Re: [PATCH 1/1] IIO: Added write function in iio_buffer_fileops

[Jonathan Cameron]

On 16/08/14 15:44, Aniroop Mathur wrote:
> 
> 
> 
> On Thu, Aug 14, 2014 at 8:08 PM, Jonathan Cameron <jic23@xxxxxxxxxx <mailto:jic23@xxxxxxxxxx>> wrote:
> 
>     On 14/08/14 10:41, Lars-Peter Clausen wrote:
>     > On 08/13/2014 06:33 PM, Aniroop Mathur wrote:
>     >> On Wed, Aug 13, 2014 at 8:18 PM, Lars-Peter Clausen <lars@xxxxxxxxxx <mailto:lars@xxxxxxxxxx>> wrote:
>     >>> On 08/13/2014 03:42 PM, Aniroop Mathur wrote:
>     >>>>
>     >>>> On Wed, Aug 13, 2014 at 2:38 PM, Lars-Peter Clausen <lars@xxxxxxxxxx <mailto:lars@xxxxxxxxxx>>
>     >>>> wrote:
>     >>>>>
>     >>>>> On 08/13/2014 08:29 AM, a.mathur@xxxxxxxxxxx <mailto:a.mathur@xxxxxxxxxxx> wrote:
>     >>>>>>
>     >>>>>>
>     >>>>>> From: Aniroop Mathur <a.mathur@xxxxxxxxxxx <mailto:a.mathur@xxxxxxxxxxx>>
>     >>>>>>
>     >>>>>> Earlier, user space can only read from iio device node but cannot write
>     >>>>>> to
>     >>>>>> it.
>     >>>>>> This patch adds write function in iio buffer file operations,
>     >>>>>> which will allow user-space applications/HAL to write the data
>     >>>>>> to iio device node.
>     >>>>>> So now there will be two way communication between IIO subsystem
>     >>>>>> and user space. (userspace <--> kernel)
>     >>>>>>
>     >>>>>> It can be used by HAL or any user-space application which wants to
>     >>>>>> write data to iio device node/buffer upon receiving some data from it.
>     >>>>>> As an example,
>     >>>>>> It is useful for iio device simulator application which need to record
>     >>>>>> the data by reading from iio device node and replay the recorded data
>     >>>>>> by writing back to iio device node.
>     >>>>>>
>     >>>>>
>     >>>>> I'm not convinced that this is something that should be added to the
>     >>>>> kernel.
>     I'm inclined to agree with Lars.  As an additional point this will cause
>     confusion when we have buffered writing for output devices (DACs).
> 
> 
>     >>>>> I'm wondering why can't this be done in userspace, e.g. by having a
>     >>>>> simulator mode for the application or by using LD_PRELOAD. Having this in
>     >>>>> userspace will be much more flexible and will be easier to implement
>     >>>>> correctly and you'll most likely want to simulate more than just buffer
>     >>>>> access, for example setting/getting properties of the device or channel.
>     >>>>> For
>     >>>>> the libiio[1] we are planning to implement a test backend that when
>     >>>>> activated will allow to simulate a whole device rather than just buffer
>     >>>>> support.
>     >>>>>
>     >>>>> [1] https://github.com/analogdevicesinc/libiio
>     >>>>>
>     >>>>>
>     >>>>
>     >>>> In normal Input Subsystem, there is two way communication between
>     >>>> kernel and userpace. It has both read and write functionality. :)
>     >>>> Why there is only one way communication in case of IIO ?
>     Why should there be?
> 
> 
> 
> Below is the use case for which write function is required:
> 
> I am working on Sensor Simulator application for android phones.
A worthy goal!
> This android application will simulate the sensor values 
> for any other already developed 3rd party/organization application/s 
> and not for simulator application itself.
> In other words, I need to check the working of other sensor applications by
> using my simulator application and along with it, checking of HAL and 
> Framework too.
> 
> For simulation, below is the data flow 
> by using user-space libiio and by directly writing to iio buffer.
> 
> Hardware--Driver--IIO--General_IIO_HAL--Framework--Other_Sensor_Application/s
>   |
>   | --> libiio <--> Simulator_Sensor_Application
>   |
>                    | <--> Simulator_Sensor_Application
> 
> 
> If we use libiio, 
> Sensor_Simulator_Application will not be able to send sensor data to 
> Other_Sensor_Application/s because it can only write and read data for itself.
> Reading and writing data for itself is not what is required. There is a need
> to send data to other application/s.
> Also, we could not check HAL and Framework here.
> 
> If we directly write to iio buffer,

This is a non starter via the approach currently suggested.  Writing
to an IIO buffer means writing to an output device (or will soon),
not shoving fake data into a driver that has been artificially hacked
to not send real data (that disabling the real device but keeping the
buffer running is a bit I really do not like.)
If you want to do this it should be via a 'fake' driver not a shim on
a real one.  To do this in a safe way would mean adding additional
locking into the fast data flow path for a fairly uncommon case.

I would happily except such a fake driver as a useful addition for
test purposes etc.  The trick here would be in developing a configuration
interface that allows a new 'fake' driver to be brought up with whatever
set of channels and interfaces is desired, along with the mechanisms for
providing data to those channels.

> Sensor_Simulator_Application will be able to send sensor data to 
> Other_Sensor_Application/s and this in turn will also check whether there
> are any problems in hal and framework for Other_Application to work properly.
> So, if some Other_Sensor_Application is behaving differently 
> for same sensor data it means there is some problem in 
> HAL/Framework/Application itself.
Sure, I understand your point here, but I would still not do it via
tricks on a normal driver.  These other programs will just as happily
connect to a 'new' device driver that offers the right facilities.

> 
> Also importantly, we can change the recorded data and check Other_Application 
> working with that new data.
> As an example,
> There is an application, which uses accel+gyro+pressure to calculate 
> amount of calories burnt of an individual.
> So once recorded the data physically by going in outside environment,
> we can change just the pressure data and measure the amount of calories burnt
> without actually physically going outside at new height for new pressure.

I'd argue that this is a userspace application problem.  If the data is not
available then userspace should cope with it, not play games with faking the
data apparently coming from the hardware.

A fake driver would work for this as well if people really want it!

> 
> Now to write directly to IIO buffer, there is a need of two things
> that should be present in IIO subsystem:
> 1. Write function in iio_buffer file_ops.

Not on the main buffer.  That has a very specific meaning and it isn't this.
The 'right' way to do this in my mind is to have a driver that can
hook up an out buffer to an in buffer (with no real hardware associated).
That would fit nicely into the interface for normal use of IIO and would be
a very handy test tool in general as we could playback 'interesting' datasets
and use it to test the various bridge drivers as well.

The only challenge is in the configuration interface to bring up such a device
with the right channels!

> 2. Multiple file descriptor support for same device node.
>    because both Simulator_Sensor_Application and Other_Sensor_Application/s
>    need to open device node for their use.

Firstly we can actually support multiple buffers just fine, but the target
for these is not repeating the same userspace interface lots of times.

Secondly, No they don't need to  This is best done in userspace as we have
emphasised in a number of previous dicussions. The requirements are quite different from
what is needed from Input and even there filtering input does adds latency to
it's data flow (and a great deal of complexity).  Input is designed primarily
for slow devices where inline meta data is reasonable.  IIO has to be designed
to handle high speed devices where that just isn't feasible.  Try adding meta
data to the multiple mega sample per second devices Lars has running.

For IIO, multiple buffers each need their own configuration
interfaces as there is no meta data in the data stream to allow idenficiation
of what is coming out.

Any change to the flow on any buffer can result in hardware changes effecting
all other buffers (typically stopping the data flow, changing the channels
being grabbed and bringing it back up again).  These changes might well slow
down the rate data is provided to another driver as various tricks that apply
with only one buffer stop working.

Userspace can mediate between multiple devices and greatly reduce the flow
of data across the kernel userspace boundary.

Now we do have the facility to do this in kernel and may perhaps support
IIO on IIO at some point (IIO drivers that are consumers of other IIO drivers)
thus providing the 'ability' to do what you ask.  However the purpose behind
this would be to allow the underlying devices userspace interface to be dropped
if for example, the adc is just feeding an input subsystem bridge driver.
We are working slowly towards this, but only to allow IIO to be stripped back
to what is required rather than to allow more complex options.

So long term, it will be possible to have multiple buffers available (and
created on the fly) but there is a fair bit of work to be done first.
>    
> Currently, both are not present in IIO subsystem.
> So first, I am hoping to add write function and 
> after that adding multiple fd support. 

I am sorry to say that I'm far from convinced that either is a good idea if
you aim to do it in a similar way to Input.

The second has been proposed a number of times before.
> 
> In input subsystem, both facilities are available.

Sorry to put it so bluntly, but IIO has some very different targets from input.
It needs to be fast.  Input does not (although this is getting a little more
interesting with mulit touch screens).  That is what IIO is fundamentally different
from input.  Yes, there are devices in IIO that are slow where perhaps it might
be nice to make userspaces job a little easier, but any changes have to take
into account what problems they will cause for the fast end of things.

Ultimately if it's really a problem, then work on the input bridge driver,
(which has been languishing in the todo heap for a long time now).  If people
want these facilities then they can use that to make any IIO device accessible
through input.

