1 /* 2 * Framework for buffer objects that can be shared across devices/subsystems. 3 * 4 * Copyright(C) 2011 Linaro Limited. All rights reserved. 5 * Author: Sumit Semwal <sumit.semwal@ti.com> 6 * 7 * Many thanks to linaro-mm-sig list, and specially 8 * Arnd Bergmann <arnd@arndb.de>, Rob Clark <rob@ti.com> and 9 * Daniel Vetter <daniel@ffwll.ch> for their support in creation and 10 * refining of this idea. 11 * 12 * This program is free software; you can redistribute it and/or modify it 13 * under the terms of the GNU General Public License version 2 as published by 14 * the Free Software Foundation. 15 * 16 * This program is distributed in the hope that it will be useful, but WITHOUT 17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 19 * more details. 20 * 21 * You should have received a copy of the GNU General Public License along with 22 * this program. If not, see <http://www.gnu.org/licenses/>. 23 */ 24 25 #include <linux/fs.h> 26 #include <linux/slab.h> 27 #include <linux/dma-buf.h> 28 #include <linux/dma-fence.h> 29 #include <linux/anon_inodes.h> 30 #include <linux/export.h> 31 #include <linux/debugfs.h> 32 #include <linux/module.h> 33 #include <linux/seq_file.h> 34 #include <linux/poll.h> 35 #include <linux/reservation.h> 36 #include <linux/mm.h> 37 38 #include <uapi/linux/dma-buf.h> 39 40 static inline int is_dma_buf_file(struct file *); 41 42 struct dma_buf_list { 43 struct list_head head; 44 struct mutex lock; 45 }; 46 47 static struct dma_buf_list db_list; 48 49 static int dma_buf_release(struct inode *inode, struct file *file) 50 { 51 struct dma_buf *dmabuf; 52 53 if (!is_dma_buf_file(file)) 54 return -EINVAL; 55 56 dmabuf = file->private_data; 57 58 BUG_ON(dmabuf->vmapping_counter); 59 60 /* 61 * Any fences that a dma-buf poll can wait on should be signaled 62 * before releasing dma-buf. This is the responsibility of each 63 * driver that uses the reservation objects. 64 * 65 * If you hit this BUG() it means someone dropped their ref to the 66 * dma-buf while still having pending operation to the buffer. 67 */ 68 BUG_ON(dmabuf->cb_shared.active || dmabuf->cb_excl.active); 69 70 dmabuf->ops->release(dmabuf); 71 72 mutex_lock(&db_list.lock); 73 list_del(&dmabuf->list_node); 74 mutex_unlock(&db_list.lock); 75 76 if (dmabuf->resv == (struct reservation_object *)&dmabuf[1]) 77 reservation_object_fini(dmabuf->resv); 78 79 module_put(dmabuf->owner); 80 kfree(dmabuf); 81 return 0; 82 } 83 84 static int dma_buf_mmap_internal(struct file *file, struct vm_area_struct *vma) 85 { 86 struct dma_buf *dmabuf; 87 88 if (!is_dma_buf_file(file)) 89 return -EINVAL; 90 91 dmabuf = file->private_data; 92 93 /* check for overflowing the buffer's size */ 94 if (vma->vm_pgoff + vma_pages(vma) > 95 dmabuf->size >> PAGE_SHIFT) 96 return -EINVAL; 97 98 return dmabuf->ops->mmap(dmabuf, vma); 99 } 100 101 static loff_t dma_buf_llseek(struct file *file, loff_t offset, int whence) 102 { 103 struct dma_buf *dmabuf; 104 loff_t base; 105 106 if (!is_dma_buf_file(file)) 107 return -EBADF; 108 109 dmabuf = file->private_data; 110 111 /* only support discovering the end of the buffer, 112 but also allow SEEK_SET to maintain the idiomatic 113 SEEK_END(0), SEEK_CUR(0) pattern */ 114 if (whence == SEEK_END) 115 base = dmabuf->size; 116 else if (whence == SEEK_SET) 117 base = 0; 118 else 119 return -EINVAL; 120 121 if (offset != 0) 122 return -EINVAL; 123 124 return base + offset; 125 } 126 127 /** 128 * DOC: fence polling 129 * 130 * To support cross-device and cross-driver synchronization of buffer access 131 * implicit fences (represented internally in the kernel with &struct fence) can 132 * be attached to a &dma_buf. The glue for that and a few related things are 133 * provided in the &reservation_object structure. 134 * 135 * Userspace can query the state of these implicitly tracked fences using poll() 136 * and related system calls: 137 * 138 * - Checking for EPOLLIN, i.e. read access, can be use to query the state of the 139 * most recent write or exclusive fence. 140 * 141 * - Checking for EPOLLOUT, i.e. write access, can be used to query the state of 142 * all attached fences, shared and exclusive ones. 143 * 144 * Note that this only signals the completion of the respective fences, i.e. the 145 * DMA transfers are complete. Cache flushing and any other necessary 146 * preparations before CPU access can begin still need to happen. 