Question about sg_count_fuse_req() in linux/fs/fuse/virtio_fs.c

Question about sg_count_fuse_req() in linux/fs/fuse/virtio_fs.c

[Connor Kuehl]

Hi,

I've been familiarizing myself with the virtiofs guest kernel module and 
I'm trying to better understand how virtiofs maps a FUSE request into 
scattergather lists.

sg_count_fuse_req() starts knowing that there will be at least one in 
header, as shown here (which makes sense):

         unsigned int size, total_sgs = 1 /* fuse_in_header */;

However, I'm confused about this snippet right beneath it:

         if (args->in_numargs - args->in_pages)
                 total_sgs += 1;

What is the significance of the sg that is needed in the cases where 
this branch is taken? I'm not sure what its relationship is with 
args->in_numargs since it will increment total_sgs regardless 
args->in_numargs is 3, 2, or even 1 if args->in_pages is false.

Especially since the block right below it counts pages if args->in_pages 
is true:

         if (args->in_pages) {
                 size = args->in_args[args->in_numargs - 1].size;
                 total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
                                                  size);
         }

The rest of the routine goes on similarly but for the 'out' components.

I doubt incrementing 'total_sgs' in the first if-statement I showed 
above is vestigial, I just think my mental model of what is happening 
here is incomplete.

Any clarification is much appreciated!

Connor

Re: Question about sg_count_fuse_req() in linux/fs/fuse/virtio_fs.c

[Vivek Goyal]

On Wed, Mar 17, 2021 at 01:12:01PM -0500, Connor Kuehl wrote:
> Hi,
> 
> I've been familiarizing myself with the virtiofs guest kernel module and I'm
> trying to better understand how virtiofs maps a FUSE request into
> scattergather lists.
> 
> sg_count_fuse_req() starts knowing that there will be at least one in
> header, as shown here (which makes sense):
> 
>         unsigned int size, total_sgs = 1 /* fuse_in_header */;
> 
> However, I'm confused about this snippet right beneath it:
> 
>         if (args->in_numargs - args->in_pages)
>                 total_sgs += 1;
> 
> What is the significance of the sg that is needed in the cases where this
> branch is taken? I'm not sure what its relationship is with args->in_numargs
> since it will increment total_sgs regardless args->in_numargs is 3, 2, or
> even 1 if args->in_pages is false.

Hi Conor,

I think all the in args are being mapped into a single scatter gather
element and that's why it does not matter whether in_numargs is 3, 2 or 1.
They will be mapped in a single element.

sg_init_fuse_args()
{
        len = fuse_len_args(numargs - argpages, args);
        if (len)
                sg_init_one(&sg[total_sgs++], argbuf, len);
}

        out_sgs += sg_init_fuse_args(&sg[out_sgs], req,
                                     (struct fuse_arg *)args->in_args,
                                     args->in_numargs, args->in_pages,
                                     req->argbuf, &argbuf_used);

When we are sending some data in some pages, then we set args->in_pages
to true. And in that case, last element of args->in_args[] contains the
total size of bytes in additional pages we are sending and is not part
of in_args being mapped to scatter gather element. That's why this
check.

	if (args->in_numargs - args->in_pages)
		total_sgs += 1;

Not sure when we will have a case where args->in_numargs = 1 and
args->in_pages=true. Do we ever hit that.

Thanks
Vivek

> 
> Especially since the block right below it counts pages if args->in_pages is
> true:
> 
>         if (args->in_pages) {
>                 size = args->in_args[args->in_numargs - 1].size;
>                 total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
>                                                  size);
>         }
> 
> The rest of the routine goes on similarly but for the 'out' components.
> 
> I doubt incrementing 'total_sgs' in the first if-statement I showed above is
> vestigial, I just think my mental model of what is happening here is
> incomplete.
> 
> Any clarification is much appreciated!

> 
> Connor
> 

Re: Question about sg_count_fuse_req() in linux/fs/fuse/virtio_fs.c

[Connor Kuehl]

On 3/18/21 8:56 AM, Vivek Goyal wrote:
> I think all the in args are being mapped into a single scatter gather
> element and that's why it does not matter whether in_numargs is 3, 2 or 1.
> They will be mapped in a single element.
> 
> sg_init_fuse_args()
> {
>          len = fuse_len_args(numargs - argpages, args);
>          if (len)
>                  sg_init_one(&sg[total_sgs++], argbuf, len);
> }
> 
>          out_sgs += sg_init_fuse_args(&sg[out_sgs], req,
>                                       (struct fuse_arg *)args->in_args,
>                                       args->in_numargs, args->in_pages,
>                                       req->argbuf, &argbuf_used);
> 
> When we are sending some data in some pages, then we set args->in_pages
> to true. And in that case, last element of args->in_args[] contains the
> total size of bytes in additional pages we are sending and is not part
> of in_args being mapped to scatter gather element. That's why this
> check.
> 
> 	if (args->in_numargs - args->in_pages)
> 		total_sgs += 1;

Aha! Thank you, Vivek. That makes sense.

> Not sure when we will have a case where args->in_numargs = 1 and
> args->in_pages=true. Do we ever hit that.

Not in my experience. My previous mail was examining this routine mainly 
in a vacuum.

Connor

> Thanks
> Vivek
> 
>>
>> Especially since the block right below it counts pages if args->in_pages is
>> true:
>>
>>          if (args->in_pages) {
>>                  size = args->in_args[args->in_numargs - 1].size;
>>                  total_sgs += sg_count_fuse_pages(ap->descs, ap->num_pages,
>>                                                   size);
>>          }
>>
>> The rest of the routine goes on similarly but for the 'out' components.
>>
>> I doubt incrementing 'total_sgs' in the first if-statement I showed above is
>> vestigial, I just think my mental model of what is happening here is
>> incomplete.
>>
>> Any clarification is much appreciated!
> 
>>
>> Connor
>>
>