Re: [RFCv4,13/21] media: videobuf2-v4l2: support for requests

From: Alexandre Courbot
Date: Thu Mar 08 2018 - 08:51:09 EST


On Thu, Mar 8, 2018 at 1:50 AM, Paul Kocialkowski
<paul.kocialkowski@xxxxxxxxxxx> wrote:
> Hi,
>
> On Tue, 2018-02-20 at 13:44 +0900, Alexandre Courbot wrote:
>> Add a new vb2_qbuf_request() (a request-aware version of vb2_qbuf())
>> that request-aware drivers can call to queue a buffer into a request
>> instead of directly into the vb2 queue if relevent.
>>
>> This function expects that drivers invoking it are using instances of
>> v4l2_request_entity and v4l2_request_entity_data to describe their
>> entity and entity data respectively.
>>
>> Also add the vb2_request_submit() helper function which drivers can
>> invoke in order to queue all the buffers previously queued into a
>> request into the regular vb2 queue.
>
> See a comment/proposal below about an issue I encountered when using
> multi-planar formats.
>
>> Signed-off-by: Alexandre Courbot <acourbot@xxxxxxxxxxxx>
>> ---
>> .../media/common/videobuf2/videobuf2-v4l2.c | 129
>> +++++++++++++++++-
>> include/media/videobuf2-v4l2.h | 59 ++++++++
>> 2 files changed, 187 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/media/common/videobuf2/videobuf2-v4l2.c
>> b/drivers/media/common/videobuf2/videobuf2-v4l2.c
>> index 6d4d184aa68e..0627c3339572 100644
>> --- a/drivers/media/common/videobuf2/videobuf2-v4l2.c
>> +++ b/drivers/media/common/videobuf2/videobuf2-v4l2.c
>
> [...]
>
>> +#if IS_ENABLED(CONFIG_MEDIA_REQUEST_API)
>> +int vb2_qbuf_request(struct vb2_queue *q, struct v4l2_buffer *b,
>> + struct media_request_entity *entity)
>> +{
>> + struct v4l2_request_entity_data *data;
>> + struct v4l2_vb2_request_buffer *qb;
>> + struct media_request *req;
>> + struct vb2_buffer *vb;
>> + int ret = 0;
>> +
>> + if (b->request_fd <= 0)
>> + return vb2_qbuf(q, b);
>> +
>> + if (!q->allow_requests)
>> + return -EINVAL;
>> +
>> + req = media_request_get_from_fd(b->request_fd);
>> + if (!req)
>> + return -EINVAL;
>> +
>> + data = to_v4l2_entity_data(media_request_get_entity_data(req,
>> entity));
>> + if (IS_ERR(data)) {
>> + ret = PTR_ERR(data);
>> + goto out;
>> + }
>> +
>> + mutex_lock(&req->lock);
>> +
>> + if (req->state != MEDIA_REQUEST_STATE_IDLE) {
>> + ret = -EINVAL;
>> + goto out;
>> + }
>> +
>> + ret = vb2_queue_or_prepare_buf(q, b, "qbuf");
>> + if (ret)
>> + goto out;
>> +
>> + vb = q->bufs[b->index];
>> + switch (vb->state) {
>> + case VB2_BUF_STATE_DEQUEUED:
>> + break;
>> + case VB2_BUF_STATE_PREPARED:
>> + break;
>> + case VB2_BUF_STATE_PREPARING:
>> + dprintk(1, "buffer still being prepared\n");
>> + ret = -EINVAL;
>> + goto out;
>> + default:
>> + dprintk(1, "invalid buffer state %d\n", vb->state);
>> + ret = -EINVAL;
>> + goto out;
>> + }
>> +
>> + /* do we already have a buffer for this request in the queue?
>> */
>> + list_for_each_entry(qb, &data->queued_buffers, node) {
>> + if (qb->queue == q) {
>> + ret = -EBUSY;
>> + goto out;
>> + }
>> + }
>> +
>> + qb = kzalloc(sizeof(*qb), GFP_KERNEL);
>> + if (!qb) {
>> + ret = -ENOMEM;
>> + goto out;
>> + }
>> +
>> + /*
>> + * TODO should be prepare the buffer here if needed, to
>> report errors
>> + * early?
>> + */
>> + qb->pre_req_state = vb->state;
>> + qb->queue = q;
>> + memcpy(&qb->v4l2_buf, b, sizeof(*b));
>
> I am getting data corruption on qb->v4l2_buf.m.planes from this when
> using planar buffers, only after exiting the ioctl handler (i.e. when
> accessing this buffer later from the queue).
>
> I initially thought this was because the planes pointer was copied as-is
> from userspace, but Maxime Ripard suggested that this would have
> automatically triggered a visible fault in the kernel.
>
> Thus, my best guess is that the data is properly copied from userspace
> but freed when leaving the ioctl handler, which doesn't work our with
> the request API.
>
> A dirty fix that I came up with consists in re-allocating the planes
> buffer here and copying its contents from "b", so that it can live
> beyond the ioctl call.
>
> I am not too sure whether this should be fixed here or in the part of
> the v4l2 common code that frees this data. What do you think?

Oh, nice catch. Copying plane information indeed requires more work
than a dumb memcpy(). I suppose this should be handled here, but let
me come back to this after a good night of sleep. :)

Thanks! I will try to fix this tomorrow and push a temporary version
somewhere so you can move forward.

Alex.