[ViewVC] Diff of: cvs/libev/ev

Comparing libev/ev_iouring.c (file contents):
Revision 1.8 by root, Fri Dec 27 16:08:24 2019 UTC vs.
Revision 1.20 by root, Sat Dec 28 08:11:20 2019 UTC

  * b) best is not necessarily very good.
  * c) it's better than the aio mess, doesn't suffer from the fork problems
  *    of linux aio or epoll and so on and so on. and you could do event stuff
  *    without any syscalls. what's not to like?
  * d) ok, it's vastly more complex, but that's ok, really.
- * e) why 3 mmaps instead of one? one would be more space-efficient,
+ * e) why two mmaps instead of one? one would be more space-efficient,
- *    and I can't see what benefit three would have (other than being
+ *    and I can't see what benefit two would have (other than being
  *    somehow resizable/relocatable, but that's apparently not possible).
- *    (FIXME: newer kernels can use 2 mmaps only, need to look into this).
- * f) hmm, it's practiclaly undebuggable (gdb can't access the memory, and
+ * f) hmm, it's practically undebuggable (gdb can't access the memory, and
  *    the bizarre way structure offsets are communicated makes it hard to
  *    just print the ring buffer heads, even *iff* the memory were visible
  *    in gdb. but then, that's also ok, really.
  * g) well, you cannot specify a timeout when waiting for events. no,
  *    seriously, the interface doesn't support a timeout. never seen _that_
  *    before. sure, you can use a timerfd, but that's another syscall
  *    you could have avoided. overall, this bizarre omission smells
  *    like a µ-optimisation by the io_uring author for his personal
  *    applications, to the detriment of everybody else who just wants
  *    an event loop. but, umm, ok, if that's all, it could be worse.
- *    (FIXME: jens mentioned timeout commands, need to investigate)
+ *    (from what I gather from the author Jens Axboe, it simply didn't
+ *    occur to him, and he made good on it by adding an unlimited nuber
+ *    of timeouts later :).
- * h) there is a hardcoded limit of 4096 outstanding events. okay,
+ * h) initially there was a hardcoded limit of 4096 outstanding events.
- *    at least there is no arbitrary low system-wide limit...
+ *    later versions not only bump this to 32k, but also can handle
- *    (FIXME: apparently, this was increased to 32768 in later kernels(
+ *    an unlimited amount of events, so this only affects the batch size.
  * i) unlike linux aio, you *can* register more then the limit
- *    of fd events, and the kernel will "gracefully" signal an
+ *    of fd events. while early verisons of io_uring signalled an overflow
- *    overflow, after which you could destroy and recreate the kernel
+ *    and you ended up getting wet. 5.5+ does not do this anymore.
- *    state, a bit bigger, or fall back to e.g. poll. thats not
- *    totally insane, but kind of questions the point a high
- *    performance I/O framework when it doesn't really work
- *    under stress.
- *    (FIXME: iouring should no longer drop events, need to investigate)
- * j) but, oh my! is has exactly the same bugs as the linux aio backend,
+ * j) but, oh my! it had exactly the same bugs as the linux aio backend,
- *    where some undocumented poll combinations just fail.
+ *    where some undocumented poll combinations just fail. fortunately,
- *    so we need epoll AGAIN as a fallback. AGAIN! epoll!! and of course,
+ *    after finally reaching the author, he was more than willing to fix
- *    this is completely undocumented, have I mantioned this already?
+ *    this probably in 5.6+.
  * k) overall, the *API* itself is, I dare to say, not a total trainwreck.
- *    the big isuess with it are the bugs requiring epoll, which might
+ *    once the bugs ae fixed (probably in 5.6+), it will be without
- *    or might not get fixed (do I hold my breath?).
+ *    competition.
  */
+/* TODO: use internal TIMEOUT */
+/* TODO: take advantage of single mmap, NODROP etc. */
+/* TODO: resize cq/sq size independently */
 #include <sys/timerfd.h>
 #include <sys/mman.h>
 #include <poll.h>
+#include <stdint.h>
 #define IOURING_INIT_ENTRIES 32
 /*****************************************************************************/
 /* syscall wrapdadoop - this section has the raw api/abi definitions */
   __u32 resv[4];
   struct io_sqring_offsets sq_off;
   struct io_cqring_offsets cq_off;
 };
+#define IORING_SETUP_CQSIZE 0x00000008
-#define IORING_OP_POLL_ADD    6
+#define IORING_OP_POLL_ADD        6
-#define IORING_OP_POLL_REMOVE 7
