… | |
… | |
511 | This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and |
511 | This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and |
512 | C<EV_WRITE> to C<POLLOUT | POLLERR | POLLHUP>. |
512 | C<EV_WRITE> to C<POLLOUT | POLLERR | POLLHUP>. |
513 | |
513 | |
514 | =item C<EVBACKEND_EPOLL> (value 4, Linux) |
514 | =item C<EVBACKEND_EPOLL> (value 4, Linux) |
515 | |
515 | |
516 | Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9 |
516 | Use the Linux-specific epoll(7) interface (for both pre- and post-2.6.9 |
517 | kernels). |
517 | kernels). |
518 | |
518 | |
519 | For few fds, this backend is a bit little slower than poll and select, but |
519 | For few fds, this backend is a bit little slower than poll and select, but |
520 | it scales phenomenally better. While poll and select usually scale like |
520 | it scales phenomenally better. While poll and select usually scale like |
521 | O(total_fds) where total_fds is the total number of fds (or the highest |
521 | O(total_fds) where total_fds is the total number of fds (or the highest |
… | |
… | |
574 | This backend maps C<EV_READ> and C<EV_WRITE> in the same way as |
574 | This backend maps C<EV_READ> and C<EV_WRITE> in the same way as |
575 | C<EVBACKEND_POLL>. |
575 | C<EVBACKEND_POLL>. |
576 | |
576 | |
577 | =item C<EVBACKEND_LINUXAIO> (value 64, Linux) |
577 | =item C<EVBACKEND_LINUXAIO> (value 64, Linux) |
578 | |
578 | |
579 | Use the linux-specific linux aio (I<not> C<< aio(7) >>) event interface |
579 | Use the Linux-specific Linux AIO (I<not> C<< aio(7) >> but C<< |
580 | available in post-4.18 kernels. |
580 | io_submit(2) >>) event interface available in post-4.18 kernels (but libev |
|
|
581 | only tries to use it in 4.19+). |
|
|
582 | |
|
|
583 | This is another Linux train wreck of an event interface. |
581 | |
584 | |
582 | If this backend works for you (as of this writing, it was very |
585 | If this backend works for you (as of this writing, it was very |
583 | experimental and only supports a subset of file types), it is the best |
586 | experimental), it is the best event interface available on Linux and might |
584 | event interface available on linux and might be well worth it enabling it |
587 | be well worth enabling it - if it isn't available in your kernel this will |
585 | - if it isn't available in your kernel this will be detected and another |
588 | be detected and this backend will be skipped. |
586 | backend will be chosen. |
|
|
587 | |
589 | |
588 | This backend can batch oneshot requests and uses a user-space ring buffer |
590 | This backend can batch oneshot requests and supports a user-space ring |
589 | to receive events. It also doesn't suffer from most of the design problems |
591 | buffer to receive events. It also doesn't suffer from most of the design |
590 | of epoll (such as not being able to remove event sources from the epoll |
592 | problems of epoll (such as not being able to remove event sources from |
591 | set), and generally sounds too good to be true. Because, this being the |
593 | the epoll set), and generally sounds too good to be true. Because, this |
592 | linux kernel, of course it suffers from a whole new set of limitations. |
594 | being the Linux kernel, of course it suffers from a whole new set of |
|
|
595 | limitations, forcing you to fall back to epoll, inheriting all its design |
|
|
596 | issues. |
593 | |
597 | |
594 | For one, it is not easily embeddable (but probably could be done using |
598 | For one, it is not easily embeddable (but probably could be done using |
595 | an event fd at some extra overhead). It also is subject to various |
599 | an event fd at some extra overhead). It also is subject to a system wide |
596 | arbitrary limits that can be configured in F</proc/sys/fs/aio-max-nr> |
600 | limit that can be configured in F</proc/sys/fs/aio-max-nr>. If no AIO |
597 | and F</proc/sys/fs/aio-nr>), which could lead to it being skipped during |
601 | requests are left, this backend will be skipped during initialisation, and |
598 | initialisation. |
602 | will switch to epoll when the loop is active. |
599 | |
603 | |
600 | Most problematic in practise, however, is that, like kqueue, it requires |
604 | Most problematic in practice, however, is that not all file descriptors |
601 | special support from drivers, and, not surprisingly, not all drivers |
|
|
602 | implement it. For example, in linux 4.19, tcp sockets, pipes, event fds, |
605 | work with it. For example, in Linux 5.1, TCP sockets, pipes, event fds, |
