Re: [patch] procfs/procps threading performance speedup, 2.5.62

[Nikita Danilov]

Ingo Molnar writes:
 > 

[...]

 > 
 > to fix this overhead i've introduced a 'lookup cursor' cookie, which is
 > cached in filp->private_data, across readdir() [getdents64()] calls. If
 > the cursor matches then we skip all the overhead of skipping threads. If
 > the cursor is not available then we fall back to the old-style skipping
 > algorithm.

Shouldn't filp->private_data be cleared on lseek? It looks like lookup
cursor is never cleared once set and so readdir will always go forward
independently of ->f_pos updates. