Hello,

On 20/07/18 02:23 PM, Vidar Holen wrote:
> I'm using noticing that `sed -i` does not do any kind of output
> buffering. While behavior is correct, this causes an excessive number of
> small writes:

> I'm seeing this on Debian with the latest sed commit (c52a676) and libc
> 2.27-2. The root cause appears to be `ck_mkstemp` using `fdopen`, which
> unlike `fopen` does not buffer by default.

Thank you for reporting this issue, and providing clear way to reproduce
it. I can confirm seeing the same behavior on Debian 9.4 with glibc 2.24 
(and latest sed).

I think the technical culprit is slightly different:
It is not due to 'ck_mkstemp/fdopen', but because sed explicitly flushes
the output after every line, except if the output is STDOUT.

This happens in execute.c:flush_output():
https://opengrok.housegordon.com/source/xref/sed/sed/execute.c#flush_output

Changing this function enables default buffering with "sed -i"
(using whatever the buffering default is for glibc).

This change does have a side-effect:
It also changes the buffering of other write commands such as 'w', 'W'
and 's///w'.
It might have other side effects I'm haven't spotted.
Based on the ChangeLog-2014 file, I see this function was added in 2003.

I wonder if there are good reasons to explicitly flush all output -
ideas anyone?

If you can give this a quick test and let me know if you also notice
improvements on your system - would be appreciated.


Comments and suggestions welcomed,
  - assaf