> 
> How can we achieve this task without directly writing to iio buffer ?

Fake device - your android kernel would then appear to have multiple similar
devices, one of which is real and one of which appears to have a matched set
of output and input channels.  This does give us the pleasing option of
having interfaces for outputting pressure, acceleration and magnetic field :)

The trick here is how to instantiate such a device.   Sounds like a job for configfs
to me though will take a bit of thinking to get the interface correct.

> How will user-space libiio help us for this task ?

Clearly it will only work if it becomes the standard option on Android :)

Note as well that a lot of IIO device drives are slow enough that the buffered
interface will never make sense and so there will always need to be a means to
simply read values from sysfs attributes.

>  
> 
>     >>>
>     >>>
>     >>> I've not seen a compelling reason yet why this must be implemented in kernel
>     >>> space. In my opinion, as outlined above, userspace if both easier and more
>     >>> flexible.
>     >>>
>     >>>
>     >>>>
>     >>>> For Input devices, I completed simulation just by reading and writing at
>     >>>> input device node /dev/input/eventX.
>     >>>> But for IIO devices, I am stuck as there is no write function available
>     >>>> for iio device node /dev/iio:device0.
>     >>>>
>     >>>> As per my understanding, if we do the simulation in userspace
>     >>>> then we need to create one more buffer in userpace. This will lead to
>     >>>> extra memory usage. With write functionality in IIO just like
>     >>>> in Input subsystem, we can avoid extra memory space. :)
>     >>>
>     >>>
>     >>> No, you don't actually have to create a additional buffer. Just return the
>     >>> data that you'd otherwise have passed to write().
>     >>>
>     >>>
>     >>
>     >> If there is no buffer, then there is clearly a chance of data loss/miss. :)
>     >> Because if one application is reading the data with frequency 5 Hz
>     >> and other application is writing the data at frequency 50 Hz (20 ms delay)
>     >> so this reading application will miss reading a lot of data.
>     >> Like in this case, after every 200 ms, 9 out of 10 data will be missed.
>     >
>     > Not if implemented correctly. Even with the current kernel implementation you'll have this issues as the buffer will
>     > simply overflow when you write faster than you read.
>     >
> 
> 
> 
> Generally, there is less difference between read and write frequency.
> But there is difference.
> So, in case of a buffer, there is very very less chance of data loss.
> But in case of no buffer, there will be data loss/miss.
Clearly any such system will need a buffer of some sort to remain efficient.
If done in userspace this could even be a file that is being read.
> 
> Generally, as you know, application requests for frequency at which
> hardware should send data for their reading.
> But driver frequency to write data in buffer 
> and application frequency to read data does not match exactly.
> So either fast or slow, we end with either data loss 
> or same data being sent again to app.
> 
> Therefore, buffer is required to avoid such cases.
> 
> 
>  
> 
>     >
>     > [...]
>     If we have a usecase for playback functionality like this I would prefer it to
>     be on a separate IIO device.  When Lars' DAC buffered code is ready it will
>     be relatively easy to string together a fake DAC with a fake ADC to get
>     this sort of functionality.   It will be rather more interesting to
>     work out how to setup an arbitary device but it could be done,
>     most likely using configfs.
> 
>     This would be more similar to uinput than the writing to the chardevs
>     directly as you are suggesting here.
> 
>     >>>>> Are you sure that this works? iio_push_to_buffer() expects a data buffer
>     >>>>> of
>     >>>>> size rb->bytes_per_datum bytes. On the other hand rb->bytes_per_datum is
>     >>>>> only valid when the buffer is enabled, so for this to work the buffer
>     >>>>> would
>     >>>>> need to be enabled. Which means you'd inject the fake data in the middle
>     >>>>> of
>     >>>>> the real data stream.
>     >>>>>
>     >>>>>
>     >>>>
>     >>>> Yes, It works :)
>     >>>> In one patch, bytes_per_datum has been removed from kifo_in.
>     >>>> Patch - iio:kfifo_buf Take advantage of the fixed record size used in IIO
>     >>>> commit c559afbfb08c7eac215ba417251225d3a8e01062
>     >>>> - ret = kfifo_in(&kf->kf, data, r->bytes_per_datum);
>     >>>> + ret = kfifo_in(&kf->kf, data, 1);
>     >>>> So, I think we can now write only one byte of data.
>     >>>
>     >>>
>     >>> No, we need to write 1 record and the size of one record is bytes_per_datum.
>     >>> If you only write one byte you'll cause a out of bounds access.
>     >>>
>     >>>
>     >>
>     >> This is the code flow below which I checked:
>     >>
>     >> kfifo_in(&kf->kf, data, 1);
>     >> so len=1
>     >> kfifo_copy_in(fifo, buf, len, fifo->in);
>     >> l = min(len, size - off);
>     >> memcpy(fifo->data + off, src, l);
>     >>
>     >> In memcpy, if l is 1, so it will copy one byte only.
>     >> So, how it is writing one record and not one byte ?
>     > You missed this part:
>     >
>     >     if (esize != 1) {
>     >         off *= esize;
>     >         size *= esize;
>     >         len *= esize;
>     >     }
>     >
>     > so len gets multiplied by the record size. len=1 means 1 record.
>     >
> 
> 
> Oh, I really missed this part.
> Thank you Mr. Lars for the correction.
> 
> So, in the write function, 
> we can just replace 1 by rb->bytes_per_datum.
> And add the check like below:
> + if(!rb || rb->bytes_per_datum==0)
> + return -1;
> 
> I initially write the code with bytes_per_datum only.
> But changed to 1 after seeing the latest kernel and
> got confused with "1" value.
> 
> Is there anything else need to be changed ?
> 
>  
> 
>     >>
>     >>>>
>     >>>> Initially, I wrote the code for write functionality in kernel version 3.6
>     >>>> using bytes_per_datum instead of fixed size of 1 byte.
>     >>>> It worked fine. :)
>     >>>> For this, we just need to replace size 1 by r->bytes_per_datum.
>     >>>>
>     >>>> We are not injecting data in middle of real data stream.
>     >>>> When we inject the recorded data, we disabled the hardware chip,
>     >>>> so no new/real data is pushed to the buffer during that time.
>     >>>>
>     >>>> To record, we enabled the buffer, read the real data and save it.
>     >>>> To replay, we disabled the hardware chip and injected saved data by
>     >>>> writing back to iio device node.
>     >>>> So, Buffer is still enabled at time of writing to iio device node. :)
>     >>>
>     >>>
>     >>> How do you disable the hardware without disabling the buffer?
>     >>>
>     >> I disabled the hardware by powering off the chip.
>     >> And after writing is complete, chip is powered on again. :)
>     >
>     > But how do you disable the device when the buffer is still active?
>     > IIO expects the device to be active when the buffer is active.
>     Agreed. This sounds like a pretty uggly hack.  I would not be happy with
>     having this in any driver.
> 
> 
> Taking example of pressure sensor hardware chip,
> 
> Case1: Normal case,
> 1. We enabled the iio software buffer.
> 2. Pressure Hardware chip generates an interrupt.
> 3. Driver receives the interrupt call and make
>    a call to push data to iio buffer
> 
> But if there is no change in pressure i.e. there is no new data
> hardware chip will not generate any interrupt and hence
> no data is pushed to iio software buffer.
That's actually unusual for an IIO device.  Unlike input which won't kick
out data if there is no change, IIO devices typically don't assess that at all
and will push data based on timing.