147 */ 148 149 static void dma_buf_poll_cb(struct dma_fence *fence, struct dma_fence_cb *cb) 150 { 151 struct dma_buf_poll_cb_t *dcb = (struct dma_buf_poll_cb_t *)cb; 152 unsigned long flags; 153 154 spin_lock_irqsave(&dcb->poll->lock, flags); 155 wake_up_locked_poll(dcb->poll, dcb->active); 156 dcb->active = 0; 157 spin_unlock_irqrestore(&dcb->poll->lock, flags); 158 } 159 160 static __poll_t dma_buf_poll(struct file *file, poll_table *poll) 161 { 162 struct dma_buf *dmabuf; 163 struct reservation_object *resv; 164 struct reservation_object_list *fobj; 165 struct dma_fence *fence_excl; 166 __poll_t events; 167 unsigned shared_count, seq; 168 169 dmabuf = file->private_data; 170 if (!dmabuf || !dmabuf->resv) 171 return EPOLLERR; 172 173 resv = dmabuf->resv; 174 175 poll_wait(file, &dmabuf->poll, poll); 176 177 events = poll_requested_events(poll) & (EPOLLIN | EPOLLOUT); 178 if (!events) 179 return 0; 180 181 retry: 182 seq = read_seqcount_begin(&resv->seq); 183 rcu_read_lock(); 184 185 fobj = rcu_dereference(resv->fence); 186 if (fobj) 187 shared_count = fobj->shared_count; 188 else 189 shared_count = 0; 190 fence_excl = rcu_dereference(resv->fence_excl); 191 if (read_seqcount_retry(&resv->seq, seq)) { 192 rcu_read_unlock(); 193 goto retry; 194 } 195 196 if (fence_excl && (!(events & EPOLLOUT) || shared_count == 0)) { 197 struct dma_buf_poll_cb_t *dcb = &dmabuf->cb_excl; 198 __poll_t pevents = EPOLLIN; 199 200 if (shared_count == 0) 201 pevents |= EPOLLOUT; 202 203 spin_lock_irq(&dmabuf->poll.lock); 204 if (dcb->active) { 205 dcb->active |= pevents; 206 events &= ~pevents; 207 } else 208 dcb->active = pevents; 209 spin_unlock_irq(&dmabuf->poll.lock); 210 211 if (events & pevents) { 212 if (!dma_fence_get_rcu(fence_excl)) { 213 /* force a recheck */ 214 events &= ~pevents; 215 dma_buf_poll_cb(NULL, &dcb->cb); 216 } else if (!dma_fence_add_callback(fence_excl, &dcb->cb, 217 dma_buf_poll_cb)) { 218 events &= ~pevents; 219 dma_fence_put(fence_excl); 220 } else { 221 /* 222 * No callback queued, wake up any additional 223 * waiters. 224 */ 225 dma_fence_put(fence_excl); 226 dma_buf_poll_cb(NULL, &dcb->cb); 227 } 228 } 229 } 230 231 if ((events & EPOLLOUT) && shared_count > 0) { 232 struct dma_buf_poll_cb_t *dcb = &dmabuf->cb_shared; 233 int i; 234 235 /* Only queue a new callback if no event has fired yet */ 236 spin_lock_irq(&dmabuf->poll.lock); 237 if (dcb->active) 238 events &= ~EPOLLOUT; 239 else 240 dcb->active = EPOLLOUT; 241 spin_unlock_irq(&dmabuf->poll.lock); 242 243 if (!(events & EPOLLOUT)) 244 goto out; 245 246 for (i = 0; i < shared_count; ++i) { 247 struct dma_fence *fence = rcu_dereference(fobj->shared[i]); 248 249 if (!dma_fence_get_rcu(fence)) { 250 /* 251 * fence refcount dropped to zero, this means 252 * that fobj has been freed 253 * 254 * call dma_buf_poll_cb and force a recheck! 255 */ 256 events &= ~EPOLLOUT; 257 dma_buf_poll_cb(NULL, &dcb->cb); 258 break; 259 } 260 if (!dma_fence_add_callback(fence, &dcb->cb, 261 dma_buf_poll_cb)) { 262 dma_fence_put(fence); 263 events &= ~EPOLLOUT; 264 break; 265 } 266 dma_fence_put(fence); 267 } 268 269 /* No callback queued, wake up any additional waiters. */ 270 if (i == shared_count) 271 dma_buf_poll_cb(NULL, &dcb->cb); 272 } 273 274 out: 275 rcu_read_unlock(); 276 return events; 277 } 278 279 static long dma_buf_ioctl(struct file *file, 280 unsigned int cmd, unsigned long arg) 281 { 282 struct dma_buf *dmabuf; 283 struct dma_buf_sync sync; 284 enum dma_data_direction direction; 285 int ret; 286 287 dmabuf = file->private_data; 288 289 switch (cmd) { 290 case DMA_BUF_IOCTL_SYNC: 291 if (copy_from_user(&sync, (void __user *) arg, sizeof(sync))) 292 return -EFAULT; 293 294 if (sync.flags & ~DMA_BUF_SYNC_VALID_FLAGS_MASK) 295 return -EINVAL; 296 297 switch (sync.flags & DMA_BUF_SYNC_RW) { 298 case DMA_BUF_SYNC_READ: 299 direction = DMA_FROM_DEVICE; 300 break; 301 case DMA_BUF_SYNC_WRITE: 302 direction = DMA_TO_DEVICE; 303 break; 304 case DMA_BUF_SYNC_RW: 305 direction = DMA_BIDIRECTIONAL; 306 break; 307 default: 308 return -EINVAL; 309 } 310 311 if (sync.flags & DMA_BUF_SYNC_END) 312 ret = dma_buf_end_cpu_access(dmabuf, direction); 313 else 314 ret = dma_buf_begin_cpu_access(dmabuf, direction); 315 316 return ret; 317 default: 318 return -ENOTTY; 319 } 320 } 321 322 static const struct file_operations dma_buf_fops = { 323 .release = dma_buf_release, 324 .mmap = dma_buf_mmap_internal, 325 .llseek = dma_buf_llseek, 326 .poll = dma_buf_poll, 327 .unlocked_ioctl = dma_buf_ioctl, 328 #ifdef CONFIG_COMPAT 329 .compat_ioctl = dma_buf_ioctl, 330 #endif 331 }; 332 333 /* 334 * is_dma_buf_file - Check if struct file* is associated with dma_buf 335 */ 336 static inline int is_dma_buf_file(struct file *file) 337 { 338 return file->f_op == &dma_buf_fops; 339 } 340 341 /** 342 * DOC: dma buf device access 343 * 344 * For device DMA access to a shared DMA buffer the usual sequence of operations 345 * is fairly simple: 346 * 347 * 1. The exporter defines his exporter instance using 348 * DEFINE_DMA_BUF_EXPORT_INFO() and calls dma_buf_export() to wrap a private 349 * buffer object into a &dma_buf. It then exports that &dma_buf to userspace 350 * as a file descriptor by calling dma_buf_fd(). 