Based on kernel version 2.6.26. Page generated on 2008-07-16 21:12 EST.
1 Dynamic DMA mapping using the generic device 2 ============================================ 3 4 James E.J. Bottomley <James.Bottomley[AT]HansenPartnership[DOT]com> 5 6 This document describes the DMA API. For a more gentle introduction 7 phrased in terms of the pci_ equivalents (and actual examples) see 8 DMA-mapping.txt 9 10 This API is split into two pieces. Part I describes the API and the 11 corresponding pci_ API. Part II describes the extensions to the API 12 for supporting non-consistent memory machines. Unless you know that 13 your driver absolutely has to support non-consistent platforms (this 14 is usually only legacy platforms) you should only use the API 15 described in part I. 16 17 Part I - pci_ and dma_ Equivalent API 18 ------------------------------------- 19 20 To get the pci_ API, you must #include <linux/pci.h> 21 To get the dma_ API, you must #include <linux/dma-mapping.h> 22 23 24 Part Ia - Using large dma-coherent buffers 25 ------------------------------------------ 26 27 void * 28 dma_alloc_coherent(struct device *dev, size_t size, 29 dma_addr_t *dma_handle, gfp_t flag) 30 void * 31 pci_alloc_consistent(struct pci_dev *dev, size_t size, 32 dma_addr_t *dma_handle) 33 34 Consistent memory is memory for which a write by either the device or 35 the processor can immediately be read by the processor or device 36 without having to worry about caching effects. (You may however need 37 to make sure to flush the processor's write buffers before telling 38 devices to read that memory.) 39 40 This routine allocates a region of <size> bytes of consistent memory. 41 It also returns a <dma_handle> which may be cast to an unsigned 42 integer the same width as the bus and used as the physical address 43 base of the region. 44 45 Returns: a pointer to the allocated region (in the processor's virtual 46 address space) or NULL if the allocation failed. 47 48 Note: consistent memory can be expensive on some platforms, and the 49 minimum allocation length may be as big as a page, so you should 50 consolidate your requests for consistent memory as much as possible. 51 The simplest way to do that is to use the dma_pool calls (see below). 52 53 The flag parameter (dma_alloc_coherent only) allows the caller to 54 specify the GFP_ flags (see kmalloc) for the allocation (the 55 implementation may choose to ignore flags that affect the location of 56 the returned memory, like GFP_DMA). For pci_alloc_consistent, you 57 must assume GFP_ATOMIC behaviour. 58 59 void 60 dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, 61 dma_addr_t dma_handle) 62 void 63 pci_free_consistent(struct pci_dev *dev, size_t size, void *cpu_addr, 64 dma_addr_t dma_handle) 65 66 Free the region of consistent memory you previously allocated. dev, 67 size and dma_handle must all be the same as those passed into the 68 consistent allocate. cpu_addr must be the virtual address returned by 69 the consistent allocate. 70 71 Note that unlike their sibling allocation calls, these routines 72 may only be called with IRQs enabled. 73 74 75 Part Ib - Using small dma-coherent buffers 76 ------------------------------------------ 77 78 To get this part of the dma_ API, you must #include <linux/dmapool.h> 79 80 Many drivers need lots of small dma-coherent memory regions for DMA 81 descriptors or I/O buffers. Rather than allocating in units of a page 82 or more using dma_alloc_coherent(), you can use DMA pools. These work 83 much like a struct kmem_cache, except that they use the dma-coherent allocator, 84 not __get_free_pages(). Also, they understand common hardware constraints 85 for alignment, like queue heads needing to be aligned on N-byte boundaries. 86 87 88 struct dma_pool * 89 dma_pool_create(const char *name, struct device *dev, 90 size_t size, size_t align, size_t alloc); 91 92 struct pci_pool * 93 pci_pool_create(const char *name, struct pci_device *dev, 94 size_t size, size_t align, size_t alloc); 95 96 The pool create() routines initialize a pool of dma-coherent buffers 97 for use with a given device. It must be called in a context which 98 can sleep. 99 100 The "name" is for diagnostics (like a struct kmem_cache name); dev and size 101 are like what you'd pass to dma_alloc_coherent(). The device's hardware 102 alignment requirement for this type of data is "align" (which is expressed 103 in bytes, and must be a power of two). If your device has no boundary 104 crossing restrictions, pass 0 for alloc; passing 4096 says memory allocated 105 from this pool must not cross 4KByte boundaries. 