D. J. Bernstein

The iopause library interface

     #include <iopause.h>


     iopause_fd *x;
     unsigned int len;
     struct taia deadline;
     struct taia stamp;
iopause checks for file descriptor readability or writability as specified by x[0].fd, x[0].events, x[1].fd, x[1].events, ..., x[len-1].fd, x[len-1].events. If x[i].events includes the bit IOPAUSE_READ, iopause checks for readability of the descriptor x[i].fd; if x[i].events includes the bit IOPAUSE_WRITE, iopause checks for writability of the descriptor x[i].fd; other bits in x[i].events have undefined effects.

iopause sets the IOPAUSE_READ bit in x[i].revents if it finds that x[i].fd is readable, and it sets the IOPAUSE_WRITE bit in x[i].revents if it finds that x[i].fd is writable. Beware that readability and writability may be destroyed at any moment by other processes with access to the same ofile that x[i].fd refers to.

If there is no readability or writability to report, iopause waits until deadline for something to happen. iopause will return before deadline if a descriptor becomes readable or writable, or an interrupting signal arrives, or some system-defined amount of time passes. iopause sets revents in any case.

You must put a current timestamp into stamp before calling iopause.

Implementation notes

The current implementation of iopause uses the poll function if that is available. On some systems, poll needs to dynamically allocate kernel memory; when not much memory is available, iopause will return immediately, and will report (often incorrectly) that no descriptors are readable or writable. This is a kernel bug, and I encourage vendors to fix it.

If poll is not available, iopause uses the select function. This function cannot see descriptor numbers past a system-defined limit, typically 256 or 1024; iopause will artificially pretend that those descriptors are never readable or writable.

Future implementations of iopause may work around these problems on some systems, at the expense of chewing up all available CPU time.

Both poll and select use relative timeouts rather than absolute deadlines. Some kernels round the timeout down to a multiple of 10 milliseconds; this can burn quite a bit of CPU time as the deadline approaches. iopause compensates for this by adding 20 milliseconds to the timeout.