Re: [PATCH v2] io_uring: let IORING_OP_FILES_UPDATE support to choose fixed file slots

[Pavel Begunkov <asml.silence@gmail.com>]

On 6/13/22 04:35, Xiaoguang Wang wrote:
> hello Pavel,
> 
>> On 5/30/22 14:15, Xiaoguang Wang wrote:
>>> One big issue with file registration feature is that it needs user
>>> space apps to maintain free slot info about io_uring's fixed file
>>> table, which really is a burden for development. Now since io_uring
>>> starts to choose free file slot for user space apps by using
>>> IORING_FILE_INDEX_ALLOC flag in accept or open operations, but they
>>> need app to uses direct accept or direct open, which as far as I know,
>>> some apps are not prepared to use direct accept or open yet.
>>>
>>> To support apps, who still need real fds, use registration feature
>>> easier, let IORING_OP_FILES_UPDATE support to choose fixed file slots,
>>> which will store picked fixed files slots in fd array and let cqe return
>>> the number of slots allocated.
>>
>> Why close bits are piggybacked in this patch without any mention
>> in the commit message?
> Sorry, it actually should be in one separate patch.
> 
>> What is error semantics of
>> IORING_CLOSE_FD_AND_FILE_SLOT. Fail if any errored or both? How
>> can it be reliably used? Why we do two separate things in one
>> request with not clear semantics?
> Yeah, indeed I know this issue when writing this patch, either normal
> fd close or file slot close may fail independently.  I put them together
> just for efficiency, one CLOSE op can do two jobs.
> 
>>
>> There is already one fix pending. Another problem on the surface
>> is that it may call io_close() twice and the second will fail, e.g.
>>
>> -> io_close()
>> ---> close_fixed()
>> ---> if (file->flush) return -EAGAIN
>> -> io_close()
>> ---> close_fixed() // fails
> Thanks for pointing this issue.
> 
>>
>>
>> I do think we need to revert the close change, and then remake
>> properly and only if someone can describe sane semantics for it.
> Revert this close change seems reasonable, I thought for a while,
> didn't figure out better solutions yet.
> Sorry for the trouble again.

No worries, let's kill it off then and return it back if and when
someone comes up with a good semantaics for error handling.