351 * 352 * 2. Userspace passes this file-descriptors to all drivers it wants this buffer 353 * to share with: First the filedescriptor is converted to a &dma_buf using 354 * dma_buf_get(). Then the buffer is attached to the device using 355 * dma_buf_attach(). 356 * 357 * Up to this stage the exporter is still free to migrate or reallocate the 358 * backing storage. 359 * 360 * 3. Once the buffer is attached to all devices userspace can initiate DMA 361 * access to the shared buffer. In the kernel this is done by calling 362 * dma_buf_map_attachment() and dma_buf_unmap_attachment(). 363 * 364 * 4. Once a driver is done with a shared buffer it needs to call 365 * dma_buf_detach() (after cleaning up any mappings) and then release the 366 * reference acquired with dma_buf_get by calling dma_buf_put(). 367 * 368 * For the detailed semantics exporters are expected to implement see 369 * &dma_buf_ops. 370 */ 371 372 /** 373 * dma_buf_export - Creates a new dma_buf, and associates an anon file 374 * with this buffer, so it can be exported. 375 * Also connect the allocator specific data and ops to the buffer. 376 * Additionally, provide a name string for exporter; useful in debugging. 377 * 378 * @exp_info: [in] holds all the export related information provided 379 * by the exporter. see &struct dma_buf_export_info 380 * for further details. 381 * 382 * Returns, on success, a newly created dma_buf object, which wraps the 383 * supplied private data and operations for dma_buf_ops. On either missing 384 * ops, or error in allocating struct dma_buf, will return negative error. 385 * 386 * For most cases the easiest way to create @exp_info is through the 387 * %DEFINE_DMA_BUF_EXPORT_INFO macro. 388 */ 389 struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info) 390 { 391 struct dma_buf *dmabuf; 392 struct reservation_object *resv = exp_info->resv; 393 struct file *file; 394 size_t alloc_size = sizeof(struct dma_buf); 395 int ret; 396 397 if (!exp_info->resv) 398 alloc_size += sizeof(struct reservation_object); 399 else 400 /* prevent &dma_buf[1] == dma_buf->resv */ 401 alloc_size += 1; 402 403 if (WARN_ON(!exp_info->priv 404 || !exp_info->ops 405 || !exp_info->ops->map_dma_buf 406 || !exp_info->ops->unmap_dma_buf 407 || !exp_info->ops->release 408 || !exp_info->ops->mmap)) { 409 return ERR_PTR(-EINVAL); 410 } 411 412 if (!try_module_get(exp_info->owner)) 413 return ERR_PTR(-ENOENT); 414 415 dmabuf = kzalloc(alloc_size, GFP_KERNEL); 416 if (!dmabuf) { 417 ret = -ENOMEM; 418 goto err_module; 419 } 420 421 dmabuf->priv = exp_info->priv; 422 dmabuf->ops = exp_info->ops; 423 dmabuf->size = exp_info->size; 424 dmabuf->exp_name = exp_info->exp_name; 425 dmabuf->owner = exp_info->owner; 426 init_waitqueue_head(&dmabuf->poll); 427 dmabuf->cb_excl.poll = dmabuf->cb_shared.poll = &dmabuf->poll; 428 dmabuf->cb_excl.active = dmabuf->cb_shared.active = 0; 429 430 if (!resv) { 431 resv = (struct reservation_object *)&dmabuf[1]; 432 reservation_object_init(resv); 433 } 434 dmabuf->resv = resv; 435 436 file = anon_inode_getfile("dmabuf", &dma_buf_fops, dmabuf, 437 exp_info->flags); 438 if (IS_ERR(file)) { 439 ret = PTR_ERR(file); 440 goto err_dmabuf; 441 } 442 443 file->f_mode |= FMODE_LSEEK; 444 dmabuf->file = file; 445 446 mutex_init(&dmabuf->lock); 447 INIT_LIST_HEAD(&dmabuf->attachments); 448 449 mutex_lock(&db_list.lock); 450 list_add(&dmabuf->list_node, &db_list.head); 451 mutex_unlock(&db_list.lock); 452 453 return dmabuf; 454 455 err_dmabuf: 456 kfree(dmabuf); 457 err_module: 458 module_put(exp_info->owner); 459 return ERR_PTR(ret); 460 } 461 EXPORT_SYMBOL_GPL(dma_buf_export); 462 463 /** 464 * dma_buf_fd - returns a file descriptor for the given dma_buf 465 * @dmabuf: [in] pointer to dma_buf for which fd is required. 466 * @flags: [in] flags to give to fd 467 * 468 * On success, returns an associated 'fd'. Else, returns error. 