| 1 | /************************************************************************** |
| 2 | * |
| 3 | * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA |
| 4 | * All Rights Reserved. |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 7 | * copy of this software and associated documentation files (the |
| 8 | * "Software"), to deal in the Software without restriction, including |
| 9 | * without limitation the rights to use, copy, modify, merge, publish, |
| 10 | * distribute, sub license, and/or sell copies of the Software, and to |
| 11 | * permit persons to whom the Software is furnished to do so, subject to |
| 12 | * the following conditions: |
| 13 | * |
| 14 | * The above copyright notice and this permission notice (including the |
| 15 | * next paragraph) shall be included in all copies or substantial portions |
| 16 | * of the Software. |
| 17 | * |
| 18 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 19 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 20 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL |
| 21 | * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, |
| 22 | * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR |
| 23 | * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE |
| 24 | * USE OR OTHER DEALINGS IN THE SOFTWARE. |
| 25 | * |
| 26 | **************************************************************************/ |
| 27 | /* |
| 28 | * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com> |
| 29 | */ |
| 30 | |
| 31 | #define pr_fmt(fmt) "[TTM] " fmt |
| 32 | |
| 33 | #include <ttm/ttm_module.h> |
| 34 | #include <ttm/ttm_bo_driver.h> |
| 35 | #include <ttm/ttm_placement.h> |
| 36 | #include <drm/drm_vma_manager.h> |
| 37 | #include <linux/mm.h> |
| 38 | #include <linux/rbtree.h> |
| 39 | #include <linux/module.h> |
| 40 | #include <linux/uaccess.h> |
| 41 | |
| 42 | #define TTM_BO_VM_NUM_PREFAULT 16 |
| 43 | |
| 44 | static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo, |
| 45 | struct vm_area_struct *vma, |
| 46 | struct vm_fault *vmf) |
| 47 | { |
| 48 | struct ttm_bo_device *bdev = bo->bdev; |
| 49 | int ret = 0; |
| 50 | |
| 51 | spin_lock(&bdev->fence_lock); |
| 52 | if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags))) |
| 53 | goto out_unlock; |
| 54 | |
| 55 | /* |
| 56 | * Quick non-stalling check for idle. |
| 57 | */ |
| 58 | ret = ttm_bo_wait(bo, false, false, true); |
| 59 | if (likely(ret == 0)) |
| 60 | goto out_unlock; |
| 61 | |
| 62 | /* |
| 63 | * If possible, avoid waiting for GPU with mmap_sem |
| 64 | * held. |
| 65 | */ |
| 66 | if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) { |
| 67 | ret = VM_FAULT_RETRY; |
| 68 | if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) |
| 69 | goto out_unlock; |
| 70 | |
| 71 | up_read(&vma->vm_mm->mmap_sem); |
| 72 | (void) ttm_bo_wait(bo, false, true, false); |
| 73 | goto out_unlock; |
| 74 | } |
| 75 | |
| 76 | /* |
| 77 | * Ordinary wait. |
| 78 | */ |
| 79 | ret = ttm_bo_wait(bo, false, true, false); |
| 80 | if (unlikely(ret != 0)) |
| 81 | ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS : |
| 82 | VM_FAULT_NOPAGE; |
| 83 | |
