Re: [PATCH v2 0/4] open/accept directly into io_uring fixed file table

On Sat, Aug 14, 2021 at 01:50:24PM +0100, Pavel Begunkov wrote:

On 8/13/21 8:00 PM, Josh Triplett wrote:

quoted

Rather than using sqe->file_index - 1, which feels like an error-prone
interface, I think it makes sense to use a dedicated flag for this, like
IOSQE_OPEN_FIXED. That flag could work for any open-like operation,
including open, accept, and in the future many other operations such as
memfd_create. (Imagine using a single ring submission to open a memfd,
write a buffer into it, seal it, send it over a UNIX socket, and then
close it.)

The only downside is that you'll need to reject that flag in all
non-open operations. One way to unify that code might be to add a flag
in io_op_def for open-like operations, and then check in common code for
the case of non-open-like operations passing IOSQE_OPEN_FIXED.

io_uring is really thin, and so I absolutely don't want any extra
overhead in the generic path, IOW anything affecting
reads/writes/sends/recvs.

There are already several checks for valid flags in io_init_req. For
instance:
        if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
            !io_op_defs[req->opcode].buffer_select)
                return -EOPNOTSUPP;
It'd be trivial to make io_op_defs have a "valid flags" byte, and one
bitwise op tells you if any invalid flags were passed. *Zero* additional
overhead for other operations.

Alternatively, since there are so few operations that open a file
descriptor, you could just add a separate opcode for those few
operations. That still seems preferable to overloading a 16-bit index
field for this.

With this new mechanism, I think we're going to want to support more
than 65535 fixed-file entries. I can easily imagine wanting to handle
hundreds of thousands of files or sockets this way.

The other reason is that there are only 2 bits left in sqe->flags,
and we may use them for something better, considering that it's
only open/accept and not much as this.

pipe, dup3, socket, socketpair, pidfds (via either pidfd_open or a
ring-based spawn mechanism), epoll_create, inotify, fanotify, signalfd,
timerfd, eventfd, memfd_create, userfaultfd, open_tree, fsopen, fsmount,
memfd_secret.

Of those, I personally would *love* to have at least pipe, socket,
pidfd, memfd_create, and fsopen/fsmount/open_tree, plus some manner of
dup-like operation for moving things between the fixed-file table and
file descriptors.

I think this is valuable and versatile enough to merit a flag. It would
also be entirely reasonable to create separate operations for these. But
either way, I don't think this should just be determined by whether a
16-bit index is non-zero.

I agree that it feels error-prone, but at least it can be wrapped
nicely enough in liburing, e.g.

void io_uring_prep_openat_direct(struct io_uring_sqe *sqe, int dfd,
				 const char *path, int flags,
				 mode_t mode, int slot_idx);

That wrapper wouldn't be able to handle more than a 16-bit slot index
though.

quoted

Also, rather than using a 16-bit index for the fixed file table and
potentially requiring expansion into a different field in the future,
what about overlapping it with the nofile field in the open and accept
requests? If they're not opening a normal file descriptor, they don't
need nofile. And in the original sqe, you can then overlap it with a
32-bit field like splice_fd_in.

There is no nofile in SQEs, though

req->open.nofile = rlimit(RLIMIT_NOFILE);

nofile isn't needed for opening into the fixed-file table, so it could
be omitted in that case, and another field unioned with it. That would
allow passing a 32-bit fixed-file index into open and accept without
growing the size of their structures. I think, with this new capability,
we're going to want a large number of fixed files available.

In the SQE, you could overlap it with the splice_fd_in field, which
isn't needed by any calls other than splice.

- Josh Triplett