ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev.pod
(Generate patch)

Comparing libev/ev.pod (file contents):
Revision 1.450 by root, Mon Jun 24 00:04:26 2019 UTC vs.
Revision 1.454 by root, Tue Jun 25 05:17:50 2019 UTC

… …
511This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and 511This backend maps C<EV_READ> to C<POLLIN | POLLERR | POLLHUP>, and
512C<EV_WRITE> to C<POLLOUT | POLLERR | POLLHUP>. 512C<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
516Use the linux-specific epoll(7) interface (for both pre- and post-2.6.9 516Use the Linux-specific epoll(7) interface (for both pre- and post-2.6.9
517kernels). 517kernels).
518 518
519For few fds, this backend is a bit little slower than poll and select, but 519For few fds, this backend is a bit little slower than poll and select, but
520it scales phenomenally better. While poll and select usually scale like 520it scales phenomenally better. While poll and select usually scale like
521O(total_fds) where total_fds is the total number of fds (or the highest 521O(total_fds) where total_fds is the total number of fds (or the highest
… …
574This backend maps C<EV_READ> and C<EV_WRITE> in the same way as 574This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
575C<EVBACKEND_POLL>. 575C<EVBACKEND_POLL>.
576 576
577=item C<EVBACKEND_LINUXAIO> (value 64, Linux) 577=item C<EVBACKEND_LINUXAIO> (value 64, Linux)
578 578
579Use the linux-specific linux aio (I<not> C<< aio(7) >> but C<< 579Use the Linux-specific Linux AIO (I<not> C<< aio(7) >> but C<<
580io_submit(2) >>) event interface available in post-4.18 kernels. 580io_submit(2) >>) event interface available in post-4.18 kernels (but libev
581only tries to use it in 4.19+).
582
583This is another Linux train wreck of an event interface.
581 584
582If this backend works for you (as of this writing, it was very 585If this backend works for you (as of this writing, it was very
583experimental), it is the best event interface available on linux and might 586experimental), it is the best event interface available on Linux and might
584be well worth enabling it - if it isn't available in your kernel this will 587be well worth enabling it - if it isn't available in your kernel this will
585be detected and this backend will be skipped. 588be detected and this backend will be skipped.
586 589
587This backend can batch oneshot requests and supports a user-space ring 590This backend can batch oneshot requests and supports a user-space ring
588buffer to receive events. It also doesn't suffer from most of the design 591buffer to receive events. It also doesn't suffer from most of the design
589problems of epoll (such as not being able to remove event sources from 592problems of epoll (such as not being able to remove event sources from
590the epoll set), and generally sounds too good to be true. Because, this 593the epoll set), and generally sounds too good to be true. Because, this
591being the linux kernel, of course it suffers from a whole new set of 594being the Linux kernel, of course it suffers from a whole new set of
592limitations. 595limitations, forcing you to fall back to epoll, inheriting all its design
596issues.
593 597
594For one, it is not easily embeddable (but probably could be done using 598For one, it is not easily embeddable (but probably could be done using
595an event fd at some extra overhead). It also is subject to a system wide 599an event fd at some extra overhead). It also is subject to a system wide
596limit that can be configured in F</proc/sys/fs/aio-max-nr> - each loop 600limit that can be configured in F</proc/sys/fs/aio-max-nr>. If no AIO
597currently requires C<61> of this number. If no aio requests are left, this
598backend will be skipped during initialisation. 601requests are left, this backend will be skipped during initialisation, and
602will switch to epoll when the loop is active.
599 603
600Most problematic in practise, however, is that not all file descriptors 604Most problematic in practice, however, is that not all file descriptors
601work with it. For example, in linux 5.1, tcp sockets, pipes, event fds, 605work with it. For example, in Linux 5.1, TCP sockets, pipes, event fds,
602files, F</dev/null> and a few others are supported, but ttys do not work 606files, F</dev/null> and many others are supported, but ttys do not work