| 84 | out_unlock: |
| 85 | spin_unlock(&bdev->fence_lock); |
| 86 | return ret; |
| 87 | } |
| 88 | |
| 89 | static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
| 90 | { |
| 91 | struct ttm_buffer_object *bo = (struct ttm_buffer_object *) |
| 92 | vma->vm_private_data; |
| 93 | struct ttm_bo_device *bdev = bo->bdev; |
| 94 | unsigned long page_offset; |
| 95 | unsigned long page_last; |
| 96 | unsigned long pfn; |
| 97 | struct ttm_tt *ttm = NULL; |
| 98 | struct page *page; |
| 99 | int ret; |
| 100 | int i; |
| 101 | unsigned long address = (unsigned long)vmf->virtual_address; |
| 102 | int retval = VM_FAULT_NOPAGE; |
| 103 | struct ttm_mem_type_manager *man = |
| 104 | &bdev->man[bo->mem.mem_type]; |
| 105 | struct vm_area_struct cvma; |
| 106 | |
| 107 | /* |
| 108 | * Work around locking order reversal in fault / nopfn |
| 109 | * between mmap_sem and bo_reserve: Perform a trylock operation |
| 110 | * for reserve, and if it fails, retry the fault after waiting |
| 111 | * for the buffer to become unreserved. |
| 112 | */ |
| 113 | ret = ttm_bo_reserve(bo, true, true, false, NULL); |
| 114 | if (unlikely(ret != 0)) { |
| 115 | if (ret != -EBUSY) |
| 116 | return VM_FAULT_NOPAGE; |
| 117 | |
| 118 | if (vmf->flags & FAULT_FLAG_ALLOW_RETRY) { |
| 119 | if (!(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) { |
| 120 | up_read(&vma->vm_mm->mmap_sem); |
| 121 | (void) ttm_bo_wait_unreserved(bo); |
| 122 | } |
| 123 | |
| 124 | return VM_FAULT_RETRY; |
| 125 | } |
| 126 | |
| 127 | /* |
| 128 | * If we'd want to change locking order to |
| 129 | * mmap_sem -> bo::reserve, we'd use a blocking reserve here |
| 130 | * instead of retrying the fault... |
| 131 | */ |
| 132 | return VM_FAULT_NOPAGE; |
| 133 | } |
| 134 | |
| 135 | /* |
| 136 | * Refuse to fault imported pages. This should be handled |
| 137 | * (if at all) by redirecting mmap to the exporter. |
| 138 | */ |
| 139 | if (bo->ttm && (bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) { |
| 140 | retval = VM_FAULT_SIGBUS; |
| 141 | goto out_unlock; |
| 142 | } |
| 143 | |
| 144 | if (bdev->driver->fault_reserve_notify) { |
| 145 | ret = bdev->driver->fault_reserve_notify(bo); |
| 146 | switch (ret) { |
| 147 | case 0: |
| 148 | break; |
| 149 | case -EBUSY: |
| 150 | case -ERESTARTSYS: |
| 151 | retval = VM_FAULT_NOPAGE; |
| 152 | goto out_unlock; |
| 153 | default: |
| 154 | retval = VM_FAULT_SIGBUS; |
| 155 | goto out_unlock; |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | /* |
| 160 | * Wait for buffer data in transit, due to a pipelined |
| 161 | * move. |
| 162 | */ |
| 163 | ret = ttm_bo_vm_fault_idle(bo, vma, vmf); |
| 164 | if (unlikely(ret != 0)) { |
| 165 | retval = ret; |
| 166 | goto out_unlock; |
| 167 | } |
| 168 | |
| 169 | ret = ttm_mem_io_lock(man, true); |
| 170 | if (unlikely(ret != 0)) { |
| 171 | retval = VM_FAULT_NOPAGE; |
| 172 | goto out_unlock; |
| 173 | } |
| 174 | ret = ttm_mem_io_reserve_vm(bo); |
| 175 | if (unlikely(ret != 0)) { |
| 176 | retval = VM_FAULT_SIGBUS; |
| 177 | goto out_io_unlock; |
| 178 | } |
| 179 | |
| 180 | page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + |
