On 2/3/19 7:56 PM, Al Viro wrote:
> On Wed, Jan 30, 2019 at 02:29:05AM +0100, Jann Horn wrote:
>> On Tue, Jan 29, 2019 at 8:27 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>> We normally have to fget/fput for each IO we do on a file. Even with
>>> the batching we do, the cost of the atomic inc/dec of the file usage
>>> count adds up.
>>>
>>> This adds IORING_REGISTER_FILES, and IORING_UNREGISTER_FILES opcodes
>>> for the io_uring_register(2) system call. The arguments passed in must
>>> be an array of __s32 holding file descriptors, and nr_args should hold
>>> the number of file descriptors the application wishes to pin for the
>>> duration of the io_uring context (or until IORING_UNREGISTER_FILES is
>>> called).
>>>
>>> When used, the application must set IOSQE_FIXED_FILE in the sqe->flags
>>> member. Then, instead of setting sqe->fd to the real fd, it sets sqe->fd
>>> to the index in the array passed in to IORING_REGISTER_FILES.
>>>
>>> Files are automatically unregistered when the io_uring context is
>>> torn down. An application need only unregister if it wishes to
>>> register a new set of fds.
>>
>> Crazy idea:
>>
>> Taking a step back, at a high level, basically this patch creates sort
>> of the same difference that you get when you compare the following
>> scenarios for normal multithreaded I/O in userspace:
>
>> This kinda makes me wonder whether this is really something that
>> should be implemented specifically for the io_uring API, or whether it
>> would make sense to somehow handle part of this in the generic VFS
>> code and give the user the ability to prepare a new files_struct that
>> can then be transferred to the worker thread, or something like
>> that... I'm not sure whether there's a particularly clean way to do
>> that though.
>
> Using files_struct for that opens a can of worms you really don't
> want to touch.
>
> Consider the following scenario with any variant of this interface:
> * create io_uring fd.
> * send an SCM_RIGHTS with that fd to AF_UNIX socket.
> * add the descriptor of that AF_UNIX socket to your fd
> * close AF_UNIX fd, close io_uring fd.
> Voila - you've got a shiny leak. No ->release() is called for
> anyone (and you really don't want to do that on ->flush(), because
> otherwise a library helper doing e.g. system("/bin/date") will tear
> down all the io_uring in your process). The socket is held by
> the reference you've stashed into io_uring (whichever way you do
> that). io_uring is held by the reference you've stashed into
> SCM_RIGHTS datagram in queue of the socket.
>
> No matter what, you need net/unix/garbage.c to be aware of that stuff.
> And getting files_struct lifetime mixed into that would be beyond
> any reason.
>
> The only reason for doing that as a descriptor table would be
> avoiding the cost of fget() in whatever uses it, right? Since
Right, the only purpose of this patch is to avoid doing fget/fput for
each IO.
> those are *not* the normal syscalls (and fdget() really should not
> be used anywhere other than the very top of syscall's call chain -
> that's another reason why tossing file_struct around like that
> is insane) and since the benefit is all due to the fact that it's
> *NOT* shared, *NOT* modified in parallel, etc., allowing us to
> treat file references as stable... why the hell use the descriptor
> tables at all?
This one is not a regular system call, since we don't do fget, then IO,
then fput. We hang on to it. But for the non-registered case, it's very
much just like a regular read/write system call, where we fget to do IO
on it, then fput when we are done.
> All you need is an array of struct file *, explicitly populated.
> With net/unix/garbage.c aware of such beasts. Guess what? We
> do have such an object already. The one net/unix/garbage.c is
> working with. SCM_RIGHTS datagrams, that is.
>
> IOW, can't we give those io_uring descriptors associated struct
> unix_sock? No socket descriptors, no struct socket (probably),
> just the AF_UNIX-specific part thereof. Then teach
> unix_inflight()/unix_notinflight() about getting unix_sock out
> of these guys (incidentally, both would seem to benefit from
> _not_ touching unix_gc_lock in case when there's no unix_sock
> attached to file we are dealing with - I might be missing
> something very subtle about barriers there, but it doesn't
> look likely).
That might be workable, though I'm not sure we currently have helpers to
just explicitly create a unix_sock by itself. Not familiar with the
networking bits at all, I'll take a look.
> And make that (i.e. registering the descriptors) mandatory.
I don't want to make it mandatory, that's very inflexible for managing
tons of files. The registration is useful for specific cases where we
have high frequency of operations on a set of files. Besides, it'd make
the use of the API cumbersome as well for the basic case of just wanting
to do async IO.
> Hell, combine that with creating io_uring fd, if we really
> care about the syscall count. Benefits:
We don't care about syscall count for setup as much. If you're doing
registration of a file set, you're expected to do a LOT of IO to those
files. Hence having an extra one for setup is not a concern. My concern
is just making it mandatory to do registration, I don't think that's a
workable alternative.
> * no file_struct refcount wanking
> * no fget()/fput() (conditional, at that) from kernel
> threads
> * no CLOEXEC-dependent anything; just the teardown
> on the final fput(), whichever way it comes.
> * no fun with duelling garbage collectors.
The fget/fput from a kernel thread can be solved by just hanging on to
the struct file * when we punt the IO. Right now we don't, which is a
little silly, that should be changed.
Getting rid of the files_struct{} is doable.
--
Jens Axboe
