| … | |
… | |
| 499 | 0.1ms) and so on. The biggest issue is fork races, however - if a program |
499 | 0.1ms) and so on. The biggest issue is fork races, however - if a program |
| 500 | forks then I<both> parent and child process have to recreate the epoll |
500 | forks then I<both> parent and child process have to recreate the epoll |
| 501 | set, which can take considerable time (one syscall per file descriptor) |
501 | set, which can take considerable time (one syscall per file descriptor) |
| 502 | and is of course hard to detect. |
502 | and is of course hard to detect. |
| 503 | |
503 | |
| 504 | Epoll is also notoriously buggy - embedding epoll fds I<should> work, but |
504 | Epoll is also notoriously buggy - embedding epoll fds I<should> work, |
| 505 | of course I<doesn't>, and epoll just loves to report events for totally |
505 | but of course I<doesn't>, and epoll just loves to report events for |
| 506 | I<different> file descriptors (even already closed ones, so one cannot |
506 | totally I<different> file descriptors (even already closed ones, so |
| 507 | even remove them from the set) than registered in the set (especially |
507 | one cannot even remove them from the set) than registered in the set |
| 508 | on SMP systems). Libev tries to counter these spurious notifications by |
508 | (especially on SMP systems). Libev tries to counter these spurious |
| 509 | employing an additional generation counter and comparing that against the |
509 | notifications by employing an additional generation counter and comparing |
| 510 | events to filter out spurious ones, recreating the set when required. Last |
510 | that against the events to filter out spurious ones, recreating the set |
|
|
511 | when required. Epoll also errornously rounds down timeouts, but gives you |
|
|
512 | no way to know when and by how much, so sometimes you have to busy-wait |
|
|
513 | because epoll returns immediately despite a nonzero timeout. And last |
| 511 | not least, it also refuses to work with some file descriptors which work |
514 | not least, it also refuses to work with some file descriptors which work |
| 512 | perfectly fine with C<select> (files, many character devices...). |
515 | perfectly fine with C<select> (files, many character devices...). |
| 513 | |
516 | |
| 514 | Epoll is truly the train wreck analog among event poll mechanisms, |
517 | Epoll is truly the train wreck among event poll mechanisms, a frankenpoll, |
| 515 | a frankenpoll, cobbled together in a hurry, no thought to design or |
518 | cobbled together in a hurry, no thought to design or interaction with |
| 516 | interaction with others. |
519 | others. Oh, the pain, will it ever stop... |
| 517 | |
520 | |
| 518 | While stopping, setting and starting an I/O watcher in the same iteration |
521 | While stopping, setting and starting an I/O watcher in the same iteration |
| 519 | will result in some caching, there is still a system call per such |
522 | will result in some caching, there is still a system call per such |
| 520 | incident (because the same I<file descriptor> could point to a different |
523 | incident (because the same I<file descriptor> could point to a different |
| 521 | I<file description> now), so its best to avoid that. Also, C<dup ()>'ed |
524 | I<file description> now), so its best to avoid that. Also, C<dup ()>'ed |
