1 ==================== 2 DMA Engine API Guide 3 ==================== 4 5 Vinod Koul <vinod dot koul at intel.com> 6 7 .. note:: For DMA Engine usage in async_tx please see: 8 ``Documentation/crypto/async-tx-api.rst`` 9 10 11 Below is a guide to device driver writers on how to use the Slave-DMA API of the 12 DMA Engine. This is applicable only for slave DMA usage only. 13 14 DMA usage 15 ========= 16 17 The slave DMA usage consists of following steps: 18 19 - Allocate a DMA slave channel 20 21 - Set slave and controller specific parameters 22 23 - Get a descriptor for transaction 24 25 - Submit the transaction 26 27 - Issue pending requests and wait for callback notification 28 29 The details of these operations are: 30 31 1. Allocate a DMA slave channel 32 33 Channel allocation is slightly different in the slave DMA context, 34 client drivers typically need a channel from a particular DMA 35 controller only and even in some cases a specific channel is desired. 36 To request a channel dma_request_chan() API is used. 37 38 Interface: 39 40 .. code-block:: c 41 42 struct dma_chan *dma_request_chan(struct device *dev, const char *name); 43 44 Which will find and return the ``name`` DMA channel associated with the 'dev' 45 device. The association is done via DT, ACPI or board file based 46 dma_slave_map matching table. 47 48 A channel allocated via this interface is exclusive to the caller, 49 until dma_release_channel() is called. 50 51 2. Set slave and controller specific parameters 52 53 Next step is always to pass some specific information to the DMA 54 driver. Most of the generic information which a slave DMA can use 55 is in struct dma_slave_config. This allows the clients to specify 56 DMA direction, DMA addresses, bus widths, DMA burst lengths etc 57 for the peripheral. 58 59 If some DMA controllers have more parameters to be sent then they 60 should try to embed struct dma_slave_config in their controller 61 specific structure. That gives flexibility to client to pass more 62 parameters, if required. 63 64 Interface: 65 66 .. code-block:: c 67 68 int dmaengine_slave_config(struct dma_chan *chan, 69 struct dma_slave_config *config) 70 71 Please see the dma_slave_config structure definition in dmaengine.h 72 for a detailed explanation of the struct members. Please note 73 that the 'direction' member will be going away as it duplicates the 74 direction given in the prepare call. 75 76 3. Get a descriptor for transaction 77 78 For slave usage the various modes of slave transfers supported by the 79 DMA-engine are: 80 81 - slave_sg: DMA a list of scatter gather buffers from/to a peripheral 82 83 - peripheral_dma_vec: DMA an array of scatter gather buffers from/to a 84 peripheral. Similar to slave_sg, but uses an array of dma_vec 85 structures instead of a scatterlist. 86 87 - dma_cyclic: Perform a cyclic DMA operation from/to a peripheral till the 88 operation is explicitly stopped. 89 90 - interleaved_dma: This is common to Slave as well as M2M clients. For slave 91 address of devices' fifo could be already known to the driver. 92 Various types of operations could be expressed by setting 93 appropriate values to the 'dma_interleaved_template' members. Cyclic 94 interleaved DMA transfers are also possible if supported by the channel by 95 setting the DMA_PREP_REPEAT transfer flag. 96 97 A non-NULL return of this transfer API represents a "descriptor" for 98 the given transaction. 99 100 Interface: 101 102 .. code-block:: c 103 104 struct dma_async_tx_descriptor *dmaengine_prep_slave_sg( 105 struct dma_chan *chan, struct scatterlist *sgl, 106 unsigned int sg_len, enum dma_data_direction direction, 107 unsigned long flags); 108 109 struct dma_async_tx_descriptor *dmaengine_prep_peripheral_dma_vec( 110 struct dma_chan *chan, const struct dma_vec *vecs, 111 size_t nents, enum dma_data_direction direction, 112 unsigned long flags); 113 114 struct dma_async_tx_descriptor *dmaengine_prep_dma_cyclic( 115 struct dma_chan *chan, dma_addr_t buf_addr, size_t buf_len, 116 size_t period_len, enum dma_data_direction direction); 117 118 struct dma_async_tx_descriptor *dmaengine_prep_interleaved_dma( 119 struct dma_chan *chan, struct dma_interleaved_template *xt, 120 unsigned long flags); 121 122 The peripheral driver is expected to have mapped the scatterlist for 123 the DMA operation prior to calling dmaengine_prep_slave_sg(), and must 124 keep the scatterlist mapped until the DMA operation has completed. 125 The scatterlist must be mapped using the DMA struct device. 