On 04/17/2011 12:15 PM, Bruno Haible wrote: > Hi Eric, > > There were indeed some mistakes in the test. Now I've debugged it and added a > test for non-blocking sockets as well.
Thanks again for tackling all of this. > > As is, the socket test fails on Cygwin 1.5.x. After putting STDOUT_FILENO in > non-blocking mode, get_nonblocking_flag reports that it is in non-blocking > mode, > but it still behaves like a blocking socket. (You can convince yourself by > setting ENABLE_DEBUGGING to 1 and doing > "make check TESTS=test-nonblocking-socket.sh".) > I don't know how to work around this, so I just put a comment in the test. I'm not sure either, and cygwin 1.5 is starting to get old enough that maybe we need not worry about it much longer. > > Another complication was that on Linux (and similarly on MacOS X and Windows), > a writer can send much more data to a socket than the values of SO_SNDBUF and > SO_RCVBUF indicate. These values are 128 KB or less, yet it is possible for > a write() call of 7 MB of data on a supposedly blocking socket to succeed > immediately, before the reader has requested to read this data. > > Yes another complication is that on Cygwin 1.5.x, writes of more than 64 MB > fail with error ENOBUFS. While POSIX allows this failure, it is not clear to > me whether it would not be better to implement in the module 'safe-write' > or 'full-write' a logic that sends the data in pieces? Indeed. Cygwin 1.7.9 ended up resorting to chunked read/write under the hood, since Windows in general has problem with pipes if you ever try to send more than 64k chunks (it would sometimes, but not always, work). I definitely think that it's better to just document that write() is not always reliable on super-large buffers, and to use safewrite() if the goal is to send large buffers with under-the-hood chunking as necessary. -- Eric Blake [email protected] +1-801-349-2682 Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