| 181 | vma->vm_pgoff - drm_vma_node_start(&bo->vma_node); |
| 182 | page_last = vma_pages(vma) + vma->vm_pgoff - |
| 183 | drm_vma_node_start(&bo->vma_node); |
| 184 | |
| 185 | if (unlikely(page_offset >= bo->num_pages)) { |
| 186 | retval = VM_FAULT_SIGBUS; |
| 187 | goto out_io_unlock; |
| 188 | } |
| 189 | |
| 190 | /* |
| 191 | * Make a local vma copy to modify the page_prot member |
| 192 | * and vm_flags if necessary. The vma parameter is protected |
| 193 | * by mmap_sem in write mode. |
| 194 | */ |
| 195 | cvma = *vma; |
| 196 | cvma.vm_page_prot = vm_get_page_prot(cvma.vm_flags); |
| 197 | |
| 198 | if (bo->mem.bus.is_iomem) { |
| 199 | cvma.vm_page_prot = ttm_io_prot(bo->mem.placement, |
| 200 | cvma.vm_page_prot); |
| 201 | } else { |
| 202 | ttm = bo->ttm; |
| 203 | if (!(bo->mem.placement & TTM_PL_FLAG_CACHED)) |
| 204 | cvma.vm_page_prot = ttm_io_prot(bo->mem.placement, |
| 205 | cvma.vm_page_prot); |
| 206 | |
| 207 | /* Allocate all page at once, most common usage */ |
| 208 | if (ttm->bdev->driver->ttm_tt_populate(ttm)) { |
| 209 | retval = VM_FAULT_OOM; |
| 210 | goto out_io_unlock; |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | /* |
| 215 | * Speculatively prefault a number of pages. Only error on |
| 216 | * first page. |
| 217 | */ |
| 218 | for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) { |
| 219 | if (bo->mem.bus.is_iomem) |
| 220 | pfn = ((bo->mem.bus.base + bo->mem.bus.offset) >> PAGE_SHIFT) + page_offset; |
| 221 | else { |
| 222 | page = ttm->pages[page_offset]; |
| 223 | if (unlikely(!page && i == 0)) { |
| 224 | retval = VM_FAULT_OOM; |
| 225 | goto out_io_unlock; |
| 226 | } else if (unlikely(!page)) { |
| 227 | break; |
| 228 | } |
| 229 | page->mapping = vma->vm_file->f_mapping; |
| 230 | page->index = drm_vma_node_start(&bo->vma_node) + |
| 231 | page_offset; |
| 232 | pfn = page_to_pfn(page); |
| 233 | } |
| 234 | |
| 235 | if (vma->vm_flags & VM_MIXEDMAP) |
| 236 | ret = vm_insert_mixed(&cvma, address, pfn); |
| 237 | else |
| 238 | ret = vm_insert_pfn(&cvma, address, pfn); |
| 239 | |
| 240 | /* |
| 241 | * Somebody beat us to this PTE or prefaulting to |
| 242 | * an already populated PTE, or prefaulting error. |
| 243 | */ |
| 244 | |
| 245 | if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0))) |
| 246 | break; |
| 247 | else if (unlikely(ret != 0)) { |
| 248 | retval = |
| 249 | (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS; |
| 250 | goto out_io_unlock; |
| 251 | } |
| 252 | |
| 253 | address += PAGE_SIZE; |
| 254 | if (unlikely(++page_offset >= page_last)) |
| 255 | break; |
| 256 | } |
| 257 | out_io_unlock: |
| 258 | ttm_mem_io_unlock(man); |
| 259 | out_unlock: |
| 260 | ttm_bo_unreserve(bo); |
| 261 | return retval; |
| 262 | } |
| 263 | |
| 264 | static void ttm_bo_vm_open(struct vm_area_struct *vma) |
| 265 | { |
| 266 | struct ttm_buffer_object *bo = |
| 267 | (struct ttm_buffer_object *)vma->vm_private_data; |
| 268 | |
| 269 | WARN_ON(bo->bdev->dev_mapping != vma->vm_file->f_mapping); |
| 270 | |
| 271 | (void)ttm_bo_reference(bo); |
| 272 | } |
| 273 | |
