"Fossies" - the Fresh Open Source Software Archive

Member "libev-4.33/ev_iouring.c" (8 Mar 2020, 21212 Bytes) of package /linux/misc/libev-4.33.tar.gz:


As a special service "Fossies" has tried to format the requested source page into HTML format using (guessed) C and C++ source code syntax highlighting (style: standard) with prefixed line numbers and code folding option. Alternatively you can here view or download the uninterpreted source code file. For more information about "ev_iouring.c" see the Fossies "Dox" file reference documentation and the latest Fossies "Diffs" side-by-side code changes report: 4.31_vs_4.33.

    1 /*
    2  * libev linux io_uring fd activity backend
    3  *
    4  * Copyright (c) 2019-2020 Marc Alexander Lehmann <libev@schmorp.de>
    5  * All rights reserved.
    6  *
    7  * Redistribution and use in source and binary forms, with or without modifica-
    8  * tion, are permitted provided that the following conditions are met:
    9  *
   10  *   1.  Redistributions of source code must retain the above copyright notice,
   11  *       this list of conditions and the following disclaimer.
   12  *
   13  *   2.  Redistributions in binary form must reproduce the above copyright
   14  *       notice, this list of conditions and the following disclaimer in the
   15  *       documentation and/or other materials provided with the distribution.
   16  *
   17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
   18  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
   19  * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO
   20  * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
   21  * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
   22  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
   23  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
   24  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
   25  * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
   26  * OF THE POSSIBILITY OF SUCH DAMAGE.
   27  *
   28  * Alternatively, the contents of this file may be used under the terms of
   29  * the GNU General Public License ("GPL") version 2 or any later version,
   30  * in which case the provisions of the GPL are applicable instead of
   31  * the above. If you wish to allow the use of your version of this file
   32  * only under the terms of the GPL and not to allow others to use your
   33  * version of this file under the BSD license, indicate your decision
   34  * by deleting the provisions above and replace them with the notice
   35  * and other provisions required by the GPL. If you do not delete the
   36  * provisions above, a recipient may use your version of this file under
   37  * either the BSD or the GPL.
   38  */
   39 
   40 /*
   41  * general notes about linux io_uring:
   42  *
   43  * a) it's the best interface I have seen so far. on linux.
   44  * b) best is not necessarily very good.
   45  * c) it's better than the aio mess, doesn't suffer from the fork problems
   46  *    of linux aio or epoll and so on and so on. and you could do event stuff
   47  *    without any syscalls. what's not to like?
   48  * d) ok, it's vastly more complex, but that's ok, really.
   49  * e) why two mmaps instead of one? one would be more space-efficient,
   50  *    and I can't see what benefit two would have (other than being
   51  *    somehow resizable/relocatable, but that's apparently not possible).
   52  * f) hmm, it's practically undebuggable (gdb can't access the memory, and
   53  *    the bizarre way structure offsets are communicated makes it hard to
   54  *    just print the ring buffer heads, even *iff* the memory were visible
   55  *    in gdb. but then, that's also ok, really.
   56  * g) well, you cannot specify a timeout when waiting for events. no,
   57  *    seriously, the interface doesn't support a timeout. never seen _that_
   58  *    before. sure, you can use a timerfd, but that's another syscall
   59  *    you could have avoided. overall, this bizarre omission smells
   60  *    like a ยต-optimisation by the io_uring author for his personal
   61  *    applications, to the detriment of everybody else who just wants
   62  *    an event loop. but, umm, ok, if that's all, it could be worse.
   63  *    (from what I gather from the author Jens Axboe, it simply didn't
   64  *    occur to him, and he made good on it by adding an unlimited nuber
   65  *    of timeouts later :).
   66  * h) initially there was a hardcoded limit of 4096 outstanding events.
   67  *    later versions not only bump this to 32k, but also can handle
   68  *    an unlimited amount of events, so this only affects the batch size.
   69  * i) unlike linux aio, you *can* register more then the limit
   70  *    of fd events. while early verisons of io_uring signalled an overflow
   71  *    and you ended up getting wet. 5.5+ does not do this anymore.
