15 static VALUE sym_EAGAIN
;
16 #define WAITALL 0x4000000
18 /* taken from haproxy */
19 #define MAX_AT_ONCE (1 << 30)
21 #ifndef F_LINUX_SPECIFIC_BASE
22 # define F_LINUX_SPECIFIC_BASE 1024
26 # define F_SETPIPE_SZ (F_LINUX_SPECIFIC_BASE + 7)
27 # define F_GETPIPE_SZ (F_LINUX_SPECIFIC_BASE + 8)
31 # define rb_io_t OpenFile
35 # define FPTR_TO_FD(fptr) (fileno(GetReadFile(fptr)))
37 # if !HAVE_RB_IO_T || (RUBY_VERSION_MAJOR == 1 && RUBY_VERSION_MINOR == 8)
38 # define FPTR_TO_FD(fptr) fileno(fptr->f)
40 # define FPTR_TO_FD(fptr) fptr->fd
45 # define SSIZET2NUM(x) LONG2NUM(x)
48 # define NUM2SSIZET(x) NUM2LONG(x)
51 # define SIZET2NUM(x) ULONG2NUM(x)
54 # define NUM2SIZET(x) NUM2ULONG(x)
57 static int my_fileno(VALUE io
)
63 case T_FIXNUM
: return FIX2INT(io
);
65 GetOpenFile(io
, fptr
);
66 return FPTR_TO_FD(fptr
);
69 io
= rb_convert_type(io
, T_FILE
, "IO", "to_io");
75 static int check_fileno(VALUE io
)
77 int saved_errno
= errno
;
78 int fd
= my_fileno(io
);
82 #ifndef HAVE_RB_THREAD_BLOCKING_REGION
83 /* partial emulation of the 1.9 rb_thread_blocking_region under 1.8 */
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 *);
89 rb_thread_blocking_region(
90 rb_blocking_function_t
*fn
, void *data1
,
91 rb_unblock_function_t
*ubf
, void *data2
)
95 assert(RUBY_UBF_IO
== ubf
&& "RUBY_UBF_IO required for emulation");
103 #endif /* ! HAVE_RB_THREAD_BLOCKING_REGION */
106 # define RSTRING_PTR(s) (RSTRING(s)->ptr)
109 # define RSTRING_LEN(s) (RSTRING(s)->len)
112 # define RARRAY_PTR(s) (RARRAY(s)->ptr)
115 # define RARRAY_LEN(s) (RARRAY(s)->len)
118 static VALUE
io_run(rb_blocking_function_t
*fn
, void *data
)
120 return rb_thread_blocking_region(fn
, data
, RUBY_UBF_IO
, 0);
132 static VALUE
nogvl_splice(void *ptr
)
134 struct splice_args
*a
= ptr
;
136 if (a
->len
> MAX_AT_ONCE
)
137 a
->len
= MAX_AT_ONCE
;
139 return (VALUE
)splice(a
->fd_in
, a
->off_in
, a
->fd_out
, a
->off_out
,
143 static ssize_t
do_splice(int argc
, VALUE
*argv
, unsigned dflags
)
146 VALUE io_in
, off_in
, io_out
, off_out
, len
, flags
;
147 struct splice_args a
;
152 rb_scan_args(argc
, argv
, "51",
153 &io_in
, &off_in
, &io_out
, &off_out
, &len
, &flags
);
155 a
.off_in
= NIL_P(off_in
) ? NULL
: (i
= NUM2OFFT(off_in
), &i
);
156 a
.off_out
= NIL_P(off_out
) ? NULL
: (o
= NUM2OFFT(off_out
), &o
);
157 a
.len
= NUM2SIZET(len
);
158 a
.flags
= NIL_P(flags
) ? dflags
: NUM2UINT(flags
) | dflags
;
159 waitall
= a
.flags
& WAITALL
;
164 a
.fd_in
= check_fileno(io_in
);
165 a
.fd_out
= check_fileno(io_out
);
166 bytes
= (ssize_t
)io_run(nogvl_splice
, &a
);
170 if (waitall
&& errno
== EAGAIN
) {
171 rb_io_wait_readable(check_fileno(io_in
));
173 rb_io_wait_writable(check_fileno(io_out
));
179 } else if (bytes
== 0) {
181 } else if (waitall
) {
183 if ((a
.len
-= bytes
) == 0)
197 * IO.splice(io_in, off_in, io_out, off_out, len) => integer
198 * IO.splice(io_in, off_in, io_out, off_out, len, flags) => integer
200 * Splice +len+ bytes from/to a pipe. Either +io_in+ or +io_out+
201 * MUST be a pipe. +io_in+ and +io_out+ may BOTH be pipes as of
202 * Linux 2.6.31 or later.
