Re: [PATCH] staging: android: ion: move map_kernel to ion_dma_buf_kmap

[Xiaqing (A)]

On 2019/1/3 6:36, Laura Abbott wrote:
> On 12/24/18 12:19 AM, Qing Xia wrote:
> > Now, as Google's user guide, if userspace need clean ION buffer's
> > cache, they should call ioctl(fd, DMA_BUF_IOCTL_SYNC, sync). Then
> > we found that ion_dma_buf_begin_cpu_access/ion_dma_buf_end_cpu_access
> > will do ION buffer's map_kernel, it's not necessary. And if usersapce
> > only need clean cache, they will call ion_dma_buf_end_cpu_access by
> > dmabuf's ioctl, ION buffer's kmap_cnt will be wrong value -1, then
> > driver could not get right kernel vaddr by dma_buf_kmap.
>
> The problem is this subtly violates how dma_buf is supposed
> to work. All setup is supposed to happen in begin_cpu_access.
> I agree calling map kernel each time isn't great but I think
> this needs to be worked out with dma_buf.
>
> Thanks,
> Laura
Thanks for your explanation.

> > Signed-off-by: Qing Xia <saberlily.xia@xxxxxxxxxxxxx>
> > ---
> > Â drivers/staging/android/ion/ion.c | 46 ++++++++++++++++++---------------------
> > Â 1 file changed, 21 insertions(+), 25 deletions(-)
> >
> > diff --git a/drivers/staging/android/ion/ion.c b/drivers/staging/android/ion/ion.c
> > index 9907332..f7e2812 100644
> > --- a/drivers/staging/android/ion/ion.c
> > +++ b/drivers/staging/android/ion/ion.c
> > @@ -303,45 +303,47 @@ static void ion_dma_buf_release(struct dma_buf *dmabuf)
> > Â static void *ion_dma_buf_kmap(struct dma_buf *dmabuf, unsigned long offset)
> > Â {
> > ÂÂÂÂÂ struct ion_buffer *buffer = dmabuf->priv;
> > +ÂÂÂ void *vaddr;
> > -ÂÂÂ return buffer->vaddr + offset * PAGE_SIZE;
> > +ÂÂÂ if (buffer->heap->ops->map_kernel) {
> > +ÂÂÂÂÂÂÂ mutex_lock(&buffer->lock);
> > +ÂÂÂÂÂÂÂ vaddr = ion_buffer_kmap_get(buffer);
> > +ÂÂÂÂÂÂÂ mutex_unlock(&buffer->lock);
> > +ÂÂÂÂÂÂÂ if (IS_ERR(vaddr))
> > +ÂÂÂÂÂÂÂÂÂÂÂ return vaddr;
> > +
> > +ÂÂÂÂÂÂÂ return vaddr + offset * PAGE_SIZE;
> > +ÂÂÂ }
> > +
> > +ÂÂÂ return NULL;
> > Â }
> > Â static void ion_dma_buf_kunmap(struct dma_buf *dmabuf, unsigned long offset,
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ void *ptr)
> > Â {
> > +ÂÂÂ struct ion_buffer *buffer = dmabuf->priv;
> > +
> > +ÂÂÂ if (buffer->heap->ops->map_kernel) {
> > +ÂÂÂÂÂÂÂ mutex_lock(&buffer->lock);
> > +ÂÂÂÂÂÂÂ ion_buffer_kmap_put(buffer);
> > +ÂÂÂÂÂÂÂ mutex_unlock(&buffer->lock);
> > +ÂÂÂ }
> > Â }
> > Â static int ion_dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ enum dma_data_direction direction)
> > Â {
> > ÂÂÂÂÂ struct ion_buffer *buffer = dmabuf->priv;
> > -ÂÂÂ void *vaddr;
> > ÂÂÂÂÂ struct ion_dma_buf_attachment *a;
> > -ÂÂÂ int ret = 0;
> > -
> > -ÂÂÂ /*
> > -ÂÂÂÂ * TODO: Move this elsewhere because we don't always need a vaddr
> > -ÂÂÂÂ */
> > -ÂÂÂ if (buffer->heap->ops->map_kernel) {
> > -ÂÂÂÂÂÂÂ mutex_lock(&buffer->lock);
> > -ÂÂÂÂÂÂÂ vaddr = ion_buffer_kmap_get(buffer);
> > -ÂÂÂÂÂÂÂ if (IS_ERR(vaddr)) {
> > -ÂÂÂÂÂÂÂÂÂÂÂ ret = PTR_ERR(vaddr);
> > -ÂÂÂÂÂÂÂÂÂÂÂ goto unlock;
> > -ÂÂÂÂÂÂÂ }
> > -ÂÂÂÂÂÂÂ mutex_unlock(&buffer->lock);
> > -ÂÂÂ }
> > ÂÂÂÂÂ mutex_lock(&buffer->lock);
> > ÂÂÂÂÂ list_for_each_entry(a, &buffer->attachments, list) {
> > ÂÂÂÂÂÂÂÂÂ dma_sync_sg_for_cpu(a->dev, a->table->sgl, a->table->nents,
> > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ direction);
> > ÂÂÂÂÂ }
> > -
> > -unlock:
> > ÂÂÂÂÂ mutex_unlock(&buffer->lock);
> > -ÂÂÂ return ret;
> > +
> > +ÂÂÂ return 0;
> > Â }
> > Â static int ion_dma_buf_end_cpu_access(struct dma_buf *dmabuf,
> > @@ -350,12 +352,6 @@ static int ion_dma_buf_end_cpu_access(struct dma_buf *dmabuf,
> > ÂÂÂÂÂ struct ion_buffer *buffer = dmabuf->priv;
> > ÂÂÂÂÂ struct ion_dma_buf_attachment *a;
> > -ÂÂÂ if (buffer->heap->ops->map_kernel) {
> > -ÂÂÂÂÂÂÂ mutex_lock(&buffer->lock);
> > -ÂÂÂÂÂÂÂ ion_buffer_kmap_put(buffer);
> > -ÂÂÂÂÂÂÂ mutex_unlock(&buffer->lock);
> > -ÂÂÂ }
> > -
> > ÂÂÂÂÂ mutex_lock(&buffer->lock);
> > ÂÂÂÂÂ list_for_each_entry(a, &buffer->attachments, list) {
> > ÂÂÂÂÂÂÂÂÂ dma_sync_sg_for_device(a->dev, a->table->sgl, a->table->nents,
>
>
> .