+#define IORING_OP_POLL_REMOVE     7
+#define IORING_OP_TIMEOUT        11
+#define IORING_OP_TIMEOUT_REMOVE 12
+/* relative or absolute, reference clock is CLOCK_MONOTONIC */
+struct iouring_kernel_timespec
+{
+  int64_t tv_sec;
+  long long tv_nsec;
+};
+#define IORING_TIMEOUT_ABS 0x00000001
 #define IORING_ENTER_GETEVENTS 0x01
 #define IORING_OFF_SQ_RING 0x00000000ULL
 #define IORING_OFF_CQ_RING 0x08000000ULL
 #define IORING_OFF_SQES	   0x10000000ULL
-#define IORING_FEAT_SINGLE_MMAP   0x1
+#define IORING_FEAT_SINGLE_MMAP   0x00000001
-#define IORING_FEAT_NODROP        0x2
+#define IORING_FEAT_NODROP        0x00000002
-#define IORING_FEAT_SUBMIT_STABLE 0x4
+#define IORING_FEAT_SUBMIT_STABLE 0x00000004
 inline_size
 int
 evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
 {
 /* the submit/completion queue entries */
 #define EV_SQES         ((struct io_uring_sqe *)         iouring_sqes)
 #define EV_CQES         ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
+inline_speed
+int
+iouring_enter (EV_P_ ev_tstamp timeout)
+{
+  int res;
+  EV_RELEASE_CB;
+  res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
+                              timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
+  assert (("libev: io_uring_enter did not consume all sqes", (res < 0 || res == iouring_to_submit)));
+  iouring_to_submit = 0;
+  EV_ACQUIRE_CB;
+  return res;
+}
+/* TODO: can we move things around so we don't need this forward-reference? */
+static void
+iouring_poll (EV_P_ ev_tstamp timeout);
 static
 struct io_uring_sqe *
 iouring_sqe_get (EV_P)
 {
+  unsigned tail;
+  for (;;)
+    {
-  unsigned tail = EV_SQ_VAR (tail);
+      tail = EV_SQ_VAR (tail);
-  if (tail + 1 - EV_SQ_VAR (head) > EV_SQ_VAR (ring_entries))
+      if (ecb_expect_true (tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)))
+        break; /* whats the problem, we have free sqes */
+      /* queue full, need to flush and possibly handle some events */
+#if EV_FEATURE_CODE
+      /* first we ask the kernel nicely, most often this frees up some sqes */
+      int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
+      ECB_MEMORY_FENCE_ACQUIRE; /* better safe than sorry */
+      if (res >= 0)
+        continue; /* yes, it worked, try again */
+#endif
+      /* some problem, possibly EBUSY - do the full poll and let it handle any issues */
+      iouring_poll (EV_A_ EV_TS_CONST (0.));
+      /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE for us */
-    {
+    }
-      /* queue full, flush */
-      evsys_io_uring_enter (iouring_fd, iouring_to_submit, 0, 0, 0, 0);
-      iouring_to_submit = 0;
-    }
-  assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));
+  /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
   return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
 }
 inline_size
   iouring_tfd     = -1;
   iouring_sq_ring = MAP_FAILED;
   iouring_cq_ring = MAP_FAILED;
   iouring_sqes    = MAP_FAILED;
+  if (!have_monotonic) /* cannot really happen, but what if11 */
+    return -1;
   for (;;)
     {
       iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
       if (iouring_fd >= 0)
         break; /* yippie */
       if (errno != EINVAL)
         return -1; /* we failed */
+#if TODO
+      if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP | IORING_FEAT_SUBMIT_STABLE))
+        return -1; /* we require the above features */
+#endif
       /* EINVAL: lots of possible reasons, but maybe
        * it is because we hit the unqueryable hardcoded size limit
        */
     {
       /* we assume the sqe's are all "properly" initialised */
       struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
       sqe->opcode    = IORING_OP_POLL_REMOVE;
       sqe->fd        = fd;
+      /* Jens Axboe notified me that user_data is not what is documented, but is
+       * some kind of unique ID that has to match, otherwise the request cannot
+       * be removed. Since we don't *really* have that, we pass in the old
+       * generation counter - if that fails, too bad, it will hopefully be removed
+       * at close time and then be ignored. */
+      sqe->addr      = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
-      sqe->user_data = -1;
+      sqe->user_data = (uint64_t)-1;
       iouring_sqe_submit (EV_A_ sqe);
       /* increment generation counter to avoid handling old events */
       ++anfds [fd].egen;
     }
   if (nev)
     {
       struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
       sqe->opcode      = IORING_OP_POLL_ADD;
