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
);
83 #if defined(HAVE_RB_THREAD_CALL_WITHOUT_GVL) && defined(HAVE_RUBY_THREAD_H)
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))
94 /* partial emulation of the 1.9 rb_thread_blocking_region under 1.8 */
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
)
104 assert(RUBY_UBF_IO
== ubf
&& "RUBY_UBF_IO required for emulation");
112 #endif /* ! HAVE_RB_THREAD_BLOCKING_REGION */
115 # define RSTRING_PTR(s) (RSTRING(s)->ptr)
118 # define RSTRING_LEN(s) (RSTRING(s)->len)
121 # define RARRAY_LEN(s) (RARRAY(s)->len)
124 #define io_run(fn,data) WITHOUT_GVL((fn),(data),RUBY_UBF_IO,0)
135 static void * nogvl_splice(void *ptr
)
137 struct splice_args
*a
= ptr
;
139 if (a
->len
> MAX_AT_ONCE
)
140 a
->len
= MAX_AT_ONCE
;
142 return (void *)splice(a
->fd_in
, a
->off_in
, a
->fd_out
, a
->off_out
,
146 static ssize_t
do_splice(int argc
, VALUE
*argv
, unsigned dflags
)
149 VALUE io_in
, off_in
, io_out
, off_out
, len
, flags
;
150 struct splice_args a
;
155 rb_scan_args(argc
, argv
, "51",
156 &io_in
, &off_in
, &io_out
, &off_out
, &len
, &flags
);
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
;
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
);
173 if (waitall
&& errno
== EAGAIN
) {
174 rb_io_wait_readable(check_fileno(io_in
));
176 rb_io_wait_writable(check_fileno(io_out
));
182 } else if (bytes
== 0) {
184 } else if (waitall
) {
186 if ((a
.len
-= bytes
) == 0)
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
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.
207 * +off_in+ and +off_out+ if non-nil may be used to
208 * specify an offset for the non-pipe file descriptor.
210 * +flags+ defaults to zero if unspecified.
211 * +flags+ may be a bitmask of the following flags:
213 * * IO::Splice::F_MOVE
214 * * IO::Splice::F_NONBLOCK
215 * * IO::Splice::F_MORE
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.
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.
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.
232 * See manpage for full documentation:
233 * http://kernel.org/doc/man-pages/online/pages/man2/splice.2.html
235 static VALUE
my_splice(int argc
, VALUE
*argv
, VALUE self
)
237 ssize_t n
= do_splice(argc
, argv
, 0);
242 rb_sys_fail("splice");
243 return SSIZET2NUM(n
);
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
251 * Exactly like IO.splice, except +:EAGAIN+ is returned when either
252 * the read or write end would block instead of raising Errno::EAGAIN.
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.
257 * See IO.splice documentation for more details.
259 * This method is recommended whenever +io_out+ is a pipe.
261 static VALUE
trysplice(int argc
, VALUE
*argv
, VALUE self
)
263 ssize_t n
= do_splice(argc
, argv
, SPLICE_F_NONBLOCK
);
270 rb_sys_fail("splice");
272 return SSIZET2NUM(n
);
282 /* runs without GVL */
283 static void * nogvl_tee(void *ptr
)
285 struct tee_args
*a
= ptr
;
287 if (a
->len
> MAX_AT_ONCE
)
288 a
->len
= MAX_AT_ONCE
;
290 return (void *)tee(a
->fd_in
, a
->fd_out
, a
->len
, a
->flags
);
293 static ssize_t
do_tee(int argc
, VALUE
*argv
, unsigned dflags
)
295 VALUE io_in
, io_out
, len
, flags
;
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
;
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
);
315 if (waitall
&& errno
== EAGAIN
) {
316 rb_io_wait_readable(check_fileno(io_in
));
318 rb_io_wait_writable(check_fileno(io_out
));
324 } else if (bytes
== 0) {
326 } else if (waitall
) {
328 if ((a
.len
-= bytes
) == 0)
340 * IO.tee(io_in, io_out, len) => integer
341 * IO.tee(io_in, io_out, len, flags) => integer
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.
347 * +flags+ may be zero (the default) or a combination of:
348 * * IO::Splice::F_NONBLOCK
350 * Other IO::Splice flags are currently unimplemented or have no effect.
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
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.
360 * See manpage for full documentation:
361 * http://kernel.org/doc/man-pages/online/pages/man2/tee.2.html
363 static VALUE
my_tee(int argc
, VALUE
*argv
, VALUE self
)
365 ssize_t n
= do_tee(argc
, argv
, 0);
372 return SSIZET2NUM(n
);
377 * IO.trytee(io_in, io_out, len) => integer
378 * IO.trytee(io_in, io_out, len, flags) => integer
380 * Exactly like IO.tee, except +:EAGAIN+ is returned when either
381 * the read or write end would block instead of raising Errno::EAGAIN.
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.
386 * See IO.tee documentation for more details.
