diff mbox series

[v3,2/5] Documentation: iio: add doc for high-speed buffer API

Message ID 20210215143234.3248-3-alexandru.ardelean@analog.com (mailing list archive)
State New
Headers show
Series iio: core: Add mmap interface infrastructure | expand

Commit Message

Alexandru Ardelean Feb. 15, 2021, 2:32 p.m. UTC
This change takes the comment from the commit that introduces the IIO
high-speed buffer API, and formats it into rst format.

Signed-off-by: Alexandru Ardelean <alexandru.ardelean@analog.com>
 Documentation/iio/iio_high_speed_buffers.rst | 100 +++++++++++++++++++
 Documentation/iio/index.rst                  |   2 +
 include/uapi/linux/iio/buffer.h              |   5 +
 3 files changed, 107 insertions(+)
 create mode 100644 Documentation/iio/iio_high_speed_buffers.rst
diff mbox series


diff --git a/Documentation/iio/iio_high_speed_buffers.rst b/Documentation/iio/iio_high_speed_buffers.rst
new file mode 100644
index 000000000000..f326e68efe49
--- /dev/null
+++ b/Documentation/iio/iio_high_speed_buffers.rst
@@ -0,0 +1,100 @@ 
+Industrial IO High-Speed Buffer API
+1. Overview
+Industrial IO supports access to buffers via an mmap interface. The
+advantage of the mmap based interface compared to the read() based
+interface is that it avoids an extra copy of the data between kernel and
+userspace. This is particular useful for high-speed devices which produce
+several megabytes or even gigabytes of data per second.
+The data for the mmap interface is managed at the granularity of so called
+blocks. A block is a contiguous region of memory (at the moment both
+physically and virtually contiguous). Reducing the granularity from byte
+level to block level is done to reduce the userspace-kernelspace
+synchronization overhead since performing syscalls for each byte at a
+data-rate of a few megabytes is not feasible.
+This of course leads to a slightly increased latency. For this reason an
+application can choose the size of the blocks as well as how many blocks it
+allocates. E.g. two blocks would be a traditional double buffering scheme.
+But using a higher number might be necessary to avoid underflow/overflow
+situations in the presence of scheduling latencies.
+A block can either be owned by kernel space or userspace. When owned by
+userspace it is safe to access the data in the block and process it. When
+owned by kernel space the block can be in one of 3 states:
+* It can be in the incoming queue where all blocks submitted from userspace
+  are placed and are waiting to be processed by the kernel driver.
+* It can be currently being processed by the kernel driver, this means it is
+  actively placing capturing data in it (usually using DMA).
+* Or it can be in the outgoing queue where all blocks that have been
+  processed by the kernel are placed. Userspace can dequeue the blocks as
+  necessary.
+2. Interface
+As part of the interface 5 IOCTLs are used to manage the blocks and exchange
+them between userspace and kernelspace. The IOCTLs can be accessed through
+a open file descriptor to a IIO device.
+* **IIO_BUFFER_BLOCK_ALLOC_IOCTL(struct iio_buffer_block_alloc_req *)**:
+    Allocates new blocks. Can be called multiple times if necessary. A newly
+    allocated block is initially owned by userspace.
+   Frees all previously allocated blocks. If the backing memory of a block is
+   still in use by a kernel driver (i.e. active DMA transfer) it will be
+   freed once the kernel driver has released it.
+* **IIO_BUFFER_BLOCK_QUERY_IOCTL(struct iio_buffer_block *)**:
+   Queries information about a block. The id of the block about which
+   information is to be queried needs to be set by userspace.
+* **IIO_BUFFER_BLOCK_ENQUEUE_IOCTL(struct iio_buffer_block *)**:
+   Places a block on the incoming queue. This transfers ownership of the
+   block from userspace to kernelspace. Userspace must populate the id field
+   of the block to indicate which block to enqueue.
+* **IIO_BUFFER_BLOCK_DEQUEUE_IOCTL(struct iio_buffer_block *)**:
+   Removes the first block from the outgoing queue. This transfers ownership
+   of the block from kernelspace to userspace. Kernelspace will populate all
+   fields of the block. If the queue is empty and the file descriptor is set
+   to blocking the IOCTL will block until a new block is available on the
+   outgoing queue.
+3. Usage
+To access the data stored in a block by userspace the block must be mapped
+to the process's memory. This is done by calling mmap() on the IIO device
+file descriptor. Each block has a unique offset assigned to it which should
+be passed to the mmap interface. E.g.
+  mmap(0, block.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
+       block.offset);
+A typical workflow for the new interface is:
+  foreach block
+     BLOCK_QUERY block
+	 mmap block.data.offset
+  enable buffer
+  while !done
+	process data
+  disable buffer
diff --git a/Documentation/iio/index.rst b/Documentation/iio/index.rst
index 58b7a4ebac51..aaba78770b47 100644
--- a/Documentation/iio/index.rst
+++ b/Documentation/iio/index.rst
@@ -9,4 +9,6 @@  Industrial I/O
+   iio_high_speed_buffers
diff --git a/include/uapi/linux/iio/buffer.h b/include/uapi/linux/iio/buffer.h
index 8c1a2f27e5a2..d8c64210c9cc 100644
--- a/include/uapi/linux/iio/buffer.h
+++ b/include/uapi/linux/iio/buffer.h
@@ -5,6 +5,11 @@ 
 #ifndef _UAPI_IIO_BUFFER_H_
 #define _UAPI_IIO_BUFFER_H_
+ * See for more details:
+ *   Documentation/iio/iio_high_speed_buffers.rst
+ */
  * struct iio_buffer_block_alloc_req - Descriptor for allocating IIO buffer blocks
  * @type:	type of block(s) to allocate (currently unused, reserved)