Quoting Jonathan Cameron (2024-10-12 17:47:32)
> On Wed, 09 Oct 2024 20:30:15 +0200
> Matteo Martelli <matteomartelli3@xxxxxxxxx> wrote:
>
> > Quoting Nuno Sá (2024-10-08 14:37:22)
> > > On Tue, 2024-10-08 at 10:03 +0200, Matteo Martelli wrote:
> > > > Quoting Nuno Sá (2024-10-08 09:29:14)
> > > > > On Tue, 2024-10-08 at 08:47 +0200, Matteo Martelli wrote:
> > > > > > Quoting Nuno Sá (2024-10-07 17:15:13)
> > > > > > > On Mon, 2024-10-07 at 10:37 +0200, Matteo Martelli wrote:
> > > > > > > > Consumers need to call the read_avail_release_resource after reading
> > > > > > > > the
> > > > > > > > available info. To call the release with info_exists locked, copy the
> > > > > > > > available info from the producer and immediately call its release
> > > > > > > > callback. With this change, users of iio_read_avail_channel_raw() and
> > > > > > > > iio_read_avail_channel_attribute() must free the copied avail info
> > > > > > > > after
> > > > > > > > calling them.
> > > > > > > >
> > > > > > > > Signed-off-by: Matteo Martelli <matteomartelli3@xxxxxxxxx>
> > > > > > > > ---
> > > > > > > > drivers/iio/inkern.c | 64 +++++++++++++++++++++++++++++++++--
> > > > > > > > ----
> > > > > > > > -----
> > > > > > > > include/linux/iio/consumer.h | 4 +--
> > > > > > > > 2 files changed, 50 insertions(+), 18 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/drivers/iio/inkern.c b/drivers/iio/inkern.c
> > > > > > > > index
> > > > > > > > 7f325b3ed08fae6674245312cf8f57bb151006c0..cc65ef79451e5aa2cea447e16800
> > > > > > > > 7a44
> > > > > > > > 7ffc0d91
> > > > > > > > 100644
> > > > > > > > --- a/drivers/iio/inkern.c
> > > > > > > > +++ b/drivers/iio/inkern.c
> > > > > > > > @@ -760,9 +760,25 @@ static int iio_channel_read_avail(struct
> > > > > > > > iio_channel
> > > > > > > > *chan,
> > > > > > > > if (!iio_channel_has_available(chan->channel, info))
> > > > > > > > return -EINVAL;
> > > > > > > >
> > > > > > > > - if (iio_info->read_avail)
> > > > > > > > - return iio_info->read_avail(chan->indio_dev, chan-
> > > > > > > > >channel,
> > > > > > > > - vals, type, length, info);
> > > > > > > > + if (iio_info->read_avail) {
> > > > > > > > + const int *vals_tmp;
> > > > > > > > + int ret;
> > > > > > > > +
> > > > > > > > + ret = iio_info->read_avail(chan->indio_dev, chan-
> > > > > > > > >channel,
> > > > > > > > + &vals_tmp, type, length,
> > > > > > > > info);
> > > > > > > > + if (ret < 0)
> > > > > > > > + return ret;
> > > > > > > > +
> > > > > > > > + *vals = kmemdup_array(vals_tmp, *length, sizeof(int),
> > > > > > > > GFP_KERNEL);
> > > > > > > > + if (!*vals)
> > > > > > > > + return -ENOMEM;
> > > > > > > > +
> > > > > > >
> > > > > > > Not a big deal but I would likely prefer to avoid yet another copy. If
> > > > > > > I'm
> > > > > > > understanding things correctly, I would rather create an inkern wrapper
> > > > > > > API
> > > > > > > like
> > > > > > > iio_channel_read_avail_release_resource() - maybe something with a
> > > > > > > smaller
> > > > > > > name :).
> > > > > > > Hence, the lifetime of the data would be only controlled by the producer
> > > > > > > of
> > > > > > > it. It
> > > > > > > would also produce a smaller diff (I think). I just find it a bit
> > > > > > > confusing
> > > > > > > that we
> > > > > > > duplicate the data in here and the producer also duplicates it on the -
> > > > > > > > read_avail()
> > > > > > > call. Another advantage I see is that often the available data is indeed
> > > > > > > const in
> > > > > > > which case no kmemdup_array() is needed at all.
> > > > > >
> > > > > >
> > > > > > If I understand correctly your suggestion you would leave the inkern
> > > > > > iio_channel_read_avail() untouched, then add a new inkern wrapper,
> > > > > > something
> > > > > > like iio_channel_read_avail_release_resource(), that would call the
> > > > > > producer's
> > > > > > read_avail_release_resource(). The consumer would invoke this new wrapper
> > > > > > in
> > > > > > its
> > > > > > own read_avail_release_resource() avoiding the additional copy. The call
> > > > > > stack
> > > > > > would look something like the following:
> > > > > >
> > > > > > iio_read_channel_info_avail() {
>
> Here you are talking about the core code that produces a string.
> But you've done that in reply to v5 which is about inkern interfaces
> that don't work in strings...
>
>
>
> > > > > > consumer->read_avail() {
> > > > > > iio_read_avail_channel_raw() {
> > > > > > iio_channel_read_avail() {
> > > > > > producer->read_avail() {
> > > > > > kmemdup_array();
> > > > > > }
> > > > > > }
> > > > > > }
> > > > > > }
> > > > > >
> > > > > > iio_format_list();
> That's effectively making the safe copy that is handed back to the
> caller. So this is fine.
