On Thu, Oct 08, 2026 at 10:45:39PM +0800, Qiliang Yuan wrote:
> The fault_inject target delays each I/O with an IORING_OP_TIMEOUT, and
> generic_06 relies on the delay to kill the server while an I/O is still
> outstanding.
>
> The timespec of the timeout is a local variable of
> ublk_fault_inject_queue_io(). The SQE only records its address, and
> io_uring reads it when the SQE is submitted, after the function has
> returned. The timeout gets whatever the stack holds by then and expires
> almost at once, so --delay_us has no effect, and generic_06 fails
> because dd completes before the server is killed.
>
> Store the timespec in the per-device fi_opts, and split the delay into
> seconds and nanoseconds so that delays of a second or more give a valid
> timespec.
>
> fio 4k random reads for 10 s at queue depth 1 on a fault_inject device
> with --delay_us 1000000:
>
> before after expected
> I/Os 446927 10 10
> mean latency 18.3 us 1.000 s 1 s
>
> Fixes: 81586652bb1f ("selftests: ublk: add generic_06 for covering fault
> inject")
> Signed-off-by: Qiliang Yuan <[email protected]>
Reviewed-by: Ming Lei <[email protected]>
Thanks,
Ming
