ext/io_splice/io_splice_ext.c

   1 #include "ruby.h"
   2 #ifdef HAVE_RUBY_IO_H
   3 #  include "ruby/io.h"
   4 #else
   5 #  include "rubyio.h"
   6 #endif
   7 #include <errno.h>
   8 #include <fcntl.h>
   9 #include <assert.h>
  10 #include <sys/uio.h>
  11 #include <limits.h>
  12 #include <alloca.h>
  13 #include <unistd.h>
  14
  15 static VALUE sym_EAGAIN;
  16 #define WAITALL 0x4000000
  17
  18 /* taken from haproxy */
  19 #define MAX_AT_ONCE (1 << 30)
  20
  21 #ifndef F_LINUX_SPECIFIC_BASE
  22 #  define F_LINUX_SPECIFIC_BASE 1024
  23 #endif
  24
  25 #ifndef F_GETPIPE_SZ
  26 #  define F_SETPIPE_SZ    (F_LINUX_SPECIFIC_BASE + 7)
  27 #  define F_GETPIPE_SZ    (F_LINUX_SPECIFIC_BASE + 8)
  28 #endif
  29
  30 #if ! HAVE_RB_IO_T
  31 #  define rb_io_t OpenFile
  32 #endif
  33
  34 #ifdef GetReadFile
  35 #  define FPTR_TO_FD(fptr) (fileno(GetReadFile(fptr)))
  36 #else
  37 #  if !HAVE_RB_IO_T || (RUBY_VERSION_MAJOR == 1 && RUBY_VERSION_MINOR == 8)
  38 #    define FPTR_TO_FD(fptr) fileno(fptr->f)
  39 #  else
  40 #    define FPTR_TO_FD(fptr) fptr->fd
  41 #  endif
  42 #endif
  43
  44 #ifndef SSIZET2NUM
  45 #  define SSIZET2NUM(x) LONG2NUM(x)
  46 #endif
  47 #ifndef NUM2SSIZET
  48 #  define NUM2SSIZET(x) NUM2LONG(x)
  49 #endif
  50 #ifndef SIZET2NUM
  51 #  define SIZET2NUM(x) ULONG2NUM(x)
  52 #endif
  53 #ifndef NUM2SIZET
  54 #  define NUM2SIZET(x) NUM2ULONG(x)
  55 #endif
  56
  57 static int my_fileno(VALUE io)
  58 {
  59         rb_io_t *fptr;
  60
  61         for (;;) {
  62                 switch (TYPE(io)) {
  63                 case T_FIXNUM: return FIX2INT(io);
  64                 case T_FILE: {
  65                         GetOpenFile(io, fptr);
  66                         return FPTR_TO_FD(fptr);
  67                 }
  68                 default:
  69                         io = rb_convert_type(io, T_FILE, "IO", "to_io");
  70                         /* retry */
  71                 }
  72         }
  73 }
  74
  75 static int check_fileno(VALUE io)
  76 {
  77         int saved_errno = errno;
  78         int fd = my_fileno(io);
  79         errno = saved_errno;
  80         return fd;
  81 }
  82
  83 #if defined(HAVE_RB_THREAD_CALL_WITHOUT_GVL) && defined(HAVE_RUBY_THREAD_H)
  84 /* Ruby 2.0+ */
  85 #  include <ruby/thread.h>
  86 #  define WITHOUT_GVL(fn,a,ubf,b) \
  87         rb_thread_call_without_gvl((fn),(a),(ubf),(b))
  88 #elif defined(HAVE_RB_THREAD_BLOCKING_REGION)
  89 typedef VALUE (*my_blocking_fn_t)(void*);
  90 #  define WITHOUT_GVL(fn,a,ubf,b) \
  91         rb_thread_blocking_region((my_blocking_fn_t)(fn),(a),(ubf),(b))
  92
  93 #else /* Ruby 1.8 */
  94 /* partial emulation of the 1.9 rb_thread_blocking_region under 1.8 */
  95 #  include <rubysig.h>
  96 #  define RUBY_UBF_IO ((rb_unblock_function_t *)-1)
  97 typedef void rb_unblock_function_t(void *);
  98 typedef void * rb_blocking_function_t(void *);
  99 static void * WITHOUT_GVL(rb_blocking_function_t *func, void *data1,
 100                         rb_unblock_function_t *ubf, void *data2)
 101 {
 102         void *rv;
 103
 104         assert(RUBY_UBF_IO == ubf && "RUBY_UBF_IO required for emulation");
 105
 106         TRAP_BEG;
 107         rv = func(data1);
 108         TRAP_END;
 109
 110         return rv;
 111 }
 112 #endif /* ! HAVE_RB_THREAD_BLOCKING_REGION */
 113
 114 #ifndef RSTRING_PTR
 115 #  define RSTRING_PTR(s) (RSTRING(s)->ptr)
 116 #endif
 117 #ifndef RSTRING_LEN
 118 #  define RSTRING_LEN(s) (RSTRING(s)->len)
 119 #endif
 120 #ifndef RARRAY_LEN
