Re: [PATCH v2 12/12] Documentation: iio: Document high-speed DMABUF based API

[Paul Cercueil <paul@crapouillou.net>]

Le mar., mars 29 2022 at 16:07:21 +0200, Daniel Vetter 
<daniel@ffwll.ch> a écrit :
> On Tue, Mar 29, 2022 at 10:47:23AM +0100, Paul Cercueil wrote:
>>  Hi Daniel,
>> 
>>  Le mar., mars 29 2022 at 10:54:43 +0200, Daniel Vetter 
>> <daniel@ffwll.ch> a
>>  écrit :
>>  > On Mon, Feb 07, 2022 at 01:01:40PM +0000, Paul Cercueil wrote:
>>  > >  Document the new DMABUF based API.
>>  > >
>>  > >  v2: - Explicitly state that the new interface is optional and 
>> is
>>  > >        not implemented by all drivers.
>>  > >      - The IOCTLs can now only be called on the buffer FD 
>> returned by
>>  > >        IIO_BUFFER_GET_FD_IOCTL.
>>  > >      - Move the page up a bit in the index since it is core 
>> stuff
>>  > > and not
>>  > >        driver-specific.
>>  > >
>>  > >  Signed-off-by: Paul Cercueil <paul@crapouillou.net>
>>  > >  ---
>>  > >   Documentation/driver-api/dma-buf.rst |  2 +
>>  > >   Documentation/iio/dmabuf_api.rst     | 94
>>  > > ++++++++++++++++++++++++++++
>>  > >   Documentation/iio/index.rst          |  2 +
>>  > >   3 files changed, 98 insertions(+)
>>  > >   create mode 100644 Documentation/iio/dmabuf_api.rst
>>  > >
>>  > >  diff --git a/Documentation/driver-api/dma-buf.rst
>>  > > b/Documentation/driver-api/dma-buf.rst
>>  > >  index 2cd7db82d9fe..d3c9b58d2706 100644
>>  > >  --- a/Documentation/driver-api/dma-buf.rst
>>  > >  +++ b/Documentation/driver-api/dma-buf.rst
>>  > >  @@ -1,3 +1,5 @@
>>  > >  +.. _dma-buf:
>>  > >  +
>>  > >   Buffer Sharing and Synchronization
>>  > >   ==================================
>>  > >
>>  > >  diff --git a/Documentation/iio/dmabuf_api.rst
>>  > > b/Documentation/iio/dmabuf_api.rst
>>  > >  new file mode 100644
>>  > >  index 000000000000..43bb2c1b9fdc
>>  > >  --- /dev/null
>>  > >  +++ b/Documentation/iio/dmabuf_api.rst
>>  > >  @@ -0,0 +1,94 @@
>>  > >  +===================================
>>  > >  +High-speed DMABUF interface for IIO
>>  > >  +===================================
>>  > >  +
>>  > >  +1. Overview
>>  > >  +===========
>>  > >  +
>>  > >  +The Industrial I/O subsystem supports access to buffers 
>> through a
>>  > > file-based
>>  > >  +interface, with read() and write() access calls through the 
>> IIO
>>  > > device's dev
>>  > >  +node.
>>  > >  +
>>  > >  +It additionally supports a DMABUF based interface, where the
>>  > > userspace
>>  > >  +application can allocate and append DMABUF objects to the 
>> buffer's
>>  > > queue.
>>  > >  +This interface is however optional and is not available in all
>>  > > drivers.
>>  > >  +
>>  > >  +The advantage of this DMABUF based interface vs. the read()
>>  > >  +interface, is that it avoids an extra copy of the data 
>> between the
>>  > >  +kernel and userspace. This is particularly useful for 
>> high-speed
>>  > >  +devices which produce several megabytes or even gigabytes of 
>> data
>>  > > per
>>  > >  +second.
>>  > >  +
>>  > >  +The data in this DMABUF interface is managed at the 
>> granularity of
>>  > >  +DMABUF objects. 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 few Mbps is just 
>> not
>>  > >  +feasible.
>>  > >  +
>>  > >  +This of course leads to a slightly increased latency. For this
>>  > > reason an
>>  > >  +application can choose the size of the DMABUFs as well as how 
>> many
>>  > > it
>>  > >  +allocates. E.g. two DMABUFs 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.
>>  >
>>  > So this reads a lot like reinventing io-uring with pre-registered
>>  > O_DIRECT
>>  > memory ranges. Except it's using dma-buf and hand-rolling a lot of
>>  > pieces
>>  > instead of io-uring and O_DIRECT.
>> 
>>  I don't see how io_uring would help us. It's an async I/O 
>> framework, does it
>>  allow us to access a kernel buffer without copying the data? Does 
>> it allow
>>  us to zero-copy the data to a network interface?
> 
> With networking, do you mean rdma, or some other kind of networking?
> Anything else than rdma doesn't support dma-buf, and I don't think it 
> will
> likely ever do so. Similar it's really tricky to glue dma-buf support 
> into
> the block layer.

By networking I mean standard sockets. If I'm not mistaken, Jonathan 
Lemon's work on zctap was to add dma-buf import/export support to 
standard sockets.

> Wrt io_uring, yes it's async, but that's not the point. The point is 
> that
> with io_uring you pre-register ranges for reads and writes to target,
> which in combination with O_DIRECT, makes it effectively (and 
> efficient!)
> zero-copy. Plus it has full integration with both networking and 
> normal
> file io, which dma-buf just doesn't have.
> 
> Like you _cannot_ do zero copy from a dma-buf into a normal file. You
> absolutely can do the same with io_uring.