106 107 108 void *dma_pool_alloc(struct dma_pool *pool, gfp_t gfp_flags, 109 dma_addr_t *dma_handle); 110 111 void *pci_pool_alloc(struct pci_pool *pool, gfp_t gfp_flags, 112 dma_addr_t *dma_handle); 113 114 This allocates memory from the pool; the returned memory will meet the size 115 and alignment requirements specified at creation time. Pass GFP_ATOMIC to 116 prevent blocking, or if it's permitted (not in_interrupt, not holding SMP locks), 117 pass GFP_KERNEL to allow blocking. Like dma_alloc_coherent(), this returns 118 two values: an address usable by the cpu, and the dma address usable by the 119 pool's device. 120 121 122 void dma_pool_free(struct dma_pool *pool, void *vaddr, 123 dma_addr_t addr); 124 125 void pci_pool_free(struct pci_pool *pool, void *vaddr, 126 dma_addr_t addr); 127 128 This puts memory back into the pool. The pool is what was passed to 129 the pool allocation routine; the cpu (vaddr) and dma addresses are what 130 were returned when that routine allocated the memory being freed. 131 132 133 void dma_pool_destroy(struct dma_pool *pool); 134 135 void pci_pool_destroy(struct pci_pool *pool); 136 137 The pool destroy() routines free the resources of the pool. They must be 138 called in a context which can sleep. Make sure you've freed all allocated 139 memory back to the pool before you destroy it. 140 141 142 Part Ic - DMA addressing limitations 143 ------------------------------------ 144 145 int 146 dma_supported(struct device *dev, u64 mask) 147 int 148 pci_dma_supported(struct device *dev, u64 mask) 149 150 Checks to see if the device can support DMA to the memory described by 151 mask. 152 153 Returns: 1 if it can and 0 if it can't. 154 155 Notes: This routine merely tests to see if the mask is possible. It 156 won't change the current mask settings. It is more intended as an 157 internal API for use by the platform than an external API for use by 158 driver writers. 159 160 int 161 dma_set_mask(struct device *dev, u64 mask) 162 int 163 pci_set_dma_mask(struct pci_device *dev, u64 mask) 164 165 Checks to see if the mask is possible and updates the device 166 parameters if it is. 167 168 Returns: 0 if successful and a negative error if not. 169 170 u64 171 dma_get_required_mask(struct device *dev) 172 173 After setting the mask with dma_set_mask(), this API returns the 174 actual mask (within that already set) that the platform actually 175 requires to operate efficiently. Usually this means the returned mask 176 is the minimum required to cover all of memory. Examining the 177 required mask gives drivers with variable descriptor sizes the 178 opportunity to use smaller descriptors as necessary. 179 180 Requesting the required mask does not alter the current mask. If you 181 wish to take advantage of it, you should issue another dma_set_mask() 182 call to lower the mask again. 183 184 185 Part Id - Streaming DMA mappings 186 -------------------------------- 187 188 dma_addr_t 189 dma_map_single(struct device *dev, void *cpu_addr, size_t size, 190 enum dma_data_direction direction) 191 dma_addr_t 192 pci_map_single(struct device *dev, void *cpu_addr, size_t size, 193 int direction) 194 195 Maps a piece of processor virtual memory so it can be accessed by the 196 device and returns the physical handle of the memory. 197 198 The direction for both api's may be converted freely by casting. 199 However the dma_ API uses a strongly typed enumerator for its 200 direction: 201 202 DMA_NONE = PCI_DMA_NONE no direction (used for 203 debugging) 204 DMA_TO_DEVICE = PCI_DMA_TODEVICE data is going from the 205 memory to the device 206 DMA_FROM_DEVICE = PCI_DMA_FROMDEVICE data is coming from 207 the device to the 208 memory 209 DMA_BIDIRECTIONAL = PCI_DMA_BIDIRECTIONAL direction isn't known 210 211 Notes: Not all memory regions in a machine can be mapped by this 212 API. Further, regions that appear to be physically contiguous in 213 kernel virtual space may not be contiguous as physical memory. Since 214 this API does not provide any scatter/gather capability, it will fail 215 if the user tries to map a non-physically contiguous piece of memory. 216 For this reason, it is recommended that memory mapped by this API be 217 obtained only from sources which guarantee it to be physically contiguous 218 (like kmalloc). 