204 * +off_in+ and +off_out+ if non-nil may be used to
205 * specify an offset for the non-pipe file descriptor.
207 * +flags+ defaults to zero if unspecified.
208 * +flags+ may be a bitmask of the following flags:
210 * * IO::Splice::F_MOVE
211 * * IO::Splice::F_NONBLOCK
212 * * IO::Splice::F_MORE
213 * * IO::Splice::WAITALL
215 * Returns the number of bytes spliced.
216 * Raises EOFError when +io_in+ has reached end of file.
217 * Raises Errno::EAGAIN if the IO::Splice::F_NONBLOCK flag is set
218 * and the pipe has no data to read from or space to write to. May
219 * also raise Errno::EAGAIN if the non-pipe descriptor has no data
220 * to read from or space to write to.
222 * As splice never exposes buffers to userspace, it will not take
223 * into account userspace buffering done by Ruby or stdio. It is
224 * also not subject to encoding/decoding filters under Ruby 1.9.
226 * Consider using IO.trysplice if +io_out+ is a pipe or if you are using
227 * non-blocking I/O on both descriptors as it avoids the cost of raising
228 * common Errno::EAGAIN exceptions.
230 * See manpage for full documentation:
231 * http://kernel.org/doc/man-pages/online/pages/man2/splice.2.html
233 static VALUE
my_splice(int argc
, VALUE
*argv
, VALUE self
)
235 ssize_t n
= do_splice(argc
, argv
, 0);
240 rb_sys_fail("splice");
241 return SSIZET2NUM(n
);
246 * IO.trysplice(io_in, off_in, io_out, off_out, len) => integer
247 * IO.trysplice(io_in, off_in, io_out, off_out, len, flags) => integer
249 * Exactly like IO.splice, except +:EAGAIN+ is returned when either
250 * the read or write end would block instead of raising Errno::EAGAIN.
252 * IO::Splice::F_NONBLOCK is always passed for the pipe descriptor,
253 * but this can still block if the non-pipe descriptor is blocking.
255 * See IO.splice documentation for more details.
257 * This method is recommended whenever +io_out+ is a pipe.
259 static VALUE
trysplice(int argc
, VALUE
*argv
, VALUE self
)
261 ssize_t n
= do_splice(argc
, argv
, SPLICE_F_NONBLOCK
);
268 rb_sys_fail("splice");
270 return SSIZET2NUM(n
);
280 /* runs without GVL */
281 static VALUE
nogvl_tee(void *ptr
)
283 struct tee_args
*a
= ptr
;
285 if (a
->len
> MAX_AT_ONCE
)
286 a
->len
= MAX_AT_ONCE
;
288 return (VALUE
)tee(a
->fd_in
, a
->fd_out
, a
->len
, a
->flags
);
291 static ssize_t
do_tee(int argc
, VALUE
*argv
, unsigned dflags
)
293 VALUE io_in
, io_out
, len
, flags
;
299 rb_scan_args(argc
, argv
, "31", &io_in
, &io_out
, &len
, &flags
);
300 a
.len
= (size_t)NUM2SIZET(len
);
301 a
.flags
= NIL_P(flags
) ? dflags
: NUM2UINT(flags
) | dflags
;
302 waitall
= a
.flags
& WAITALL
;
307 a
.fd_in
= check_fileno(io_in
);
308 a
.fd_out
= check_fileno(io_out
);
309 bytes
= (ssize_t
)io_run(nogvl_tee
, &a
);
313 if (waitall
&& errno
== EAGAIN
) {
314 rb_io_wait_readable(check_fileno(io_in
));
316 rb_io_wait_writable(check_fileno(io_out
));
322 } else if (bytes
== 0) {
324 } else if (waitall
) {
326 if ((a
.len
-= bytes
) == 0)
338 * IO.tee(io_in, io_out, len) => integer
339 * IO.tee(io_in, io_out, len, flags) => integer
341 * Copies up to +len+ bytes of data from +io_in+ to +io_out+. +io_in+
342 * and +io_out+ must both refer to pipe descriptors. +io_in+ and +io_out+
343 * may not be endpoints of the same pipe.
