Thomas Hellstrom | ba4e7d9 | 2009-06-10 15:20:19 +0200 | [diff] [blame] | 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 | #include <ttm/ttm_module.h> |
| 32 | #include <ttm/ttm_bo_driver.h> |
| 33 | #include <ttm/ttm_placement.h> |
| 34 | #include <linux/mm.h> |
Thomas Hellstrom | ba4e7d9 | 2009-06-10 15:20:19 +0200 | [diff] [blame] | 35 | #include <linux/rbtree.h> |
| 36 | #include <linux/module.h> |
| 37 | #include <linux/uaccess.h> |
| 38 | |
| 39 | #define TTM_BO_VM_NUM_PREFAULT 16 |
| 40 | |
| 41 | static struct ttm_buffer_object *ttm_bo_vm_lookup_rb(struct ttm_bo_device *bdev, |
| 42 | unsigned long page_start, |
| 43 | unsigned long num_pages) |
| 44 | { |
| 45 | struct rb_node *cur = bdev->addr_space_rb.rb_node; |
| 46 | unsigned long cur_offset; |
| 47 | struct ttm_buffer_object *bo; |
| 48 | struct ttm_buffer_object *best_bo = NULL; |
| 49 | |
| 50 | while (likely(cur != NULL)) { |
| 51 | bo = rb_entry(cur, struct ttm_buffer_object, vm_rb); |
| 52 | cur_offset = bo->vm_node->start; |
| 53 | if (page_start >= cur_offset) { |
| 54 | cur = cur->rb_right; |
| 55 | best_bo = bo; |
| 56 | if (page_start == cur_offset) |
| 57 | break; |
| 58 | } else |
| 59 | cur = cur->rb_left; |
| 60 | } |
| 61 | |
| 62 | if (unlikely(best_bo == NULL)) |
| 63 | return NULL; |
| 64 | |
| 65 | if (unlikely((best_bo->vm_node->start + best_bo->num_pages) < |
| 66 | (page_start + num_pages))) |
| 67 | return NULL; |
| 68 | |
| 69 | return best_bo; |
| 70 | } |
| 71 | |
| 72 | static int ttm_bo_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
| 73 | { |
| 74 | struct ttm_buffer_object *bo = (struct ttm_buffer_object *) |
| 75 | vma->vm_private_data; |
| 76 | struct ttm_bo_device *bdev = bo->bdev; |
| 77 | unsigned long bus_base; |
| 78 | unsigned long bus_offset; |
| 79 | unsigned long bus_size; |
| 80 | unsigned long page_offset; |
| 81 | unsigned long page_last; |
| 82 | unsigned long pfn; |
| 83 | struct ttm_tt *ttm = NULL; |
| 84 | struct page *page; |
| 85 | int ret; |
| 86 | int i; |
| 87 | bool is_iomem; |
| 88 | unsigned long address = (unsigned long)vmf->virtual_address; |
| 89 | int retval = VM_FAULT_NOPAGE; |
| 90 | |
| 91 | /* |
| 92 | * Work around locking order reversal in fault / nopfn |
| 93 | * between mmap_sem and bo_reserve: Perform a trylock operation |
| 94 | * for reserve, and if it fails, retry the fault after scheduling. |
| 95 | */ |
| 96 | |
| 97 | ret = ttm_bo_reserve(bo, true, true, false, 0); |
| 98 | if (unlikely(ret != 0)) { |
| 99 | if (ret == -EBUSY) |
| 100 | set_need_resched(); |
| 101 | return VM_FAULT_NOPAGE; |
| 102 | } |
| 103 | |
Dave Airlie | e024e11 | 2009-06-24 09:48:08 +1000 | [diff] [blame^] | 104 | if (bdev->driver->fault_reserve_notify) |
| 105 | bdev->driver->fault_reserve_notify(bo); |
| 106 | |
Thomas Hellstrom | ba4e7d9 | 2009-06-10 15:20:19 +0200 | [diff] [blame] | 107 | /* |
| 108 | * Wait for buffer data in transit, due to a pipelined |
| 109 | * move. |
| 110 | */ |
| 111 | |
| 112 | spin_lock(&bo->lock); |
| 113 | if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) { |
| 114 | ret = ttm_bo_wait(bo, false, true, false); |
| 115 | spin_unlock(&bo->lock); |
| 116 | if (unlikely(ret != 0)) { |
| 117 | retval = (ret != -ERESTART) ? |
| 118 | VM_FAULT_SIGBUS : VM_FAULT_NOPAGE; |
| 119 | goto out_unlock; |
| 120 | } |
| 121 | } else |
| 122 | spin_unlock(&bo->lock); |
| 123 | |
| 124 | |
