On Tue, Mar 12, 2019 at 5:48 PM Bingbu Cao <bingbu.cao@xxxxxxxxxxxxxxx> wrote: > > > > On 03/12/2019 03:43 PM, Tomasz Figa wrote: > > On Tue, Mar 12, 2019 at 3:48 PM Bingbu Cao <bingbu.cao@xxxxxxxxxxxxxxx> wrote: > >> > >> > >> On 03/12/2019 01:33 PM, Tomasz Figa wrote: > >>> Hi Bingbu, > >>> > >>> On Fri, Feb 15, 2019 at 6:02 PM <bingbu.cao@xxxxxxxxx> wrote: > >>>> From: Bingbu Cao <bingbu.cao@xxxxxxxxx> > >>>> > >>>> Current ImgU driver processes and releases the parameter buffer > >>>> immediately after queued from user. This does not align with other > >>>> image buffers which are grouped in sets and used for the same frame. > >>>> If user queues multiple parameter buffers continuously, only the last > >>>> one will take effect. > >>>> To make consistent buffers usage, this patch changes the parameter > >>>> buffer handling and group parameter buffer with other image buffers > >>>> for each frame. > >>> Thanks for the patch. Please see my comments inline. > >>> > >>>> Signed-off-by: Tianshu Qiu <tian.shu.qiu@xxxxxxxxx> > >>>> Signed-off-by: Bingbu Cao <bingbu.cao@xxxxxxxxx> > >>>> --- > >>>> drivers/staging/media/ipu3/ipu3-css.c | 5 ----- > >>>> drivers/staging/media/ipu3/ipu3-v4l2.c | 41 ++++++++-------------------------- > >>>> drivers/staging/media/ipu3/ipu3.c | 24 ++++++++++++++++++++ > >>>> 3 files changed, 33 insertions(+), 37 deletions(-) > >>>> > >>>> diff --git a/drivers/staging/media/ipu3/ipu3-css.c b/drivers/staging/media/ipu3/ipu3-css.c > >>>> index b9354d2bb692..bcb1d436bc98 100644 > >>>> --- a/drivers/staging/media/ipu3/ipu3-css.c > >>>> +++ b/drivers/staging/media/ipu3/ipu3-css.c > >>>> @@ -2160,11 +2160,6 @@ int ipu3_css_set_parameters(struct ipu3_css *css, unsigned int pipe, > >>>> obgrid_size = ipu3_css_fw_obgrid_size(bi); > >>>> stripes = bi->info.isp.sp.iterator.num_stripes ? : 1; > >>>> > >>>> - /* > >>>> - * TODO(b/118782861): If userspace queues more than 4 buffers, the > >>>> - * parameters from previous buffers will be overwritten. Fix the driver > >>>> - * not to allow this. > >>>> - */ > >>> Wouldn't this still happen even with current patch? > >>> imgu_queue_buffers() supposedly queues "as many buffers to CSS as > >>> possible". This means that if the userspace queues more than 4 > >>> complete frames, we still end up overwriting the parameter buffers in > >>> the pool. Please correct me if I'm wrong. > >> The parameter buffers are queued to CSS sequentially and queue one > >> parameter along with one input buffer once ready, all the data and > >> parameter buffers are tied together to queue to the CSS. If userspace > >> queue more parameter buffers then input buffer, they are pending on the > >> buffer list. > > It doesn't seem to be what the code does. I'm talking about the > > following example: > > > > Queue OUT buffer 1 > > Queue PARAM buffer 1 > > Queue IN buffer 1 > > Queue OUT buffer 2 > > Queue PARAM buffer 2 > > Queue IN buffer 2 > > Queue OUT buffer 3 > > Queue PARAM buffer 3 > > Queue IN buffer 3 > > Queue OUT buffer 4 > > Queue PARAM buffer 4 > > Queue IN buffer 4 > > Queue OUT buffer 5 > > Queue PARAM buffer 5 > > Queue IN buffer 5 > > > > All the operations happening exactly one after each other. How would > > the code prevent the 5th PARAM buffer to be queued to the IMGU, after > > the 5th IN buffer is queued? As I said, imgu_queue_buffers() just > > queues as many buffers of each type as there are IN buffers available. > So the parameter pool now is only used as record last valid parameter not > used as a list or cached, all the parameters will be queued to CSS as soon as > possible(if queue for CSS is not full). > As the size of pool now is a bit confusing, I think we can shrink the its value > for each pipe to 2. I don't follow. Don't we take one buffer from the pool, fill in the parameters in hardware format there and then queue that buffer from the pool to the ISP? The ISP wouldn't read those parameters from the buffer until the previous frame is processed, would it? > > The buffer queue size is limited for CSS, if massive buffers from > user pass down and the css queue is full, driver will get -EBUSY > and return the buffers to user with error. > That's not a correct behavior. If the VB2 queue has N buffers, the userspace needs to be able to queue all the N buffers. The driver should account the queued buffers internally and queue them to the hardware/firmware only as much as the hardware/firmware allows. Best regards, Tomasz
