Description of problem:
 Although there is no documentation on splice() I assume that SPLICE_F_NONBLOCK
is supposed to be the same as having a nonblocking flag on the pipe ... this
isn't so, with the SPLICE_F_NONBLOCK flag splice() can return -1 EAGAIN when
poll() says the pipe is ready to accept data.
 strace:

poll([{fd=4, events=POLLOUT, revents=POLLOUT}, {fd=6, events=POLLOUT,
revents=POLLOUT}], 2, -1) = 2
SYS_313(0x7, 0, 0x4, 0, 0x3d060000)     = -1 EAGAIN (Resource temporarily
unavailable)
SYS_313(0x9, 0, 0x6, 0, 0x618e000)      = -1 EAGAIN (Resource temporarily
unavailable)

...I've included a test program.
 Note that even if I'm wrong in my assumption of what SPLICE_F_NONBLOCK does,
the only other thing I could imagine is that it is doing a real nonblock on the
page cache too ... and I've included readahead() calls which should "fix" that
(although not having a working poll() would make that definition much less useful).

Version-Release number of selected component (if applicable):
% uname -a                                        ~/glibc-tst-rpms
Linux code.and.org 2.6.18-1.2849.fc6 #1 SMP Fri Nov 10 12:45:28 EST 2006 i686
athlon i386 GNU/Linux

...note that although the syscall is in glibc FC6, it's broken (fixed in
rawhide) if you have a rawhide/test glibc and want to make sure it's not that
just change the LOCAL_SPLICE define -- I've done this in my testing.

How reproducible:
 Always

Additional info:
 See the attached splice test application. Note that if you change
USE_O_NONBLOCK_R from 0 to 1 and USE_SPLICE_NONBLOCK_R from 1 to 0, the test
program will use fcntl() O_NONBLOCK instead of the SPLICE_F_NONBLOCK flag which
does the exected thing.

 Use it like:

./splicetst /tmp/big_1GB /tmp/o1 /tmp/big_100MB /tmp/o2