[libcamera-devel] [PATCH v5 2/4] libcamera: V4L2Device: Use std::vector for v4l2_ext_control in getControls()

[Hirokazu Honda]

On Mon, Apr 26, 2021 at 11:46 AM Laurent Pinchart
<laurent.pinchart at ideasonboard.com> wrote:
>
> Hi Hiro,
>
> On Mon, Apr 26, 2021 at 11:01:11AM +0900, Hirokazu Honda wrote:
> > On Mon, Apr 26, 2021 at 9:12 AM Laurent Pinchart wrote:
> > > On Fri, Apr 23, 2021 at 06:36:51PM +0900, Hirokazu Honda wrote:
> > > > The original code uses Variable-Length-Array, which is not
> > > > officially supported in C++. This replaces the array with
> > > > std::vector.
> > > >
> > > > Signed-off-by: Hirokazu Honda <hiroh at chromium.org>
> > > > ---
> > > >  src/libcamera/v4l2_device.cpp | 41 +++++++++++++++++------------------
> > > >  1 file changed, 20 insertions(+), 21 deletions(-)
> > > >
> > > > diff --git a/src/libcamera/v4l2_device.cpp b/src/libcamera/v4l2_device.cpp
> > > > index ce2860c4..bbe8f154 100644
> > > > --- a/src/libcamera/v4l2_device.cpp
> > > > +++ b/src/libcamera/v4l2_device.cpp
> > > > @@ -173,13 +173,18 @@ void V4L2Device::close()
> > > >   */
> > > >  ControlList V4L2Device::getControls(const std::vector<uint32_t> &ids)
> > > >  {
> > > > -     unsigned int count = ids.size();
> > > > -     if (count == 0)
> > > > +     if (ids.empty())
> > > >               return {};
> > > >
> > > >       ControlList ctrls{ controls_ };
> > > > +     std::vector<v4l2_ext_control> v4l2Ctrls(ids.size());
> > > > +     memset(v4l2Ctrls.data(), 0, sizeof(v4l2_ext_control) * ctrls.size());
> > > > +
> > > > +     for (size_t i = 0, j = 0; j < ids.size(); ++j) {
> > > > +             const unsigned int id = ids[j];
> > > > +             if (ctrls.contains(id))
> > > > +                     continue;
> > > >
> > > > -     for (uint32_t id : ids) {
> > > >               const auto iter = controls_.find(id);
> > > >               if (iter == controls_.end()) {
> > > >                       LOG(V4L2, Error)
> > > > @@ -187,20 +192,12 @@ ControlList V4L2Device::getControls(const std::vector<uint32_t> &ids)
> > > >                       return {};
> > > >               }
> > > >
> > > > -             ctrls.set(id, {});
> > > > -     }
> > > > -
> > > > -     struct v4l2_ext_control v4l2Ctrls[count];
> > > > -     memset(v4l2Ctrls, 0, sizeof(v4l2Ctrls));
> > > > -
> > > > -     unsigned int i = 0;
> > > > -     for (auto &ctrl : ctrls) {
> > > > -             unsigned int id = ctrl.first;
> > > > +             v4l2_ext_control &v4l2Ctrl = v4l2Ctrls[i++];
> > >
> > > You increase i here, ...
> > >
> > > >               const struct v4l2_query_ext_ctrl &info = controlInfo_[id];
> > > > +             ControlValue value{};
> > >
> > > ControlValue has a default constructor, no need for {}.
> > >
> > > >
> > > >               if (info.flags & V4L2_CTRL_FLAG_HAS_PAYLOAD) {
> > > >                       ControlType type;
> > > > -
> > > >                       switch (info.type) {
> > > >                       case V4L2_CTRL_TYPE_U8:
> > > >                               type = ControlTypeByte;
> > > > @@ -213,7 +210,6 @@ ControlList V4L2Device::getControls(const std::vector<uint32_t> &ids)
> > > >                               return {};
> > > >                       }
> > > >
> > > > -                     ControlValue &value = ctrl.second;
> > > >                       value.reserve(type, true, info.elems);
> > > >                       Span<uint8_t> data = value.data();
> > > >
> > > > @@ -221,21 +217,23 @@ ControlList V4L2Device::getControls(const std::vector<uint32_t> &ids)
> > > >                       v4l2Ctrls[i].size = data.size();
> > >
> > > ... and use it here. I don't think that's correct.
> > >
> > > >               }
> > > >
> > > > -             v4l2Ctrls[i].id = id;
> > > > -             i++;
> > > > +             v4l2Ctrl.id = id;
> > >
> > > Should we move this just after the declaration of v4l2Ctrl above ?
> > >
> > > > +             ctrls.set(id, std::move(value));
> > >
> > > v4l2Ctrls contains pointers to data stored in the ControlValue
> > > instances. As far as I can tell the pointers will remain valid, but
> > > that's very dependent on the internals of ControlList.
> > >
> > > To be honest, I'm not very fond of patches 1/4 and 2/4 in this series.
> > > They make the code more fragile and possibly a bit less efficient
> > > (although that's likely not significant, as there shouldn't be thousands
> > > of controls in the list). The VLA removal is fine, but the rest doesn't
> > > bring much value in my opinion.
> >
> > Can you clarify more how the new code is fragile and less efficient?
> > You're right, I am sorry that this implementation has some problems.
> > But I think 1/4 is worth doing and correct, and 2/4 is good if I fix the bugs?
>
> The efficiency is related to the additional lookup in updateControls()
> (the O(N*log(N)) we've discussed previously). I think it's a theoretical
> problem only, given the expected size of the control list, I expect the
> difference to be completely negligible in practice. As for the
> fragility, I was referring to the fact that we store ub v4l2Ctrls
> pointers to data from the ControlValue, stored in the ControlList. Any
> reallocation in the underlying container would make those pointers
> invalid. It's an existing issue, but with the current implementation, we
> don't touch the ControlList after adding controls to it at the beginning
> of getControls(), while with this patch we rely on
>
>         ctrls.set(id, std::move(value));
>

