1 /* 2 * Copyright (c) 2015, NVIDIA CORPORATION. All rights reserved. 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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 20 * DEALINGS IN THE SOFTWARE. 21 */ 22 23 /* 24 * GK20A does not have dedicated video memory, and to accurately represent this 25 * fact Nouveau will not create a RAM device for it. Therefore its instmem 26 * implementation must be done directly on top of system memory, while providing 27 * coherent read and write operations. 28 * 29 * Instmem can be allocated through two means: 30 * 1) If an IOMMU mapping has been probed, the IOMMU API is used to make memory 31 * pages contiguous to the GPU. This is the preferred way. 32 * 2) If no IOMMU mapping is probed, the DMA API is used to allocate physically 33 * contiguous memory. 34 * 35 * In both cases CPU read and writes are performed using PRAMIN (i.e. using the 36 * GPU path) to ensure these operations are coherent for the GPU. This allows us 37 * to use more "relaxed" allocation parameters when using the DMA API, since we 38 * never need a kernel mapping. 39 */ 40 #define gk20a_instmem(p) container_of((p), struct gk20a_instmem, base) 41 #include "priv.h" 42 43 #include <core/memory.h> 44 #include <core/mm.h> 45 #include <subdev/fb.h> 46 47 #ifdef __KERNEL__ 48 #include <linux/dma-attrs.h> 49 #include <linux/iommu.h> 50 #include <nouveau_platform.h> 51 #endif 52 53 #define gk20a_instobj(p) container_of((p), struct gk20a_instobj, memory) 54 55 struct gk20a_instobj { 56 struct nvkm_memory memory; 57 struct gk20a_instmem *imem; 58 struct nvkm_mem mem; 59 }; 60 61 /* 62 * Used for objects allocated using the DMA API 63 */ 64 struct gk20a_instobj_dma { 65 struct gk20a_instobj base; 66 67 void *cpuaddr; 68 dma_addr_t handle; 69 struct nvkm_mm_node r; 70 }; 71 72 /* 73 * Used for objects flattened using the IOMMU API 74 */ 75 struct gk20a_instobj_iommu { 76 struct gk20a_instobj base; 77 78 /* array of base.mem->size pages */ 79 struct page *pages[]; 80 }; 81 82 struct gk20a_instmem { 83 struct nvkm_instmem base; 84 unsigned long lock_flags; 85 spinlock_t lock; 86 u64 addr; 87 88 /* Only used if IOMMU if present */ 89 struct mutex *mm_mutex; 90 struct nvkm_mm *mm; 91 struct iommu_domain *domain; 92 unsigned long iommu_pgshift; 93 94 /* Only used by DMA API */ 95 struct dma_attrs attrs; 96 }; 97 98 static enum nvkm_memory_target 99 gk20a_instobj_target(struct nvkm_memory *memory) 100 { 101 return NVKM_MEM_TARGET_HOST; 102 } 103 104 static u64 105 gk20a_instobj_addr(struct nvkm_memory *memory) 106 { 107 return gk20a_instobj(memory)->mem.offset; 108 109 } 110 111 static u64 112 gk20a_instobj_size(struct nvkm_memory *memory) 113 { 114 return (u64)gk20a_instobj(memory)->mem.size << 12; 115 } 116 117 static void __iomem * 118 gk20a_instobj_acquire(struct nvkm_memory *memory) 119 { 120 struct gk20a_instmem *imem = gk20a_instobj(memory)->imem; 121 unsigned long flags; 122 spin_lock_irqsave(&imem->lock, flags); 123 imem->lock_flags = flags; 124 return NULL; 125 } 126 127 static void 128 gk20a_instobj_release(struct nvkm_memory *memory) 129 { 130 struct gk20a_instmem *imem = gk20a_instobj(memory)->imem; 131 spin_unlock_irqrestore(&imem->lock, imem->lock_flags); 132 } 133 134 /* 135 * Use PRAMIN to read/write data and avoid coherency issues. 