345 * +flags+ may be zero (the default) or a combination of:
346 * * IO::Splice::F_NONBLOCK
347 * * IO::Splice::WAITALL
349 * Other IO::Splice flags are currently unimplemented or have no effect.
351 * Returns the number of bytes duplicated if successful.
352 * Raises EOFError when +io_in+ is closed and emptied.
353 * Raises Errno::EAGAIN when +io_in+ is empty and/or +io_out+ is full
354 * and +flags+ contains IO::Splice::F_NONBLOCK
356 * Consider using IO.trytee if you are using IO::Splice::F_NONBLOCK
357 * as it avoids the cost of raising common Errno::EAGAIN exceptions.
359 * See manpage for full documentation:
360 * http://kernel.org/doc/man-pages/online/pages/man2/tee.2.html
362 static VALUE
my_tee(int argc
, VALUE
*argv
, VALUE self
)
364 ssize_t n
= do_tee(argc
, argv
, 0);
371 return SSIZET2NUM(n
);
376 * IO.trytee(io_in, io_out, len) => integer
377 * IO.trytee(io_in, io_out, len, flags) => integer
379 * Exactly like IO.tee, except +:EAGAIN+ is returned when either
380 * the read or write end would block instead of raising Errno::EAGAIN.
382 * IO::Splice::F_NONBLOCK is always passed for the pipe descriptor,
383 * but this can still block if the non-pipe descriptor is blocking.
385 * See IO.tee documentation for more details.
387 static VALUE
trytee(int argc
, VALUE
*argv
, VALUE self
)
389 ssize_t n
= do_tee(argc
, argv
, SPLICE_F_NONBLOCK
);
399 return SSIZET2NUM(n
);
402 struct vmsplice_args
{
405 unsigned long nr_segs
;
409 static VALUE
nogvl_vmsplice(void *ptr
)
411 struct vmsplice_args
*a
= ptr
;
413 return (VALUE
)vmsplice(a
->fd
, a
->iov
, a
->nr_segs
, a
->flags
);
416 /* this can't be a function since we use alloca() */
417 #define ARY2IOVEC(iov,iovcnt,expect,ary) \
422 cur = RARRAY_PTR(ary); \
423 n = RARRAY_LEN(ary); \
425 rb_raise(rb_eArgError, "array is larger than IOV_MAX"); \
426 iov = tmp = alloca(sizeof(struct iovec) * n); \
429 for (; --n >= 0; tmp++, cur++) { \
430 Check_Type(*cur, T_STRING); \
431 tmp->iov_base = RSTRING_PTR(*cur); \
432 tmp->iov_len = RSTRING_LEN(*cur); \
433 expect += tmp->iov_len; \
437 static void advance_vmsplice_args(struct vmsplice_args
*a
, long n
)
439 struct iovec
*new_iov
= a
->iov
;
442 /* skip over iovecs we've already written completely */
443 for (i
= 0; i
< a
->nr_segs
; i
++, new_iov
++) {
447 * partially written iov,
448 * modify and retry with current iovec in
451 if (new_iov
->iov_len
> (size_t)n
) {
452 VALUE base
= (VALUE
)new_iov
->iov_base
;
454 new_iov
->iov_len
-= n
;
455 new_iov
->iov_base
= (void *)(base
+ n
);
459 n
-= new_iov
->iov_len
;
462 /* setup to retry without the already-written iovecs */
469 * IO.vmsplice(io, string_array) => integer
470 * IO.vmsplice(io, string_array, flags) => integer
471 * IO.vmsplice(io, string) => integer
472 * IO.vmsplice(io, string, flags) => integer
474 * Transfers an array of strings into the pipe descriptor given by io.
475 * +io+ must be the writable end of a pipe.
477 * This may allow the kernel to avoid data copies in some cases.
478 * but is (probably) of limited usefulness in Ruby. If you have
479 * use cases or ideas for making this more useful for Ruby users,
480 * please tell us at ruby.io.splice@librelist.org!