I believe io_uring does zero-copy the same way as splice(), by 
duplicating/moving pages? Because that wouldn't work with DMA coherent 
memory, which is contiguous and not backed by pages.

>>  > At least if the entire justification for dma-buf support is 
>> zero-copy
>>  > support between the driver and userspace it's _really_ not the 
>> right
>>  > tool
>>  > for the job. dma-buf is for zero-copy between devices, with cpu 
>> access
>>  > from userpace (or kernel fwiw) being very much the exception (and 
>> often
>>  > flat-out not supported at all).
>> 
>>  We want both. Using dma-bufs for the driver/userspace interface is a
>>  convenience as we then have a unique API instead of two distinct 
>> ones.
>> 
>>  Why should CPU access from userspace be the exception? It works 
>> fine for IIO
>>  dma-bufs. You keep warning about this being a terrible design, but 
>> I simply
>>  don't see it.
> 
> It depends really on what you're trying to do, and there's extremely 
> high
> chances it will simply not work.

Well it does work though. The userspace interface is stupidly simple 
here - one dma-buf, backed by DMA coherent memory, is enqueued for 
processing by the DMA. The userspace calling the "sync" ioctl on the 
dma-buf will block until the transfer is complete, and then userspace 
can access it again.


> Unless you want to do zero copy with a gpu, or something which is in 
> that
> ecosystem of accelerators and devices, then dma-buf is probably not 
> what
> you're looking for.
> -Daniel

I want to do zero-copy between a IIO device and the network/USB, and 
right now there is absolutely nothing in place that allows me to do 
that. So I have to get creative.

Cheers,
-Paul

>> 
>>  > >  +
>>  > >  +2. User API
>>  > >  +===========
>>  > >  +
>>  > >  +``IIO_BUFFER_DMABUF_ALLOC_IOCTL(struct iio_dmabuf_alloc_req 
>> *)``
>>  > >  
>> +----------------------------------------------------------------
>>  > >  +
>>  > >  +Each call will allocate a new DMABUF object. The return value 
>> (if
>>  > > not
>>  > >  +a negative errno value as error) will be the file descriptor 
>> of
>>  > > the new
>>  > >  +DMABUF.
>>  > >  +
>>  > >  +``IIO_BUFFER_DMABUF_ENQUEUE_IOCTL(struct iio_dmabuf *)``
>>  > >  +--------------------------------------------------------
>>  > >  +
>>  > >  +Place the DMABUF object into the queue pending for hardware
>>  > > process.
>>  > >  +
>>  > >  +These two IOCTLs have to be performed on the IIO buffer's file
>>  > >  +descriptor, obtained using the `IIO_BUFFER_GET_FD_IOCTL` 
>> ioctl.
>>  > >  +
>>  > >  +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
>>  > >  +DMABUF's file descriptor.
>>  > >  +
>>  > >  +Before accessing the data through the map, you must use the
>>  > >  +DMA_BUF_IOCTL_SYNC(struct dma_buf_sync *) ioctl, with the
>>  > >  +DMA_BUF_SYNC_START flag, to make sure that the data is 
>> available.
>>  > >  +This call may block until the hardware is done with this 
>> block.
>>  > > Once
>>  > >  +you are done reading or writing the data, you must use this 
>> ioctl
>>  > > again
>>  > >  +with the DMA_BUF_SYNC_END flag, before enqueueing the DMABUF 
>> to the
>>  > >  +kernel's queue.
>>  > >  +
>>  > >  +If you need to know when the hardware is done with a DMABUF, 
>> you
>>  > > can
>>  > >  +poll its file descriptor for the EPOLLOUT event.
>>  > >  +
>>  > >  +Finally, to destroy a DMABUF object, simply call close() on 
>> its
>>  > > file
>>  > >  +descriptor.
>>  > >  +
>>  > >  +For more information about manipulating DMABUF objects, see:
>>  > > :ref:`dma-buf`.
>>  > >  +
>>  > >  +A typical workflow for the new interface is:
>>  > >  +
>>  > >  +    for block in blocks:
>>  > >  +      DMABUF_ALLOC block
>>  > >  +      mmap block
>>  > >  +
>>  > >  +    enable buffer
>>  > >  +
>>  > >  +    while !done
>>  > >  +      for block in blocks:
>>  > >  +        DMABUF_ENQUEUE block
>>  > >  +
>>  > >  +        DMABUF_SYNC_START block
>>  > >  +        process data
>>  > >  +        DMABUF_SYNC_END block
>>  > >  +
>>  > >  +    disable buffer
>>  > >  +
>>  > >  +    for block in blocks:
>>  > >  +      close block
>>  > >  diff --git a/Documentation/iio/index.rst
>>  > > b/Documentation/iio/index.rst
>>  > >  index 58b7a4ebac51..669deb67ddee 100644
>>  > >  --- a/Documentation/iio/index.rst
>>  > >  +++ b/Documentation/iio/index.rst
>>  > >  @@ -9,4 +9,6 @@ Industrial I/O
>>  > >
>>  > >      iio_configfs
>>  > >
>>  > >  +   dmabuf_api
>>  > >  +
>>  > >      ep93xx_adc
>>  > >  --
>>  > >  2.34.1
>>  > >
>>  >
>>  > --
>>  > Daniel Vetter
>>  > Software Engineer, Intel Corporation
>>  > http://blog.ffwll.ch
>> 
>> 
> 
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch