blob: c1e9545c59bd2366cfdb76dfa6946a24a9b4c547 [file] [log] [blame]
Dan Williamsc5d2b9f2007-09-20 15:49:08 -07001 Asynchronous Transfers/Transforms API
2
31 INTRODUCTION
4
52 GENEALOGY
6
73 USAGE
83.1 General format of the API
93.2 Supported operations
103.3 Descriptor management
113.4 When does the operation execute?
123.5 When does the operation complete?
133.6 Constraints
143.7 Example
15
164 DRIVER DEVELOPER NOTES
174.1 Conformance points
184.2 "My application needs finer control of hardware channels"
19
205 SOURCE
21
22---
23
241 INTRODUCTION
25
26The async_tx API provides methods for describing a chain of asynchronous
27bulk memory transfers/transforms with support for inter-transactional
28dependencies. It is implemented as a dmaengine client that smooths over
29the details of different hardware offload engine implementations. Code
30that is written to the API can optimize for asynchronous operation and
31the API will fit the chain of operations to the available offload
32resources.
33
342 GENEALOGY
35
36The API was initially designed to offload the memory copy and
37xor-parity-calculations of the md-raid5 driver using the offload engines
38present in the Intel(R) Xscale series of I/O processors. It also built
39on the 'dmaengine' layer developed for offloading memory copies in the
40network stack using Intel(R) I/OAT engines. The following design
41features surfaced as a result:
421/ implicit synchronous path: users of the API do not need to know if
43 the platform they are running on has offload capabilities. The
44 operation will be offloaded when an engine is available and carried out
45 in software otherwise.
462/ cross channel dependency chains: the API allows a chain of dependent
47 operations to be submitted, like xor->copy->xor in the raid5 case. The
48 API automatically handles cases where the transition from one operation
49 to another implies a hardware channel switch.
503/ dmaengine extensions to support multiple clients and operation types
51 beyond 'memcpy'
52
533 USAGE
54
553.1 General format of the API:
56struct dma_async_tx_descriptor *
57async_<operation>(<op specific parameters>,
58 enum async_tx_flags flags,
59 struct dma_async_tx_descriptor *dependency,
60 dma_async_tx_callback callback_routine,
61 void *callback_parameter);
62
633.2 Supported operations:
64memcpy - memory copy between a source and a destination buffer
65memset - fill a destination buffer with a byte value
66xor - xor a series of source buffers and write the result to a
67 destination buffer
68xor_zero_sum - xor a series of source buffers and set a flag if the
69 result is zero. The implementation attempts to prevent
70 writes to memory
71
723.3 Descriptor management:
73The return value is non-NULL and points to a 'descriptor' when the operation
74has been queued to execute asynchronously. Descriptors are recycled
75resources, under control of the offload engine driver, to be reused as
76operations complete. When an application needs to submit a chain of
77operations it must guarantee that the descriptor is not automatically recycled
78before the dependency is submitted. This requires that all descriptors be
79acknowledged by the application before the offload engine driver is allowed to
80recycle (or free) the descriptor. A descriptor can be acked by one of the
81following methods:
821/ setting the ASYNC_TX_ACK flag if no child operations are to be submitted
832/ setting the ASYNC_TX_DEP_ACK flag to acknowledge the parent
84 descriptor of a new operation.
853/ calling async_tx_ack() on the descriptor.
86
873.4 When does the operation execute?
88Operations do not immediately issue after return from the
89async_<operation> call. Offload engine drivers batch operations to
90improve performance by reducing the number of mmio cycles needed to
91manage the channel. Once a driver-specific threshold is met the driver
92automatically issues pending operations. An application can force this
93event by calling async_tx_issue_pending_all(). This operates on all
94channels since the application has no knowledge of channel to operation
95mapping.
96
973.5 When does the operation complete?
98There are two methods for an application to learn about the completion
99of an operation.
1001/ Call dma_wait_for_async_tx(). This call causes the CPU to spin while
101 it polls for the completion of the operation. It handles dependency
102 chains and issuing pending operations.
1032/ Specify a completion callback. The callback routine runs in tasklet
104 context if the offload engine driver supports interrupts, or it is
105 called in application context if the operation is carried out
106 synchronously in software. The callback can be set in the call to
107 async_<operation>, or when the application needs to submit a chain of
108 unknown length it can use the async_trigger_callback() routine to set a
109 completion interrupt/callback at the end of the chain.