219 220 Further, the physical address of the memory must be within the 221 dma_mask of the device (the dma_mask represents a bit mask of the 222 addressable region for the device. I.e., if the physical address of 223 the memory anded with the dma_mask is still equal to the physical 224 address, then the device can perform DMA to the memory). In order to 225 ensure that the memory allocated by kmalloc is within the dma_mask, 226 the driver may specify various platform-dependent flags to restrict 227 the physical memory range of the allocation (e.g. on x86, GFP_DMA 228 guarantees to be within the first 16Mb of available physical memory, 229 as required by ISA devices). 230 231 Note also that the above constraints on physical contiguity and 232 dma_mask may not apply if the platform has an IOMMU (a device which 233 supplies a physical to virtual mapping between the I/O memory bus and 234 the device). However, to be portable, device driver writers may *not* 235 assume that such an IOMMU exists. 236 237 Warnings: Memory coherency operates at a granularity called the cache 238 line width. In order for memory mapped by this API to operate 239 correctly, the mapped region must begin exactly on a cache line 240 boundary and end exactly on one (to prevent two separately mapped 241 regions from sharing a single cache line). Since the cache line size 242 may not be known at compile time, the API will not enforce this 243 requirement. Therefore, it is recommended that driver writers who 244 don't take special care to determine the cache line size at run time 245 only map virtual regions that begin and end on page boundaries (which 246 are guaranteed also to be cache line boundaries). 247 248 DMA_TO_DEVICE synchronisation must be done after the last modification 249 of the memory region by the software and before it is handed off to 250 the driver. Once this primitive is used, memory covered by this 251 primitive should be treated as read-only by the device. If the device 252 may write to it at any point, it should be DMA_BIDIRECTIONAL (see 253 below). 254 255 DMA_FROM_DEVICE synchronisation must be done before the driver 256 accesses data that may be changed by the device. This memory should 257 be treated as read-only by the driver. If the driver needs to write 258 to it at any point, it should be DMA_BIDIRECTIONAL (see below). 259 260 DMA_BIDIRECTIONAL requires special handling: it means that the driver 261 isn't sure if the memory was modified before being handed off to the 262 device and also isn't sure if the device will also modify it. Thus, 263 you must always sync bidirectional memory twice: once before the 264 memory is handed off to the device (to make sure all memory changes 265 are flushed from the processor) and once before the data may be 266 accessed after being used by the device (to make sure any processor 267 cache lines are updated with data that the device may have changed). 268 269 void 270 dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size, 271 enum dma_data_direction direction) 272 void 273 pci_unmap_single(struct pci_dev *hwdev, dma_addr_t dma_addr, 274 size_t size, int direction) 275 276 Unmaps the region previously mapped. All the parameters passed in 277 must be identical to those passed in (and returned) by the mapping 278 API. 279 280 dma_addr_t 281 dma_map_page(struct device *dev, struct page *page, 282 unsigned long offset, size_t size, 283 enum dma_data_direction direction) 284 dma_addr_t 285 pci_map_page(struct pci_dev *hwdev, struct page *page, 286 unsigned long offset, size_t size, int direction) 287 void 288 dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size, 289 enum dma_data_direction direction) 290 void 291 pci_unmap_page(struct pci_dev *hwdev, dma_addr_t dma_address, 292 size_t size, int direction) 293 294 API for mapping and unmapping for pages. All the notes and warnings 295 for the other mapping APIs apply here. Also, although the <offset> 296 and <size> parameters are provided to do partial page mapping, it is 297 recommended that you never use these unless you really know what the 298 cache width is. 299 300 int 301 dma_mapping_error(dma_addr_t dma_addr) 302 303 int 304 pci_dma_mapping_error(dma_addr_t dma_addr) 305 306 In some circumstances dma_map_single and dma_map_page will fail to create 307 a mapping. A driver can check for these errors by testing the returned 308 dma address with dma_mapping_error(). A non-zero return value means the mapping 309 could not be created and the driver should take appropriate action (e.g. 310 reduce current DMA mapping usage or delay and try again later). 311 312 int 313 dma_map_sg(struct device *dev, struct scatterlist *sg, 314 int nents, enum dma_data_direction direction) 315 int 316 pci_map_sg(struct pci_dev *hwdev, struct scatterlist *sg, 317 int nents, int direction) 318 319 Maps a scatter gather list from the block layer. 