126 If a mapping needs to be synchronized later, dma_sync_*_for_*() must be 127 called using the DMA struct device, too. 128 So, normal setup should look like this: 129 130 .. code-block:: c 131 132 struct device *dma_dev = dmaengine_get_dma_device(chan); 133 134 nr_sg = dma_map_sg(dma_dev, sgl, sg_len); 135 if (nr_sg == 0) 136 /* error */ 137 138 desc = dmaengine_prep_slave_sg(chan, sgl, nr_sg, direction, flags); 139 140 Once a descriptor has been obtained, the callback information can be 141 added and the descriptor must then be submitted. Some DMA engine 142 drivers may hold a spinlock between a successful preparation and 143 submission so it is important that these two operations are closely 144 paired. 145 146 .. note:: 147 148 Although the async_tx API specifies that completion callback 149 routines cannot submit any new operations, this is not the 150 case for slave/cyclic DMA. 151 152 For slave DMA, the subsequent transaction may not be available 153 for submission prior to callback function being invoked, so 154 slave DMA callbacks are permitted to prepare and submit a new 155 transaction. 156 157 For cyclic DMA, a callback function may wish to terminate the 158 DMA via dmaengine_terminate_async(). 159 160 Therefore, it is important that DMA engine drivers drop any 161 locks before calling the callback function which may cause a 162 deadlock. 163 164 Note that callbacks will always be invoked from the DMA 165 engines tasklet, never from interrupt context. 166 167 **Optional: per descriptor metadata** 168 169 DMAengine provides two ways for metadata support. 170 171 DESC_METADATA_CLIENT 172 173 The metadata buffer is allocated/provided by the client driver and it is 174 attached to the descriptor. 175 176 .. code-block:: c 177 178 int dmaengine_desc_attach_metadata(struct dma_async_tx_descriptor *desc, 179 void *data, size_t len); 180 181 DESC_METADATA_ENGINE 182 183 The metadata buffer is allocated/managed by the DMA driver. The client 184 driver can ask for the pointer, maximum size and the currently used size of 185 the metadata and can directly update or read it. 186 187 Because the DMA driver manages the memory area containing the metadata, 188 clients must make sure that they do not try to access or get the pointer 189 after their transfer completion callback has run for the descriptor. 190 If no completion callback has been defined for the transfer, then the 191 metadata must not be accessed after issue_pending. 192 In other words: if the aim is to read back metadata after the transfer is 193 completed, then the client must use completion callback. 194 195 .. code-block:: c 196 197 void *dmaengine_desc_get_metadata_ptr(struct dma_async_tx_descriptor *desc, 198 size_t *payload_len, size_t *max_len); 199 200 int dmaengine_desc_set_metadata_len(struct dma_async_tx_descriptor *desc, 201 size_t payload_len); 202 203 Client drivers can query if a given mode is supported with: 204 205 .. code-block:: c 206 207 bool dmaengine_is_metadata_mode_supported(struct dma_chan *chan, 208 enum dma_desc_metadata_mode mode); 209 210 Depending on the used mode client drivers must follow different flow. 211 212 DESC_METADATA_CLIENT 213 214 - DMA_MEM_TO_DEV / DEV_MEM_TO_MEM: 215 216 1. prepare the descriptor (dmaengine_prep_*) 217 construct the metadata in the client's buffer 218 2. use dmaengine_desc_attach_metadata() to attach the buffer to the 219 descriptor 220 3. submit the transfer 221 222 - DMA_DEV_TO_MEM: 223 224 1. prepare the descriptor (dmaengine_prep_*) 225 2. use dmaengine_desc_attach_metadata() to attach the buffer to the 226 descriptor 227 3. submit the transfer 228 4. when the transfer is completed, the metadata should be available in the 229 attached buffer 230 231 DESC_METADATA_ENGINE 232 233 - DMA_MEM_TO_DEV / DEV_MEM_TO_MEM: 234 235 1. prepare the descriptor (dmaengine_prep_*) 236 2. use dmaengine_desc_get_metadata_ptr() to get the pointer to the 237 engine's metadata area 238 3. update the metadata at the pointer 239 4. use dmaengine_desc_set_metadata_len() to tell the DMA engine the 240 amount of data the client has placed into the metadata buffer 241 5. submit the transfer 242 243 - DMA_DEV_TO_MEM: 244 245 1. prepare the descriptor (dmaengine_prep_*) 246 2. submit the transfer 247 3. on transfer completion, use dmaengine_desc_get_metadata_ptr() to get 248 the pointer to the engine's metadata area 249 4. read out the metadata from the pointer 250 251 .. note:: 252 253 When DESC_METADATA_ENGINE mode is used the metadata area for the descriptor 254 is no longer valid after the transfer has been completed (valid up to the 255 point when the completion callback returns if used). 256 257 Mixed use of DESC_METADATA_CLIENT / DESC_METADATA_ENGINE is not allowed, 258 client drivers must use either of the modes per descriptor. 259 260 4. Submit the transaction 261 262 Once the descriptor has been prepared and the callback information 263 added, it must be placed on the DMA engine drivers pending queue. 