On 9/12/19 5:02 AM, Richard W.M. Jones wrote:
The test interop-qemu-nbd-tls-certs frequently fails on slow (32 bit) machines in Fedora Koji. (Is crypto slow on these already overloaded machines?)
As we cannot wait for a signal when qemu-nbd is ready start serving, we have to use a sleep. The current sleep is 5 seconds, which is not long enough. Making the sleep longer would work but is inconsiderate for people using faster machines. Therefore replace this with a retry loop with exponential backoff.
I tested this with a simple wrapper around qemu-nbd which did:
sleep 5; exec /usr/bin/qemu-nbd "$@" --- interop/interop.c | 19 +++++++++++++------ 1 file changed, 13 insertions(+), 6 deletions(-)
/* Unfortunately there's no good way to wait for qemu-nbd to start - * serving, so ... + * serving, so we need to retry here. */
qemu 4.1 added --pid-file, where you can test for that file existing to know when the server is up and running. But we still want an adaptive exponential backoff whether we test for a file or just try connections, so this patch is fine.
- sleep (5); - - if (nbd_connect_tcp (nbd, "localhost", port_str) == -1) { - fprintf (stderr, "%s\n", nbd_get_error ()); - goto out; + for (retry = 0; retry < 5; ++retry) { + sleep (1 << retry); + if (nbd_connect_tcp (nbd, "localhost", port_str) == -1) { + fprintf (stderr, "%s\n", nbd_get_error ()); + if (nbd_get_errno () != ECONNREFUSED) + goto out; + } + else break; } + if (retry == 5) + goto out;
#else /* !SERVE_OVER_TCP */
ACK series -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org