       sqe->fd          = fd;
+      sqe->addr        = 0;
       sqe->user_data   = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
       sqe->poll_events =
         (nev & EV_READ ? POLLIN : 0)
         | (nev & EV_WRITE ? POLLOUT : 0);
       iouring_sqe_submit (EV_A_ sqe);
 {
   int      fd  = cqe->user_data & 0xffffffffU;
   uint32_t gen = cqe->user_data >> 32;
   int      res = cqe->res;
-  /* ignore fd removal events, if there are any. TODO: verify */
+  /* user_data -1 is a remove that we are not atm. interested in */
-  if (cqe->user_data == (__u64)-1)
+  if (cqe->user_data == (uint64_t)-1)
-    abort ();//D
+    return;
   assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
   /* documentation lies, of course. the result value is NOT like
    * normal syscalls, but like linux raw syscalls, i.e. negative
    * error numbers. fortunate, as otherwise there would be no way
    * to get error codes at all. still, why not document this?
    */
   /* ignore event if generation doesn't match */
+  /* other than skipping removal events, */
   /* this should actually be very rare */
   if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
     return;
   if (ecb_expect_false (res < 0))
     {
-      //TODO: EINVAL handling (was something failed with this fd)
+      /*TODO: EINVAL handling (was something failed with this fd)*/
-      //TODO: EBUSY happens when?
       if (res == -EBADF)
         {
           assert (("libev: event loop rejected bad fd", res != -EBADF));
           fd_kill (EV_A_ fd);
 iouring_overflow (EV_P)
 {
   /* we have two options, resize the queue (by tearing down
    * everything and recreating it, or living with it
    * and polling.
-   * we implement this by resizing tghe queue, and, if that fails,
+   * we implement this by resizing the queue, and, if that fails,
    * we just recreate the state on every failure, which
    * kind of is a very inefficient poll.
    * one danger is, due to the bios toward lower fds,
    * we will only really get events for those, so
    * maybe we need a poll() fallback, after all.
   else
     {
       /* we hit the kernel limit, we should fall back to something else.
        * we can either poll() a few times and hope for the best,
        * poll always, or switch to epoll.
-       * since we use epoll anyways, go epoll.
+       * TODO: is this necessary with newer kernels?
        */
       iouring_internal_destroy (EV_A);
-      /* this should make it so that on return, we don'T call any uring functions */
+      /* this should make it so that on return, we don't call any uring functions */
       iouring_to_submit = 0;
       for (;;)
         {
           backend = epoll_init (EV_A_ 0);
 static void
 iouring_poll (EV_P_ ev_tstamp timeout)
 {
   /* if we have events, no need for extra syscalls, but we might have to queue events */
+  /* we also clar the timeout if there are outstanding fdchanges */
+  /* the latter should only happen if both the sq and cq are full, most likely */
+  /* because we have a lot of event sources that immediately complete */
+  /* TODO: fdchacngecnt is always 0 because fd_reify does not have two buffers yet */
-  if (iouring_handle_cq (EV_A))
+  if (iouring_handle_cq (EV_A) || fdchangecnt)
     timeout = EV_TS_CONST (0.);
   else
     /* no events, so maybe wait for some */
     iouring_tfd_update (EV_A_ timeout);
   /* only enter the kernel if we have something to submit, or we need to wait */
   if (timeout || iouring_to_submit)
     {
-      int res;
+      int res = iouring_enter (EV_A_ timeout);
-      EV_RELEASE_CB;
-      res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
-                                  timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
-      iouring_to_submit = 0;
-      EV_ACQUIRE_CB;
       if (ecb_expect_false (res < 0))
         if (errno == EINTR)
           /* ignore */;
+        else if (errno == EBUSY)
+          /* cq full, cannot submit - should be rare because we flush the cq first, so simply ignore */;
         else
           ev_syserr ("(libev) iouring setup");
       else
         iouring_handle_cq (EV_A);
     }

Diff Legend

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

Comparing libev/ev_iouring.c (file contents): Revision 1.8 by root, Fri Dec 27 16:08:24 2019 UTC vs. Revision 1.20 by root, Sat Dec 28 08:11:20 2019 UTC

Diff Legend

Comparing libev/ev_iouring.c (file contents):
Revision 1.8 by root, Fri Dec 27 16:08:24 2019 UTC vs.
Revision 1.20 by root, Sat Dec 28 08:11:20 2019 UTC