> > > > > >
> > > > > > consumer->read_avail_release_resource() {
> > > > > > iio_read_avail_channel_release_resource() {
> > > > > > producer->read_avail_release_resource() {
> > > > > > kfree();
> > > > > > }
> > > > > > }
> > > > > > }
> > > > > > }
> > > > >
> > > > > Yeah, exactly what came to mind...
>
> I'm very confused what scope of comments here is. Maybe the easiest thing is to send the code.
>
The function call graph example and the related comments above were just
to show the full flow of a read_avail() call on a consumer driver in
case an inkern release_resource() wrapper was used instead of an
additional consumer copy. This just to confirm that I was on the same
page as what Nuno suggested.
>
> > > > >
> > > > > >
> > > > > >
> > > > > > I was going with the simpler solution you described, but my concern with
> > > > > > it
> > > > > > was
> > > > > > that the info_exists_lock mutex would be unlocked between a
> > > > > > iio_channel_read_avail()
> > > > > > call and its corresponding iio_channel_read_avail_release_resource() call.
> > > > > > To my understanding, this could potentially allow for the device to be
> > > > > > unregistered between the two calls and result in a memleak of the avail
> > > > > > buffer
> > > > > > allocated by the producer.
>
> Yes. That's why we have to free the produced copy under the exist_lock
> (and take yet another copy).
>
> > > > > >
> > > > > > However, I have been trying to reproduce a similar case by adding a delay
> > > > > > between the consumer->read_avail() and the
> > > > > > consumer->read_avail_release_resources(), and by unbinding the driver
> > > > > > during
> > > > > > that delay, thus with the info_exists_lock mutex unlocked. In this case
> > > > > > the
> > > > > > driver is not unregistered until the iio_read_channel_info_avail()
> > > > > > function
> > > > > > completes, likely because of some other lock on the sysfs file after the
> > > > > > call
> > > > > > of
> > > > > > cdev_device_del() in iio_device_unregister().
> > > > > >
> > > > >
> > > > > Yes, you need to have some sync point at the kernfs level otherwise we could
> > > > > always be handling a sysfs attr while the device is being removed under our
> > > > > feet. But I'm not sure what you're trying to do... IIUC, the problem might
> > > > > come
> > > > > if have:
> > > > >
> > > > > consumer->read_avail_channel_attribute()
> > > > > producer->info_lock()
> > > > > producer->read_avail()
> > > > > producer->kmalloc()
> > > > >
> > > > > ...
> > > > > // producer unbound
> > > > > ...
> > > > > consumer->read_avail_release()
> > > > > return -ENODEV;
> > > > >
> > > > > // producer->kmalloc() never get's freed...
> > > > >
> > > > > The above is your problem right? And I think it should be a valid one since
> > > > > between ->read_avail_channel_attribute() and read_avail_release() there's
> > > > > nothing preventing the producer from being unregistered...
> > > >
> > > > Yes, that's the problem.
> > > >
> > > > >
> > > > > If I'm not missing nothing one solution would be for the producer to do
> > > > > devm_kmalloc() and devm_kfree() on read_avail() and release_resources() but
> > > > > at
> > > > > that point I'm not sure it's better than what you have since it's odd enough
> > > > > for
> > > > > being missed in reviews...
>
> If it's an allocation to keep a copy for an active consumer then that is nasty
> as the lifetime is completely untidied. Effectively you are garbage collecting.
>
> > > >
> > > > I honestly didn't think of this and it would in fact prevent the
> > > > additional copy. But I agree that it could be missed in new drivers,
> > > > maybe a comment in the iio_info read_avail_release_resource() callback
> > > > declaration would help?
> > > > >
> > > At this point I would say whatever you or Jonathan prefer :)
> > >
> >
> > I run some quick tests with this approach and haven't found any issue so
> > far.
>
> I can't see what is preventing the race you describe where the
> release callback is swept out by a driver unbind of the provider.
> So unless we can show that is safe I don't see a way to avoid a consumer copy.
>
> Long shot, Lars-Peter I think you fixed up most of the previous races in this
> code, don't suppose you remember how it works?
>
> > I would personally switch to this approach as it would be much
> > simpler and easier to understand, and since the avail lists are const
> > for most of the current drivers I would not expect many new drivers
> > needing a dynamic available list. However, I will wait Jonathan feedback
> > first.
> The idea here is almost no one actually makes a copy in the producer.
> The consumer copy is a necessity to my thinking because we are effectively
> passing the ownership of the data. Unfortunately we have no idea how
> the producer would free it so we need to create our own copy.
>
You are indeed right. During the last comment reviews I focused on the
fact that using a devm_kfree() in the producer's read_avail_release()
would prevent the memleak but I completely missed that it would not
prevent the race condition: without the consumer copy, the consumer or
core code could still try to access a freed buffer after the producer
driver gets unregistered.
Thus I will keep this patch with the consumer copy as it is. I am soon
going to send a next version of this patch series to address the other
findings during this version.
> Key here is this a very rare operation. No one polls available
> data at high frequency, it's a read once kind of thing typically.
> + these are normally really short lists in the cases we actually use
> (so far I 'think' they have always been the 3 value range form, not
> a list of potential values).
>
> Jonathan
>
>
Thanks,
Matteo Martelli