> 
> Regards,
> Xiaoguang Wang
>>
>>
>>
>>> Suggested-by: Hao Xu <howeyxu@tencent.com>
>>> Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
>>> ---
>>> V2:
>>>     Add illegal flags check in io_close_prep().
>>> ---
>>>    fs/io_uring.c                 | 75 +++++++++++++++++++++++++++++++++++++------
>>>    include/uapi/linux/io_uring.h |  1 +
>>>    2 files changed, 66 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>> index 6d91148e9679..18a7459fb6e7 100644
>>> --- a/fs/io_uring.c
>>> +++ b/fs/io_uring.c
>>> @@ -574,6 +574,7 @@ struct io_close {
>>>        struct file            *file;
>>>        int                fd;
>>>        u32                file_slot;
>>> +    u32                flags;
>>>    };
>>>      struct io_timeout_data {
>>> @@ -1366,7 +1367,9 @@ static int io_req_prep_async(struct io_kiocb *req);
>>>      static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
>>>                     unsigned int issue_flags, u32 slot_index);
>>> -static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
>>> +static int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
>>> +                unsigned int offset);
>>> +static inline int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
>>>      static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
>>>    static void io_eventfd_signal(struct io_ring_ctx *ctx);
>>> @@ -5945,16 +5948,22 @@ static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
>>>        return 0;
>>>    }
>>>    +#define IORING_CLOSE_FD_AND_FILE_SLOT 1
>>> +
>>>    static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>>    {
>>> -    if (sqe->off || sqe->addr || sqe->len || sqe->rw_flags || sqe->buf_index)
>>> +    if (sqe->off || sqe->addr || sqe->len || sqe->buf_index)
>>>            return -EINVAL;
>>>        if (req->flags & REQ_F_FIXED_FILE)
>>>            return -EBADF;
>>>          req->close.fd = READ_ONCE(sqe->fd);
>>>        req->close.file_slot = READ_ONCE(sqe->file_index);
>>> -    if (req->close.file_slot && req->close.fd)
>>> +    req->close.flags = READ_ONCE(sqe->close_flags);
>>> +    if (req->close.flags & ~IORING_CLOSE_FD_AND_FILE_SLOT)
>>> +        return -EINVAL;
>>> +    if (!(req->close.flags & IORING_CLOSE_FD_AND_FILE_SLOT) &&
>>> +        req->close.file_slot && req->close.fd)
>>>            return -EINVAL;
>>>          return 0;
>>> @@ -5970,7 +5979,8 @@ static int io_close(struct io_kiocb *req, unsigned int issue_flags)
>>>          if (req->close.file_slot) {
>>>            ret = io_close_fixed(req, issue_flags);
>>> -        goto err;
>>> +        if (ret || !(req->close.flags & IORING_CLOSE_FD_AND_FILE_SLOT))
>>> +            goto err;
>>>        }
>>>          spin_lock(&files->file_lock);
>>> @@ -8003,6 +8013,42 @@ static int io_files_update_prep(struct io_kiocb *req,
>>>        return 0;
>>>    }
>>>    +static int io_files_update_with_index_alloc(struct io_kiocb *req,
>>> +                        unsigned int issue_flags)
>>> +{
>>> +    __s32 __user *fds = u64_to_user_ptr(req->rsrc_update.arg);
>>> +    struct file *file;
>>> +    unsigned int done, nr_fds = req->rsrc_update.nr_args;
>>> +    int ret, fd;
>>> +
>>> +    for (done = 0; done < nr_fds; done++) {
>>> +        if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
>>> +            ret = -EFAULT;
>>> +            break;
>>> +        }
>>> +
>>> +        file = fget(fd);
>>> +        if (!file) {
>>> +            ret = -EBADF;
>>> +            goto out;
>>> +        }
>>> +        ret = io_fixed_fd_install(req, issue_flags, file,
>>> +                      IORING_FILE_INDEX_ALLOC);
>>> +        if (ret < 0)
>>> +            goto out;
>>> +        if (copy_to_user(&fds[done], &ret, sizeof(ret))) {
>>> +            ret = -EFAULT;
>>> +            __io_close_fixed(req, issue_flags, ret);
>>> +            break;
>>> +        }
>>> +    }
>>> +
>>> +out:
>>> +    if (done)
>>> +        return done;
>>> +    return ret;
>>> +}
>>> +
>>>    static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
>>>    {
>>>        struct io_ring_ctx *ctx = req->ctx;
>>> @@ -8016,10 +8062,14 @@ static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
>>>        up.resv = 0;
>>>        up.resv2 = 0;
>>>    -    io_ring_submit_lock(ctx, issue_flags);
>>> -    ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
>>> -                    &up, req->rsrc_update.nr_args);
>>> -    io_ring_submit_unlock(ctx, issue_flags);
>>> +    if (req->rsrc_update.offset == IORING_FILE_INDEX_ALLOC) {
>>> +        ret = io_files_update_with_index_alloc(req, issue_flags);
>>> +    } else {
>>> +        io_ring_submit_lock(ctx, issue_flags);
>>> +        ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
>>> +                &up, req->rsrc_update.nr_args);
>>> +        io_ring_submit_unlock(ctx, issue_flags);
>>> +    }
>>>          if (ret < 0)
>>>            req_set_fail(req);
>>> @@ -10183,9 +10233,9 @@ static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
>>>        return ret;
>>>    }
>>>    -static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
>>> +static int __io_close_fixed(struct io_kiocb *req, unsigned int issue_flags,
>>> +                unsigned int offset)
>>>    {
>>> -    unsigned int offset = req->close.file_slot - 1;
>>>        struct io_ring_ctx *ctx = req->ctx;
>>>        struct io_fixed_file *file_slot;
>>>        struct file *file;
>>> @@ -10222,6 +10272,11 @@ static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
>>>        return ret;
>>>    }
>>>    +static inline int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
>>> +{
>>> +    return __io_close_fixed(req, issue_flags, req->close.file_slot - 1);
>>> +}
>>> +
>>>    static int __io_sqe_files_update(struct io_ring_ctx *ctx,
>>>                     struct io_uring_rsrc_update2 *up,
>>>                     unsigned nr_args)
>>> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
>>> index 53e7dae92e42..e347b3fea4e4 100644
>>> --- a/include/uapi/linux/io_uring.h
>>> +++ b/include/uapi/linux/io_uring.h
>>> @@ -47,6 +47,7 @@ struct io_uring_sqe {
>>>            __u32        unlink_flags;
>>>            __u32        hardlink_flags;
>>>            __u32        xattr_flags;
>>> +        __u32        close_flags;
>>>        };
>>>        __u64    user_data;    /* data to be passed back at completion time */
>>>        /* pack this to avoid bogus arm OABI complaints */
>>
> 

-- 
Pavel Begunkov