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