469 */ 470 int dma_buf_fd(struct dma_buf *dmabuf, int flags) 471 { 472 int fd; 473 474 if (!dmabuf || !dmabuf->file) 475 return -EINVAL; 476 477 fd = get_unused_fd_flags(flags); 478 if (fd < 0) 479 return fd; 480 481 fd_install(fd, dmabuf->file); 482 483 return fd; 484 } 485 EXPORT_SYMBOL_GPL(dma_buf_fd); 486 487 /** 488 * dma_buf_get - returns the dma_buf structure related to an fd 489 * @fd: [in] fd associated with the dma_buf to be returned 490 * 491 * On success, returns the dma_buf structure associated with an fd; uses 492 * file's refcounting done by fget to increase refcount. returns ERR_PTR 493 * otherwise. 494 */ 495 struct dma_buf *dma_buf_get(int fd) 496 { 497 struct file *file; 498 499 file = fget(fd); 500 501 if (!file) 502 return ERR_PTR(-EBADF); 503 504 if (!is_dma_buf_file(file)) { 505 fput(file); 506 return ERR_PTR(-EINVAL); 507 } 508 509 return file->private_data; 510 } 511 EXPORT_SYMBOL_GPL(dma_buf_get); 512 513 /** 514 * dma_buf_put - decreases refcount of the buffer 515 * @dmabuf: [in] buffer to reduce refcount of 516 * 517 * Uses file's refcounting done implicitly by fput(). 518 * 519 * If, as a result of this call, the refcount becomes 0, the 'release' file 520 * operation related to this fd is called. It calls &dma_buf_ops.release vfunc 521 * in turn, and frees the memory allocated for dmabuf when exported. 522 */ 523 void dma_buf_put(struct dma_buf *dmabuf) 524 { 525 if (WARN_ON(!dmabuf || !dmabuf->file)) 526 return; 527 528 fput(dmabuf->file); 529 } 530 EXPORT_SYMBOL_GPL(dma_buf_put); 531 532 /** 533 * dma_buf_attach - Add the device to dma_buf's attachments list; optionally, 534 * calls attach() of dma_buf_ops to allow device-specific attach functionality 535 * @dmabuf: [in] buffer to attach device to. 536 * @dev: [in] device to be attached. 537 * 538 * Returns struct dma_buf_attachment pointer for this attachment. Attachments 539 * must be cleaned up by calling dma_buf_detach(). 540 * 541 * Returns: 542 * 543 * A pointer to newly created &dma_buf_attachment on success, or a negative 544 * error code wrapped into a pointer on failure. 545 * 546 * Note that this can fail if the backing storage of @dmabuf is in a place not 547 * accessible to @dev, and cannot be moved to a more suitable place. This is 548 * indicated with the error code -EBUSY. 549 */ 550 struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf, 551 struct device *dev) 552 { 553 struct dma_buf_attachment *attach; 554 int ret; 555 556 if (WARN_ON(!dmabuf || !dev)) 557 return ERR_PTR(-EINVAL); 558 559 attach = kzalloc(sizeof(*attach), GFP_KERNEL); 560 if (!attach) 561 return ERR_PTR(-ENOMEM); 562 563 attach->dev = dev; 564 attach->dmabuf = dmabuf; 565 566 mutex_lock(&dmabuf->lock); 567 568 if (dmabuf->ops->attach) { 569 ret = dmabuf->ops->attach(dmabuf, attach); 570 if (ret) 571 goto err_attach; 572 } 573 list_add(&attach->node, &dmabuf->attachments); 574 575 mutex_unlock(&dmabuf->lock); 576 return attach; 577 578 err_attach: 579 kfree(attach); 580 mutex_unlock(&dmabuf->lock); 581 return ERR_PTR(ret); 582 } 583 EXPORT_SYMBOL_GPL(dma_buf_attach); 584 585 /** 586 * dma_buf_detach - Remove the given attachment from dmabuf's attachments list; 587 * optionally calls detach() of dma_buf_ops for device-specific detach 588 * @dmabuf: [in] buffer to detach from. 589 * @attach: [in] attachment to be detached; is free'd after this call. 590 * 591 * Clean up a device attachment obtained by calling dma_buf_attach(). 592 */ 593 void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach) 594 { 595 if (WARN_ON(!dmabuf || !attach)) 596 return; 597 598 mutex_lock(&dmabuf->lock); 599 list_del(&attach->node); 600 if (dmabuf->ops->detach) 601 dmabuf->ops->detach(dmabuf, attach); 602 603 mutex_unlock(&dmabuf->lock); 604 kfree(attach); 605 } 606 EXPORT_SYMBOL_GPL(dma_buf_detach); 607 608 /** 609 * dma_buf_map_attachment - Returns the scatterlist table of the attachment; 610 * mapped into _device_ address space. Is a wrapper for map_dma_buf() of the 611 * dma_buf_ops. 612 * @attach: [in] attachment whose scatterlist is to be returned 613 * @direction: [in] direction of DMA transfer 614 * 615 * Returns sg_table containing the scatterlist to be returned; returns ERR_PTR 616 * on error. May return -EINTR if it is interrupted by a signal. 