264 265 Interface: 266 267 .. code-block:: c 268 269 dma_cookie_t dmaengine_submit(struct dma_async_tx_descriptor *desc) 270 271 This returns a cookie can be used to check the progress of DMA engine 272 activity via other DMA engine calls not covered in this document. 273 274 dmaengine_submit() will not start the DMA operation, it merely adds 275 it to the pending queue. For this, see step 5, dma_async_issue_pending. 276 277 .. note:: 278 279 After calling ``dmaengine_submit()`` the submitted transfer descriptor 280 (``struct dma_async_tx_descriptor``) belongs to the DMA engine. 281 Consequently, the client must consider invalid the pointer to that 282 descriptor. 283 284 5. Issue pending DMA requests and wait for callback notification 285 286 The transactions in the pending queue can be activated by calling the 287 issue_pending API. If channel is idle then the first transaction in 288 queue is started and subsequent ones queued up. 289 290 On completion of each DMA operation, the next in queue is started and 291 a tasklet triggered. The tasklet will then call the client driver 292 completion callback routine for notification, if set. 293 294 Interface: 295 296 .. code-block:: c 297 298 void dma_async_issue_pending(struct dma_chan *chan); 299 300 Further APIs 301 ------------ 302 303 1. Terminate APIs 304 305 .. code-block:: c 306 307 int dmaengine_terminate_sync(struct dma_chan *chan) 308 int dmaengine_terminate_async(struct dma_chan *chan) 309 int dmaengine_terminate_all(struct dma_chan *chan) /* DEPRECATED */ 310 311 This causes all activity for the DMA channel to be stopped, and may 312 discard data in the DMA FIFO which hasn't been fully transferred. 313 No callback functions will be called for any incomplete transfers. 314 315 Two variants of this function are available. 316 317 dmaengine_terminate_async() might not wait until the DMA has been fully 318 stopped or until any running complete callbacks have finished. But it is 319 possible to call dmaengine_terminate_async() from atomic context or from 320 within a complete callback. dmaengine_synchronize() must be called before it 321 is safe to free the memory accessed by the DMA transfer or free resources 322 accessed from within the complete callback. 323 324 dmaengine_terminate_sync() will wait for the transfer and any running 325 complete callbacks to finish before it returns. But the function must not be 326 called from atomic context or from within a complete callback. 327 328 dmaengine_terminate_all() is deprecated and should not be used in new code. 329 330 2. Pause API 331 332 .. code-block:: c 333 334 int dmaengine_pause(struct dma_chan *chan) 335 336 This pauses activity on the DMA channel without data loss. 337 338 3. Resume API 339 340 .. code-block:: c 341 342 int dmaengine_resume(struct dma_chan *chan) 343 344 Resume a previously paused DMA channel. It is invalid to resume a 345 channel which is not currently paused. 346 347 4. Check Txn complete 348 349 .. code-block:: c 350 351 enum dma_status dma_async_is_tx_complete(struct dma_chan *chan, 352 dma_cookie_t cookie, dma_cookie_t *last, dma_cookie_t *used) 353 354 This can be used to check the status of the channel. Please see 355 the documentation in include/linux/dmaengine.h for a more complete 356 description of this API. 357 358 This can be used in conjunction with dma_async_is_complete() and 359 the cookie returned from dmaengine_submit() to check for 360 completion of a specific DMA transaction. 361 362 .. note:: 363 364 Not all DMA engine drivers can return reliable information for 365 a running DMA channel. It is recommended that DMA engine users 366 pause or stop (via dmaengine_terminate_all()) the channel before 367 using this API. 368 369 5. Synchronize termination API 370 371 .. code-block:: c 372 373 void dmaengine_synchronize(struct dma_chan *chan) 374 375 Synchronize the termination of the DMA channel to the current context. 376 377 This function should be used after dmaengine_terminate_async() to synchronize 378 the termination of the DMA channel to the current context. The function will 379 wait for the transfer and any running complete callbacks to finish before it 380 returns. 381 382 If dmaengine_terminate_async() is used to stop the DMA channel this function 383 must be called before it is safe to free memory accessed by previously 384 submitted descriptors or to free any resources accessed within the complete 385 callback of previously submitted descriptors. 386 387 The behavior of this function is undefined if dma_async_issue_pending() has 388 been called between dmaengine_terminate_async() and this function.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.