| 274 | static void ttm_bo_vm_close(struct vm_area_struct *vma) |
| 275 | { |
| 276 | struct ttm_buffer_object *bo = (struct ttm_buffer_object *)vma->vm_private_data; |
| 277 | |
| 278 | ttm_bo_unref(&bo); |
| 279 | vma->vm_private_data = NULL; |
| 280 | } |
| 281 | |
| 282 | static const struct vm_operations_struct ttm_bo_vm_ops = { |
| 283 | .fault = ttm_bo_vm_fault, |
| 284 | .open = ttm_bo_vm_open, |
| 285 | .close = ttm_bo_vm_close |
| 286 | }; |
| 287 | |
| 288 | static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev, |
| 289 | unsigned long offset, |
| 290 | unsigned long pages) |
| 291 | { |
| 292 | struct drm_vma_offset_node *node; |
| 293 | struct ttm_buffer_object *bo = NULL; |
| 294 | |
| 295 | drm_vma_offset_lock_lookup(&bdev->vma_manager); |
| 296 | |
| 297 | node = drm_vma_offset_lookup_locked(&bdev->vma_manager, offset, pages); |
| 298 | if (likely(node)) { |
| 299 | bo = container_of(node, struct ttm_buffer_object, vma_node); |
| 300 | if (!kref_get_unless_zero(&bo->kref)) |
| 301 | bo = NULL; |
| 302 | } |
| 303 | |
| 304 | drm_vma_offset_unlock_lookup(&bdev->vma_manager); |
| 305 | |
| 306 | if (!bo) |
| 307 | pr_err("Could not find buffer object to map\n"); |
| 308 | |
| 309 | return bo; |
| 310 | } |
| 311 | |
| 312 | int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma, |
| 313 | struct ttm_bo_device *bdev) |
| 314 | { |
| 315 | struct ttm_bo_driver *driver; |
| 316 | struct ttm_buffer_object *bo; |
| 317 | int ret; |
| 318 | |
| 319 | bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma)); |
| 320 | if (unlikely(!bo)) |
| 321 | return -EINVAL; |
| 322 | |
| 323 | driver = bo->bdev->driver; |
| 324 | if (unlikely(!driver->verify_access)) { |
| 325 | ret = -EPERM; |
| 326 | goto out_unref; |
| 327 | } |
| 328 | ret = driver->verify_access(bo, filp); |
| 329 | if (unlikely(ret != 0)) |
| 330 | goto out_unref; |
| 331 | |
| 332 | vma->vm_ops = &ttm_bo_vm_ops; |
| 333 | |
| 334 | /* |
| 335 | * Note: We're transferring the bo reference to |
| 336 | * vma->vm_private_data here. |
| 337 | */ |
| 338 | |
| 339 | vma->vm_private_data = bo; |
| 340 | |
| 341 | /* |
| 342 | * We'd like to use VM_PFNMAP on shared mappings, where |
| 343 | * (vma->vm_flags & VM_SHARED) != 0, for performance reasons, |
| 344 | * but for some reason VM_PFNMAP + x86 PAT + write-combine is very |
| 345 | * bad for performance. Until that has been sorted out, use |
| 346 | * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719 |
| 347 | */ |
| 348 | vma->vm_flags |= VM_MIXEDMAP; |
| 349 | vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP; |
| 350 | return 0; |
| 351 | out_unref: |
| 352 | ttm_bo_unref(&bo); |
| 353 | return ret; |
| 354 | } |
| 355 | EXPORT_SYMBOL(ttm_bo_mmap); |
| 356 | |
| 357 | int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo) |
| 358 | { |
| 359 | if (vma->vm_pgoff != 0) |
| 360 | return -EACCES; |
| 361 | |
| 362 | vma->vm_ops = &ttm_bo_vm_ops; |
| 363 | vma->vm_private_data = ttm_bo_reference(bo); |
| 364 | vma->vm_flags |= VM_MIXEDMAP; |
| 365 | vma->vm_flags |= VM_IO | VM_DONTEXPAND; |
| 366 | return 0; |
| 367 | } |
| 368 | EXPORT_SYMBOL(ttm_fbdev_mmap); |