Disk-full errors are not correctly reported

[nh2]

Throughout the code, PoissonRecon does not check the return values of write()/fwrite() correctly.

While in some parts of the code there's certainly awareness that writes can fail, e.g. when the disk is full, such as here with a bool being returned:

PoissonRecon/Src/Streams.h

Lines 70 to 77 in d4e8715

	bool write( const std::vector< std::vector< C > > &c )
	{
	size_t sz = c.size();
	if( !write( sz ) ) return false;
	bool ret = true;
	for( size_t i=0 ; i<sz && ret ; i++ ) ret &= write( c[i] );
	return ret;
	}

much of the code simply calls write(), ignoring the returned bool and continuing:

PoissonRecon/Src/PoissonRecon.client.inl

Lines 1456 to 1457 in d4e8715

	stream.write( _index );
	stream.write( _range );

---

Because of this, you can get the client failing with:

[RECEIVE 1]  288   B: 0.0 (s)
[PROCESS 1]         : 1652.8 (s), 132842.0 (MB) / 132842.0 (MB)
[SEND    1]   96   B: 0.0 (s)
[CACHE   1]         : 490.9 (s) , 214887 (MB)
[ERROR] Src/FEMTree.inl (Line 372)
        FEMTree
        Failed to read transform

and the server saying:

[RECEIVE 2]   96   B: 2066.2 (s)
[PROCESS 2]         : 0.0 (s), 5.2 (MB) / 5.2 (MB)
[SEND    2]   32   B: 0.0 (s)
[ERROR] Src/Socket.h (Line 59)
        socket_receive
        Failed to read from socket

when in fact e.g. the --tempDir is full and the error should really be no space left on device.

[nh2]

I recommend the following approach:

• If there are no cases where you actually want to handle write failing, throw an exception.
• In the long run, additionally move off fwrite() (C stdlib) and use write() (POSIX) instead, because the FILE based functions do not report what the error was (errno), see e.g. here.
  C++ IO streams are also notoriously bad for not reporting errno, so POSIX functions are the simplest solution to get good error messages like no space left on device (with perror() or strerror()).

[nh2]

Same issue for read() functions, which also do not do failure checks (errors can happen e.g. if the disk is bad, the network disk is disconnected, and so on).

[nh2]

Some improvements for this in my branch: master...nh2:PoissonRecon:nh2fixes-v15.03

[nh2]

• move off fwrite() (C stdlib)

Another motivation for that:

https://stackoverflow.com/questions/2329842/checking-for-success-of-fwrite-in-c-perror#comment5694186_2329882

fwrite cannot handle EAGAIN or EINTR at all. They will set the error indicator for the stream, and there's no way to resume writing without possibly getting lost or duplicate data (due to not knowing the state of the buffer). If you want to use interrupting signals or want to setup a file as nonblocking, stdio is generally not usable.

[nh2]

Some improvements for this in my branch: master...nh2:PoissonRecon:nh2fixes-v15.03

I'm furhter improving this situation in that branch with commit nh2@e155f45