On Thu, Jul 13, 2023 at 05:06:32PM +0800, Hao Xu wrote:
> Hi Christian,
>
> On 7/13/23 15:10, Christian Brauner wrote:
> > On Thu, Jul 13, 2023 at 12:35:07PM +0800, Hao Xu wrote:
> > > On 7/12/23 23:27, Christian Brauner wrote:
> > > > On Tue, Jul 11, 2023 at 07:40:27PM +0800, Hao Xu wrote:
> > > > > From: Hao Xu <howeyxu@xxxxxxxxxxx>
> > > > >
> > > > > This add support for getdents64 to io_uring, acting exactly like the
> > > > > syscall: the directory is iterated from it's current's position as
> > > > > stored in the file struct, and the file's position is updated exactly as
> > > > > if getdents64 had been called.
> > > > >
> > > > > For filesystems that support NOWAIT in iterate_shared(), try to use it
> > > > > first; if a user already knows the filesystem they use do not support
> > > > > nowait they can force async through IOSQE_ASYNC in the sqe flags,
> > > > > avoiding the need to bounce back through a useless EAGAIN return.
> > > > >
> > > > > Co-developed-by: Dominique Martinet <asmadeus@xxxxxxxxxxxxx>
> > > > > Signed-off-by: Dominique Martinet <asmadeus@xxxxxxxxxxxxx>
> > > > > Signed-off-by: Hao Xu <howeyxu@xxxxxxxxxxx>
> > > > > ---
> > > > > include/uapi/linux/io_uring.h | 7 ++++
> > > > > io_uring/fs.c | 60 +++++++++++++++++++++++++++++++++++
> > > > > io_uring/fs.h | 3 ++
> > > > > io_uring/opdef.c | 8 +++++
> > > > > 4 files changed, 78 insertions(+)
> > > > >
> > > > > diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> > > > > index 08720c7bd92f..6c0d521135a6 100644
> > > > > --- a/include/uapi/linux/io_uring.h
> > > > > +++ b/include/uapi/linux/io_uring.h
> > > > > @@ -65,6 +65,7 @@ struct io_uring_sqe {
> > > > > __u32 xattr_flags;
> > > > > __u32 msg_ring_flags;
> > > > > __u32 uring_cmd_flags;
> > > > > + __u32 getdents_flags;
> > > > > };
> > > > > __u64 user_data; /* data to be passed back at completion time */
> > > > > /* pack this to avoid bogus arm OABI complaints */
> > > > > @@ -235,6 +236,7 @@ enum io_uring_op {
> > > > > IORING_OP_URING_CMD,
> > > > > IORING_OP_SEND_ZC,
> > > > > IORING_OP_SENDMSG_ZC,
> > > > > + IORING_OP_GETDENTS,
> > > > > /* this goes last, obviously */
> > > > > IORING_OP_LAST,
> > > > > @@ -273,6 +275,11 @@ enum io_uring_op {
> > > > > */
> > > > > #define SPLICE_F_FD_IN_FIXED (1U << 31) /* the last bit of __u32 */
> > > > > +/*
> > > > > + * sqe->getdents_flags
> > > > > + */
> > > > > +#define IORING_GETDENTS_REWIND (1U << 0)
> > > > > +
> > > > > /*
> > > > > * POLL_ADD flags. Note that since sqe->poll_events is the flag space, the
> > > > > * command flags for POLL_ADD are stored in sqe->len.
> > > > > diff --git a/io_uring/fs.c b/io_uring/fs.c
> > > > > index f6a69a549fd4..77f00577e09c 100644
> > > > > --- a/io_uring/fs.c
> > > > > +++ b/io_uring/fs.c
> > > > > @@ -47,6 +47,13 @@ struct io_link {
> > > > > int flags;
> > > > > };
> > > > > +struct io_getdents {
> > > > > + struct file *file;
> > > > > + struct linux_dirent64 __user *dirent;
> > > > > + unsigned int count;
> > > > > + int flags;
> > > > > +};
> > > > > +
> > > > > int io_renameat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
> > > > > {
> > > > > struct io_rename *ren = io_kiocb_to_cmd(req, struct io_rename);
> > > > > @@ -291,3 +298,56 @@ void io_link_cleanup(struct io_kiocb *req)
> > > > > putname(sl->oldpath);
> > > > > putname(sl->newpath);
> > > > > }
> > > > > +
> > > > > +int io_getdents_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
> > > > > +{
> > > > > + struct io_getdents *gd = io_kiocb_to_cmd(req, struct io_getdents);
> > > > > +
> > > > > + if (READ_ONCE(sqe->off) != 0)
> > > > > + return -EINVAL;
> > > > > +
> > > > > + gd->dirent = u64_to_user_ptr(READ_ONCE(sqe->addr));
> > > > > + gd->count = READ_ONCE(sqe->len);
> > > > > +
> > > > > + return 0;
> > > > > +}
> > > > > +
> > > > > +int io_getdents(struct io_kiocb *req, unsigned int issue_flags)
> > > > > +{
> > > > > + struct io_getdents *gd = io_kiocb_to_cmd(req, struct io_getdents);
> > > > > + struct file *file;
> > > > > + unsigned long getdents_flags = 0;
> > > > > + bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
> > > > > + bool should_lock = false;
> > > > > + int ret;
> > > > > +
> > > > > + if (force_nonblock) {
> > > > > + if (!(req->file->f_mode & FMODE_NOWAIT))
> > > > > + return -EAGAIN;
> > > > > +
> > > > > + getdents_flags = DIR_CONTEXT_F_NOWAIT;
> > > >
> > > > I mentioned this on the other patch but it seems really pointless to
> > > > have that extra flag. I would really like to hear a good reason for
> > > > this.