603properly (a known bug that the kernel developers don't care about, see 607properly (a known bug that the kernel developers don't care about, see
604L<https://lore.kernel.org/patchwork/patch/1047453/>), so this is not 608L<https://lore.kernel.org/patchwork/patch/1047453/>), so this is not
605(yet?) a generic event polling interface. 609(yet?) a generic event polling interface.
606 610
611Overall, it seems the Linux developers just don't want it to have a
612generic event handling mechanism other than C<select> or C<poll>.
613
607To work around this latter problem, the current version of libev uses 614To work around all these problem, the current version of libev uses its
608epoll as a fallback for file deescriptor types that do not work. Epoll 615epoll backend as a fallback for file descriptor types that do not work. Or
609is used in, kind of, slow mode that hopefully avoids most of its design 616falls back completely to epoll if the kernel acts up.
610problems and requires 1-3 extra syscalls per active fd every iteration.
611 617
612This backend maps C<EV_READ> and C<EV_WRITE> in the same way as 618This backend maps C<EV_READ> and C<EV_WRITE> in the same way as
613C<EVBACKEND_POLL>. 619C<EVBACKEND_POLL>.
614 620
615=item C<EVBACKEND_KQUEUE> (value 8, most BSD clones) 621=item C<EVBACKEND_KQUEUE> (value 8, most BSD clones)
616 622
617Kqueue deserves special mention, as at the time of this writing, it 623Kqueue deserves special mention, as at the time this backend was
618was broken on all BSDs except NetBSD (usually it doesn't work reliably 624implemented, it was broken on all BSDs except NetBSD (usually it doesn't
619with anything but sockets and pipes, except on Darwin, where of course 625work reliably with anything but sockets and pipes, except on Darwin,
620it's completely useless). Unlike epoll, however, whose brokenness 626where of course it's completely useless). Unlike epoll, however, whose
621is by design, these kqueue bugs can (and eventually will) be fixed 627brokenness is by design, these kqueue bugs can be (and mostly have been)
622without API changes to existing programs. For this reason it's not being 628fixed without API changes to existing programs. For this reason it's not
623"auto-detected" unless you explicitly specify it in the flags (i.e. using 629being "auto-detected" on all platforms unless you explicitly specify it
624C<EVBACKEND_KQUEUE>) or libev was compiled on a known-to-be-good (-enough) 630in the flags (i.e. using C<EVBACKEND_KQUEUE>) or libev was compiled on a
625system like NetBSD. 631known-to-be-good (-enough) system like NetBSD.
626 632
627You still can embed kqueue into a normal poll or select backend and use it 633You still can embed kqueue into a normal poll or select backend and use it
628only for sockets (after having made sure that sockets work with kqueue on 634only for sockets (after having made sure that sockets work with kqueue on
629the target platform). See C<ev_embed> watchers for more info. 635the target platform). See C<ev_embed> watchers for more info.
630 636
631It scales in the same way as the epoll backend, but the interface to the 637It scales in the same way as the epoll backend, but the interface to the
632kernel is more efficient (which says nothing about its actual speed, of 638kernel is more efficient (which says nothing about its actual speed, of
633course). While stopping, setting and starting an I/O watcher does never 639course). While stopping, setting and starting an I/O watcher does never
634cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to 640cause an extra system call as with C<EVBACKEND_EPOLL>, it still adds up to
635two event changes per incident. Support for C<fork ()> is very bad (you 641two event changes per incident. Support for C<fork ()> is very bad (you
636might have to leak fd's on fork, but it's more sane than epoll) and it 642might have to leak fds on fork, but it's more sane than epoll) and it
637drops fds silently in similarly hard-to-detect cases. 643drops fds silently in similarly hard-to-detect cases.
638 644
639This backend usually performs well under most conditions. 645This backend usually performs well under most conditions.
640 646
641While nominally embeddable in other event loops, this doesn't work 647While nominally embeddable in other event loops, this doesn't work

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines