pa_write() knows two types of operation:
calling send() and calling write()
there is a flag (a pointer to an int) passed to pa_write()
which can remember which write type was successful
if the pointer is NULL or the int is 0, send() is tried first,
with a fallback to write() if send() resulted in ENOTSOCK
pa_fdsem_post() calls pa_write() with a NULL pointer;
unfortunately (at least with HAVE_SYS_EVENTFD_H #define'd) send()
always fails here and write() is called -- causing an extra syscall
quite frequently
strace:
send(17, "\1\0\0\0\0\0\0\0", 8, MSG_NOSIGNAL) = -1 ENOTSOCK (Socket operation on non-socket)
write(17, "\1\0\0\0\0\0\0\0", 8) = 8
the patch adds a write_type field to pa_fdsem to the successful
pa_write() type can be remembered and unnecessary send() calls are
avoided
Signed-off-by: Peter Meerwald <p.meerwald@bct-electronic.com>
Apperently reading from an eventfd can fail, which results in an assert
to be hit. I am not sure about the reason for the failure, but in
attempt to track down the issue the next time is hit this prints a more
useful log message.
https://bugzilla.redhat.com/attachment.cgi?id=386380
