Re: write() on pipe blocking due to in-page fragmentation?

[Chris Friesen <chris.friesen@genband.com>]

On 09/23/2011 01:42 PM, Ricardo Nabinger Sanchez wrote:
> Hello,
>
> The simple program attached allocates a pipe, perform a number of
> writes in it in order to fill the pipe, and then reads that data to
> empty the pipe.  The argument is used to determine how much data to
> write per write iteration.
>
> Values that are power of 2 up to PIPE_BUF work without any issues.
> Other values may cause the write() call to block.

> Intuitively, it seems that pages in the pipe are getting fragmented,
> and eventually it will reach the limit of 16 pages and, if the data is
> not consumed, will cause writers to block --- even though the data
> would fit nicely otherwise.

I suggest reading "man 7 pipe" carefully, looking at the pipe capacity 
and pipe_buf sections.

I suspect that what you're seeing is that due to the atomicity 
requirements the kernel will not spread a single write over multiple 
pages, so that when writing 3 bytes at a time each page in the queue has 
a byte of free space.

Thus, you succeed in writing up to byte 65520 (out of 65536) but 
anything after that blocks.  Note that 65536-65520=16.

>
> Is this understanding correct?  If so, is it something that should be
> fixed in the Linux kernel?
>
> Or should the application ensure that data written to the pipe will be
> done carefully as to not block a writer?

 From the man page:

"Applications should not rely on a particular capacity: an application 
should be designed so that a reading process consumes data as soon as it 
is available, so that a writing process does not remain blocked."

Chris


-- 
Chris Friesen
Software Developer
GENBAND
chris.friesen@genband.com
www.genband.com