 121 #  define RARRAY_LEN(s) (RARRAY(s)->len)
 122 #endif
 123
 124 #define io_run(fn,data) WITHOUT_GVL((fn),(data),RUBY_UBF_IO,0)
 125
 126 struct splice_args {
 127         int fd_in;
 128         int fd_out;
 129         off_t *off_in;
 130         off_t *off_out;
 131         size_t len;
 132         unsigned flags;
 133 };
 134
 135 static void * nogvl_splice(void *ptr)
 136 {
 137         struct splice_args *a = ptr;
 138
 139         if (a->len > MAX_AT_ONCE)
 140                 a->len = MAX_AT_ONCE;
 141
 142         return (void *)splice(a->fd_in, a->off_in, a->fd_out, a->off_out,
 143                              a->len, a->flags);
 144 }
 145
 146 static ssize_t do_splice(int argc, VALUE *argv, unsigned dflags)
 147 {
 148         off_t i = 0, o = 0;
 149         VALUE io_in, off_in, io_out, off_out, len, flags;
 150         struct splice_args a;
 151         ssize_t bytes;
 152         ssize_t total = 0;
 153         unsigned waitall;
 154
 155         rb_scan_args(argc, argv, "51",
 156                      &io_in, &off_in, &io_out, &off_out, &len, &flags);
 157
 158         a.off_in = NIL_P(off_in) ? NULL : (i = NUM2OFFT(off_in), &i);
 159         a.off_out = NIL_P(off_out) ? NULL : (o = NUM2OFFT(off_out), &o);
 160         a.len = NUM2SIZET(len);
 161         a.flags = NIL_P(flags) ? dflags : NUM2UINT(flags) | dflags;
 162         waitall = a.flags & WAITALL;
 163         if (waitall)
 164                 a.flags ^= WAITALL;
 165
 166         for (;;) {
 167                 a.fd_in = check_fileno(io_in);
 168                 a.fd_out = check_fileno(io_out);
 169                 bytes = (ssize_t)io_run(nogvl_splice, &a);
 170                 if (bytes < 0) {
 171                         if (errno == EINTR)
 172                                 continue;
 173                         if (waitall && errno == EAGAIN) {
 174                                 rb_io_wait_readable(check_fileno(io_in));
 175                                 errno = EAGAIN;
 176                                 rb_io_wait_writable(check_fileno(io_out));
 177                                 continue;
 178                         }
 179                         if (total > 0)
 180                                 return total;
 181                         return bytes;
 182                 } else if (bytes == 0) {
 183                         break;
 184                 } else if (waitall) {
 185                         total += bytes;
 186                         if ((a.len -= bytes) == 0)
 187                                 return total;
 188                         i += bytes;
 189                         o += bytes;
 190                 } else {
 191                         return bytes;
 192                 }
 193         }
 194
 195         return total;
 196 }
 197
 198 /*
 199  * call-seq:
 200  *    IO.splice(io_in, off_in, io_out, off_out, len) => integer
 201  *    IO.splice(io_in, off_in, io_out, off_out, len, flags) => integer
 202  *
 203  * Splice +len+ bytes from/to a pipe.  Either +io_in+ or +io_out+
 204  * MUST be a pipe.  +io_in+ and +io_out+ may BOTH be pipes as of
 205  * Linux 2.6.31 or later.
 206  *
 207  * +off_in+ and +off_out+ if non-nil may be used to
 208  * specify an offset for the non-pipe file descriptor.
 209  *
 210  * +flags+ defaults to zero if unspecified.
 211  * +flags+ may be a bitmask of the following flags:
 212  *
 213  * * IO::Splice::F_MOVE
 214  * * IO::Splice::F_NONBLOCK
 215  * * IO::Splice::F_MORE
 216  *
 217  * Returns the number of bytes spliced.