320 321 Returns: the number of physical segments mapped (this may be shorter 322 than <nents> passed in if the block layer determines that some 323 elements of the scatter/gather list are physically adjacent and thus 324 may be mapped with a single entry). 325 326 Please note that the sg cannot be mapped again if it has been mapped once. 327 The mapping process is allowed to destroy information in the sg. 328 329 As with the other mapping interfaces, dma_map_sg can fail. When it 330 does, 0 is returned and a driver must take appropriate action. It is 331 critical that the driver do something, in the case of a block driver 332 aborting the request or even oopsing is better than doing nothing and 333 corrupting the filesystem. 334 335 With scatterlists, you use the resulting mapping like this: 336 337 int i, count = dma_map_sg(dev, sglist, nents, direction); 338 struct scatterlist *sg; 339 340 for (i = 0, sg = sglist; i < count; i++, sg++) { 341 hw_address[i] = sg_dma_address(sg); 342 hw_len[i] = sg_dma_len(sg); 343 } 344 345 where nents is the number of entries in the sglist. 346 347 The implementation is free to merge several consecutive sglist entries 348 into one (e.g. with an IOMMU, or if several pages just happen to be 349 physically contiguous) and returns the actual number of sg entries it 350 mapped them to. On failure 0, is returned. 351 352 Then you should loop count times (note: this can be less than nents times) 353 and use sg_dma_address() and sg_dma_len() macros where you previously 354 accessed sg->address and sg->length as shown above. 355 356 void 357 dma_unmap_sg(struct device *dev, struct scatterlist *sg, 358 int nhwentries, enum dma_data_direction direction) 359 void 360 pci_unmap_sg(struct pci_dev *hwdev, struct scatterlist *sg, 361 int nents, int direction) 362 363 Unmap the previously mapped scatter/gather list. All the parameters 364 must be the same as those and passed in to the scatter/gather mapping 365 API. 366 367 Note: <nents> must be the number you passed in, *not* the number of 368 physical entries returned. 369 370 void 371 dma_sync_single(struct device *dev, dma_addr_t dma_handle, size_t size, 372 enum dma_data_direction direction) 373 void 374 pci_dma_sync_single(struct pci_dev *hwdev, dma_addr_t dma_handle, 375 size_t size, int direction) 376 void 377 dma_sync_sg(struct device *dev, struct scatterlist *sg, int nelems, 378 enum dma_data_direction direction) 379 void 380 pci_dma_sync_sg(struct pci_dev *hwdev, struct scatterlist *sg, 381 int nelems, int direction) 382 383 Synchronise a single contiguous or scatter/gather mapping. All the 384 parameters must be the same as those passed into the single mapping 385 API. 386 387 Notes: You must do this: 388 389 - Before reading values that have been written by DMA from the device 390 (use the DMA_FROM_DEVICE direction) 391 - After writing values that will be written to the device using DMA 392 (use the DMA_TO_DEVICE) direction 393 - before *and* after handing memory to the device if the memory is 394 DMA_BIDIRECTIONAL 395 396 See also dma_map_single(). 397 398 399 Part II - Advanced dma_ usage 400 ----------------------------- 401 402 Warning: These pieces of the DMA API have no PCI equivalent. They 403 should also not be used in the majority of cases, since they cater for 404 unlikely corner cases that don't belong in usual drivers. 405 406 If you don't understand how cache line coherency works between a 407 processor and an I/O device, you should not be using this part of the 408 API at all. 409 410 void * 411 dma_alloc_noncoherent(struct device *dev, size_t size, 412 dma_addr_t *dma_handle, gfp_t flag) 413 414 Identical to dma_alloc_coherent() except that the platform will 415 choose to return either consistent or non-consistent memory as it sees 416 fit. By using this API, you are guaranteeing to the platform that you 417 have all the correct and necessary sync points for this memory in the 418 driver should it choose to return non-consistent memory. 419 420 Note: where the platform can return consistent memory, it will 421 guarantee that the sync points become nops. 422 423 Warning: Handling non-consistent memory is a real pain. You should 424 only ever use this API if you positively know your driver will be 425 required to work on one of the rare (usually non-PCI) architectures 426 that simply cannot make consistent memory. 427 428 void 429 dma_free_noncoherent(struct device *dev, size_t size, void *cpu_addr, 430 dma_addr_t dma_handle) 431 432 Free memory allocated by the nonconsistent API. All parameters must 433 be identical to those passed in (and returned by 434 dma_alloc_noncoherent()). 