> > > >
> > > > > + }
> > > > > +
> > > > > + file = req->file;
> > > > > + if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
> > > > > + if (file_count(file) > 1)
> > > >
> > > > Assume we have a regular non-threaded process that just opens an fd to a
> > > > file. The process registers an async readdir request via that fd for the
> > > > file with io_uring and goes to do other stuff while waiting for the
> > > > result.
> > > >
> > > > Some time later, io_uring gets to io_getdents() and the task is still
> > > > single threaded and the file hasn't been shared in the meantime. So
> > > > io_getdents() doesn't take the lock and starts the readdir() call.
> > > >
> > > > Concurrently, the process that registered the io_uring request was free
> > > > to do other stuff and issued a synchronous readdir() system call which
> > > > calls fdget_pos(). Since the fdtable still isn't shared it doesn't
> > > > increment f_count and doesn't acquire the mutex. Now there's another
> > > > concurrent readdir() going on.
> > > >
> > > > (Similar thing can happen if the process creates a thread for example.)
> > > >
> > > > Two readdir() requests now proceed concurrently which is not intended.
> > > > Now to verify that this race can't happen with io_uring:
> > > >
> > > > * regular fds:
> > > > It seems that io_uring calls fget() on each regular file descriptor
> > > > when an async request is registered. So that means that io_uring
> > > > always hold its own explicit reference here.
> > > > So as long as the original task is alive or another thread is alive
> > > > f_count is guaranteed to be > 1 and so the mutex would always be
> > > > acquired.
> > > >
> > > > If the registering process dies right before io_uring gets to the
> > > > io_getdents() request no other process can steal the fd anymore and in
> > > > that case the readdir call would not lock. But that's fine.
> > > >
> > > > * fixed fds:
> > > > I don't know the reference counting rules here. io_uring would need to
> > > > ensure that it's impossible for two async readdir requests via a fixed
> > > > fd to race because f_count is == 1.
> > > >
> > > > Iiuc, if a process registers a file it opened as a fixed file and
> > > > immediately closes the fd afterwards - without anyone else holding a
> > > > reference to that file - and only uses the fixed fd going forward, the
> > > > f_count of that file in io_uring's fixed file table is always 1.
> > > >
> > > > So one could issue any number of concurrent readdir requests with no
> > > > mutual exclusion. So for fixed files there definitely is a race, no?
> > >
> > > Hi Christian,
> > > The ref logic for fixed file is that it does fdget() when registering
> >
> > It absolutely can't be the case that io_uring uses fdget()/fdput() for
> > long-term file references. fdget() internally use __fget_light() which
> > avoids taking a reference on the file if the file table isn't shared. So
> > should that file be stashed anywhere for async work its a UAF waiting to
> > happen.
> >
>
> Yes, I typed the wrong name, should be fget() not fdget().
>
> > > the file, and fdput() when unregistering it. So the ref in between is
> > > always > 1. The fixed file feature is to reduce frequent fdget/fdput,
> > > but it does call them at the register/unregister time.
> >
> > So consider:
> >
> > // Caller opens some file.
> > fd_register = open("/some/file", ...); // f_count == 1
> >
> > // Caller registers that file as a fixed file
> > IORING_REGISTER_FILES
> > -> io_sqe_files_register()
> > -> fget(fd_register) // f_count == 2
> > -> io_fixed_file_set()
> >
> > // Caller trades regular fd reference for fixed file reference completely.
> > close(fd_register);
> > -> close_fd(fd_register)
> > -> file = pick_file()
> > -> filp_close(file)
> > -> fput(file) // f_count == 1
> >
> >
> > // Caller spawns a second thread. Both treads issue async getdents via
> > // fixed file.
> > T1 T2
> > IORING_OP_GETDENTS IORING_OP_GETDENTS
> >
> > // At some point io_assign_file() must be called which has:
> >
> > if (req->flags & REQ_F_FIXED_FILE)
> > req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
> > else
> > req->file = io_file_get_normal(req, req->cqe.fd);
> >
> > // Since this is REQ_F_FIXED_FILE f_count == 1
> >
> > if (file && (file->f_mode & FMODE_ATOMIC_POS)) {
> > if (file_count(file) > 1)
> >
> > // No lock is taken; T1 and T2 issue getdents concurrently without any
> > // locking. -> race on f_pos
> >
> > I'm happy to be convinced that this is safe, but please someone explain
> > in detail why this can't happen and where that extra f_count reference
> > for fixed files that this code wants to rely on is coming from.
> >
> > Afaik, the whole point is that fixed files don't ever call fget()/fput()
> > after having been registered anymore. Consequently, f_count should be 1
> > once io_uring has taken full ownership of the file and the file can only
> > be addressed via a fixed file reference.
>
> Thanks for explanation, I now realize it's an issue, even for non-fixed
> files when io_uring takes full ownership. for example:
>
> io_uring submit a getdents --> f_count == 2, get the lock
> nowait submission fails --> f_count == 2, release the lock
> punt it to io-wq thread and return to userspace
> close(fd) --> f_count == 1
> call sync getdents64 --> doing getdents without lock
> the io-wq thread begins to run --> f_count == 1, doing getdents
> without lock.
>
> Though this looks like a silly use case but users can do that anyway.
>
> How about remove this f_count > 1 small optimization in io_uring and always
> get the lock, looks like it makes big trouble for async
> situation. and there may often be parallel io_uring getdents in the
> same time for a file [1], it may be not very meaningful to do this
> file count optimization.
>
> [1] I believe users will issue multiple async getdents at same time rather
> than issue them one by one to get better performance.
Could someone with perf experience try and remove that f_count == 1
optimization from __fdget_pos() completely and make it always acquire
the mutex? I wonder what the performance impact of that is.
