Re: [PATCH RESEND V12 2/8] fuse: 32-bit user space ioctl compat for fuse device

[Peng Tao]

On Thu, Jan 28, 2021 at 10:15 PM Alessio Balsini <balsini@xxxxxxxxxxx> wrote:
>
> Hi all,
>
> I'm more than happy to change the interface into something that is
> objectively better and accepted by everyone.
> I would really love to reach the point at which we have a "stable-ish"
> UAPI as soon as possible.
>
> I've been thinking about a few possible approaches to fix the issue, yet
> to preserve its flexibility. These are mentioned below.
>
>
>   Solution 1: Size
>
> As mentioned in my previous email, one solution could be to introduce
> the "size" field to allow the structure to grow in the future.
>
> struct fuse_passthrough_out {
>     uint32_t        size;   // Size of this data structure
>     uint32_t        fd;
> };
>
> The problem here is that we are making the promise that all the upcoming
> fields are going to be maintained forever and at the offsets they were
> originally defined.
>
>
>   Solution 2: Version
>
> Another solution could be to s/size/version, where for every version of
> FUSE passthrough we reserve the right to modifying the fields over time,
> casting them to the right data structure according to the version.
>
>
>   Solution 3: Type
>
> Using an enumerator to define the data structure content and purpose is
> the most flexible solution I can think of.  This would for example allow
> us to substitute FUSE_DEV_IOC_PASSTHROUGH_OPEN with the generic
> FUSE_DEV_IOC_PASSTHROUGH and having a single ioctl for any eventually
> upcoming passthrough requests.
>
> enum fuse_passthrough_type {
>     FUSE_PASSTHROUGH_OPEN
> };
>
> struct fuse_passthrough_out {
>     uint32_t type; /* as defined by enum fuse_passthrough_type */
>     union {
>         uint32_t fd;
>     };
> };
>
> This last is my favorite, as regardless the minimal logic required to
> detect the size and content of the struct (not required now as we only
> have a single option), it would also allow to do some kind of interface
> versioning (e.g., in case we want to implement
> FUSE_PASSTHROUGH_OPEN_V2).
>
Usually a new type of ioctl will be added in such cases. If we want to
stick to the same ioctl number, it might be easier to add a `flags`
field to differentiate compatible passthrough ioctls. So in future, if
a new interface is compatible with the existing one, we can use flags
to tell it. If it is not compatible, we still need to add a new ioctl.
wdyt?

struct fuse_passthrough_out {
    uint32_t flags;
    union {
        uint32_t fd;
    };
};

This somehow follows the "Flags as a system call API design pattern"
(https://lwn.net/Articles/585415/).

Just my two cents.

Cheers,
Tao
-- 
Into Sth. Rich & Strange