110
1113.6 Constraints:
1121/ Calls to async_<operation> are not permitted in IRQ context. Other
113 contexts are permitted provided constraint #2 is not violated.
1142/ Completion callback routines cannot submit new operations. This
115 results in recursion in the synchronous case and spin_locks being
116 acquired twice in the asynchronous case.
117
1183.7 Example:
119Perform a xor->copy->xor operation where each operation depends on the
120result from the previous operation:
121
122void complete_xor_copy_xor(void *param)
123{
124 printk("complete\n");
125}
126
127int run_xor_copy_xor(struct page **xor_srcs,
128 int xor_src_cnt,
129 struct page *xor_dest,
130 size_t xor_len,
131 struct page *copy_src,
132 struct page *copy_dest,
133 size_t copy_len)
134{
135 struct dma_async_tx_descriptor *tx;
136
137 tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
138 ASYNC_TX_XOR_DROP_DST, NULL, NULL, NULL);
139 tx = async_memcpy(copy_dest, copy_src, 0, 0, copy_len,
140 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
141 tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
142 ASYNC_TX_XOR_DROP_DST | ASYNC_TX_DEP_ACK | ASYNC_TX_ACK,
143 tx, complete_xor_copy_xor, NULL);
144
145 async_tx_issue_pending_all();
146}
147
148See include/linux/async_tx.h for more information on the flags. See the
149ops_run_* and ops_complete_* routines in drivers/md/raid5.c for more
150implementation examples.
151
1524 DRIVER DEVELOPMENT NOTES
1534.1 Conformance points:
154There are a few conformance points required in dmaengine drivers to
155accommodate assumptions made by applications using the async_tx API:
1561/ Completion callbacks are expected to happen in tasklet context
1572/ dma_async_tx_descriptor fields are never manipulated in IRQ context
1583/ Use async_tx_run_dependencies() in the descriptor clean up path to
159 handle submission of dependent operations
160
1614.2 "My application needs finer control of hardware channels"
162This requirement seems to arise from cases where a DMA engine driver is
163trying to support device-to-memory DMA. The dmaengine and async_tx
164implementations were designed for offloading memory-to-memory
165operations; however, there are some capabilities of the dmaengine layer
166that can be used for platform-specific channel management.
167Platform-specific constraints can be handled by registering the
168application as a 'dma_client' and implementing a 'dma_event_callback' to
169apply a filter to the available channels in the system. Before showing
170how to implement a custom dma_event callback some background of
171dmaengine's client support is required.
172
173The following routines in dmaengine support multiple clients requesting
174use of a channel:
175- dma_async_client_register(struct dma_client *client)
176- dma_async_client_chan_request(struct dma_client *client)
177
178dma_async_client_register takes a pointer to an initialized dma_client
179structure. It expects that the 'event_callback' and 'cap_mask' fields
180are already initialized.
181
182dma_async_client_chan_request triggers dmaengine to notify the client of
183all channels that satisfy the capability mask. It is up to the client's
184event_callback routine to track how many channels the client needs and
185how many it is currently using. The dma_event_callback routine returns a
186dma_state_client code to let dmaengine know the status of the
187allocation.
188
189Below is the example of how to extend this functionality for
190platform-specific filtering of the available channels beyond the
191standard capability mask:
192
193static enum dma_state_client
194my_dma_client_callback(struct dma_client *client,
195 struct dma_chan *chan, enum dma_state state)
196{
197 struct dma_device *dma_dev;
198 struct my_platform_specific_dma *plat_dma_dev;
199
200 dma_dev = chan->device;
201 plat_dma_dev = container_of(dma_dev,
202 struct my_platform_specific_dma,
203 dma_dev);
204
205 if (!plat_dma_dev->platform_specific_capability)
206 return DMA_DUP;
207
208 . . .
209}
210
2115 SOURCE
212include/linux/dmaengine.h: core header file for DMA drivers and clients
213drivers/dma/dmaengine.c: offload engine channel management routines
214drivers/dma/: location for offload engine drivers
215include/linux/async_tx.h: core header file for the async_tx api
216crypto/async_tx/async_tx.c: async_tx interface to dmaengine and common code
217crypto/async_tx/async_memcpy.c: copy offload
218crypto/async_tx/async_memset.c: memory fill offload
219crypto/async_tx/async_xor.c: xor and xor zero sum offload