_vm_update_params definition 71 * @sync: amdgpu_sync object with fences to wait for 72 * 73 * Returns: 74 * Negativ errno, 0 for success. 75 */ 76 static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p, 77 struct amdgpu_sync *sync) 78 { 79 int r; 80 81 r = amdgpu_vm_sdma_alloc_job(p, 0); 82 if (r) 83 return r; 84 85 if (!sync) 86 return 0; 87 88 r = amdgpu_sync_push_to_job(sync, p->job); 89 if (r) { 90 p->num_dw_left = 0; 91 amdgpu_job_free(p->job); 92 } 93 return r; 94 } 95 96 /** 97 * amdgpu_vm_sdma_commit - commit SDMA command submission 98 * 99 * @p: see amdgpu_vm_update_params definition 100 * @fence: resulting fence 101 * 102 * Returns: 103 * Negativ errno, 0 for success. 104 */ 105 static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p, 106 struct dma_fence **fence) 107 { 108 struct amdgpu_ib *ib = p->job->ibs; 109 struct amdgpu_ring *ring; 110 struct dma_fence *f; 111 112 ring = container_of(p->vm->delayed.rq->sched, struct amdgpu_ring, 113 sched); 114 115 WARN_ON(ib->length_dw == 0); 116 amdgpu_ring_pad_ib(ring, ib); 117 118 if (p->needs_flush) 119 atomic64_inc(&p->vm->tlb_seq); 120 121 WARN_ON(ib->length_dw > p->num_dw_left); 122 f = amdgpu_job_submit(p->job); 123 124 if (p->unlocked) { 125 struct dma_fence *tmp = dma_fence_get(f); 126 127 swap(p->vm->last_unlocked, tmp); 128 dma_fence_put(tmp); 129 } else { 130 dma_resv_add_fence(p->vm->root.bo->tbo.base.resv, f, 131 DMA_RESV_USAGE_BOOKKEEP); 132 } 133 134 if (fence && !p->immediate) { 135 /* 136 * Most hw generations now have a separate queue for page table 137 * updates, but when the queue is shared with userspace we need 138 * the extra CPU round trip to correctly flush the TLB. 139 */ 140 set_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &f->flags); 141 swap(*fence, f); 142 } 143 dma_fence_put(f); 144 return 0; 145 } 146 147 /** 148 * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping 149 * 150 * @p: see amdgpu_vm_update_params definition 151 * @bo: PD/PT to update 152 * @pe: addr of the page entry 153 * @count: number of page entries to copy 154 * 155 * Traces the parameters and calls the DMA function to copy the PTEs. 156 */ 157 static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p, 158 struct amdgpu_bo *bo, uint64_t pe, 159 unsigned count) 160 { 161 struct amdgpu_ib *ib = p->job->ibs; 162 uint64_t src = ib->gpu_addr; 163 164 src += p->num_dw_left * 4; 165 166 pe += amdgpu_bo_gpu_offset_no_check(bo); 167 trace_amdgpu_vm_copy_ptes(pe, src, count, p->immediate); 168 169 amdgpu_vm_copy_pte(p->adev, ib, pe, src, count); 170 } 171 172 /** 173 * amdgpu_vm_sdma_set_ptes - helper to call the right asic function 174 * 175 * @p: see amdgpu_vm_update_params definition 176 * @bo: PD/PT to update 177 * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB 178 * @addr: dst addr to write into pe 179 * @count: number of page entries to update 180 * @incr: increase next addr by incr bytes 181 * @flags: hw access flags 182 * 183 * Traces the parameters and calls the right asic functions 184 * to setup the page table using the DMA. 185 */ 186 static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p, 187 struct amdgpu_bo *bo, uint64_t pe, 188 uint64_t addr, unsigned count, 189 uint32_t incr, uint64_t flags) 190 { 191 struct amdgpu_ib *ib = p->job->ibs; 192 193 pe += amdgpu_bo_gpu_offset_no_check(bo); 194 trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags, p->immediate); 195 if (count < 3) { 196 amdgpu_vm_write_pte(p->adev, ib, pe, addr | flags, 197 count, incr); 198 } else { 199 amdgpu_vm_set_pte_pde(p->adev, ib, pe, addr, 200 count, incr, flags); 201 } 202 } 203 204 /** 205 * amdgpu_vm_sdma_update - execute VM update 206 * 207 * @p: see amdgpu_vm_update_params definition 208 * @vmbo: PD/PT to update 209 * @pe: byte offset of the PDE/PTE, relative to start of PDB/PTB 210 * @addr: dst addr to write into pe 211 * @count: number of page entries to update 212 * @incr: increase next addr by incr bytes 213 * @flags: hw access flags 214 * 215 * Reserve space in the IB, setup mapping buffer on demand and write commands to 216 * the IB. 217 */ 218 static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p, 219 struct amdgpu_bo_vm *vmbo, uint64_t pe, 220 uint64_t addr, unsigned count, uint32_t incr, 221 uint64_t flags) 222 { 223 struct amdgpu_bo *bo = &vmbo->bo; 224 struct dma_resv_iter cursor; 225 unsigned int i, ndw, nptes; 226 struct dma_fence *fence; 227 uint64_t *pte; 228 int r; 229 230 /* Wait for PD/PT moves to be completed */ 231 dma_resv_iter_begin(&cursor, bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL); 232 dma_resv_for_each_fence_unlocked(&cursor, fence) { 233 dma_fence_get(fence); 234 r = drm_sched_job_add_dependency(&p->job->base, fence); 235 if (r) { 236 dma_fence_put(fence); 237 dma_resv_iter_end(&cursor); 238 return r; 239 } 240 } 241 dma_resv_iter_end(&cursor); 242 243 do { 244 ndw = p->num_dw_left; 245 ndw -= p->job->ibs->length_dw; 246 247 if (ndw < 32) { 248 r = amdgpu_vm_sdma_commit(p, NULL); 249 if (r) 250 return r; 251 252 r = amdgpu_vm_sdma_alloc_job(p, count); 253 if (r) 254 return r; 255 } 256 257 if (!p->pages_addr) { 258 /* set page commands needed */ 259 amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count, 260 incr, flags); 261 return 0; 262 } 263 264 /* copy commands needed */ 265 ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw; 266 267 /* for padding */ 268 ndw -= 7; 269 270 nptes = min(count, ndw / 2); 271 272 /* Put the PTEs at the end of the IB. */ 273 p->num_dw_left -= nptes * 2; 274 pte = (uint64_t *)&(p->job->ibs->ptr[p->num_dw_left]); 275 for (i = 0; i < nptes; ++i, addr += incr) { 276 pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr); 277 pte[i] |= flags; 278 } 279 280 amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes); 281 282 pe += nptes * 8; 283 count -= nptes; 284 } while (count); 285 286 return 0; 287 } 288 289 const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = { 290 .map_table = amdgpu_vm_sdma_map_table, 291 .prepare = amdgpu_vm_sdma_prepare, 292 .update = amdgpu_vm_sdma_update, 293 .commit = amdgpu_vm_sdma_commit 294 }; 295