   72  * j) but, oh my! it had exactly the same bugs as the linux aio backend,
   73  *    where some undocumented poll combinations just fail. fortunately,
   74  *    after finally reaching the author, he was more than willing to fix
   75  *    this probably in 5.6+.
   76  * k) overall, the *API* itself is, I dare to say, not a total trainwreck.
   77  *    once the bugs ae fixed (probably in 5.6+), it will be without
   78  *    competition.
   79  */
   80 
   81 /* TODO: use internal TIMEOUT */
   82 /* TODO: take advantage of single mmap, NODROP etc. */
   83 /* TODO: resize cq/sq size independently */
   84 
   85 #include <sys/timerfd.h>
   86 #include <sys/mman.h>
   87 #include <poll.h>
   88 #include <stdint.h>
   89 
   90 #define IOURING_INIT_ENTRIES 32
   91 
   92 /*****************************************************************************/
   93 /* syscall wrapdadoop - this section has the raw api/abi definitions */
   94 
   95 #include <linux/fs.h>
   96 #include <linux/types.h>
   97 
   98 /* mostly directly taken from the kernel or documentation */
   99 
  100 struct io_uring_sqe
  101 {
  102   __u8 opcode;
  103   __u8 flags;
  104   __u16 ioprio;
  105   __s32 fd;
  106   union {
  107     __u64 off;
  108     __u64 addr2;
  109   };
  110   __u64 addr;
  111   __u32 len;
  112   union {
  113     __kernel_rwf_t rw_flags;
  114     __u32 fsync_flags;
  115     __u16 poll_events;
  116     __u32 sync_range_flags;
  117     __u32 msg_flags;
  118     __u32 timeout_flags;
  119     __u32 accept_flags;
  120     __u32 cancel_flags;
  121     __u32 open_flags;
  122     __u32 statx_flags;
  123   };
  124   __u64 user_data;
  125   union {
  126     __u16 buf_index;
  127     __u64 __pad2[3];
  128   };
  129 };
  130 
  131 struct io_uring_cqe
  132 {
  133   __u64 user_data;
  134   __s32 res;
  135   __u32 flags;
  136 };
  137 
  138 struct io_sqring_offsets
  139 {
  140   __u32 head;
  141   __u32 tail;
  142   __u32 ring_mask;
  143   __u32 ring_entries;
  144   __u32 flags;
  145   __u32 dropped;
  146   __u32 array;
  147   __u32 resv1;
  148   __u64 resv2;
  149 };
  150 
  151 struct io_cqring_offsets
  152 {
  153   __u32 head;
  154   __u32 tail;
  155   __u32 ring_mask;
  156   __u32 ring_entries;
  157   __u32 overflow;
  158   __u32 cqes;
  159   __u64 resv[2];
  160 };
  161 
  162 struct io_uring_params
  163 {
  164   __u32 sq_entries;
  165   __u32 cq_entries;
  166   __u32 flags;
  167   __u32 sq_thread_cpu;
  168   __u32 sq_thread_idle;
  169   __u32 features;
  170   __u32 resv[4];
  171   struct io_sqring_offsets sq_off;
  172   struct io_cqring_offsets cq_off;
  173 };
  174 
  175 #define IORING_SETUP_CQSIZE 0x00000008
  176 
  177 #define IORING_OP_POLL_ADD        6
  178 #define IORING_OP_POLL_REMOVE     7
  179 #define IORING_OP_TIMEOUT        11
  180 #define IORING_OP_TIMEOUT_REMOVE 12
  181 
  182 /* relative or absolute, reference clock is CLOCK_MONOTONIC */
  183 struct iouring_kernel_timespec
  184 {
  185   int64_t tv_sec;
  186   long long tv_nsec;
  187 };
  188 
  189 #define IORING_TIMEOUT_ABS 0x00000001
  190 
  191 #define IORING_ENTER_GETEVENTS 0x01
  192 
  193 #define IORING_OFF_SQ_RING 0x00000000ULL
  194 #define IORING_OFF_CQ_RING 0x08000000ULL
  195 #define IORING_OFF_SQES    0x10000000ULL
  196 
  197 #define IORING_FEAT_SINGLE_MMAP   0x00000001
  198 #define IORING_FEAT_NODROP        0x00000002
  199 #define IORING_FEAT_SUBMIT_STABLE 0x00000004
  200 
  201 inline_size
  202 int
  203 evsys_io_uring_setup (unsigned entries, struct io_uring_params *params)
  204 {
  205   return ev_syscall2 (SYS_io_uring_setup, entries, params);