388 static VALUE
trytee(int argc
, VALUE
*argv
, VALUE self
)
390 ssize_t n
= do_tee(argc
, argv
, SPLICE_F_NONBLOCK
);
400 return SSIZET2NUM(n
);
403 struct vmsplice_args
{
406 unsigned long nr_segs
;
410 static void * nogvl_vmsplice(void *ptr
)
412 struct vmsplice_args
*a
= ptr
;
414 return (void *)vmsplice(a
->fd
, a
->iov
, a
->nr_segs
, a
->flags
);
417 /* this can't be a function since we use alloca() */
418 #define ARY2IOVEC(iov,iovcnt,expect,ary) \
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); \
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; \
436 static void advance_vmsplice_args(struct vmsplice_args
*a
, long n
)
438 struct iovec
*new_iov
= a
->iov
;
441 /* skip over iovecs we've already written completely */
442 for (i
= 0; i
< a
->nr_segs
; i
++, new_iov
++) {
446 * partially written iov,
447 * modify and retry with current iovec in
450 if (new_iov
->iov_len
> (size_t)n
) {
451 VALUE base
= (VALUE
)new_iov
->iov_base
;
453 new_iov
->iov_len
-= n
;
454 new_iov
->iov_base
= (void *)(base
+ n
);
458 n
-= new_iov
->iov_len
;
461 /* setup to retry without the already-written iovecs */
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
473 * Transfers an array of strings into the pipe descriptor given by io.
474 * +io+ must be the writable end of a pipe.
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@librelist.org!
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,
486 * See manpage for full documentation:
487 * http://kernel.org/doc/man-pages/online/pages/man2/vmsplice.2.html
489 static VALUE
my_vmsplice(int argc
, VALUE
* argv
, VALUE self
)
493 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
));
517 a
.flags
= NIL_P(flags
) ? 0 : NUM2UINT(flags
);
520 a
.fd
= check_fileno(io
);
521 n
= (ssize_t
)io_run(nogvl_vmsplice
, &a
);
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
)))
529 /* fall through on error */
532 * unlikely to hit this case, return the
533 * already written bytes, we'll let the next
534 * write (or close) fail instead
540 rb_sys_fail("vmsplice");
547 advance_vmsplice_args(&a
, n
);
550 return SSIZET2NUM(rv
);
555 * reader, writer = IO.pipe
556 * reader.pipe_size => integer
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.
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.
565 * This method is only exposed on Linux 2.6.35 or later.
567 static VALUE
pipe_size(VALUE self
)
569 int size
= fcntl(my_fileno(self
), F_GETPIPE_SZ
);
572 rb_sys_fail("fcntl(F_GETPIPE_SZ)");
574 return INT2NUM(size
);
579 * reader, writer = IO.pipe
580 * reader.pipe_size = integer
582 * Sets and returns the pipe capacity of the underlying pipe in bytes.
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.
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.
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.
596 * Raises Errno::EBUSY if the assigned value is less than
597 * the currently filled portion of the pipe.
599 * This method is only exposed on Linux 2.6.35 or later.
601 static VALUE
set_pipe_size(VALUE self
, VALUE size
)
603 int fd
= my_fileno(self
);
604 int bytes
= NUM2INT(size
);
605 int rv
= fcntl(fd
, F_SETPIPE_SZ
, bytes
);
608 if (errno
== ENOMEM
) {
610 rv
= fcntl(fd
, F_SETPIPE_SZ
, bytes
);
613 rb_sys_fail("fcntl(F_SETPIPE_SZ)");
619 static int can_mod_pipe_size(void)
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
628 int rc
= pipe2(fds
, O_CLOEXEC
);
631 rc
= fcntl(fds
[0], F_GETPIPE_SZ
);
638 * weird error, but don't raise during init, this could be
645 #else /* ! HAVE_PIPE2 */
647 #endif /* ! HAVE_PIPE2 */
650 #define NODOC_CONST(klass,name,value) \
651 rb_define_const((klass),(name),(value))
653 void Init_io_splice_ext(void)
655 VALUE mSplice
= rb_define_module_under(rb_cIO
, "Splice");
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);
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.
668 rb_define_const(mSplice
, "F_MOVE", UINT2NUM(SPLICE_F_MOVE
));
669 assert(WAITALL
!= SPLICE_F_MOVE
&& "WAITALL == F_MOVE");
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
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
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.
688 rb_define_const(mSplice
, "F_NONBLOCK", UINT2NUM(SPLICE_F_NONBLOCK
));
689 assert(WAITALL
!= SPLICE_F_NONBLOCK
&& "WAITALL == F_NONBLOCK");
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.
696 rb_define_const(mSplice
, "F_MORE", UINT2NUM(SPLICE_F_MORE
));
697 assert(WAITALL
!= SPLICE_F_MORE
&& "WAITALL == F_MORE");
700 * Only usable by vmsplice. This flag probably not useful in the
701 * context of Ruby applications which cannot control alignment.
703 rb_define_const(mSplice
, "F_GIFT", UINT2NUM(SPLICE_F_GIFT
));
704 assert(WAITALL
!= SPLICE_F_GIFT
&& "WAITALL == F_GIFT");
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.
711 * IO.vmsplice always defaults to this behavior.
713 NODOC_CONST(mSplice
, "WAITALL", UINT2NUM(WAITALL
));
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.
720 rb_define_const(mSplice
, "PIPE_BUF", UINT2NUM(PIPE_BUF
));
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.
727 rb_define_const(mSplice
, "MAX_AT_ONCE", SIZET2NUM(MAX_AT_ONCE
));
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);
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.
738 rb_define_const(mSplice
, "F_GETPIPE_SZ",
739 UINT2NUM(F_GETPIPE_SZ
));
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.
746 rb_define_const(mSplice
, "F_SETPIPE_SZ",
747 UINT2NUM(F_SETPIPE_SZ
));
750 sym_EAGAIN
= ID2SYM(rb_intern("EAGAIN"));