 218  * Raises EOFError when +io_in+ has reached end of file.
 219  * Raises Errno::EAGAIN if the IO::Splice::F_NONBLOCK flag is set
 220  * and the pipe has no data to read from or space to write to.  May
 221  * also raise Errno::EAGAIN if the non-pipe descriptor has no data
 222  * to read from or space to write to.
 223  *
 224  * As splice never exposes buffers to userspace, it will not take
 225  * into account userspace buffering done by Ruby or stdio.  It is
 226  * also not subject to encoding/decoding filters under Ruby 1.9.
 227  *
 228  * Consider using IO.trysplice if +io_out+ is a pipe or if you are using
 229  * non-blocking I/O on both descriptors as it avoids the cost of raising
 230  * common Errno::EAGAIN exceptions.
 231  *
 232  * See manpage for full documentation:
 233  * http://kernel.org/doc/man-pages/online/pages/man2/splice.2.html
 234  */
 235 static VALUE my_splice(int argc, VALUE *argv, VALUE self)
 236 {
 237         ssize_t n = do_splice(argc, argv, 0);
 238
 239         if (n == 0)
 240                 rb_eof_error();
 241         if (n < 0)
 242                 rb_sys_fail("splice");
 243         return SSIZET2NUM(n);
 244 }
 245
 246 /*
 247  * call-seq:
 248  *    IO.trysplice(io_in, off_in, io_out, off_out, len) => integer
 249  *    IO.trysplice(io_in, off_in, io_out, off_out, len, flags) => integer
 250  *
 251  * Exactly like IO.splice, except +:EAGAIN+ is returned when either
 252  * the read or write end would block instead of raising Errno::EAGAIN.
 253  *
 254  * IO::Splice::F_NONBLOCK is always passed for the pipe descriptor,
 255  * but this can still block if the non-pipe descriptor is blocking.
 256  *
 257  * See IO.splice documentation for more details.
 258  *
 259  * This method is recommended whenever +io_out+ is a pipe.
 260  */
 261 static VALUE trysplice(int argc, VALUE *argv, VALUE self)
 262 {
 263         ssize_t n = do_splice(argc, argv, SPLICE_F_NONBLOCK);
 264
 265         if (n == 0)
 266                 return Qnil;
 267         if (n < 0) {
 268                 if (errno == EAGAIN)
 269                         return sym_EAGAIN;
 270                 rb_sys_fail("splice");
 271         }
 272         return SSIZET2NUM(n);
 273 }
 274
 275 struct tee_args {
 276         int fd_in;
 277         int fd_out;
 278         size_t len;
 279         unsigned flags;
 280 };
 281
 282 /* runs without GVL */
 283 static void * nogvl_tee(void *ptr)
 284 {
 285         struct tee_args *a = ptr;
 286
 287         if (a->len > MAX_AT_ONCE)
 288                 a->len = MAX_AT_ONCE;
 289
 290         return (void *)tee(a->fd_in, a->fd_out, a->len, a->flags);
 291 }
 292
 293 static ssize_t do_tee(int argc, VALUE *argv, unsigned dflags)
 294 {
 295         VALUE io_in, io_out, len, flags;
 296         struct tee_args a;
 297         ssize_t bytes;
 298         ssize_t total = 0;
 299         unsigned waitall;
 300
 301         rb_scan_args(argc, argv, "31", &io_in, &io_out, &len, &flags);
 302         a.len = (size_t)NUM2SIZET(len);
 303         a.flags = NIL_P(flags) ? dflags : NUM2UINT(flags) | dflags;
 304         waitall = a.flags & WAITALL;
 305         if (waitall)
 306                 a.flags ^= WAITALL;
 307
 308         for (;;) {
 309                 a.fd_in = check_fileno(io_in);
 310                 a.fd_out = check_fileno(io_out);
 311                 bytes = (ssize_t)io_run(nogvl_tee, &a);
 312                 if (bytes < 0) {
 313                         if (errno == EINTR)
 314                                 continue;
 315                         if (waitall && errno == EAGAIN) {
 316                                 rb_io_wait_readable(check_fileno(io_in));
 317                                 errno = EAGAIN;
 318                                 rb_io_wait_writable(check_fileno(io_out));
 319                                 continue;
 320                         }
 321                         if (total > 0)
 322                                 return total;
 323                         return bytes;
 324                 } else if (bytes == 0) {
 325                         break;
 326                 } else if (waitall) {
 327                         total += bytes;
 328                         if ((a.len -= bytes) == 0)
 329                                 return total;
 330                 } else {
 331                         return bytes;
 332                 }
 333         }
 334
 335         return total;
 336 }
 337
 338 /*
 339  * call-seq:
 340  *   IO.tee(io_in, io_out, len) => integer
 341  *   IO.tee(io_in, io_out, len, flags) => integer
 342  *
 343  * Copies up to +len+ bytes of data from +io_in+ to +io_out+.  +io_in+
 344  * and +io_out+ must both refer to pipe descriptors.  +io_in+ and +io_out+
 345  * may not be endpoints of the same pipe.