435 436 int 437 dma_is_consistent(struct device *dev, dma_addr_t dma_handle) 438 439 Returns true if the device dev is performing consistent DMA on the memory 440 area pointed to by the dma_handle. 441 442 int 443 dma_get_cache_alignment(void) 444 445 Returns the processor cache alignment. This is the absolute minimum 446 alignment *and* width that you must observe when either mapping 447 memory or doing partial flushes. 448 449 Notes: This API may return a number *larger* than the actual cache 450 line, but it will guarantee that one or more cache lines fit exactly 451 into the width returned by this call. It will also always be a power 452 of two for easy alignment. 453 454 void 455 dma_sync_single_range(struct device *dev, dma_addr_t dma_handle, 456 unsigned long offset, size_t size, 457 enum dma_data_direction direction) 458 459 Does a partial sync, starting at offset and continuing for size. You 460 must be careful to observe the cache alignment and width when doing 461 anything like this. You must also be extra careful about accessing 462 memory you intend to sync partially. 463 464 void 465 dma_cache_sync(struct device *dev, void *vaddr, size_t size, 466 enum dma_data_direction direction) 467 468 Do a partial sync of memory that was allocated by 469 dma_alloc_noncoherent(), starting at virtual address vaddr and 470 continuing on for size. Again, you *must* observe the cache line 471 boundaries when doing this. 472 473 int 474 dma_declare_coherent_memory(struct device *dev, dma_addr_t bus_addr, 475 dma_addr_t device_addr, size_t size, int 476 flags) 477 478 Declare region of memory to be handed out by dma_alloc_coherent when 479 it's asked for coherent memory for this device. 480 481 bus_addr is the physical address to which the memory is currently 482 assigned in the bus responding region (this will be used by the 483 platform to perform the mapping). 484 485 device_addr is the physical address the device needs to be programmed 486 with actually to address this memory (this will be handed out as the 487 dma_addr_t in dma_alloc_coherent()). 488 489 size is the size of the area (must be multiples of PAGE_SIZE). 490 491 flags can be or'd together and are: 492 493 DMA_MEMORY_MAP - request that the memory returned from 494 dma_alloc_coherent() be directly writable. 495 496 DMA_MEMORY_IO - request that the memory returned from 497 dma_alloc_coherent() be addressable using read/write/memcpy_toio etc. 498 499 One or both of these flags must be present. 500 501 DMA_MEMORY_INCLUDES_CHILDREN - make the declared memory be allocated by 502 dma_alloc_coherent of any child devices of this one (for memory residing 503 on a bridge). 504 505 DMA_MEMORY_EXCLUSIVE - only allocate memory from the declared regions. 506 Do not allow dma_alloc_coherent() to fall back to system memory when 507 it's out of memory in the declared region. 508 509 The return value will be either DMA_MEMORY_MAP or DMA_MEMORY_IO and 510 must correspond to a passed in flag (i.e. no returning DMA_MEMORY_IO 511 if only DMA_MEMORY_MAP were passed in) for success or zero for 512 failure. 513 514 Note, for DMA_MEMORY_IO returns, all subsequent memory returned by 515 dma_alloc_coherent() may no longer be accessed directly, but instead 516 must be accessed using the correct bus functions. If your driver 517 isn't prepared to handle this contingency, it should not specify 518 DMA_MEMORY_IO in the input flags. 519 520 As a simplification for the platforms, only *one* such region of 521 memory may be declared per device. 522 523 For reasons of efficiency, most platforms choose to track the declared 524 region only at the granularity of a page. For smaller allocations, 525 you should use the dma_pool() API. 526 527 void 528 dma_release_declared_memory(struct device *dev) 529 530 Remove the memory region previously declared from the system. This 531 API performs *no* in-use checking for this region and will return 532 unconditionally having removed all the required structures. It is the 533 driver's job to ensure that no parts of this memory region are 534 currently in use. 535 536 void * 537 dma_mark_declared_memory_occupied(struct device *dev, 538 dma_addr_t device_addr, size_t size) 539 540 This is used to occupy specific regions of the declared space 541 (dma_alloc_coherent() will hand out the first free region it finds). 542 543 device_addr is the *device* address of the region requested. 544 545 size is the size (and should be a page-sized multiple). 546 547 The return value will be either a pointer to the processor virtual 548 address of the memory, or an error (via PTR_ERR()) if any part of the 549 region is occupied.