617 * 618 * A mapping must be unmapped by using dma_buf_unmap_attachment(). Note that 619 * the underlying backing storage is pinned for as long as a mapping exists, 620 * therefore users/importers should not hold onto a mapping for undue amounts of 621 * time. 622 */ 623 struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach, 624 enum dma_data_direction direction) 625 { 626 struct sg_table *sg_table; 627 628 might_sleep(); 629 630 if (WARN_ON(!attach || !attach->dmabuf)) 631 return ERR_PTR(-EINVAL); 632 633 sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction); 634 if (!sg_table) 635 sg_table = ERR_PTR(-ENOMEM); 636 637 return sg_table; 638 } 639 EXPORT_SYMBOL_GPL(dma_buf_map_attachment); 640 641 /** 642 * dma_buf_unmap_attachment - unmaps and decreases usecount of the buffer;might 643 * deallocate the scatterlist associated. Is a wrapper for unmap_dma_buf() of 644 * dma_buf_ops. 645 * @attach: [in] attachment to unmap buffer from 646 * @sg_table: [in] scatterlist info of the buffer to unmap 647 * @direction: [in] direction of DMA transfer 648 * 649 * This unmaps a DMA mapping for @attached obtained by dma_buf_map_attachment(). 650 */ 651 void dma_buf_unmap_attachment(struct dma_buf_attachment *attach, 652 struct sg_table *sg_table, 653 enum dma_data_direction direction) 654 { 655 might_sleep(); 656 657 if (WARN_ON(!attach || !attach->dmabuf || !sg_table)) 658 return; 659 660 attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, 661 direction); 662 } 663 EXPORT_SYMBOL_GPL(dma_buf_unmap_attachment); 664 665 /** 666 * DOC: cpu access 667 * 668 * There are mutliple reasons for supporting CPU access to a dma buffer object: 669 * 670 * - Fallback operations in the kernel, for example when a device is connected 671 * over USB and the kernel needs to shuffle the data around first before 672 * sending it away. Cache coherency is handled by braketing any transactions 673 * with calls to dma_buf_begin_cpu_access() and dma_buf_end_cpu_access() 674 * access. 675 * 676 * To support dma_buf objects residing in highmem cpu access is page-based 677 * using an api similar to kmap. Accessing a dma_buf is done in aligned chunks 678 * of PAGE_SIZE size. Before accessing a chunk it needs to be mapped, which 679 * returns a pointer in kernel virtual address space. Afterwards the chunk 680 * needs to be unmapped again. There is no limit on how often a given chunk 681 * can be mapped and unmapped, i.e. the importer does not need to call 682 * begin_cpu_access again before mapping the same chunk again. 683 * 684 * Interfaces:: 685 * void \*dma_buf_kmap(struct dma_buf \*, unsigned long); 686 * void dma_buf_kunmap(struct dma_buf \*, unsigned long, void \*); 687 * 688 * Implementing the functions is optional for exporters and for importers all 689 * the restrictions of using kmap apply. 690 * 691 * dma_buf kmap calls outside of the range specified in begin_cpu_access are 692 * undefined. If the range is not PAGE_SIZE aligned, kmap needs to succeed on 693 * the partial chunks at the beginning and end but may return stale or bogus 694 * data outside of the range (in these partial chunks). 695 * 696 * For some cases the overhead of kmap can be too high, a vmap interface 697 * is introduced. This interface should be used very carefully, as vmalloc 698 * space is a limited resources on many architectures. 699 * 700 * Interfaces:: 701 * void \*dma_buf_vmap(struct dma_buf \*dmabuf) 702 * void dma_buf_vunmap(struct dma_buf \*dmabuf, void \*vaddr) 703 * 704 * The vmap call can fail if there is no vmap support in the exporter, or if 705 * it runs out of vmalloc space. Fallback to kmap should be implemented. Note 706 * that the dma-buf layer keeps a reference count for all vmap access and 707 * calls down into the exporter's vmap function only when no vmapping exists, 708 * and only unmaps it once. Protection against concurrent vmap/vunmap calls is 709 * provided by taking the dma_buf->lock mutex. 710 * 711 * - For full compatibility on the importer side with existing userspace 712 * interfaces, which might already support mmap'ing buffers. This is needed in 713 * many processing pipelines (e.g. feeding a software rendered image into a 714 * hardware pipeline, thumbnail creation, snapshots, ...). Also, Android's ION 715 * framework already supported this and for DMA buffer file descriptors to 716 * replace ION buffers mmap support was needed. 717 * 718 * There is no special interfaces, userspace simply calls mmap on the dma-buf 719 * fd. But like for CPU access there's a need to braket the actual access, 720 * which is handled by the ioctl (DMA_BUF_IOCTL_SYNC). Note that 721 * DMA_BUF_IOCTL_SYNC can fail with -EAGAIN or -EINTR, in which case it must 722 * be restarted. 