 346  *
 347  * +flags+ may be zero (the default) or a combination of:
 348  * * IO::Splice::F_NONBLOCK
 349  *
 350  * Other IO::Splice flags are currently unimplemented or have no effect.
 351  *
 352  * Returns the number of bytes duplicated if successful.
 353  * Raises EOFError when +io_in+ is closed and emptied.
 354  * Raises Errno::EAGAIN when +io_in+ is empty and/or +io_out+ is full
 355  * and +flags+ contains IO::Splice::F_NONBLOCK
 356  *
 357  * Consider using IO.trytee if you are using IO::Splice::F_NONBLOCK
 358  * as it avoids the cost of raising common Errno::EAGAIN exceptions.
 359  *
 360  * See manpage for full documentation:
 361  * http://kernel.org/doc/man-pages/online/pages/man2/tee.2.html
 362  */
 363 static VALUE my_tee(int argc, VALUE *argv, VALUE self)
 364 {
 365         ssize_t n = do_tee(argc, argv, 0);
 366
 367         if (n == 0)
 368                 rb_eof_error();
 369         if (n < 0)
 370                 rb_sys_fail("tee");
 371
 372         return SSIZET2NUM(n);
 373 }
 374
 375 /*
 376  * call-seq:
 377  *    IO.trytee(io_in, io_out, len) => integer
 378  *    IO.trytee(io_in, io_out, len, flags) => integer
 379  *
 380  * Exactly like IO.tee, except +:EAGAIN+ is returned when either
 381  * the read or write end would block instead of raising Errno::EAGAIN.
 382  *
 383  * IO::Splice::F_NONBLOCK is always passed for the pipe descriptor,
 384  * but this can still block if the non-pipe descriptor is blocking.
 385  *
 386  * See IO.tee documentation for more details.
 387  */
 388 static VALUE trytee(int argc, VALUE *argv, VALUE self)
 389 {
 390         ssize_t n = do_tee(argc, argv, SPLICE_F_NONBLOCK);
 391
 392         if (n == 0)
 393                 return Qnil;
 394         if (n < 0) {
 395                 if (errno == EAGAIN)
 396                         return sym_EAGAIN;
 397                 rb_sys_fail("tee");
 398         }
 399
 400         return SSIZET2NUM(n);
 401 }
 402
 403 struct vmsplice_args {
 404         int fd;
 405         unsigned flags;
 406         struct iovec *iov;
 407         unsigned long nr_segs;
 408 };
 409
 410 static void * nogvl_vmsplice(void *ptr)
 411 {
 412         struct vmsplice_args *a = ptr;
 413
 414         return (void *)vmsplice(a->fd, a->iov, a->nr_segs, a->flags);
 415 }
 416
 417 /* this can't be a function since we use alloca() */
 418 #define ARY2IOVEC(iov,iovcnt,expect,ary) \
 419 do { \
 420         struct iovec *tmp; \
 421         unsigned long i; \
 422         iovcnt = (unsigned long)RARRAY_LEN(ary); \
 423         if (iovcnt > IOV_MAX) \
 424                 rb_raise(rb_eArgError, "array is larger than IOV_MAX"); \
 425         iov = tmp = alloca(sizeof(struct iovec) * iovcnt); \
 426         expect = 0; \
 427         for (i = 0; i < iovcnt; tmp++, i++) { \
 428                 VALUE cur = rb_ary_entry(ary, (long)i); \
 429                 Check_Type(cur, T_STRING); \
 430                 tmp->iov_base = RSTRING_PTR(cur); \
 431                 tmp->iov_len = RSTRING_LEN(cur); \
 432                 expect += tmp->iov_len; \
 433         } \
 434 } while (0)
 435
 436 static void advance_vmsplice_args(struct vmsplice_args *a, long n)
 437 {
 438         struct iovec *new_iov = a->iov;
 439         unsigned long i;
 440
 441         /* skip over iovecs we've already written completely */
 442         for (i = 0; i < a->nr_segs; i++, new_iov++) {
 443                 if (n == 0)
 444                         break;
 445                 /*
 446                  * partially written iov,
 447                  * modify and retry with current iovec in
 448                  * front
 449                  */
 450                 if (new_iov->iov_len > (size_t)n) {
 451                         VALUE base = (VALUE)new_iov->iov_base;
 452
 453                         new_iov->iov_len -= n;
 454                         new_iov->iov_base = (void *)(base + n);
 455                         break;
 456                 }
