Re: sw_ring.c poll problem

[Jonathan Cameron <jic23@kernel.org>]

On 05/19/2012 11:43 AM, Lars-Peter Clausen wrote:
> On 05/19/2012 11:50 AM, Jonathan Cameron wrote:
>> On 05/19/2012 10:41 AM, Lars-Peter Clausen wrote:
>>> On 05/19/2012 11:16 AM, Jonathan Cameron wrote:
>>>> On 05/18/2012 07:26 PM, Ge Gao wrote:
>>>>> Thanks for the patch. There could be one problem with the implementation.
>>>>> Consider the situation:
>>>>> 1. there are 10 packets in the kfifo.
>>>>> 2. Read 1 packet. Then poll again.
>>>>> 3. the poll will stuck at poll_wait(filp, &rb->pollq, wait) because it
>>>>> needs the next store to wake it up, while in reality, it should return
>>>>> immediately if there is data already in the kfifo.
>>>> Agreed that is what should happen.  I'm just getting my head around why
>>>> it doesn't (or more specifically what is different about how we are
>>>> doing it from other similar cases).
>>>
>>> I'm not sure that this is what should happen. We most certainly want to have a
>>> threshold for poll. Otherwise we'd switch back and forth between userspace and
>>> kernel space for each sample individual sample captured. This will decrease the
>>> overall system performance (because of all the context switching). So userspace
>>> should be aware of that threshold and that poll might not wake the process up
>>> even if there are samples in the buffer. ALSA basically does the same.
>> When this originally came up, one suggestion was to use one of the other
>> types of poll for this (can't recall, but maybe POLLRDBAND)
>>
>> For Ge's case it actually makes sense to do it for every scan as the
>> scans can be very large and at relatively low frequency.
> 
> Well, for the sw_ring buffer the threshold was buffer_size/2, so if you'd set
> your buffer size to 2 you'd get woken up for each sample written.
True. But the chances of loosing stuff are extremely high.  Obviously
a variable watershed would have done the trick (and not been that hard
in that implementation).  Right now I have two kfifo targets.
1) Work out how to use the fixed length record form without defining a
datatype for the record.
2) Watersheads.

Do both of those and we have covered the main stuff sw_ring had going
for it.

> 
>>
>> What you are asking for is one of the primary things that we loose in
>> moving from ring_sw to kfifo.   I keep meaning to take time to look
>> at adding watersheads to the kfifo implementation. Right now we just
>> have to dead recon in userspace and read slightly faster than data is
>> being generated.  We had some bells and whistles in that sw_ring that
>> aren't anywhere else at the moment but we'll work on that going forward.
>>
> 
> I think it is better to get rid of that stufftoread flag and add a
> samples_available() callback. It'd serve a similar purpose, but instead of just
> telling you that something is available or not it will tell you the exact
> number of samples. Which then allows us to implement stuff like watersheads in
> the generic code.
I'm warey of doing this at this stage for a couple of reasons.

a) Some buffer types allow setting a watershead but do not give easy
access to the number of elements currently there.
b) Buffers also get used in forms where this stuff has no meaning.

So lets do it in an implementation first then think about moving it out
to the general code.
> 
>>
>> I don't think there is any harm in Ge's fix of adding a check before
>> poll_wait. Worst thing it does is tell userspace there is data when
>> there isn't.  Slight overhead due to the race, but trivial given we
>> can argue the read and the poll should be in the same thread anyway
>> in almost all cases.
> 
> 
> I suppose it makes sense, but I'm still a bit confused as to why it works for
> everybody else (all the other subsystems) without this extra check.
I think this is because convention is to read everything present if
using poll.  There are cases that do the equivalent of Ge's fix to
be found though.

Certainly can't see how it work as poll man page says it should for
evdev etc. (can't find any poll code evdev as it's buried in Xorg
somewhere not directly associated with evdev)

