async_tx: usage documentation and developer notes (v2)

Changes in v2:
* cleanups from Randy and Shannon

Reviewed-by: Randy Dunlap <randy.dunlap@oracle.com>
Reviewed-by: Shannon Nelson <shannon.nelson@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>

diff --git a/Documentation/crypto/async-tx-api.txt b/Documentation/crypto/async-tx-api.txt
new file mode 100644 (file)
index 0000000..c1e9545
--- /dev/null
+++ b/Documentation/crypto/async-tx-api.txt
@@ -0,0 +1,219 @@
+                Asynchronous Transfers/Transforms API
+
+1 INTRODUCTION
+
+2 GENEALOGY
+
+3 USAGE
+3.1 General format of the API
+3.2 Supported operations
+3.3 Descriptor management
+3.4 When does the operation execute?
+3.5 When does the operation complete?
+3.6 Constraints
+3.7 Example
+
+4 DRIVER DEVELOPER NOTES
+4.1 Conformance points
+4.2 "My application needs finer control of hardware channels"
+
+5 SOURCE
+
+---
+
+1 INTRODUCTION
+
+The async_tx API provides methods for describing a chain of asynchronous
+bulk memory transfers/transforms with support for inter-transactional
+dependencies.  It is implemented as a dmaengine client that smooths over
+the details of different hardware offload engine implementations.  Code
+that is written to the API can optimize for asynchronous operation and
+the API will fit the chain of operations to the available offload
+resources.
+
+2 GENEALOGY
+
+The API was initially designed to offload the memory copy and
+xor-parity-calculations of the md-raid5 driver using the offload engines
+present in the Intel(R) Xscale series of I/O processors.  It also built
+on the 'dmaengine' layer developed for offloading memory copies in the
+network stack using Intel(R) I/OAT engines.  The following design
+features surfaced as a result:
+1/ implicit synchronous path: users of the API do not need to know if
+   the platform they are running on has offload capabilities.  The
+   operation will be offloaded when an engine is available and carried out
+   in software otherwise.
+2/ cross channel dependency chains: the API allows a chain of dependent
+   operations to be submitted, like xor->copy->xor in the raid5 case.  The
+   API automatically handles cases where the transition from one operation
+   to another implies a hardware channel switch.
+3/ dmaengine extensions to support multiple clients and operation types
+   beyond 'memcpy'
+
+3 USAGE
+
+3.1 General format of the API:
+struct dma_async_tx_descriptor *
+async_<operation>(<op specific parameters>,
+                 enum async_tx_flags flags,
+                 struct dma_async_tx_descriptor *dependency,
+                 dma_async_tx_callback callback_routine,
+                 void *callback_parameter);
+
+3.2 Supported operations:
+memcpy       - memory copy between a source and a destination buffer
+memset       - fill a destination buffer with a byte value
+xor          - xor a series of source buffers and write the result to a
+              destination buffer
+xor_zero_sum - xor a series of source buffers and set a flag if the
+              result is zero.  The implementation attempts to prevent
+              writes to memory
+
+3.3 Descriptor management:
+The return value is non-NULL and points to a 'descriptor' when the operation
+has been queued to execute asynchronously.  Descriptors are recycled
+resources, under control of the offload engine driver, to be reused as
+operations complete.  When an application needs to submit a chain of
+operations it must guarantee that the descriptor is not automatically recycled
+before the dependency is submitted.  This requires that all descriptors be
+acknowledged by the application before the offload engine driver is allowed to
+recycle (or free) the descriptor.  A descriptor can be acked by one of the
+following methods:
+1/ setting the ASYNC_TX_ACK flag if no child operations are to be submitted
+2/ setting the ASYNC_TX_DEP_ACK flag to acknowledge the parent
+   descriptor of a new operation.
+3/ calling async_tx_ack() on the descriptor.
+
+3.4 When does the operation execute?
+Operations do not immediately issue after return from the
+async_<operation> call.  Offload engine drivers batch operations to
+improve performance by reducing the number of mmio cycles needed to
+manage the channel.  Once a driver-specific threshold is met the driver
+automatically issues pending operations.  An application can force this
+event by calling async_tx_issue_pending_all().  This operates on all
+channels since the application has no knowledge of channel to operation
+mapping.
+
+3.5 When does the operation complete?
+There are two methods for an application to learn about the completion
+of an operation.
+1/ Call dma_wait_for_async_tx().  This call causes the CPU to spin while
+   it polls for the completion of the operation.  It handles dependency
+   chains and issuing pending operations.
+2/ Specify a completion callback.  The callback routine runs in tasklet
+   context if the offload engine driver supports interrupts, or it is