 457
 458                 n -= new_iov->iov_len;
 459         }
 460
 461         /* setup to retry without the already-written iovecs */
 462         a->nr_segs -= i;
 463         a->iov = new_iov;
 464 }
 465
 466 /*
 467  * call-seq:
 468  *   IO.vmsplice(io, string_array) => integer
 469  *   IO.vmsplice(io, string_array, flags) => integer
 470  *   IO.vmsplice(io, string) => integer
 471  *   IO.vmsplice(io, string, flags) => integer
 472  *
 473  * Transfers an array of strings into the pipe descriptor given by io.
 474  * +io+ must be the writable end of a pipe.
 475  *
 476  * This may allow the kernel to avoid data copies in some cases.
 477  * but is (probably) of limited usefulness in Ruby.  If you have
 478  * use cases or ideas for making this more useful for Ruby users,
 479  * please tell us at ruby-io-splice@yhbt.net!
 480  *
 481  * Also consider the "sendfile" RubyGem or IO.copy_stream in Ruby 1.9
 482  * if you want to do zero-copy file transfers to pipes or sockets.  As
 483  * of Linux 2.6.33, sendfile(2) can copy to any output descriptor,
 484  * not just sockets.
 485  *
 486  * See manpage for full documentation:
 487  * http://kernel.org/doc/man-pages/online/pages/man2/vmsplice.2.html
 488  */
 489 static VALUE my_vmsplice(int argc, VALUE * argv, VALUE self)
 490 {
 491         ssize_t rv = 0;
 492         ssize_t left;
 493         struct vmsplice_args a;
 494         struct iovec iov;
 495         ssize_t n;
 496
 497         VALUE io, data, flags;
 498
 499         rb_scan_args(argc, argv, "21", &io, &data, &flags);
 500
 501         switch (TYPE(data)) {
 502         case T_STRING:
 503                 iov.iov_base = RSTRING_PTR(data);
 504                 iov.iov_len = (size_t)(left = (ssize_t)RSTRING_LEN(data));
 505                 a.iov = &iov;
 506                 a.nr_segs = 1;
 507                 break;
 508         case T_ARRAY:
 509                 ARY2IOVEC(a.iov, a.nr_segs, left, data);
 510                 break;
 511         default:
 512                 rb_raise(rb_eTypeError, "wrong argument type %s "
 513                          "(expected a String or Array of strings)",
 514                          rb_obj_classname(data));
 515         }
 516
 517         a.flags = NIL_P(flags) ? 0 : NUM2UINT(flags);
 518
 519         for (;;) {
 520                 a.fd = check_fileno(io);
 521                 n = (ssize_t)io_run(nogvl_vmsplice, &a);
 522
 523                 if (n < 0) {
 524                         if (errno == EAGAIN) {
 525                                 if (a.flags & SPLICE_F_NONBLOCK)
 526                                         rb_sys_fail("vmsplice");
 527                                 if (rb_io_wait_writable(check_fileno(io)))
 528                                         continue;
 529                                 /* fall through on error */
 530                         }
 531                         /*
 532                          * unlikely to hit this case, return the
 533                          * already written bytes, we'll let the next
 534                          * write (or close) fail instead
 535                          */
 536                         if (rv > 0)
 537                                 break;
 538                         if (errno == EINTR)
 539                                 continue;
 540                         rb_sys_fail("vmsplice");
 541                 }
 542
 543                 rv += n;
 544                 left -= n;
 545                 if (left == 0)
 546                         break;
 547                 advance_vmsplice_args(&a, n);
 548         }
 549
 550         return SSIZET2NUM(rv);
 551 }
 552
 553 /*
 554  * call-seq:
 555  *   reader, writer = IO.pipe
 556  *   reader.pipe_size => integer
 557  *
 558  * Returns the pipe capacity of the underlying pipe in bytes.  The
 559  * default capacity is 65536 bytes since Linux 2.6.11, and 4096 bytes
 560  * in previous kernels.