| 125 | ret = ttm_bo_pci_offset(bdev, &bo->mem, &bus_base, &bus_offset, |
| 126 | &bus_size); |
| 127 | if (unlikely(ret != 0)) { |
| 128 | retval = VM_FAULT_SIGBUS; |
| 129 | goto out_unlock; |
| 130 | } |
| 131 | |
| 132 | is_iomem = (bus_size != 0); |
| 133 | |
| 134 | page_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + |
| 135 | bo->vm_node->start - vma->vm_pgoff; |
| 136 | page_last = ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) + |
| 137 | bo->vm_node->start - vma->vm_pgoff; |
| 138 | |
| 139 | if (unlikely(page_offset >= bo->num_pages)) { |
| 140 | retval = VM_FAULT_SIGBUS; |
| 141 | goto out_unlock; |
| 142 | } |
| 143 | |
| 144 | /* |
| 145 | * Strictly, we're not allowed to modify vma->vm_page_prot here, |
| 146 | * since the mmap_sem is only held in read mode. However, we |
| 147 | * modify only the caching bits of vma->vm_page_prot and |
| 148 | * consider those bits protected by |
| 149 | * the bo->mutex, as we should be the only writers. |
| 150 | * There shouldn't really be any readers of these bits except |
| 151 | * within vm_insert_mixed()? fork? |
| 152 | * |
| 153 | * TODO: Add a list of vmas to the bo, and change the |
| 154 | * vma->vm_page_prot when the object changes caching policy, with |
| 155 | * the correct locks held. |
| 156 | */ |
| 157 | |
| 158 | if (is_iomem) { |
| 159 | vma->vm_page_prot = ttm_io_prot(bo->mem.placement, |
| 160 | vma->vm_page_prot); |
| 161 | } else { |
| 162 | ttm = bo->ttm; |
| 163 | vma->vm_page_prot = (bo->mem.placement & TTM_PL_FLAG_CACHED) ? |
| 164 | vm_get_page_prot(vma->vm_flags) : |
| 165 | ttm_io_prot(bo->mem.placement, vma->vm_page_prot); |
| 166 | } |
| 167 | |
| 168 | /* |
| 169 | * Speculatively prefault a number of pages. Only error on |
| 170 | * first page. |
| 171 | */ |
| 172 | |
| 173 | for (i = 0; i < TTM_BO_VM_NUM_PREFAULT; ++i) { |
| 174 | |
| 175 | if (is_iomem) |
| 176 | pfn = ((bus_base + bus_offset) >> PAGE_SHIFT) + |
| 177 | page_offset; |
| 178 | else { |
| 179 | page = ttm_tt_get_page(ttm, page_offset); |
| 180 | if (unlikely(!page && i == 0)) { |
| 181 | retval = VM_FAULT_OOM; |
| 182 | goto out_unlock; |
| 183 | } else if (unlikely(!page)) { |
| 184 | break; |
| 185 | } |
| 186 | pfn = page_to_pfn(page); |
| 187 | } |
| 188 | |
| 189 | ret = vm_insert_mixed(vma, address, pfn); |
| 190 | /* |
| 191 | * Somebody beat us to this PTE or prefaulting to |
| 192 | * an already populated PTE, or prefaulting error. |
| 193 | */ |
| 194 | |
| 195 | if (unlikely((ret == -EBUSY) || (ret != 0 && i > 0))) |
| 196 | break; |
| 197 | else if (unlikely(ret != 0)) { |
| 198 | retval = |
| 199 | (ret == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS; |
| 200 | goto out_unlock; |
| 201 | |
| 202 | } |
| 203 | |
| 204 | address += PAGE_SIZE; |
| 205 | if (unlikely(++page_offset >= page_last)) |
| 206 | break; |
| 207 | } |
| 208 | |
| 209 | out_unlock: |
| 210 | ttm_bo_unreserve(bo); |
| 211 | return retval; |
| 212 | } |
| 213 | |
| 214 | static void ttm_bo_vm_open(struct vm_area_struct *vma) |
| 215 | { |
| 216 | struct ttm_buffer_object *bo = |
| 217 | (struct ttm_buffer_object *)vma->vm_private_data; |
| 218 | |
| 219 | (void)ttm_bo_reference(bo); |
| 220 | } |
| 221 | |
| 222 | static void ttm_bo_vm_close(struct vm_area_struct *vma) |
| 223 | { |
| 224 | struct ttm_buffer_object *bo = |
| 225 | (struct ttm_buffer_object *)vma->vm_private_data; |
| 226 | |
| 227 | ttm_bo_unref(&bo); |