723 * 724 * Some systems might need some sort of cache coherency management e.g. when 725 * CPU and GPU domains are being accessed through dma-buf at the same time. 726 * To circumvent this problem there are begin/end coherency markers, that 727 * forward directly to existing dma-buf device drivers vfunc hooks. Userspace 728 * can make use of those markers through the DMA_BUF_IOCTL_SYNC ioctl. The 729 * sequence would be used like following: 730 * 731 * - mmap dma-buf fd 732 * - for each drawing/upload cycle in CPU 1. SYNC_START ioctl, 2. read/write 733 * to mmap area 3. SYNC_END ioctl. This can be repeated as often as you 734 * want (with the new data being consumed by say the GPU or the scanout 735 * device) 736 * - munmap once you don't need the buffer any more 737 * 738 * For correctness and optimal performance, it is always required to use 739 * SYNC_START and SYNC_END before and after, respectively, when accessing the 740 * mapped address. Userspace cannot rely on coherent access, even when there 741 * are systems where it just works without calling these ioctls. 742 * 743 * - And as a CPU fallback in userspace processing pipelines. 744 * 745 * Similar to the motivation for kernel cpu access it is again important that 746 * the userspace code of a given importing subsystem can use the same 747 * interfaces with a imported dma-buf buffer object as with a native buffer 748 * object. This is especially important for drm where the userspace part of 749 * contemporary OpenGL, X, and other drivers is huge, and reworking them to 750 * use a different way to mmap a buffer rather invasive. 751 * 752 * The assumption in the current dma-buf interfaces is that redirecting the 753 * initial mmap is all that's needed. A survey of some of the existing 754 * subsystems shows that no driver seems to do any nefarious thing like 755 * syncing up with outstanding asynchronous processing on the device or 756 * allocating special resources at fault time. So hopefully this is good 757 * enough, since adding interfaces to intercept pagefaults and allow pte 758 * shootdowns would increase the complexity quite a bit. 759 * 760 * Interface:: 761 * int dma_buf_mmap(struct dma_buf \*, struct vm_area_struct \*, 762 * unsigned long); 763 * 764 * If the importing subsystem simply provides a special-purpose mmap call to 765 * set up a mapping in userspace, calling do_mmap with dma_buf->file will 766 * equally achieve that for a dma-buf object. 767 */ 768 769 static int __dma_buf_begin_cpu_access(struct dma_buf *dmabuf, 770 enum dma_data_direction direction) 771 { 772 bool write = (direction == DMA_BIDIRECTIONAL || 773 direction == DMA_TO_DEVICE); 774 struct reservation_object *resv = dmabuf->resv; 775 long ret; 776 777 /* Wait on any implicit rendering fences */ 778 ret = reservation_object_wait_timeout_rcu(resv, write, true, 779 MAX_SCHEDULE_TIMEOUT); 780 if (ret < 0) 781 return ret; 782 783 return 0; 784 } 785 786 /** 787 * dma_buf_begin_cpu_access - Must be called before accessing a dma_buf from the 788 * cpu in the kernel context. Calls begin_cpu_access to allow exporter-specific 789 * preparations. Coherency is only guaranteed in the specified range for the 790 * specified access direction. 791 * @dmabuf: [in] buffer to prepare cpu access for. 792 * @direction: [in] length of range for cpu access. 793 * 794 * After the cpu access is complete the caller should call 795 * dma_buf_end_cpu_access(). Only when cpu access is braketed by both calls is 796 * it guaranteed to be coherent with other DMA access. 797 * 798 * Can return negative error values, returns 0 on success. 799 */ 800 int dma_buf_begin_cpu_access(struct dma_buf *dmabuf, 801 enum dma_data_direction direction) 802 { 803 int ret = 0; 804 805 if (WARN_ON(!dmabuf)) 806 return -EINVAL; 807 808 if (dmabuf->ops->begin_cpu_access) 809 ret = dmabuf->ops->begin_cpu_access(dmabuf, direction); 810 811 /* Ensure that all fences are waited upon - but we first allow 812 * the native handler the chance to do so more efficiently if it 813 * chooses. A double invocation here will be reasonably cheap no-op. 814 */ 815 if (ret == 0) 816 ret = __dma_buf_begin_cpu_access(dmabuf, direction); 817 818 return ret; 819 } 820 EXPORT_SYMBOL_GPL(dma_buf_begin_cpu_access); 821 822 /** 823 * dma_buf_end_cpu_access - Must be called after accessing a dma_buf from the 824 * cpu in the kernel context. Calls end_cpu_access to allow exporter-specific 825 * actions. Coherency is only guaranteed in the specified range for the 826 * specified access direction. 827 * @dmabuf: [in] buffer to complete cpu access for. 828 * @direction: [in] length of range for cpu access. 