Yes, this should be ctrls.set(id, {}) or ctrls.set(id, value).
Sorry this std::move() is just confusing, which is actually equivalent
to ctrls.set(id, value).

> to not cause a reallocation. I think that's guaranteed by the current
> implementation of ControlList, but if that changes later, we'll possibly
> have hard to debug bugs.
>
> These two points are not blocker by themselves, but the gains should
> outweight the risks. It will be easier to check that after rebasing
> patches 1/4 and 2/4 on top of the VLA removal.
>

Okay, I would like to merge patch 1/4 at least.
I think our life will be easier if we have utils::enumerate().

> > > Let's split this in two parts, with the fixes first, and the rework on
> > > top, so both can be discussed separately. I've posted the former as a
> > > v6.
> > >
> > > >       }
> > > >
> > > > +     v4l2Ctrls.resize(ctrls.size());
> > > > +
> > > >       struct v4l2_ext_controls v4l2ExtCtrls = {};
> > > >       v4l2ExtCtrls.which = V4L2_CTRL_WHICH_CUR_VAL;
> > > > -     v4l2ExtCtrls.controls = v4l2Ctrls;
> > > > -     v4l2ExtCtrls.count = count;
> > > > +     v4l2ExtCtrls.controls = v4l2Ctrls.data();
> > > > +     v4l2ExtCtrls.count = v4l2Ctrls.size();
> > > >
> > > >       int ret = ioctl(VIDIOC_G_EXT_CTRLS, &v4l2ExtCtrls);
> > > >       if (ret) {
> > > >               unsigned int errorIdx = v4l2ExtCtrls.error_idx;
> > > >
> > > >               /* Generic validation error. */
> > > > -             if (errorIdx == 0 || errorIdx >= count) {
> > > > +             if (errorIdx == 0 || errorIdx >= v4l2Ctrls.size()) {
> > > >                       LOG(V4L2, Error) << "Unable to read controls: "
> > > >                                        << strerror(-ret);
> > > >                       return {};
> > > > @@ -244,10 +242,11 @@ ControlList V4L2Device::getControls(const std::vector<uint32_t> &ids)
> > > >               /* A specific control failed. */
> > > >               LOG(V4L2, Error) << "Unable to read control " << errorIdx
> > > >                                << ": " << strerror(-ret);
> > > > -             count = errorIdx - 1;
> > > > +
> > > > +             v4l2Ctrls.resize(errorIdx);
> > > >       }
> > > >
> > > > -     updateControls(&ctrls, v4l2Ctrls, count);
> > > > +     updateControls(&ctrls, v4l2Ctrls.data(), v4l2Ctrls.size());
> > > >
> > > >       return ctrls;
> > > >  }
>
> --
> Regards,
>
> Laurent Pinchart