| 228 | vma->vm_private_data = NULL; |
| 229 | } |
| 230 | |
| 231 | static struct vm_operations_struct ttm_bo_vm_ops = { |
| 232 | .fault = ttm_bo_vm_fault, |
| 233 | .open = ttm_bo_vm_open, |
| 234 | .close = ttm_bo_vm_close |
| 235 | }; |
| 236 | |
| 237 | int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma, |
| 238 | struct ttm_bo_device *bdev) |
| 239 | { |
| 240 | struct ttm_bo_driver *driver; |
| 241 | struct ttm_buffer_object *bo; |
| 242 | int ret; |
| 243 | |
| 244 | read_lock(&bdev->vm_lock); |
| 245 | bo = ttm_bo_vm_lookup_rb(bdev, vma->vm_pgoff, |
| 246 | (vma->vm_end - vma->vm_start) >> PAGE_SHIFT); |
| 247 | if (likely(bo != NULL)) |
| 248 | ttm_bo_reference(bo); |
| 249 | read_unlock(&bdev->vm_lock); |
| 250 | |
| 251 | if (unlikely(bo == NULL)) { |
| 252 | printk(KERN_ERR TTM_PFX |
| 253 | "Could not find buffer object to map.\n"); |
| 254 | return -EINVAL; |
| 255 | } |
| 256 | |
| 257 | driver = bo->bdev->driver; |
| 258 | if (unlikely(!driver->verify_access)) { |
| 259 | ret = -EPERM; |
| 260 | goto out_unref; |
| 261 | } |
| 262 | ret = driver->verify_access(bo, filp); |
| 263 | if (unlikely(ret != 0)) |
| 264 | goto out_unref; |
| 265 | |
| 266 | vma->vm_ops = &ttm_bo_vm_ops; |
| 267 | |
| 268 | /* |
| 269 | * Note: We're transferring the bo reference to |
| 270 | * vma->vm_private_data here. |
| 271 | */ |
| 272 | |
| 273 | vma->vm_private_data = bo; |
| 274 | vma->vm_flags |= VM_RESERVED | VM_IO | VM_MIXEDMAP | VM_DONTEXPAND; |
| 275 | return 0; |
| 276 | out_unref: |
| 277 | ttm_bo_unref(&bo); |
| 278 | return ret; |
| 279 | } |
| 280 | EXPORT_SYMBOL(ttm_bo_mmap); |
| 281 | |
| 282 | int ttm_fbdev_mmap(struct vm_area_struct *vma, struct ttm_buffer_object *bo) |
| 283 | { |
| 284 | if (vma->vm_pgoff != 0) |
| 285 | return -EACCES; |
| 286 | |
| 287 | vma->vm_ops = &ttm_bo_vm_ops; |
| 288 | vma->vm_private_data = ttm_bo_reference(bo); |
| 289 | vma->vm_flags |= VM_RESERVED | VM_IO | VM_MIXEDMAP | VM_DONTEXPAND; |
| 290 | return 0; |
| 291 | } |
| 292 | EXPORT_SYMBOL(ttm_fbdev_mmap); |
| 293 | |
| 294 | |
| 295 | ssize_t ttm_bo_io(struct ttm_bo_device *bdev, struct file *filp, |
| 296 | const char __user *wbuf, char __user *rbuf, size_t count, |
| 297 | loff_t *f_pos, bool write) |
| 298 | { |
| 299 | struct ttm_buffer_object *bo; |
| 300 | struct ttm_bo_driver *driver; |
| 301 | struct ttm_bo_kmap_obj map; |
| 302 | unsigned long dev_offset = (*f_pos >> PAGE_SHIFT); |
| 303 | unsigned long kmap_offset; |
| 304 | unsigned long kmap_end; |
| 305 | unsigned long kmap_num; |
| 306 | size_t io_size; |
| 307 | unsigned int page_offset; |
| 308 | char *virtual; |
| 309 | int ret; |
| 310 | bool no_wait = false; |
| 311 | bool dummy; |
| 312 | |
| 313 | read_lock(&bdev->vm_lock); |
| 314 | bo = ttm_bo_vm_lookup_rb(bdev, dev_offset, 1); |
| 315 | if (likely(bo != NULL)) |
| 316 | ttm_bo_reference(bo); |
| 317 | read_unlock(&bdev->vm_lock); |
| 318 | |
| 319 | if (unlikely(bo == NULL)) |
| 320 | return -EFAULT; |
| 321 | |
| 322 | driver = bo->bdev->driver; |
| 323 | if (unlikely(driver->verify_access)) { |
| 324 | ret = -EPERM; |
| 325 | goto out_unref; |
| 326 | } |
| 327 | |
| 328 | ret = driver->verify_access(bo, filp); |
| 329 | if (unlikely(ret != 0)) |
| 330 | goto out_unref; |
| 331 | |
| 332 | kmap_offset = dev_offset - bo->vm_node->start; |
| 333 | if (unlikely(kmap_offset) >= bo->num_pages) { |
| 334 | ret = -EFBIG; |
| 335 | goto out_unref; |
