| … | |
… | |
| 463 | epoll scales either O(1) or O(active_fds). |
463 | epoll scales either O(1) or O(active_fds). |
| 464 | |
464 | |
| 465 | The epoll mechanism deserves honorable mention as the most misdesigned |
465 | The epoll mechanism deserves honorable mention as the most misdesigned |
| 466 | of the more advanced event mechanisms: mere annoyances include silently |
466 | of the more advanced event mechanisms: mere annoyances include silently |
| 467 | dropping file descriptors, requiring a system call per change per file |
467 | dropping file descriptors, requiring a system call per change per file |
| 468 | descriptor (and unnecessary guessing of parameters), problems with dup and |
468 | descriptor (and unnecessary guessing of parameters), problems with dup, |
|
|
469 | returning before the timeout value, resulting in additional iterations |
|
|
470 | (and only giving 5ms accuracy while select on the same platform gives |
| 469 | so on. The biggest issue is fork races, however - if a program forks then |
471 | 0.1ms) and so on. The biggest issue is fork races, however - if a program |
| 470 | I<both> parent and child process have to recreate the epoll set, which can |
472 | forks then I<both> parent and child process have to recreate the epoll |
| 471 | take considerable time (one syscall per file descriptor) and is of course |
473 | set, which can take considerable time (one syscall per file descriptor) |
| 472 | hard to detect. |
474 | and is of course hard to detect. |
| 473 | |
475 | |
| 474 | Epoll is also notoriously buggy - embedding epoll fds I<should> work, but |
476 | Epoll is also notoriously buggy - embedding epoll fds I<should> work, but |
| 475 | of course I<doesn't>, and epoll just loves to report events for totally |
477 | of course I<doesn't>, and epoll just loves to report events for totally |
| 476 | I<different> file descriptors (even already closed ones, so one cannot |
478 | I<different> file descriptors (even already closed ones, so one cannot |
| 477 | even remove them from the set) than registered in the set (especially |
479 | even remove them from the set) than registered in the set (especially |
| 478 | on SMP systems). Libev tries to counter these spurious notifications by |
480 | on SMP systems). Libev tries to counter these spurious notifications by |
| 479 | employing an additional generation counter and comparing that against the |
481 | employing an additional generation counter and comparing that against the |
| 480 | events to filter out spurious ones, recreating the set when required. Last |
482 | events to filter out spurious ones, recreating the set when required. Last |
| 481 | not least, it also refuses to work with some file descriptors which work |
483 | not least, it also refuses to work with some file descriptors which work |
| 482 | perfectly fine with C<select> (files, many character devices...). |
484 | perfectly fine with C<select> (files, many character devices...). |
|
|
485 | |
|
|
486 | Epoll is truly the train wreck analog among event poll mechanisms. |
| 483 | |
487 | |
| 484 | While stopping, setting and starting an I/O watcher in the same iteration |
488 | While stopping, setting and starting an I/O watcher in the same iteration |
| 485 | will result in some caching, there is still a system call per such |
489 | will result in some caching, there is still a system call per such |
| 486 | incident (because the same I<file descriptor> could point to a different |
490 | incident (because the same I<file descriptor> could point to a different |
| 487 | I<file description> now), so its best to avoid that. Also, C<dup ()>'ed |
491 | I<file description> now), so its best to avoid that. Also, C<dup ()>'ed |