829 * 830 * This terminates CPU access started with dma_buf_begin_cpu_access(). 831 * 832 * Can return negative error values, returns 0 on success. 833 */ 834 int dma_buf_end_cpu_access(struct dma_buf *dmabuf, 835 enum dma_data_direction direction) 836 { 837 int ret = 0; 838 839 WARN_ON(!dmabuf); 840 841 if (dmabuf->ops->end_cpu_access) 842 ret = dmabuf->ops->end_cpu_access(dmabuf, direction); 843 844 return ret; 845 } 846 EXPORT_SYMBOL_GPL(dma_buf_end_cpu_access); 847 848 /** 849 * dma_buf_kmap - Map a page of the buffer object into kernel address space. The 850 * same restrictions as for kmap and friends apply. 851 * @dmabuf: [in] buffer to map page from. 852 * @page_num: [in] page in PAGE_SIZE units to map. 853 * 854 * This call must always succeed, any necessary preparations that might fail 855 * need to be done in begin_cpu_access. 856 */ 857 void *dma_buf_kmap(struct dma_buf *dmabuf, unsigned long page_num) 858 { 859 WARN_ON(!dmabuf); 860 861 if (!dmabuf->ops->map) 862 return NULL; 863 return dmabuf->ops->map(dmabuf, page_num); 864 } 865 EXPORT_SYMBOL_GPL(dma_buf_kmap); 866 867 /** 868 * dma_buf_kunmap - Unmap a page obtained by dma_buf_kmap. 869 * @dmabuf: [in] buffer to unmap page from. 870 * @page_num: [in] page in PAGE_SIZE units to unmap. 871 * @vaddr: [in] kernel space pointer obtained from dma_buf_kmap. 872 * 873 * This call must always succeed. 874 */ 875 void dma_buf_kunmap(struct dma_buf *dmabuf, unsigned long page_num, 876 void *vaddr) 877 { 878 WARN_ON(!dmabuf); 879 880 if (dmabuf->ops->unmap) 881 dmabuf->ops->unmap(dmabuf, page_num, vaddr); 882 } 883 EXPORT_SYMBOL_GPL(dma_buf_kunmap); 884 885 886 /** 887 * dma_buf_mmap - Setup up a userspace mmap with the given vma 888 * @dmabuf: [in] buffer that should back the vma 889 * @vma: [in] vma for the mmap 890 * @pgoff: [in] offset in pages where this mmap should start within the 891 * dma-buf buffer. 892 * 893 * This function adjusts the passed in vma so that it points at the file of the 894 * dma_buf operation. It also adjusts the starting pgoff and does bounds 895 * checking on the size of the vma. Then it calls the exporters mmap function to 896 * set up the mapping. 897 * 898 * Can return negative error values, returns 0 on success. 899 */ 900 int dma_buf_mmap(struct dma_buf *dmabuf, struct vm_area_struct *vma, 901 unsigned long pgoff) 902 { 903 struct file *oldfile; 904 int ret; 905 906 if (WARN_ON(!dmabuf || !vma)) 907 return -EINVAL; 908 909 /* check for offset overflow */ 910 if (pgoff + vma_pages(vma) < pgoff) 911 return -EOVERFLOW; 912 913 /* check for overflowing the buffer's size */ 914 if (pgoff + vma_pages(vma) > 915 dmabuf->size >> PAGE_SHIFT) 916 return -EINVAL; 917 918 /* readjust the vma */ 919 get_file(dmabuf->file); 920 oldfile = vma->vm_file; 921 vma->vm_file = dmabuf->file; 922 vma->vm_pgoff = pgoff; 923 924 ret = dmabuf->ops->mmap(dmabuf, vma); 925 if (ret) { 926 /* restore old parameters on failure */ 927 vma->vm_file = oldfile; 928 fput(dmabuf->file); 929 } else { 930 if (oldfile) 931 fput(oldfile); 932 } 933 return ret; 934 935 } 936 EXPORT_SYMBOL_GPL(dma_buf_mmap); 937 938 /** 939 * dma_buf_vmap - Create virtual mapping for the buffer object into kernel 940 * address space. Same restrictions as for vmap and friends apply. 941 * @dmabuf: [in] buffer to vmap 942 * 943 * This call may fail due to lack of virtual mapping address space. 944 * These calls are optional in drivers. The intended use for them 945 * is for mapping objects linear in kernel space for high use objects. 946 * Please attempt to use kmap/kunmap before thinking about these interfaces. 947 * 948 * Returns NULL on error. 949 */ 950 void *dma_buf_vmap(struct dma_buf *dmabuf) 951 { 952 void *ptr; 953 954 if (WARN_ON(!dmabuf)) 955 return NULL; 956 957 if (!dmabuf->ops->vmap) 958 return NULL; 959 960 mutex_lock(&dmabuf->lock); 961 if (dmabuf->vmapping_counter) { 962 dmabuf->vmapping_counter++; 963 BUG_ON(!dmabuf->vmap_ptr); 964 ptr = dmabuf->vmap_ptr; 965 goto out_unlock; 966 } 967 968 BUG_ON(dmabuf->vmap_ptr); 969 970 ptr = dmabuf->ops->vmap(dmabuf); 971 if (WARN_ON_ONCE(IS_ERR(ptr))) 972 ptr = NULL; 973 if (!ptr) 974 goto out_unlock; 975 976 dmabuf->vmap_ptr = ptr; 977 dmabuf->vmapping_counter = 1; 978 979 out_unlock: 980 mutex_unlock(&dmabuf->lock); 981 return ptr; 982 } 983 EXPORT_SYMBOL_GPL(dma_buf_vmap); 984 985 /** 986 * dma_buf_vunmap - Unmap a vmap obtained by dma_buf_vmap. 987 * @dmabuf: [in] buffer to vunmap 988 * @vaddr: [in] vmap to vunmap 989 */ 990 void dma_buf_vunmap(struct dma_buf *dmabuf, void *vaddr) 991 { 992 if (WARN_ON(!dmabuf)) 993 return; 994 995 BUG_ON(!dmabuf->vmap_ptr); 996 BUG_ON(dmabuf->vmapping_counter == 0); 997 BUG_ON(dmabuf->vmap_ptr != vaddr); 998 999 mutex_lock(&dmabuf->lock); 1000 if (--dmabuf->vmapping_counter == 0) { 1001 if (dmabuf->ops->vunmap) 1002 dmabuf->ops->vunmap(dmabuf, vaddr); 1003 dmabuf->vmap_ptr = NULL; 1004 } 1005 mutex_unlock(&dmabuf->lock); 1006 } 1007 EXPORT_SYMBOL_GPL(dma_buf_vunmap); 1008 1009 #ifdef CONFIG_DEBUG_FS 1010 static int dma_buf_debug_show(struct seq_file *s, void *unused) 1011 { 1012 int ret; 1013 struct dma_buf *buf_obj; 1014 struct dma_buf_attachment *attach_obj; 1015 struct reservation_object *robj; 1016 struct reservation_object_list *fobj; 1017 struct dma_fence *fence; 1018 unsigned seq; 1019 int count = 0, attach_count, shared_count, i; 1020 size_t size = 0; 1021 1022 ret = mutex_lock_interruptible(&db_list.lock); 1023 1024 if (ret) 1025 return ret; 1026 1027 seq_puts(s, "\nDma-buf Objects:\n"); 1028 seq_printf(s, "%-8s\t%-8s\t%-8s\t%-8s\texp_name\n", 1029 "size", "flags", "mode", "count"); 1030 1031 list_for_each_entry(buf_obj, &db_list.head, list_node) { 1032 ret = mutex_lock_interruptible(&buf_obj->lock); 1033 1034 if (ret) { 1035 seq_puts(s, 1036 "\tERROR locking buffer object: skipping\n"); 1037 continue; 1038 } 1039 1040 seq_printf(s, "%08zu\t%08x\t%08x\t%08ld\t%s\n", 1041 buf_obj->size, 1042 buf_obj->file->f_flags, buf_obj->file->f_mode, 1043 file_count(buf_obj->file), 1044 buf_obj->exp_name); 1045 1046 robj = buf_obj->resv; 1047 while (true) { 1048 seq = read_seqcount_begin(&robj->seq); 1049 rcu_read_lock(); 1050 fobj = rcu_dereference(robj->fence); 1051 shared_count = fobj ? fobj->shared_count : 0; 1052 fence = rcu_dereference(robj->fence_excl); 1053 if (!read_seqcount_retry(&robj->seq, seq)) 1054 break; 1055 rcu_read_unlock(); 1056 } 1057 1058 if (fence) 1059 seq_printf(s, "\tExclusive fence: %s %s %ssignalled\n", 1060 fence->ops->get_driver_name(fence), 1061 fence->ops->get_timeline_name(fence), 1062 dma_fence_is_signaled(fence) ? "" : "un"); 1063 for (i = 0; i < shared_count; i++) { 1064 fence = rcu_dereference(fobj->shared[i]); 1065 if (!dma_fence_get_rcu(fence)) 1066 continue; 1067 seq_printf(s, "\tShared fence: %s %s %ssignalled\n", 1068 fence->ops->get_driver_name(fence), 1069 fence->ops->get_timeline_name(fence), 1070 dma_fence_is_signaled(fence) ? "" : "un"); 1071 } 1072 rcu_read_unlock(); 1073 1074 seq_puts(s, "\tAttached Devices:\n"); 1075 attach_count = 0; 1076 1077 list_for_each_entry(attach_obj, &buf_obj->attachments, node) { 1078 seq_printf(s, "\t%s\n", dev_name(attach_obj->dev)); 1079 attach_count++; 1080 } 1081 1082 seq_printf(s, "Total %d devices attached\n\n", 1083 attach_count); 1084 1085 count++; 1086 size += buf_obj->size; 1087 mutex_unlock(&buf_obj->lock); 1088 } 1089 1090 seq_printf(s, "\nTotal %d objects, %zu bytes\n", count, size); 1091 1092 mutex_unlock(&db_list.lock); 1093 return 0; 1094 } 1095 1096 static int dma_buf_debug_open(struct inode *inode, struct file *file) 1097 { 1098 return single_open(file, dma_buf_debug_show, NULL); 1099 } 1100 1101 static const struct file_operations dma_buf_debug_fops = { 1102 .open = dma_buf_debug_open, 1103 .read = seq_read, 1104 .llseek = seq_lseek, 1105 .release = single_release, 1106 }; 1107 1108 static struct dentry *dma_buf_debugfs_dir; 1109 1110 static int dma_buf_init_debugfs(void) 1111 { 1112 struct dentry *d; 1113 int err = 0; 1114 1115 d = debugfs_create_dir("dma_buf", NULL); 1116 if (IS_ERR(d)) 1117 return PTR_ERR(d); 1118 1119 dma_buf_debugfs_dir = d; 1120 1121 d = debugfs_create_file("bufinfo", S_IRUGO, dma_buf_debugfs_dir, 1122 NULL, &dma_buf_debug_fops); 1123 if (IS_ERR(d)) { 1124 pr_debug("dma_buf: debugfs: failed to create node bufinfo\n"); 1125 debugfs_remove_recursive(dma_buf_debugfs_dir); 1126 dma_buf_debugfs_dir = NULL; 1127 err = PTR_ERR(d); 1128 } 1129 1130 return err; 1131 } 1132 1133 static void dma_buf_uninit_debugfs(void) 1134 { 1135 debugfs_remove_recursive(dma_buf_debugfs_dir); 1136 } 1137 #else 1138 static inline int dma_buf_init_debugfs(void) 1139 { 1140 return 0; 1141 } 1142 static inline void dma_buf_uninit_debugfs(void) 1143 { 1144 } 1145 #endif 1146 1147 static int __init dma_buf_init(void) 1148 { 1149 mutex_init(&db_list.lock); 1150 INIT_LIST_HEAD(&db_list.head); 1151 dma_buf_init_debugfs(); 1152 return 0; 1153 } 1154 subsys_initcall(dma_buf_init); 1155 1156 static void __exit dma_buf_deinit(void) 1157 { 1158 dma_buf_uninit_debugfs(); 1159 } 1160 __exitcall(dma_buf_deinit); 1161