136 * PRAMIN uses the GPU path and ensures data will always be coherent. 137 * 138 * A dynamic mapping based solution would be desirable in the future, but 139 * the issue remains of how to maintain coherency efficiently. On ARM it is 140 * not easy (if possible at all?) to create uncached temporary mappings. 141 */ 142 143 static u32 144 gk20a_instobj_rd32(struct nvkm_memory *memory, u64 offset) 145 { 146 struct gk20a_instobj *node = gk20a_instobj(memory); 147 struct gk20a_instmem *imem = node->imem; 148 struct nvkm_device *device = imem->base.subdev.device; 149 u64 base = (node->mem.offset + offset) & 0xffffff00000ULL; 150 u64 addr = (node->mem.offset + offset) & 0x000000fffffULL; 151 u32 data; 152 153 if (unlikely(imem->addr != base)) { 154 nvkm_wr32(device, 0x001700, base >> 16); 155 imem->addr = base; 156 } 157 data = nvkm_rd32(device, 0x700000 + addr); 158 return data; 159 } 160 161 static void 162 gk20a_instobj_wr32(struct nvkm_memory *memory, u64 offset, u32 data) 163 { 164 struct gk20a_instobj *node = gk20a_instobj(memory); 165 struct gk20a_instmem *imem = node->imem; 166 struct nvkm_device *device = imem->base.subdev.device; 167 u64 base = (node->mem.offset + offset) & 0xffffff00000ULL; 168 u64 addr = (node->mem.offset + offset) & 0x000000fffffULL; 169 170 if (unlikely(imem->addr != base)) { 171 nvkm_wr32(device, 0x001700, base >> 16); 172 imem->addr = base; 173 } 174 nvkm_wr32(device, 0x700000 + addr, data); 175 } 176 177 static void 178 gk20a_instobj_map(struct nvkm_memory *memory, struct nvkm_vma *vma, u64 offset) 179 { 180 struct gk20a_instobj *node = gk20a_instobj(memory); 181 nvkm_vm_map_at(vma, offset, &node->mem); 182 } 183 184 static void 185 gk20a_instobj_dtor_dma(struct gk20a_instobj *_node) 186 { 187 struct gk20a_instobj_dma *node = (void *)_node; 188 struct gk20a_instmem *imem = _node->imem; 189 struct device *dev = nv_device_base(nv_device(imem)); 190 191 if (unlikely(!node->cpuaddr)) 192 return; 193 194 dma_free_attrs(dev, _node->mem.size << PAGE_SHIFT, node->cpuaddr, 195 node->handle, &imem->attrs); 196 } 197 198 static void 199 gk20a_instobj_dtor_iommu(struct gk20a_instobj *_node) 200 { 201 struct gk20a_instobj_iommu *node = (void *)_node; 202 struct gk20a_instmem *imem = _node->imem; 203 struct nvkm_mm_node *r; 204 int i; 205 206 if (unlikely(list_empty(&_node->mem.regions))) 207 return; 208 209 r = list_first_entry(&_node->mem.regions, struct nvkm_mm_node, 210 rl_entry); 211 212 /* clear bit 34 to unmap pages */ 213 r->offset &= ~BIT(34 - imem->iommu_pgshift); 214 215 /* Unmap pages from GPU address space and free them */ 216 for (i = 0; i < _node->mem.size; i++) { 217 iommu_unmap(imem->domain, 218 (r->offset + i) << imem->iommu_pgshift, PAGE_SIZE); 219 __free_page(node->pages[i]); 220 } 221 222 /* Release area from GPU address space */ 223 mutex_lock(imem->mm_mutex); 224 nvkm_mm_free(imem->mm, &r); 225 mutex_unlock(imem->mm_mutex); 226 } 227 228 static void * 229 gk20a_instobj_dtor(struct nvkm_memory *memory) 230 { 231 struct gk20a_instobj *node = gk20a_instobj(memory); 232 struct gk20a_instmem *imem = node->imem; 233 234 if (imem->domain) 235 gk20a_instobj_dtor_iommu(node); 236 else 237 gk20a_instobj_dtor_dma(node); 238 239 return node; 240 } 241 242 static const struct