482 * Also consider the "sendfile" RubyGem or IO.copy_stream in Ruby 1.9
483 * if you want to do zero-copy file transfers to pipes or sockets. As
484 * of Linux 2.6.33, sendfile(2) can copy to any output descriptor,
487 * See manpage for full documentation:
488 * http://kernel.org/doc/man-pages/online/pages/man2/vmsplice.2.html
490 static VALUE
my_vmsplice(int argc
, VALUE
* argv
, VALUE self
)
494 struct vmsplice_args a
;
497 VALUE io
, data
, flags
;
499 rb_scan_args(argc
, argv
, "21", &io
, &data
, &flags
);
501 switch (TYPE(data
)) {
503 iov
.iov_base
= RSTRING_PTR(data
);
504 iov
.iov_len
= (size_t)(left
= (ssize_t
)RSTRING_LEN(data
));
509 ARY2IOVEC(a
.iov
, a
.nr_segs
, left
, data
);
512 rb_raise(rb_eTypeError
, "wrong argument type %s "
513 "(expected a String or Array of strings)",
514 rb_obj_classname(data
));
516 a
.fd
= my_fileno(io
);
517 a
.flags
= NIL_P(flags
) ? 0 : NUM2UINT(flags
);
520 ssize_t n
= (ssize_t
)io_run(nogvl_vmsplice
, &a
);
523 if (errno
== EAGAIN
) {
524 if (a
.flags
& SPLICE_F_NONBLOCK
) {
525 rb_sys_fail("vmsplice");
527 a
.fd
= check_fileno(io
);
528 if (rb_io_wait_writable(a
.fd
))
531 /* fall through on error */
534 * unlikely to hit this case, return the
535 * already written bytes, we'll let the next
536 * write (or close) fail instead
540 if (errno
== EINTR
) {
541 a
.fd
= check_fileno(io
);
544 rb_sys_fail("vmsplice");
551 advance_vmsplice_args(&a
, n
);
554 return SSIZET2NUM(rv
);
559 * reader, writer = IO.pipe
560 * reader.pipe_size => integer
562 * Returns the pipe capacity of the underlying pipe in bytes. The
563 * default capacity is 65536 bytes since Linux 2.6.11, and 4096 bytes
564 * in previous kernels.
566 * Since the pipe is a circular buffer in the same kernel, the size
567 * of the reader is exactly the same as the size of the writer.
569 * This method is only exposed on Linux 2.6.35 or later.
571 static VALUE
pipe_size(VALUE self
)
573 int size
= fcntl(my_fileno(self
), F_GETPIPE_SZ
);
576 rb_sys_fail("fcntl(F_GETPIPE_SZ)");
578 return INT2NUM(size
);
583 * reader, writer = IO.pipe
584 * reader.pipe_size = integer
586 * Sets and returns the pipe capacity of the underlying pipe in bytes.
588 * This MUST be a power-of-two, or Errno::EINVAL will be raised.
589 * Linux will silently increase this to be equal to the page size
590 * (4096 bytes on most architectures) if the specified value is
591 * less than the size of a page.
593 * For users without CAP_SYS_RESOURCE, this raises Errno::EPERM when
594 * attempting to specify a value greater than the value in
595 * /proc/sys/fs/pipe-max-size.
597 * Since the pipe is a circular buffer in the same kernel, the size
598 * of the reader is exactly the same as the size of the writer.
600 * Raises Errno::EBUSY if the assigned value is less than
601 * the currently filled portion of the pipe.
603 * This method is only exposed on Linux 2.6.35 or later.
605 static VALUE
set_pipe_size(VALUE self
, VALUE size
)
607 int fd
= my_fileno(self
);
608 int bytes
= NUM2INT(size
);
609 int rv
= fcntl(fd
, F_SETPIPE_SZ
, bytes
);
612 if (errno
== ENOMEM
) {
614 rv
= fcntl(fd
, F_SETPIPE_SZ
, bytes
);
617 rb_sys_fail("fcntl(F_SETPIPE_SZ)");
623 static int can_mod_pipe_size(void)
629 rc
= fcntl(fds
[0], F_GETPIPE_SZ
);
635 /* weird error, but don't raise during init */
642 void Init_io_splice_ext(void)
644 VALUE mSplice
= rb_define_module_under(rb_cIO
, "Splice");
646 rb_define_singleton_method(rb_cIO
, "splice", my_splice
, -1);
647 rb_define_singleton_method(rb_cIO
, "trysplice", trysplice
, -1);
648 rb_define_singleton_method(rb_cIO
, "tee", my_tee
, -1);
649 rb_define_singleton_method(rb_cIO
, "trytee", trytee
, -1);
650 rb_define_singleton_method(rb_cIO
, "vmsplice", my_vmsplice
, -1);
653 * Attempt to move pages instead of copying. This is only a hint
654 * and support for it was removed in Linux 2.6.21. It will be
655 * re-added for FUSE filesystems only in Linux 2.6.35.