  206 }
  207 
  208 inline_size
  209 int
  210 evsys_io_uring_enter (int fd, unsigned to_submit, unsigned min_complete, unsigned flags, const sigset_t *sig, size_t sigsz)
  211 {
  212   return ev_syscall6 (SYS_io_uring_enter, fd, to_submit, min_complete, flags, sig, sigsz);
  213 }
  214 
  215 /*****************************************************************************/
  216 /* actual backed implementation */
  217 
  218 /* we hope that volatile will make the compiler access this variables only once */
  219 #define EV_SQ_VAR(name) *(volatile unsigned *)((char *)iouring_sq_ring + iouring_sq_ ## name)
  220 #define EV_CQ_VAR(name) *(volatile unsigned *)((char *)iouring_cq_ring + iouring_cq_ ## name)
  221 
  222 /* the index array */
  223 #define EV_SQ_ARRAY     ((unsigned *)((char *)iouring_sq_ring + iouring_sq_array))
  224 
  225 /* the submit/completion queue entries */
  226 #define EV_SQES         ((struct io_uring_sqe *)         iouring_sqes)
  227 #define EV_CQES         ((struct io_uring_cqe *)((char *)iouring_cq_ring + iouring_cq_cqes))
  228 
  229 inline_speed
  230 int
  231 iouring_enter (EV_P_ ev_tstamp timeout)
  232 {
  233   int res;
  234 
  235   EV_RELEASE_CB;
  236 
  237   res = evsys_io_uring_enter (iouring_fd, iouring_to_submit, 1,
  238                               timeout > EV_TS_CONST (0.) ? IORING_ENTER_GETEVENTS : 0, 0, 0);
  239 
  240   assert (("libev: io_uring_enter did not consume all sqes", (res < 0 || res == iouring_to_submit)));
  241 
  242   iouring_to_submit = 0;
  243 
  244   EV_ACQUIRE_CB;
  245 
  246   return res;
  247 }
  248 
  249 /* TODO: can we move things around so we don't need this forward-reference? */
  250 static void
  251 iouring_poll (EV_P_ ev_tstamp timeout);
  252 
  253 static
  254 struct io_uring_sqe *
  255 iouring_sqe_get (EV_P)
  256 {
  257   unsigned tail;
  258   
  259   for (;;)
  260     {
  261       tail = EV_SQ_VAR (tail);
  262 
  263       if (ecb_expect_true (tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)))
  264         break; /* whats the problem, we have free sqes */
  265 
  266       /* queue full, need to flush and possibly handle some events */
  267 
  268 #if EV_FEATURE_CODE
  269       /* first we ask the kernel nicely, most often this frees up some sqes */
  270       int res = iouring_enter (EV_A_ EV_TS_CONST (0.));
  271 
  272       ECB_MEMORY_FENCE_ACQUIRE; /* better safe than sorry */
  273 
  274       if (res >= 0)
  275         continue; /* yes, it worked, try again */
  276 #endif
  277 
  278       /* some problem, possibly EBUSY - do the full poll and let it handle any issues */
  279 
  280       iouring_poll (EV_A_ EV_TS_CONST (0.));
  281       /* iouring_poll should have done ECB_MEMORY_FENCE_ACQUIRE for us */
  282     }
  283 
  284   /*assert (("libev: io_uring queue full after flush", tail + 1 - EV_SQ_VAR (head) <= EV_SQ_VAR (ring_entries)));*/
  285 
  286   return EV_SQES + (tail & EV_SQ_VAR (ring_mask));
  287 }
  288 
  289 inline_size
  290 struct io_uring_sqe *
  291 iouring_sqe_submit (EV_P_ struct io_uring_sqe *sqe)
  292 {
  293   unsigned idx = sqe - EV_SQES;
  294 
  295   EV_SQ_ARRAY [idx] = idx;
  296   ECB_MEMORY_FENCE_RELEASE;
  297   ++EV_SQ_VAR (tail);
  298   /*ECB_MEMORY_FENCE_RELEASE; /* for the time being we assume this is not needed */
  299   ++iouring_to_submit;
  300 }
  301 
  302 /*****************************************************************************/
  303 
  304 /* when the timerfd expires we simply note the fact,
  305  * as the purpose of the timerfd is to wake us up, nothing else.
  306  * the next iteration should re-set it.
  307  */
  308 static void
  309 iouring_tfd_cb (EV_P_ struct ev_io *w, int revents)
  310 {
  311   iouring_tfd_to = EV_TSTAMP_HUGE;
  312 }
  313 
  314 /* called for full and partial cleanup */
  315 ecb_cold
  316 static int
  317 iouring_internal_destroy (EV_P)
  318 {
  319   close (iouring_tfd);
  320   close (iouring_fd);
  321 
  322   if (iouring_sq_ring != MAP_FAILED) munmap (iouring_sq_ring, iouring_sq_ring_size);
  323   if (iouring_cq_ring != MAP_FAILED) munmap (iouring_cq_ring, iouring_cq_ring_size);
  324   if (iouring_sqes    != MAP_FAILED) munmap (iouring_sqes   , iouring_sqes_size   );
  325 
  326   if (ev_is_active (&iouring_tfd_w))
  327     {
  328       ev_ref (EV_A);
  329       ev_io_stop (EV_A_ &iouring_tfd_w);
  330     }
  331 }
  332 
  333 ecb_cold
  334 static int
  335 iouring_internal_init (EV_P)
  336 {
  337   struct io_uring_params params = { 0 };
  338 
  339   iouring_to_submit = 0;
  340 
  341   iouring_tfd     = -1;
  342   iouring_sq_ring = MAP_FAILED;
  343   iouring_cq_ring = MAP_FAILED;
  344   iouring_sqes    = MAP_FAILED;
  345 
  346   if (!have_monotonic) /* cannot really happen, but what if11 */
  347     return -1;
  348 
  349   for (;;)
  350     {
  351       iouring_fd = evsys_io_uring_setup (iouring_entries, &params);
  352 
  353       if (iouring_fd >= 0)
  354         break; /* yippie */
  355 
  356       if (errno != EINVAL)
  357         return -1; /* we failed */
  358 
  359 #if TODO
  360       if ((~params.features) & (IORING_FEAT_NODROP | IORING_FEATURE_SINGLE_MMAP | IORING_FEAT_SUBMIT_STABLE))
  361         return -1; /* we require the above features */
  362 #endif
  363 
  364       /* EINVAL: lots of possible reasons, but maybe
  365        * it is because we hit the unqueryable hardcoded size limit
  366        */
  367 
  368       /* we hit the limit already, give up */
  369       if (iouring_max_entries)
  370         return -1;
  371 
  372       /* first time we hit EINVAL? assume we hit the limit, so go back and retry */
  373       iouring_entries >>= 1;
  374       iouring_max_entries = iouring_entries;
  375     }
  376 
  377   iouring_sq_ring_size = params.sq_off.array + params.sq_entries * sizeof (unsigned);
  378   iouring_cq_ring_size = params.cq_off.cqes  + params.cq_entries * sizeof (struct io_uring_cqe);
  379   iouring_sqes_size    =                       params.sq_entries * sizeof (struct io_uring_sqe);
  380 
  381   iouring_sq_ring = mmap (0, iouring_sq_ring_size, PROT_READ | PROT_WRITE,
  382                           MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQ_RING);
  383   iouring_cq_ring = mmap (0, iouring_cq_ring_size, PROT_READ | PROT_WRITE,
  384                           MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_CQ_RING);
  385   iouring_sqes    = mmap (0, iouring_sqes_size, PROT_READ | PROT_WRITE,
  386                           MAP_SHARED | MAP_POPULATE, iouring_fd, IORING_OFF_SQES);
  387 
  388   if (iouring_sq_ring == MAP_FAILED || iouring_cq_ring == MAP_FAILED || iouring_sqes == MAP_FAILED)
  389     return -1;
  390 
  391   iouring_sq_head         = params.sq_off.head;
  392   iouring_sq_tail         = params.sq_off.tail;
