ext/unicorn_http/epollexclusive.h

   1 /*
   2  * This is only intended for use inside a unicorn worker, nowhere else.
   3  * EPOLLEXCLUSIVE somewhat mitigates the thundering herd problem for
   4  * mostly idle processes since we can't use blocking accept4.
   5  * This is NOT intended for use with multi-threaded servers, nor
   6  * single-threaded multi-client ("C10K") servers or anything advanced
   7  * like that.  This use of epoll is only appropriate for a primitive,
   8  * single-client, single-threaded servers like unicorn that need to
   9  * support SIGKILL timeouts and parent death detection.
  10  */
  11 #if defined(HAVE_EPOLL_CREATE1)
  12 #  include <sys/epoll.h>
  13 #  include <errno.h>
  14 #  include <ruby/io.h>
  15 #  include <ruby/thread.h>
  16 #endif /* __linux__ */
  17
  18 #if defined(EPOLLEXCLUSIVE) && defined(HAVE_EPOLL_CREATE1)
  19 #  define USE_EPOLL (1)
  20 #else
  21 #  define USE_EPOLL (0)
  22 #endif
  23
  24 #if USE_EPOLL
  25 /*
  26  * :nodoc:
  27  * returns IO object if EPOLLEXCLUSIVE works and arms readers
  28  */
  29 static VALUE prep_readers(VALUE cls, VALUE readers)
  30 {
  31         long i;
  32         int epfd = epoll_create1(EPOLL_CLOEXEC);
  33         VALUE epio;
  34
  35         if (epfd < 0) rb_sys_fail("epoll_create1");
  36
  37         epio = rb_funcall(cls, rb_intern("for_fd"), 1, INT2NUM(epfd));
  38
  39         Check_Type(readers, T_ARRAY);
  40         for (i = 0; i < RARRAY_LEN(readers); i++) {
  41                 int rc;
  42                 struct epoll_event e;
  43                 rb_io_t *fptr;
  44                 VALUE io = rb_ary_entry(readers, i);
  45
  46                 e.data.u64 = i; /* the reason readers shouldn't change */
  47
  48                 /*
  49                  * I wanted to use EPOLLET here, but maintaining our own
  50                  * equivalent of ep->rdllist in Ruby-space doesn't fit
  51                  * our design at all (and the kernel already has it's own
  52                  * code path for doing it).  So let the kernel spend
  53                  * cycles on maintaining level-triggering.
  54                  */
  55                 e.events = EPOLLEXCLUSIVE | EPOLLIN;
  56                 io = rb_io_get_io(io);
  57                 GetOpenFile(io, fptr);
  58                 rc = epoll_ctl(epfd, EPOLL_CTL_ADD, fptr->fd, &e);
  59                 if (rc < 0) rb_sys_fail("epoll_ctl");
  60         }
  61         return epio;
  62 }
  63 #endif /* USE_EPOLL */
  64
  65 #if USE_EPOLL
  66 struct ep_wait {
  67         struct epoll_event event;
  68         rb_io_t *fptr;
  69         int timeout_msec;
  70 };
  71
  72 static void *do_wait(void *ptr) /* runs w/o GVL */
  73 {
  74         struct ep_wait *epw = ptr;
  75         /*
  76          * Linux delivers epoll events in the order received, and using
  77          * maxevents=1 ensures we pluck one item off ep->rdllist
  78          * at-a-time (c.f. fs/eventpoll.c in linux.git, it's quite
  79          * easy-to-understand for anybody familiar with Ruby C).
  80          */
  81         return (void *)(long)epoll_wait(epw->fptr->fd, &epw->event, 1,
  82                                         epw->timeout_msec);
  83 }
  84
  85 /* :nodoc: */
  86 /* readers must not change between prepare_readers and get_readers */
  87 static VALUE
  88 get_readers(VALUE epio, VALUE ready, VALUE readers, VALUE timeout_msec)
  89 {
  90         struct ep_wait epw;
  91         long n;
  92
  93         Check_Type(ready, T_ARRAY);
  94         Check_Type(readers, T_ARRAY);
  95         epio = rb_io_get_io(epio);
  96         GetOpenFile(epio, epw.fptr);
  97
  98         epw.timeout_msec = NUM2INT(timeout_msec);
  99         n = (long)rb_thread_call_without_gvl(do_wait, &epw, RUBY_UBF_IO, NULL);
 100         if (n < 0) {
 101                 if (errno != EINTR) rb_sys_fail("epoll_wait");
 102         } else if (n > 0) { /* maxevents is hardcoded to 1 */
 103                 VALUE obj = rb_ary_entry(readers, epw.event.data.u64);
 104
 105                 if (RTEST(obj))
 106                         rb_ary_push(ready, obj);
 107         } /* n == 0 : timeout */
 108         return Qfalse;
 109 }
 110 #endif /* USE_EPOLL */
 111
 112 static void init_epollexclusive(VALUE mUnicorn)
 113 {
 114 #if USE_EPOLL
 115         VALUE cWaiter = rb_define_class_under(mUnicorn, "Waiter", rb_cIO);
 116         rb_define_singleton_method(cWaiter, "prep_readers", prep_readers, 1);
 117         rb_define_method(cWaiter, "get_readers", get_readers, 3);
 118 #endif
 119 }