> -----Original Message-----
> From: Christian König <christian.koe...@amd.com>
> Sent: Tuesday, June 3, 2025 6:42 PM
> To: wangtao <tao.wang...@honor.com>; sumit.sem...@linaro.org;
> kra...@redhat.com; vivek.kasire...@intel.com; v...@zeniv.linux.org.uk;
> brau...@kernel.org; hu...@google.com; a...@linux-foundation.org;
> amir7...@gmail.com
> Cc: benjamin.gaign...@collabora.com; brian.star...@arm.com;
> jstu...@google.com; tjmerc...@google.com; j...@suse.cz;
> baolin.w...@linux.alibaba.com; linux-me...@vger.kernel.org; dri-
> de...@lists.freedesktop.org; linaro-mm-...@lists.linaro.org; linux-
> ker...@vger.kernel.org; linux-fsde...@vger.kernel.org; linux-
> m...@kvack.org; wangbintian(BintianWang) <bintian.w...@honor.com>;
> yipengxiang <yipengxi...@honor.com>; liulu 00013167
> <liulu....@honor.com>; hanfeng 00012985 <feng....@honor.com>
> Subject: Re: [PATCH v4 2/4] dmabuf: Implement copy_file_range callback for
> dmabuf direct I/O prep
>
>
>
> On 6/3/25 11:52, wangtao wrote:
> > First determine if dmabuf reads from or writes to the file.
> > Then call exporter's rw_file callback function.
> >
> > Signed-off-by: wangtao <tao.wang...@honor.com>
> > ---
> > drivers/dma-buf/dma-buf.c | 32 ++++++++++++++++++++++++++++++++
> > include/linux/dma-buf.h | 16 ++++++++++++++++
> > 2 files changed, 48 insertions(+)
> >
> > diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
> > index 5baa83b85515..fc9bf54c921a 100644
> > --- a/drivers/dma-buf/dma-buf.c
> > +++ b/drivers/dma-buf/dma-buf.c
> > @@ -523,7 +523,38 @@ static void dma_buf_show_fdinfo(struct seq_file
> *m, struct file *file)
> > spin_unlock(&dmabuf->name_lock);
> > }
> >
> > +static ssize_t dma_buf_rw_file(struct dma_buf *dmabuf, loff_t my_pos,
> > + struct file *file, loff_t pos, size_t count, bool is_write) {
> > + if (!dmabuf->ops->rw_file)
> > + return -EINVAL;
> > +
> > + if (my_pos >= dmabuf->size)
> > + count = 0;
> > + else
> > + count = min_t(size_t, count, dmabuf->size - my_pos);
> > + if (!count)
> > + return 0;
> > +
> > + return dmabuf->ops->rw_file(dmabuf, my_pos, file, pos, count,
> > +is_write); }
> > +
> > +static ssize_t dma_buf_copy_file_range(struct file *file_in, loff_t pos_in,
> > + struct file *file_out, loff_t pos_out,
> > + size_t count, unsigned int flags)
> > +{
> > + if (is_dma_buf_file(file_in) && file_out->f_op->write_iter)
> > + return dma_buf_rw_file(file_in->private_data, pos_in,
> > + file_out, pos_out, count, true);
> > + else if (is_dma_buf_file(file_out) && file_in->f_op->read_iter)
> > + return dma_buf_rw_file(file_out->private_data, pos_out,
> > + file_in, pos_in, count, false);
> > + else
> > + return -EINVAL;
> > +}
> > +
> > static const struct file_operations dma_buf_fops = {
> > + .fop_flags = FOP_MEMORY_FILE,
> > .release = dma_buf_file_release,
> > .mmap = dma_buf_mmap_internal,
> > .llseek = dma_buf_llseek,
> > @@ -531,6 +562,7 @@ static const struct file_operations dma_buf_fops = {
> > .unlocked_ioctl = dma_buf_ioctl,
> > .compat_ioctl = compat_ptr_ioctl,
> > .show_fdinfo = dma_buf_show_fdinfo,
> > + .copy_file_range = dma_buf_copy_file_range,
> > };
> >
> > /*
> > diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h index
> > 36216d28d8bd..d3636e985399 100644
> > --- a/include/linux/dma-buf.h
> > +++ b/include/linux/dma-buf.h
> > @@ -22,6 +22,7 @@
> > #include <linux/fs.h>
> > #include <linux/dma-fence.h>
> > #include <linux/wait.h>
> > +#include <uapi/linux/dma-buf.h>
> >
> > struct device;
> > struct dma_buf;
> > @@ -285,6 +286,21 @@ struct dma_buf_ops {
> >
> > int (*vmap)(struct dma_buf *dmabuf, struct iosys_map *map);
> > void (*vunmap)(struct dma_buf *dmabuf, struct iosys_map *map);
> > +
> > + /**
> > + * @rw_file:
> > + *
> > + * If an Exporter needs to support Direct I/O file operations, it can
> > + * implement this optional callback. The exporter must verify that no
> > + * other objects hold the sg_table, ensure exclusive access to the
> > + * dmabuf's sg_table, and only then proceed with the I/O operation.
>
> Explain why and not what. E.g. something like "Allows direct I/O between
> this DMA-buf and the file".
>
> Completely drop mentioning the sg_table, that is irrelevant. Exclusive access
> depends on how the exporter implements the whole thing.
>
Will fix this shortly.
Regards,
Wangtao.
> Regards,
> Christian.
>
> > + *
> > + * Returns:
> > + *
> > + * 0 on success or a negative error code on failure.
> > + */
> > + ssize_t (*rw_file)(struct dma_buf *dmabuf, loff_t my_pos,
> > + struct file *file, loff_t pos, size_t count, bool is_write);
> > };
> >
> > /**