  393   iouring_sq_ring_mask    = params.sq_off.ring_mask;
  394   iouring_sq_ring_entries = params.sq_off.ring_entries;
  395   iouring_sq_flags        = params.sq_off.flags;
  396   iouring_sq_dropped      = params.sq_off.dropped;
  397   iouring_sq_array        = params.sq_off.array;
  398 
  399   iouring_cq_head         = params.cq_off.head;
  400   iouring_cq_tail         = params.cq_off.tail;
  401   iouring_cq_ring_mask    = params.cq_off.ring_mask;
  402   iouring_cq_ring_entries = params.cq_off.ring_entries;
  403   iouring_cq_overflow     = params.cq_off.overflow;
  404   iouring_cq_cqes         = params.cq_off.cqes;
  405 
  406   iouring_tfd = timerfd_create (CLOCK_MONOTONIC, TFD_CLOEXEC);
  407 
  408   if (iouring_tfd < 0)
  409     return iouring_tfd;
  410 
  411   iouring_tfd_to = EV_TSTAMP_HUGE;
  412 
  413   return 0;
  414 }
  415 
  416 ecb_cold
  417 static void
  418 iouring_fork (EV_P)
  419 {
  420   iouring_internal_destroy (EV_A);
  421 
  422   while (iouring_internal_init (EV_A) < 0)
  423     ev_syserr ("(libev) io_uring_setup");
  424 
  425   fd_rearm_all (EV_A);
  426 
  427   ev_io_stop  (EV_A_ &iouring_tfd_w);
  428   ev_io_set   (EV_A_ &iouring_tfd_w, iouring_tfd, EV_READ);
  429   ev_io_start (EV_A_ &iouring_tfd_w);
  430 }
  431 
  432 /*****************************************************************************/
  433 
  434 static void
  435 iouring_modify (EV_P_ int fd, int oev, int nev)
  436 {
  437   if (oev)
  438     {
  439       /* we assume the sqe's are all "properly" initialised */
  440       struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
  441       sqe->opcode    = IORING_OP_POLL_REMOVE;
  442       sqe->fd        = fd;
  443       /* Jens Axboe notified me that user_data is not what is documented, but is
  444        * some kind of unique ID that has to match, otherwise the request cannot
  445        * be removed. Since we don't *really* have that, we pass in the old
  446        * generation counter - if that fails, too bad, it will hopefully be removed
  447        * at close time and then be ignored. */
  448       sqe->addr      = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
  449       sqe->user_data = (uint64_t)-1;
  450       iouring_sqe_submit (EV_A_ sqe);
  451 
  452       /* increment generation counter to avoid handling old events */
  453       ++anfds [fd].egen;
  454     }
  455 
  456   if (nev)
  457     {
  458       struct io_uring_sqe *sqe = iouring_sqe_get (EV_A);
  459       sqe->opcode      = IORING_OP_POLL_ADD;
  460       sqe->fd          = fd;
  461       sqe->addr        = 0;
  462       sqe->user_data   = (uint32_t)fd | ((__u64)(uint32_t)anfds [fd].egen << 32);
  463       sqe->poll_events =
  464         (nev & EV_READ ? POLLIN : 0)
  465         | (nev & EV_WRITE ? POLLOUT : 0);
  466       iouring_sqe_submit (EV_A_ sqe);
  467     }
  468 }
  469 
  470 inline_size
  471 void
  472 iouring_tfd_update (EV_P_ ev_tstamp timeout)
  473 {
  474   ev_tstamp tfd_to = mn_now + timeout;
  475 
  476   /* we assume there will be many iterations per timer change, so
  477    * we only re-set the timerfd when we have to because its expiry
  478    * is too late.
