[Virtio-fs] Few queries about virtiofsd read implementation

[Dr. David Alan Gilbert]

* Vivek Goyal (vgoyal at redhat.com) wrote:
> Hi David/Stefan,
> 
> I am browsing through the code of read requests (FUSE_READ) in virtiofsd
> (and in virtiofs) and I have few questions. You folks probably know the
> answers.
> 
> 1. virtio_send_data_iov(), reads the data from file into the scatter list.
>   Some of the code looks strange.
> 
>   We seem to be retrying read if we read less number of bytes than what
>   client asked for. I am wondering shoudl this really be our
>   responsibility or client should deal with it. I am assuming that client
>   should be ready to deal with less number of bytes read.
> 
>   So what was the thought process behind retrying.
> 
>           if (ret < len && ret) {
>             fuse_log(FUSE_LOG_DEBUG, "%s: ret < len\n", __func__);
>             /* Skip over this much next time around */
>             skip_size = ret;
>             buf->buf[0].pos += ret;
>             len -= ret;
> 
>             /* Lets do another read */
>             continue;
>         }
> 
> - After this we have code where if number of bytes read are not same
>   as we expect to, then we return EIO.
> 
>           if (ret != len) {
>             fuse_log(FUSE_LOG_DEBUG, "%s: ret!=len\n", __func__);
>             ret = EIO;
>             free(in_sg_cpy);
>             goto err;
>         }
> 
>   When do we hit this. IIUC, preadv() will return.
> 
>   A. Either number of bytes we expected (no issues)
>   B. 0 in case of EOF (We break out of loop and just return to client with
>   		      number of bytes we have read so far).
>   C. <0 (This is error case and we return error to client)
>   D. X bytes which is less than len. 
> 
> To handle D we have code to retry. So when do we hit the above if
> condition where "ret !=len). Is this a dead code. Or I missed something.

I think you're right, that's dead.
And oyu're probably also right that we could just take it easy and
return less data to the client if preadv just gives us part of it.

> 2. When client sent FUSE_READ, we put pointer to pages into sglist. IIUC,
>    we put pointer to "struct page *" and not actual page. So who converts
>    these struct page pointer to memory belong to page.
> 
>    sg_init_fuse_pages() {
>    	sg_set_page(&sg[i], pages[i], this_len, page_descs[i].offset);
>    }

Hmm don't know.

> 3. Who converts guest memory address (when) into qemu process address which is
>    accessible by virtiofsd.

I think that's libvhost-user doing that for us; in vu_queue_pop it calls
vu_queue_map_desc.

Dave

> Thanks
> Vivek
-- 
Dr. David Alan Gilbert / dgilbert at redhat.com / Manchester, UK