public inbox for [email protected]
 help / color / mirror / Atom feed
* IORING_OP_FILES_UPDATE fail with ECANCELED when IOSQE_IO_LINK is used
@ 2020-07-13 22:55 Daniele Salvatore Albano
  2020-07-14  2:20 ` Jens Axboe
  0 siblings, 1 reply; 3+ messages in thread
From: Daniele Salvatore Albano @ 2020-07-13 22:55 UTC (permalink / raw)
  To: io-uring

Hi everyone,

I am trying to use IORING_OP_FILES_UPDATE  in combination with
IOSQE_IO_LINK and IORING_OP_RECV / IORING_OP_RECV as linked op but I
keep getting ECANCELED (errno 125).

I am using io_uring (kernel 5.8.0-rc4 built 3 days ago) and liburing (tag 0.7).

I went through the test cases and I wasn't able to find any
combination of the OP and the flag and I can't find any related docs
so I am not sure if the combo isn't allowed.

Although I have found
https://github.com/torvalds/linux/blob/v5.8-rc5/fs/io_uring.c#L4926

 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
    return -EINVAL;

Not sure if this is the reason for which the linked operation is
failing, I don't see in the other *_prep sqe->flags being broadly
checked in general.

I wrote two simple test cases that perform the following sequence of operations:
- open a local file (for the two test cases below /proc/cmdline)
- IORING_OP_FILES_UPDATE +  IOSQE_IO_LINK (only in the first test case)
- IORING_OP_READ + IOSQE_FIXED_FILE

Here a small test case to trigger the issue I am facing

int main() {
    struct io_uring ring = {0};
    uint32_t head, count = 0;
    struct io_uring_sqe *sqe = NULL;
    struct io_uring_cqe *cqe = NULL;
    uint32_t files_map_count = 16;
    const int *files_map_registered = malloc(sizeof(int) * files_map_count);
    memset((void*)files_map_registered, 0, sizeof(int) * files_map_count);

    io_uring_queue_init(16, &ring, 0);
    io_uring_register_files(&ring, files_map_registered, files_map_count);

    int fd = open("/proc/cmdline", O_RDONLY);
    int fd_index = 10;

    sqe = io_uring_get_sqe(&ring);
    io_uring_prep_files_update(sqe, &fd, 1, fd_index);
    io_uring_sqe_set_flags(sqe, IOSQE_IO_LINK);
    sqe->user_data = 1;

    char buffer[512] = {0};
    sqe = io_uring_get_sqe(&ring);
    io_uring_prep_read(sqe, fd_index, &buffer, sizeof(buffer), 0);
    io_uring_sqe_set_flags(sqe, IOSQE_FIXED_FILE);
    sqe->user_data = 2;

    io_uring_submit_and_wait(&ring, 2);

    io_uring_for_each_cqe(&ring, head, cqe) {
        count++;

        fprintf(stdout, "count = %d\n", count);
        fprintf(stdout, "cqe->res = %d\n", cqe->res);
        fprintf(stdout, "cqe->user_data = %llu\n", cqe->user_data);
        fprintf(stdout, "cqe->flags = %u\n", cqe->flags);
    }

    io_uring_cq_advance(&ring, count);

    io_uring_unregister_files(&ring);
    io_uring_queue_exit(&ring);

    return 0;
}

It will report for both the cqes res = -125

Instead if the code doesn't link and wait for the read it works as I
am expecting.

int main() {
    struct io_uring ring = {0};
    uint32_t head, count = 0;
    char buffer[512] = {0};
    struct io_uring_sqe *sqe = NULL;
    struct io_uring_cqe *cqe = NULL;
    uint32_t files_map_count = 16;
    const int *files_map_registered = malloc(sizeof(int) * files_map_count);
    memset((void*)files_map_registered, 0, sizeof(int) * files_map_count);

    io_uring_queue_init(16, &ring, 0);
    io_uring_register_files(&ring, files_map_registered, files_map_count);

    int fd = open("/proc/cmdline", O_RDONLY);
    int fd_index = 10;

    sqe = io_uring_get_sqe(&ring);
    io_uring_prep_files_update(sqe, &fd, 1, fd_index);
    io_uring_sqe_set_flags(sqe, 0);
    sqe->user_data = 1;

    int exit_loop = 0;
    do {
        io_uring_submit_and_wait(&ring, 1);

        io_uring_for_each_cqe(&ring, head, cqe) {
            count++;

            fprintf(stdout, "count = %d\n", count);
            fprintf(stdout, "cqe->res = %d\n", cqe->res);
            fprintf(stdout, "cqe->user_data = %llu\n", cqe->user_data);
            fprintf(stdout, "cqe->flags = %u\n", cqe->flags);

            if (cqe->user_data == 1) {
                sqe = io_uring_get_sqe(&ring);
                io_uring_prep_read(sqe, fd_index, &buffer, sizeof(buffer), 0);
                io_uring_sqe_set_flags(sqe, IOSQE_FIXED_FILE);
                sqe->user_data = 2;
            } else {
                if (cqe->res >= 0) {
                    fprintf(stdout, "buffer = <");
                    fwrite(buffer, cqe->res, 1, stdout);
                    fprintf(stdout, ">\n");
                }

                exit_loop = 1;
            }
        }

        io_uring_cq_advance(&ring, count);
    } while(exit_loop == 0);

    io_uring_unregister_files(&ring);
    io_uring_queue_exit(&ring);

    return 0;
}

The output here is
count = 1
cqe->res = 1
cqe->user_data = 1
cqe->flags = 0
count = 2
cqe->res = 58
cqe->user_data = 2
cqe->flags = 0
buffer = <initrd=....yada yada yada...>

Is this the expected behaviour? If no, any hint? What am I doing wrong?

If the expected behaviour is that IORING_OP_FILES_UPDATE can't be
linked, how can I use it properly to issue a read or another op that
requires to work with fds in a chain of operations?

Sorry for the silly questions!

Thanks!
Daniele

^ permalink raw reply	[flat|nested] 3+ messages in thread
end of thread, other threads:[~2020-07-14 16:50 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2020-07-13 22:55 IORING_OP_FILES_UPDATE fail with ECANCELED when IOSQE_IO_LINK is used Daniele Salvatore Albano
2020-07-14  2:20 ` Jens Axboe
     [not found]   ` <CAKq9yRhAQ_zyE8k5kDYvdh8qAwZZOANFvwaR_T_n66MkJ-=+3w@mail.gmail.com>
2020-07-14 16:50     ` Daniele Salvatore Albano

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox