Re: [PATCH fifo 0/2] Add support for duplex FIFOs

[Ken Brown <kbrown at cornell dot edu>]

On 3/25/2019 7:06 PM, Ken Brown wrote:
> The second patch in this series enables opening a FIFO with O_RDWR
> access.  The underlying Windows named pipe is created with duplex
> access, and its handle is made the I/O handle of the first client.
> 
> I tested the patch in two ways.
> 
> First,

[...]

> The second test was the following sequence of commands in a bash
> shell:
> 
> $ mkfifo foo
> 
> $ exec 7<>foo
> 
> $ echo blah > foo
> 
> $ read bar <&7
> 
> $ echo $bar
> blah

I just realized that this doesn't test *writing* to the fd of a FIFO opened with 
O_RDWR.  If I change the third command to "echo blah >&7", it does test this, 
and the write fails with ECOMM.  It turns out that the call to NtWriteFile in 
fhandler_fifo::raw_write fails with STATUS_PIPE_LISTENING.

Corinna, I'll try to debug this, but since I know you're about to be AFK for a 
month, I thought I'd check to see if you have any idea why this would happen.

The Windows named pipe that we're writing to was created by a call to 
NtCreateNamedPipeFile with access = GENERIC_READ |GENERIC_WRITE | 
FILE_READ_ATTRIBUTES | FILE_WRITE_ATTRIBUTES | SYNCHRONIZE, and it was put in 
nonblocking mode by a call to NtSetInformationFile with fpi.CompletionMode = 
FILE_PIPE_COMPLETE_OPERATION.

Thanks for any suggestions.

Ken