603 | files, F</dev/null> and a few others are supported, but ttys are not, so |
606 | files, F</dev/null> and many others are supported, but ttys do not work |
604 | this is not (yet?) a generic event polling interface but is probably still |
607 | properly (a known bug that the kernel developers don't care about, see |
605 | be very useful in a web server or similar program. |
608 | L<https://lore.kernel.org/patchwork/patch/1047453/>), so this is not |
|
|
609 | (yet?) a generic event polling interface. |
|
|
610 | |
|
|
611 | Overall, it seems the Linux developers just don't want it to have a |
|
|
612 | generic event handling mechanism other than C<select> or C<poll>. |
|
|
613 | |
|
|
614 | To work around all these problem, the current version of libev uses its |
|
|
615 | epoll backend as a fallback for file descriptor types that do not work. Or |
|
|
616 | falls back completely to epoll if the kernel acts up. |
606 | |
617 | |
607 | This backend maps C<EV_READ> and C<EV_WRITE> in the same way as |
618 | This backend maps C<EV_READ> and C<EV_WRITE> in the same way as |
608 | C<EVBACKEND_POLL>. |
619 | C<EVBACKEND_POLL>. |
609 | |
620 | |
610 | =item C<EVBACKEND_KQUEUE> (value 8, most BSD clones) |
621 | =item C<EVBACKEND_KQUEUE> (value 8, most BSD clones) |
611 | |
622 | |
612 | Kqueue deserves special mention, as at the time of this writing, it |
623 | Kqueue deserves special mention, as at the time this backend was |
613 | was broken on all BSDs except NetBSD (usually it doesn't work reliably |
624 | implemented, it was broken on all BSDs except NetBSD (usually it doesn't |
614 | with anything but sockets and pipes, except on Darwin, where of course |
625 | work reliably with anything but sockets and pipes, except on Darwin, |
615 | it's completely useless). Unlike epoll, however, whose brokenness |
626 | where of course it's completely useless). Unlike epoll, however, whose |
616 | is by design, these kqueue bugs can (and eventually will) be fixed |
627 | brokenness is by design, these kqueue bugs can be (and mostly have been) |
617 | without API changes to existing programs. For this reason it's not being |
628 | fixed without API changes to existing programs. For this reason it's not |
618 | "auto-detected" unless you explicitly specify it in the flags (i.e. using |
629 | being "auto-detected" on all platforms unless you explicitly specify it |
619 | C<EVBACKEND_KQUEUE>) or libev was compiled on a known-to-be-good (-enough) |
630 | in the flags (i.e. using C<EVBACKEND_KQUEUE>) or libev was compiled on a |
620 | system like NetBSD. |
631 | known-to-be-good (-enough) system like NetBSD. |
621 | |
632 | |
622 | You still can embed kqueue into a normal poll or select backend and use it |
633 | You still can embed kqueue into a normal poll or select backend and use it |
623 | only for sockets (after having made sure that sockets work with kqueue on |
634 | only for sockets (after having made sure that sockets work with kqueue on |
624 | the target platform). See C<ev_embed> watchers for more info. |
635 | the target platform). See C<ev_embed> watchers for more info. |
625 | |
636 | |
626 | It scales in the same way as the epoll backend, but the interface to the |
637 | It scales in the same way as the epoll backend, but the interface to the |
627 | kernel is more efficient (which says nothing about its actual speed, of |
638 | kernel is more efficient (which says nothing about its actual speed, of |
628 | course). While stopping, setting and starting an I/O watcher does never |
639 | course). While stopping, setting and starting an I/O watcher does never |
629 | cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to |
640 | cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to |
630 | two event changes per incident. Support for C<fork ()> is very bad (you |
641 | two event changes per incident. Support for C<fork ()> is very bad (you |
631 | might have to leak fd's on fork, but it's more sane than epoll) and it |
642 | might have to leak fds on fork, but it's more sane than epoll) and it |
632 | drops fds silently in similarly hard-to-detect cases. |
643 | drops fds silently in similarly hard-to-detect cases. |
633 | |
644 | |
634 | This backend usually performs well under most conditions. |
645 | This backend usually performs well under most conditions. |
635 | |
646 | |
636 | While nominally embeddable in other event loops, this doesn't work |
647 | While nominally embeddable in other event loops, this doesn't work |