nvkm_memory_func 243 gk20a_instobj_func = { 244 .dtor = gk20a_instobj_dtor, 245 .target = gk20a_instobj_target, 246 .addr = gk20a_instobj_addr, 247 .size = gk20a_instobj_size, 248 .acquire = gk20a_instobj_acquire, 249 .release = gk20a_instobj_release, 250 .rd32 = gk20a_instobj_rd32, 251 .wr32 = gk20a_instobj_wr32, 252 .map = gk20a_instobj_map, 253 }; 254 255 static int 256 gk20a_instobj_ctor_dma(struct gk20a_instmem *imem, u32 npages, u32 align, 257 struct gk20a_instobj **_node) 258 { 259 struct gk20a_instobj_dma *node; 260 struct nvkm_subdev *subdev = &imem->base.subdev; 261 struct device *dev = subdev->device->dev; 262 263 if (!(node = kzalloc(sizeof(*node), GFP_KERNEL))) 264 return -ENOMEM; 265 *_node = &node->base; 266 267 node->cpuaddr = dma_alloc_attrs(dev, npages << PAGE_SHIFT, 268 &node->handle, GFP_KERNEL, 269 &imem->attrs); 270 if (!node->cpuaddr) { 271 nvkm_error(subdev, "cannot allocate DMA memory\n"); 272 return -ENOMEM; 273 } 274 275 /* alignment check */ 276 if (unlikely(node->handle & (align - 1))) 277 nvkm_warn(subdev, 278 "memory not aligned as requested: %pad (0x%x)\n", 279 &node->handle, align); 280 281 /* present memory for being mapped using small pages */ 282 node->r.type = 12; 283 node->r.offset = node->handle >> 12; 284 node->r.length = (npages << PAGE_SHIFT) >> 12; 285 286 node->base.mem.offset = node->handle; 287 288 INIT_LIST_HEAD(&node->base.mem.regions); 289 list_add_tail(&node->r.rl_entry, &node->base.mem.regions); 290 291 return 0; 292 } 293 294 static int 295 gk20a_instobj_ctor_iommu(struct gk20a_instmem *imem, u32 npages, u32 align, 296 struct gk20a_instobj **_node) 297 { 298 struct gk20a_instobj_iommu *node; 299 struct nvkm_subdev *subdev = &imem->base.subdev; 300 struct nvkm_mm_node *r; 301 int ret; 302 int i; 303 304 if (!(node = kzalloc(sizeof(*node) + 305 sizeof( node->pages[0]) * npages, GFP_KERNEL))) 306 return -ENOMEM; 307 *_node = &node->base; 308 309 /* Allocate backing memory */ 310 for (i = 0; i < npages; i++) { 311 struct page *p = alloc_page(GFP_KERNEL); 312 313 if (p == NULL) { 314 ret = -ENOMEM; 315 goto free_pages; 316 } 317 node->pages[i] = p; 318 } 319 320 mutex_lock(imem->mm_mutex); 321 /* Reserve area from GPU address space */ 322 ret = nvkm_mm_head(imem->mm, 0, 1, npages, npages, 323 align >> imem->iommu_pgshift, &r); 324 mutex_unlock(imem->mm_mutex); 325 if (ret) { 326 nvkm_error(subdev, "virtual space is full!\n"); 327 goto free_pages; 328 } 329 330 /* Map into GPU address space */ 331 for (i = 0; i < npages; i++) { 332 struct page *p = node->pages[i]; 333 u32 offset = (r->offset + i) << imem->iommu_pgshift; 334 335 ret = iommu_map(imem->domain, offset, page_to_phys(p), 336 PAGE_SIZE, IOMMU_READ | IOMMU_WRITE); 337 if (ret < 0) { 338 nvkm_error(subdev, "IOMMU mapping failure: %d\n", ret); 339 340 while (i-- > 0) { 341 offset -= PAGE_SIZE; 342 iommu_unmap(imem->domain, offset, PAGE_SIZE); 343 } 344 goto release_area; 345 } 346 } 347 348 /* Bit 34 tells that an address is to be resolved through the IOMMU */ 349 r->offset |= BIT(34 - imem->iommu_pgshift); 350 351 node->base.mem.offset = ((u64)r->offset) << imem->iommu_pgshift; 352 353 INIT_LIST_HEAD(&node->base.mem.regions); 354 list_add_tail(&r->rl_entry, &node->base.mem.regions); 355 356 return 0; 357 358 release_area: 