There are devices that do support hysteresis on their triggers (thus requiring
a significant change in value) but they are pretty unusual.

> And iio software buffer is still active and waiting if there is any data comes.
> 
> Case 2: My case,
> 1. We enable the iio software buffer.
> 2. We record the pressure real data if there is.
> 3. We power off the pressure hardware chip 
>    or second way, we disable the irq only.

Nasty and hacky.   Putting aside that I think this is generally a bad idea as
explained above.  If you do this, you need to allow pushing of data with the interrupt
still running with appropriate locking to prevent anything nasty occuring.
Otherwise, you've just killed off the real reading.  The snag here is that you
have no way of identifying the real reading from fake ones?  Which is right?

I'm not going to be easily persuaded to take any driver that does this sort of
faking without it being readily apparent to userspace that this is what is going
on.

> 
> In this case too, there will be no interrupt call received by driver,
> and hence no data is pushed to iio buffer.
> And iio software buffer is still active and waiting if there is any data comes.
> 
> In both cases, from iio buffer perspective,
> iio software buffer is active and waiting for some data.
> 
> I cannot see any difference for IIO buffer in both cases.
> So, I am not able to understand the problem here with IIO buffer ?
> 
> Later in case 2, as buffer is active,
> Sensor_Simulator_Application can write data to it
> and Other_Sensor_Application/s can read data from it.
> And after simulation is over,
> we power on the hardware chip again or second way, enable the irq again,
> everything gets back to normal as before.
> 
Of course you are welcome to do what you like, but I'm only going to be
interested in a true 'fake' driver to provide this functionality without
ever pretending to be real.  Such a driver should be easy to userspace
to identify and elect to ignore if it wants to.

Sorry it has taken me so long to get back - been a busy few weeks.
Also if this appears overly negative, remember that with my maintainers
hat on I have to look at every proposal to change the IIO core and consider
how it effects all our current use cases.

Jonathan

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/