Jonathan

> 
>>
>> Jonathan
>>>
>>> - Lars
>>>
>>>>
>>>>
>>>>> 4. It might be better to put poll inside ring buffer implementation itself
>>>>> rather than industrial-buffer.c, such as providing a new method like poll
>>>>> to do it.
>>>> Certainly could do that, but I'm not sure how it will help.
>>>>
>>>>
>>>>
>>>>> What do you think?
>>>>> Thanks.
>>>>>
>>>>> Best Regards,
>>>>>
>>>>> GE GAO
>>>>>
>>>>>
>>>>>
>>>>> -----Original Message-----
>>>>> From: Jonathan Cameron [mailto:jic23@cam.ac.uk]
>>>>> Sent: Friday, May 18, 2012 10:31 AM
>>>>> To: Ge Gao
>>>>> Cc: Jonathan Cameron; linux-iio@vger.kernel.org
>>>>> Subject: Re: sw_ring.c poll problem
>>>>>
>>>>>
>>>>>
>>>>> On 05/18/2012 02:08 AM, Ge Gao wrote:
>>>>>> Dear Jonathan,
>>>>>> 	I check the kfifo buffer implementation under IIO directory.
>>>>> However,
>>>>>> I didn't find the "pollq" is released anywhere as sw_ring does.
>>>>>> Without it, how can kfifo be polled if it is used by
>>>>> industrial-buffer.c? Thanks.
>>>>> Sorry for not getting back to you sooner.  Totally manic at the day job
>>>>> today tracking two delightful bugs.
>>>>>
>>>>> Anyhow, looks like I'd imagined we'd actually added poll support to kfifo
>>>>> when it never got implemented.  Should be easy enough though.
>>>>>
>>>>> diff --git a/drivers/iio/kfifo_buf.c b/drivers/iio/kfifo_buf.c index
>>>>> 6bf9d05..e69d6ce 100644
>>>>> --- a/drivers/iio/kfifo_buf.c
>>>>> +++ b/drivers/iio/kfifo_buf.c
>>>>> @@ -6,6 +6,7 @@
>>>>>  #include <linux/kfifo.h>
>>>>>  #include <linux/mutex.h>
>>>>>  #include <linux/iio/kfifo_buf.h>
>>>>> +#include <linux/sched.h>
>>>>>
>>>>>  struct iio_kfifo {
>>>>>  	struct iio_buffer buffer;
>>>>> @@ -35,6 +36,7 @@ static int iio_request_update_kfifo(struct iio_buffer
>>>>> *r)
>>>>>  	kfifo_free(&buf->kf);
>>>>>  	ret = __iio_allocate_kfifo(buf, buf->buffer.bytes_per_datum,
>>>>>  				   buf->buffer.length);
>>>>> +	r->stufftoread = false;
>>>>>  error_ret:
>>>>>  	return ret;
>>>>>  }
>>>>> @@ -97,6 +99,10 @@ static int iio_store_to_kfifo(struct iio_buffer *r,
>>>>>  	ret = kfifo_in(&kf->kf, data, r->bytes_per_datum);
>>>>>  	if (ret != r->bytes_per_datum)
>>>>>  		return -EBUSY;
>>>>> +	
>>>>> +	r->stufftoread = true;
>>>>> +	wake_up_interruptible(&r->pollq);
>>>>> +
>>>>>  	return 0;
>>>>>  }
>>>>>
>>>>> @@ -112,6 +118,12 @@ static int iio_read_first_n_kfifo(struct iio_buffer
>>>>> *r,
>>>>>  	n = rounddown(n, r->bytes_per_datum);
>>>>>  	ret = kfifo_to_user(&kf->kf, buf, n, &copied);
>>>>>
>>>>> +	if (kfifo_is_empty(&kf->kf))
>>>>> +		r->stufftoread = false;
>>>>> +	/* verify it is still empty to avoid race */
>>>>> +	if (!kfifo_is_empty(&kf->kf))
>>>>> +		r->stufftoread = true;
>>>>> +
>>>>>  	return copied;
>>>>>  }
>>>>>
>>>>> .. is a totally untested patch to add poll support to it.
>>>>> The little dance at the end is to avoid a write having occured between the
>>>>> kfifo_is_empty and setting the flag false as that could wipe out the
>>>>> existence of some data in the buffer. Given we only ever have one writer
>>>>> and one reader that should work I think.
>>>>>
>>>>> Jonathan
>>>>>
>>>>>>
>>>>>> Hi Ge,
>>>>>>
>>>>>> I realised after sending that message that I was being rather
>>>>>> dismissive of your query.  Got up far too early this morning (as every
>>>>>> morning ;)
>>>>>>
>>>>>> Anyhow, to give more details. sw_ring is probably never going to make
>>>>>> it out of staging, hence the move to kfifo_buf. At somepoint we need
>>>>>> to work out how to do equivalent functionality of sw_ring but I've not
>>>>>> had time to more than start looking into this.
>>>>>>
>>>>>> As you saw, poll on sw_ring is a watershead signal indicating (in
>>>>>> theory and last I checked it worked) that the ring is more than half
>>>>> full.
>>>>>> Any read that takes the fill level below half (test code just reads
>>>>>> half the size of the buffer), should allow a new passing of the
>>>>>> watershead to resignal poll. It's entirely possible there is a bug in
>>>>>> there though I know it is been getting a fair bit of testing with some
>>>>>> other drivers so could be todo with the precise way you are reading it
>>>>>> hitting some corner case? (I'm
>>>>>> stretching...)
>>>>>>
>>>>>> Right now I'd just move over to kfifo_buf if I were you. It's much
>>>>>> more 'standard' in that it's a fifo and poll indicates if there is
>>>>>> anything there at all.
>>>>>>>> 	Thanks.
>>>>>>>>
>>>>>>>> Best regards,
>>>>>>>>
>>>>>>>> Ge GAO
>>>>>>>> --
>>>>>>>> To unsubscribe from this list: send the line "unsubscribe linux-iio"
>>>>>>>> in the body of a message to majordomo@vger.kernel.org More majordomo
>>>>>>>> info at  http://vger.kernel.org/majordomo-info.html
>>>>>>>
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-iio" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>