  479    */
  480   if (ecb_expect_false (tfd_to < iouring_tfd_to))
  481     {
  482        struct itimerspec its;
  483 
  484        iouring_tfd_to = tfd_to;
  485        EV_TS_SET (its.it_interval, 0.);
  486        EV_TS_SET (its.it_value, tfd_to);
  487 
  488        if (timerfd_settime (iouring_tfd, TFD_TIMER_ABSTIME, &its, 0) < 0)
  489          assert (("libev: iouring timerfd_settime failed", 0));
  490     }
  491 }
  492 
  493 inline_size
  494 void
  495 iouring_process_cqe (EV_P_ struct io_uring_cqe *cqe)
  496 {
  497   int      fd  = cqe->user_data & 0xffffffffU;
  498   uint32_t gen = cqe->user_data >> 32;
  499   int      res = cqe->res;
  500 
  501   /* user_data -1 is a remove that we are not atm. interested in */
  502   if (cqe->user_data == (uint64_t)-1)
  503     return;
  504 
  505   assert (("libev: io_uring fd must be in-bounds", fd >= 0 && fd < anfdmax));
  506 
  507   /* documentation lies, of course. the result value is NOT like
  508    * normal syscalls, but like linux raw syscalls, i.e. negative
  509    * error numbers. fortunate, as otherwise there would be no way
  510    * to get error codes at all. still, why not document this?
  511    */
  512 
  513   /* ignore event if generation doesn't match */
  514   /* other than skipping removal events, */
  515   /* this should actually be very rare */
  516   if (ecb_expect_false (gen != (uint32_t)anfds [fd].egen))
  517     return;
  518 
  519   if (ecb_expect_false (res < 0))
  520     {
  521       /*TODO: EINVAL handling (was something failed with this fd)*/
  522 
  523       if (res == -EBADF)
  524         {
  525           assert (("libev: event loop rejected bad fd", res != -EBADF));
  526           fd_kill (EV_A_ fd);
  527         }
  528       else
  529         {
  530           errno = -res;
  531           ev_syserr ("(libev) IORING_OP_POLL_ADD");
  532         }
  533 
  534       return;
  535     }
  536 
  537   /* feed events, we do not expect or handle POLLNVAL */
  538   fd_event (
  539     EV_A_
  540     fd,
  541     (res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0)
  542     | (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0)
  543   );
  544 
  545   /* io_uring is oneshot, so we need to re-arm the fd next iteration */
  546   /* this also means we usually have to do at least one syscall per iteration */
  547   anfds [fd].events = 0;
  548   fd_change (EV_A_ fd, EV_ANFD_REIFY);
  549 }
  550 
  551 /* called when the event queue overflows */
  552 ecb_cold
  553 static void
  554 iouring_overflow (EV_P)
  555 {
  556   /* we have two options, resize the queue (by tearing down
  557    * everything and recreating it, or living with it
  558    * and polling.
  559    * we implement this by resizing the queue, and, if that fails,
  560    * we just recreate the state on every failure, which
  561    * kind of is a very inefficient poll.
  562    * one danger is, due to the bios toward lower fds,
  563    * we will only really get events for those, so
  564    * maybe we need a poll() fallback, after all.
  565    */
  566   /*EV_CQ_VAR (overflow) = 0;*/ /* need to do this if we keep the state and poll manually */
  567 
  568   fd_rearm_all (EV_A);
  569 
  570   /* we double the size until we hit the hard-to-probe maximum */
  571   if (!iouring_max_entries)
  572     {
  573       iouring_entries <<= 1;
  574       iouring_fork (EV_A);
  575     }
  576   else
  577     {
  578       /* we hit the kernel limit, we should fall back to something else.