 561  *
 562  * Since the pipe is a circular buffer in the same kernel, the size
 563  * of the reader is exactly the same as the size of the writer.
 564  *
 565  * This method is only exposed on Linux 2.6.35 or later.
 566  */
 567 static VALUE pipe_size(VALUE self)
 568 {
 569         int size = fcntl(my_fileno(self),  F_GETPIPE_SZ);
 570
 571         if (size < 0)
 572                 rb_sys_fail("fcntl(F_GETPIPE_SZ)");
 573
 574         return INT2NUM(size);
 575 }
 576
 577 /*
 578  * call-seq:
 579  *   reader, writer = IO.pipe
 580  *   reader.pipe_size = integer
 581  *
 582  * Sets and returns the pipe capacity of the underlying pipe in bytes.
 583  *
 584  * This MUST be a power-of-two, or Errno::EINVAL will be raised.
 585  * Linux will silently increase this to be equal to the page size
 586  * (4096 bytes on most architectures) if the specified value is
 587  * less than the size of a page.
 588  *
 589  * For users without CAP_SYS_RESOURCE, this raises Errno::EPERM when
 590  * attempting to specify a value greater than the value in
 591  * /proc/sys/fs/pipe-max-size.
 592  *
 593  * Since the pipe is a circular buffer in the same kernel, the size
 594  * of the reader is exactly the same as the size of the writer.
 595  *
 596  * Raises Errno::EBUSY if the assigned value is less than
 597  * the currently filled portion of the pipe.
 598  *
 599  * This method is only exposed on Linux 2.6.35 or later.
 600  */
 601 static VALUE set_pipe_size(VALUE self, VALUE size)
 602 {
 603         int fd = my_fileno(self);
 604         int bytes = NUM2INT(size);
 605         int rv = fcntl(fd, F_SETPIPE_SZ, bytes);
 606
 607         if (rv < 0) {
 608                 if (errno == ENOMEM) {
 609                         rb_gc();
 610                         rv = fcntl(fd, F_SETPIPE_SZ, bytes);
 611                 }
 612                 if (rv < 0)
 613                         rb_sys_fail("fcntl(F_SETPIPE_SZ)");
 614         }
 615
 616         return size;
 617 }
 618
 619 static int can_mod_pipe_size(void)
 620 {
 621         /*
 622          * pipe2 appeared in Linux 2.6.27, F_*PIPE_SZ appeared in 2.6.35,
 623          * thus not having pipe2 automatically disqualifies us from having
 624          * F_*PIPE_SZ support
 625          */
 626 #ifdef HAVE_PIPE2
 627         int fds[2];
 628         int rc = pipe2(fds, O_CLOEXEC);
 629
 630         if (rc == 0) {
 631                 rc = fcntl(fds[0], F_GETPIPE_SZ);
 632                 rc = rc < 0 ? 0 : 1;
 633
 634                 (void)close(fds[0]);
 635                 (void)close(fds[1]);
 636         } else {
 637                 /*
 638                  * weird error, but don't raise during init, this could be
 639                  * ENOSYS, even..
 640                  */
 641                 rc = 0;
 642         }
 643         errno = 0;
 644         return rc;
 645 #else  /* ! HAVE_PIPE2 */
 646         return 0;
 647 #endif /* ! HAVE_PIPE2 */
 648 }
 649
 650 #define NODOC_CONST(klass,name,value) \
 651   rb_define_const((klass),(name),(value))
 652
 653 void Init_io_splice_ext(void)
 654 {
 655         VALUE mSplice = rb_define_module_under(rb_cIO, "Splice");
 656
 657         rb_define_singleton_method(rb_cIO, "splice", my_splice, -1);
 658         rb_define_singleton_method(rb_cIO, "trysplice", trysplice, -1);
 659         rb_define_singleton_method(rb_cIO, "tee", my_tee, -1);
 660         rb_define_singleton_method(rb_cIO, "trytee", trytee, -1);
 661         rb_define_singleton_method(rb_cIO, "vmsplice", my_vmsplice, -1);
 662
 663         /*
 664          * Attempt to move pages instead of copying.  This is only a hint
 665          * and support for it was removed in Linux 2.6.21.  It will be
 666          * re-added for FUSE filesystems only in Linux 2.6.35.