| 336 | } |
| 337 | |
| 338 | page_offset = *f_pos & ~PAGE_MASK; |
| 339 | io_size = bo->num_pages - kmap_offset; |
| 340 | io_size = (io_size << PAGE_SHIFT) - page_offset; |
| 341 | if (count < io_size) |
| 342 | io_size = count; |
| 343 | |
| 344 | kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; |
| 345 | kmap_num = kmap_end - kmap_offset + 1; |
| 346 | |
| 347 | ret = ttm_bo_reserve(bo, true, no_wait, false, 0); |
| 348 | |
| 349 | switch (ret) { |
| 350 | case 0: |
| 351 | break; |
| 352 | case -ERESTART: |
| 353 | ret = -EINTR; |
| 354 | goto out_unref; |
| 355 | case -EBUSY: |
| 356 | ret = -EAGAIN; |
| 357 | goto out_unref; |
| 358 | default: |
| 359 | goto out_unref; |
| 360 | } |
| 361 | |
| 362 | ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); |
| 363 | if (unlikely(ret != 0)) { |
| 364 | ttm_bo_unreserve(bo); |
| 365 | goto out_unref; |
| 366 | } |
| 367 | |
| 368 | virtual = ttm_kmap_obj_virtual(&map, &dummy); |
| 369 | virtual += page_offset; |
| 370 | |
| 371 | if (write) |
| 372 | ret = copy_from_user(virtual, wbuf, io_size); |
| 373 | else |
| 374 | ret = copy_to_user(rbuf, virtual, io_size); |
| 375 | |
| 376 | ttm_bo_kunmap(&map); |
| 377 | ttm_bo_unreserve(bo); |
| 378 | ttm_bo_unref(&bo); |
| 379 | |
| 380 | if (unlikely(ret != 0)) |
| 381 | return -EFBIG; |
| 382 | |
| 383 | *f_pos += io_size; |
| 384 | |
| 385 | return io_size; |
| 386 | out_unref: |
| 387 | ttm_bo_unref(&bo); |
| 388 | return ret; |
| 389 | } |
| 390 | |
| 391 | ssize_t ttm_bo_fbdev_io(struct ttm_buffer_object *bo, const char __user *wbuf, |
| 392 | char __user *rbuf, size_t count, loff_t *f_pos, |
| 393 | bool write) |
| 394 | { |
| 395 | struct ttm_bo_kmap_obj map; |
| 396 | unsigned long kmap_offset; |
| 397 | unsigned long kmap_end; |
| 398 | unsigned long kmap_num; |
| 399 | size_t io_size; |
| 400 | unsigned int page_offset; |
| 401 | char *virtual; |
| 402 | int ret; |
| 403 | bool no_wait = false; |
| 404 | bool dummy; |
| 405 | |
| 406 | kmap_offset = (*f_pos >> PAGE_SHIFT); |
| 407 | if (unlikely(kmap_offset) >= bo->num_pages) |
| 408 | return -EFBIG; |
| 409 | |
| 410 | page_offset = *f_pos & ~PAGE_MASK; |
| 411 | io_size = bo->num_pages - kmap_offset; |
| 412 | io_size = (io_size << PAGE_SHIFT) - page_offset; |
| 413 | if (count < io_size) |
| 414 | io_size = count; |
| 415 | |
| 416 | kmap_end = (*f_pos + count - 1) >> PAGE_SHIFT; |
| 417 | kmap_num = kmap_end - kmap_offset + 1; |
| 418 | |
| 419 | ret = ttm_bo_reserve(bo, true, no_wait, false, 0); |
| 420 | |
| 421 | switch (ret) { |
| 422 | case 0: |
| 423 | break; |
| 424 | case -ERESTART: |
| 425 | return -EINTR; |
| 426 | case -EBUSY: |
| 427 | return -EAGAIN; |
| 428 | default: |
| 429 | return ret; |
| 430 | } |
| 431 | |
| 432 | ret = ttm_bo_kmap(bo, kmap_offset, kmap_num, &map); |
| 433 | if (unlikely(ret != 0)) { |
| 434 | ttm_bo_unreserve(bo); |
| 435 | return ret; |
| 436 | } |
| 437 | |
| 438 | virtual = ttm_kmap_obj_virtual(&map, &dummy); |
| 439 | virtual += page_offset; |
| 440 | |
| 441 | if (write) |
| 442 | ret = copy_from_user(virtual, wbuf, io_size); |
| 443 | else |
| 444 | ret = copy_to_user(rbuf, virtual, io_size); |
| 445 | |
| 446 | ttm_bo_kunmap(&map); |
| 447 | ttm_bo_unreserve(bo); |
| 448 | ttm_bo_unref(&bo); |
| 449 | |
| 450 | if (unlikely(ret != 0)) |
| 451 | return ret; |
| 452 | |
| 453 | *f_pos += io_size; |
| 454 | |
| 455 | return io_size; |
| 456 | } |