  579        * we can either poll() a few times and hope for the best,
  580        * poll always, or switch to epoll.
  581        * TODO: is this necessary with newer kernels?
  582        */
  583 
  584       iouring_internal_destroy (EV_A);
  585 
  586       /* this should make it so that on return, we don't call any uring functions */
  587       iouring_to_submit = 0;
  588 
  589       for (;;)
  590         {
  591           backend = epoll_init (EV_A_ 0);
  592 
  593           if (backend)
  594             break;
  595 
  596           ev_syserr ("(libev) iouring switch to epoll");
  597         }
  598     }
  599 }
  600 
  601 /* handle any events in the completion queue, return true if there were any */
  602 static int
  603 iouring_handle_cq (EV_P)
  604 {
  605   unsigned head, tail, mask;
  606   
  607   head = EV_CQ_VAR (head);
  608   ECB_MEMORY_FENCE_ACQUIRE;
  609   tail = EV_CQ_VAR (tail);
  610 
  611   if (head == tail)
  612     return 0;
  613 
  614   /* it can only overflow if we have events, yes, yes? */
  615   if (ecb_expect_false (EV_CQ_VAR (overflow)))
  616     {
  617       iouring_overflow (EV_A);
  618       return 1;
  619     }
  620 
  621   mask = EV_CQ_VAR (ring_mask);
  622 
  623   do
  624     iouring_process_cqe (EV_A_ &EV_CQES [head++ & mask]);
  625   while (head != tail);
  626 
  627   EV_CQ_VAR (head) = head;
  628   ECB_MEMORY_FENCE_RELEASE;
  629 
  630   return 1;
  631 }
  632 
  633 static void
  634 iouring_poll (EV_P_ ev_tstamp timeout)
  635 {
  636   /* if we have events, no need for extra syscalls, but we might have to queue events */
  637   /* we also clar the timeout if there are outstanding fdchanges */
  638   /* the latter should only happen if both the sq and cq are full, most likely */
  639   /* because we have a lot of event sources that immediately complete */
  640   /* TODO: fdchacngecnt is always 0 because fd_reify does not have two buffers yet */
  641   if (iouring_handle_cq (EV_A) || fdchangecnt)
  642     timeout = EV_TS_CONST (0.);
  643   else
  644     /* no events, so maybe wait for some */
  645     iouring_tfd_update (EV_A_ timeout);
  646 
  647   /* only enter the kernel if we have something to submit, or we need to wait */
  648   if (timeout || iouring_to_submit)
  649     {
  650       int res = iouring_enter (EV_A_ timeout);
  651 
  652       if (ecb_expect_false (res < 0))
  653         if (errno == EINTR)
  654           /* ignore */;
  655         else if (errno == EBUSY)
  656           /* cq full, cannot submit - should be rare because we flush the cq first, so simply ignore */;
  657         else
  658           ev_syserr ("(libev) iouring setup");
  659       else
  660         iouring_handle_cq (EV_A);
  661     }
  662 }
  663 
  664 inline_size
  665 int
  666 iouring_init (EV_P_ int flags)
  667 {
  668   iouring_entries     = IOURING_INIT_ENTRIES;
  669   iouring_max_entries = 0;
  670 
  671   if (iouring_internal_init (EV_A) < 0)
  672     {
  673       iouring_internal_destroy (EV_A);
  674       return 0;
  675     }
  676 
  677   ev_io_init  (&iouring_tfd_w, iouring_tfd_cb, iouring_tfd, EV_READ);
  678   ev_set_priority (&iouring_tfd_w, EV_MINPRI);
  679   ev_io_start (EV_A_ &iouring_tfd_w);
  680   ev_unref (EV_A); /* watcher should not keep loop alive */
  681 
  682   backend_modify = iouring_modify;
  683   backend_poll   = iouring_poll;
  684 
  685   return EVBACKEND_IOURING;
  686 }
  687 
  688 inline_size
  689 void
  690 iouring_destroy (EV_P)
  691 {
  692   iouring_internal_destroy (EV_A);
  693 }
  694