 667          */
 668         rb_define_const(mSplice, "F_MOVE", UINT2NUM(SPLICE_F_MOVE));
 669         assert(WAITALL != SPLICE_F_MOVE && "WAITALL == F_MOVE");
 670
 671         /*
 672          * Do not block on pipe I/O.  This flag only affects the pipe(s)
 673          * being spliced from/to and has no effect on the non-pipe
 674          * descriptor (which requires non-blocking operation to be set
 675          * explicitly).
 676          *
 677          * The non-blocking flag (O_NONBLOCK) on the pipe descriptors
 678          * themselves are ignored by this family of functions, and
 679          * using this flag is the only way to get non-blocking operation
 680          * out of them.
 681          *
 682          * It is highly recommended this flag be set (or IO.trysplice used)
 683          * whenever splicing from a socket into a pipe unless there is
 684          * another (native) thread or process doing a blocking read on that
 685          * pipe.  Otherwise it is possible to block a single-threaded process
 686          * if the socket buffers are larger than the pipe buffers.
 687          */
 688         rb_define_const(mSplice, "F_NONBLOCK", UINT2NUM(SPLICE_F_NONBLOCK));
 689         assert(WAITALL != SPLICE_F_NONBLOCK && "WAITALL == F_NONBLOCK");
 690
 691         /*
 692          * Indicate that there may be more data coming into the outbound
 693          * descriptor.  This can allow the kernel to avoid sending partial
 694          * frames from sockets.  Currently only used with splice.
 695          */
 696         rb_define_const(mSplice, "F_MORE", UINT2NUM(SPLICE_F_MORE));
 697         assert(WAITALL != SPLICE_F_MORE && "WAITALL == F_MORE");
 698
 699         /*
 700          * Only usable by vmsplice.  This flag probably not useful in the
 701          * context of Ruby applications which cannot control alignment.
 702          */
 703         rb_define_const(mSplice, "F_GIFT", UINT2NUM(SPLICE_F_GIFT));
 704         assert(WAITALL != SPLICE_F_GIFT && "WAITALL == F_GIFT");
 705
 706         /*
 707          * Retry until the requested transfer is complete, this will
 708          * cause IO.splice/IO.tee to never return less than the requested
 709          * transfer size unless an error occored.
 710          *
 711          * IO.vmsplice always defaults to this behavior.
 712          */
 713         NODOC_CONST(mSplice, "WAITALL", UINT2NUM(WAITALL));
 714
 715         /*
 716          * The maximum size of an atomic write to a pipe
 717          * POSIX requires this to be at least 512 bytes.
 718          * Under Linux, this is 4096 bytes.
 719          */
 720         rb_define_const(mSplice, "PIPE_BUF", UINT2NUM(PIPE_BUF));
 721
 722         /*
 723          * The maximum size we're allowed to splice at once.  Larger
 724          * sizes will be broken up and retried if the WAITALL flag or
 725          * IO::Splice.copy_stream is used.
 726          */
 727         rb_define_const(mSplice, "MAX_AT_ONCE", SIZET2NUM(MAX_AT_ONCE));
 728
 729         if (can_mod_pipe_size()) {
 730                 rb_define_method(rb_cIO, "pipe_size", pipe_size, 0);
 731                 rb_define_method(rb_cIO, "pipe_size=", set_pipe_size, 1);
 732
 733                 /*
 734                  * fcntl() command constant used to return the size of a pipe.
 735                  * This constant is only defined when running Linux 2.6.35
 736                  * or later.  For convenience, use IO#pipe_size instead.
 737                  */
 738                 rb_define_const(mSplice, "F_GETPIPE_SZ",
 739                                 UINT2NUM(F_GETPIPE_SZ));
 740
 741                 /*
 742                  * fcntl() command constant used to set the size of a pipe.
 743                  * This constant is only defined when running Linux 2.6.35
 744                  * or later.  For convenience, use IO#pipe_size= instead.
 745                  */
 746                 rb_define_const(mSplice, "F_SETPIPE_SZ",
 747                                 UINT2NUM(F_SETPIPE_SZ));
 748         }
 749
 750         sym_EAGAIN = ID2SYM(rb_intern("EAGAIN"));
 751 }