657 rb_define_const(mSplice
, "F_MOVE", UINT2NUM(SPLICE_F_MOVE
));
658 assert(WAITALL
!= SPLICE_F_MOVE
&& "WAITALL == F_MOVE");
661 * Do not block on pipe I/O. This flag only affects the pipe(s)
662 * being spliced from/to and has no effect on the non-pipe
663 * descriptor (which requires non-blocking operation to be set
666 * The non-blocking flag (O_NONBLOCK) on the pipe descriptors
667 * themselves are ignored by this family of functions, and
668 * using this flag is the only way to get non-blocking operation
671 * It is highly recommended this flag be set (or IO.trysplice used)
672 * whenever splicing from a socket into a pipe unless there is
673 * another (native) thread or process doing a blocking read on that
674 * pipe. Otherwise it is possible to block a single-threaded process
675 * if the socket buffers are larger than the pipe buffers.
677 rb_define_const(mSplice
, "F_NONBLOCK", UINT2NUM(SPLICE_F_NONBLOCK
));
678 assert(WAITALL
!= SPLICE_F_NONBLOCK
&& "WAITALL == F_NONBLOCK");
681 * Indicate that there may be more data coming into the outbound
682 * descriptor. This can allow the kernel to avoid sending partial
683 * frames from sockets. Currently only used with splice.
685 rb_define_const(mSplice
, "F_MORE", UINT2NUM(SPLICE_F_MORE
));
686 assert(WAITALL
!= SPLICE_F_MORE
&& "WAITALL == F_MORE");
689 * Only usable by vmsplice. This flag probably not useful in the
690 * context of Ruby applications which cannot control alignment.
692 rb_define_const(mSplice
, "F_GIFT", UINT2NUM(SPLICE_F_GIFT
));
693 assert(WAITALL
!= SPLICE_F_GIFT
&& "WAITALL == F_GIFT");
696 * Retry until the requested transfer is complete, this will
697 * cause IO.splice/IO.tee to never return less than the requested
698 * transfer size unless an error occored.
700 * IO.vmsplice always defaults to this behavior.
702 rb_define_const(mSplice
, "WAITALL", UINT2NUM(WAITALL
));
705 * The maximum size of an atomic write to a pipe
706 * POSIX requires this to be at least 512 bytes.
707 * Under Linux, this is 4096 bytes.
709 rb_define_const(mSplice
, "PIPE_BUF", UINT2NUM(PIPE_BUF
));
712 * The maximum size we're allowed to splice at once. Larger
713 * sizes will be broken up and retried if the WAITALL flag or
714 * IO::Splice.copy_stream is used.
716 rb_define_const(mSplice
, "MAX_AT_ONCE", SIZET2NUM(MAX_AT_ONCE
));
718 if (can_mod_pipe_size()) {
719 rb_define_method(rb_cIO
, "pipe_size", pipe_size
, 0);
720 rb_define_method(rb_cIO
, "pipe_size=", set_pipe_size
, 1);
723 * fcntl() command constant used to return the size of a pipe.
724 * This constant is only defined when running Linux 2.6.35
725 * or later. For convenience, use IO#pipe_size instead.
727 rb_define_const(mSplice
, "F_GETPIPE_SZ",
728 UINT2NUM(F_GETPIPE_SZ
));
731 * fcntl() command constant used to set the size of a pipe.
732 * This constant is only defined when running Linux 2.6.35
733 * or later. For convenience, use IO#pipe_size= instead.
735 rb_define_const(mSplice
, "F_SETPIPE_SZ",
736 UINT2NUM(F_SETPIPE_SZ
));
739 sym_EAGAIN
= ID2SYM(rb_intern("EAGAIN"));