359 mutex_lock(imem->mm_mutex); 360 nvkm_mm_free(imem->mm, &r); 361 mutex_unlock(imem->mm_mutex); 362 363 free_pages: 364 for (i = 0; i < npages && node->pages[i] != NULL; i++) 365 __free_page(node->pages[i]); 366 367 return ret; 368 } 369 370 static int 371 gk20a_instobj_new(struct nvkm_instmem *base, u32 size, u32 align, bool zero, 372 struct nvkm_memory **pmemory) 373 { 374 struct gk20a_instmem *imem = gk20a_instmem(base); 375 struct gk20a_instobj *node; 376 struct nvkm_subdev *subdev = &imem->base.subdev; 377 int ret; 378 379 nvkm_debug(subdev, "%s (%s): size: %x align: %x\n", __func__, 380 imem->domain ? "IOMMU" : "DMA", size, align); 381 382 /* Round size and align to page bounds */ 383 size = max(roundup(size, PAGE_SIZE), PAGE_SIZE); 384 align = max(roundup(align, PAGE_SIZE), PAGE_SIZE); 385 386 if (imem->domain) 387 ret = gk20a_instobj_ctor_iommu(imem, size >> PAGE_SHIFT, 388 align, &node); 389 else 390 ret = gk20a_instobj_ctor_dma(imem, size >> PAGE_SHIFT, 391 align, &node); 392 if (ret) 393 return ret; 394 *pmemory = &node->memory; 395 396 nvkm_memory_ctor(&gk20a_instobj_func, &node->memory); 397 node->imem = imem; 398 399 /* present memory for being mapped using small pages */ 400 node->mem.size = size >> 12; 401 node->mem.memtype = 0; 402 node->mem.page_shift = 12; 403 404 nvkm_debug(subdev, "alloc size: 0x%x, align: 0x%x, gaddr: 0x%llx\n", 405 size, align, node->mem.offset); 406 407 return 0; 408 } 409 410 static int 411 gk20a_instmem_fini(struct nvkm_object *object, bool suspend) 412 { 413 struct gk20a_instmem *imem = (void *)object; 414 imem->addr = ~0ULL; 415 return nvkm_instmem_fini(&imem->base, suspend); 416 } 417 418 static int 419 gk20a_instmem_ctor(struct nvkm_object *parent, struct nvkm_object *engine, 420 struct nvkm_oclass *oclass, void *data, u32 size, 421 struct nvkm_object **pobject) 422 { 423 struct nvkm_device *device = (void *)parent; 424 struct gk20a_instmem *imem; 425 int ret; 426 427 ret = nvkm_instmem_create(parent, engine, oclass, &imem); 428 *pobject = nv_object(imem); 429 if (ret) 430 return ret; 431 432 spin_lock_init(&imem->lock); 433 434 if (device->gpu->iommu.domain) { 435 imem->domain = device->gpu->iommu.domain; 436 imem->mm = device->gpu->iommu.mm; 437 imem->iommu_pgshift = device->gpu->iommu.pgshift; 438 imem->mm_mutex = &device->gpu->iommu.mutex; 439 440 nvkm_info(&imem->base.subdev, "using IOMMU\n"); 441 } else { 442 init_dma_attrs(&imem->attrs); 443 /* 444 * We will access instmem through PRAMIN and thus do not need a 445 * consistent CPU pointer or kernel mapping 446 */ 447 dma_set_attr(DMA_ATTR_NON_CONSISTENT, &imem->attrs); 448 dma_set_attr(DMA_ATTR_WEAK_ORDERING, &imem->attrs); 449 dma_set_attr(DMA_ATTR_WRITE_COMBINE, &imem->attrs); 450 dma_set_attr(DMA_ATTR_NO_KERNEL_MAPPING, &imem->attrs); 451 452 nvkm_info(&imem->base.subdev, "using DMA API\n"); 453 } 454 455 return 0; 456 } 457 458 struct nvkm_oclass * 459 gk20a_instmem_oclass = &(struct nvkm_instmem_impl) { 460 .base.handle = NV_SUBDEV(INSTMEM, 0xea), 461 .base.ofuncs = &(struct nvkm_ofuncs) { 462 .ctor = gk20a_instmem_ctor, 463 .dtor = _nvkm_instmem_dtor, 464 .init = _nvkm_instmem_init, 465 .fini = gk20a_instmem_fini, 466 }, 467 .memory_new = gk20a_instobj_new, 468 .persistent = true, 469 .zero = false, 470 }.base; 471