+   called in application context if the operation is carried out
+   synchronously in software.  The callback can be set in the call to
+   async_<operation>, or when the application needs to submit a chain of
+   unknown length it can use the async_trigger_callback() routine to set a
+   completion interrupt/callback at the end of the chain.
+
+3.6 Constraints:
+1/ Calls to async_<operation> are not permitted in IRQ context.  Other
+   contexts are permitted provided constraint #2 is not violated.
+2/ Completion callback routines cannot submit new operations.  This
+   results in recursion in the synchronous case and spin_locks being
+   acquired twice in the asynchronous case.
+
+3.7 Example:
+Perform a xor->copy->xor operation where each operation depends on the
+result from the previous operation:
+
+void complete_xor_copy_xor(void *param)
+{
+       printk("complete\n");
+}
+
+int run_xor_copy_xor(struct page **xor_srcs,
+                    int xor_src_cnt,
+                    struct page *xor_dest,
+                    size_t xor_len,
+                    struct page *copy_src,
+                    struct page *copy_dest,
+                    size_t copy_len)
+{
+       struct dma_async_tx_descriptor *tx;
+
+       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
+                      ASYNC_TX_XOR_DROP_DST, NULL, NULL, NULL);
+       tx = async_memcpy(copy_dest, copy_src, 0, 0, copy_len,
+                         ASYNC_TX_DEP_ACK, tx, NULL, NULL);
+       tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len,
+                      ASYNC_TX_XOR_DROP_DST | ASYNC_TX_DEP_ACK | ASYNC_TX_ACK,
+                      tx, complete_xor_copy_xor, NULL);
+
+       async_tx_issue_pending_all();
+}
+
+See include/linux/async_tx.h for more information on the flags.  See the
+ops_run_* and ops_complete_* routines in drivers/md/raid5.c for more
+implementation examples.
+
+4 DRIVER DEVELOPMENT NOTES
+4.1 Conformance points:
+There are a few conformance points required in dmaengine drivers to
+accommodate assumptions made by applications using the async_tx API:
+1/ Completion callbacks are expected to happen in tasklet context
+2/ dma_async_tx_descriptor fields are never manipulated in IRQ context
+3/ Use async_tx_run_dependencies() in the descriptor clean up path to
+   handle submission of dependent operations
+
+4.2 "My application needs finer control of hardware channels"
+This requirement seems to arise from cases where a DMA engine driver is
+trying to support device-to-memory DMA.  The dmaengine and async_tx
+implementations were designed for offloading memory-to-memory
+operations; however, there are some capabilities of the dmaengine layer
+that can be used for platform-specific channel management.
+Platform-specific constraints can be handled by registering the
+application as a 'dma_client' and implementing a 'dma_event_callback' to
+apply a filter to the available channels in the system.  Before showing
+how to implement a custom dma_event callback some background of
+dmaengine's client support is required.
+
+The following routines in dmaengine support multiple clients requesting
+use of a channel:
+- dma_async_client_register(struct dma_client *client)
+- dma_async_client_chan_request(struct dma_client *client)
+
+dma_async_client_register takes a pointer to an initialized dma_client
+structure.  It expects that the 'event_callback' and 'cap_mask' fields
+are already initialized.
+
+dma_async_client_chan_request triggers dmaengine to notify the client of
+all channels that satisfy the capability mask.  It is up to the client's
+event_callback routine to track how many channels the client needs and
+how many it is currently using.  The dma_event_callback routine returns a
+dma_state_client code to let dmaengine know the status of the
+allocation.
+
+Below is the example of how to extend this functionality for
+platform-specific filtering of the available channels beyond the
+standard capability mask:
+
+static enum dma_state_client
+my_dma_client_callback(struct dma_client *client,
+                       struct dma_chan *chan, enum dma_state state)
+{
+       struct dma_device *dma_dev;
+       struct my_platform_specific_dma *plat_dma_dev;
+       
+       dma_dev = chan->device;
+       plat_dma_dev = container_of(dma_dev,
+                                   struct my_platform_specific_dma,
+                                   dma_dev);
+
+       if (!plat_dma_dev->platform_specific_capability)
+               return DMA_DUP;
+
+       . . .
+}
+
+5 SOURCE
+include/linux/dmaengine.h: core header file for DMA drivers and clients
+drivers/dma/dmaengine.c: offload engine channel management routines
+drivers/dma/: location for offload engine drivers
+include/linux/async_tx.h: core header file for the async_tx api
+crypto/async_tx/async_tx.c: async_tx interface to dmaengine and common code
+crypto/async_tx/async_memcpy.c: copy offload
+crypto/async_tx/async_memset.c: memory fill offload
+crypto/async_tx/async_xor.c: xor and xor zero sum offload