On Thu, Jun 13, 2024 at 08:35:37AM +0200, Ruediger Pluem wrote: > > > On 6/11/24 6:05 PM, jor...@apache.org wrote: > > Author: jorton > > Date: Tue Jun 11 16:05:22 2024 > > New Revision: 1918258 > > > > URL: http://svn.apache.org/viewvc?rev=1918258&view=rev > > Log: > > * file_io/unix/pipe.c (file_pipe_create): Use pipe2(,O_NONBLOCK) by > > default unless APR_FULL_BLOCK was used; unconditionally set the > > blocking state later. Saves two syscalls per invocation for both > > APR_READ_BLOCK and APR_WRITE_BLOCK, no [intended] functional change. > > I guess it is me being blind , but how do we save two syscalls for the > APR_READ_BLOCK and APR_WRITE_BLOCK case each? With the code before the patch > we started with a blocking pipe and needed to set one end of the pipe to non > blocking. > Now we start with a non blocking pipe and need to set the other end to > blocking. > Where did I get lost and miss the point?
No, you got it correct, I obviously got confused counting syscalls at some point there, sorry! APR_FULL_BLOCK is the only case where using pipe2() saves syscalls, and for every other case it costs three syscalls regardless of whether pipe() or pipe2() is used. So r1918258 is a noop. I'll revert this, thanks for the review. Regards, Joe
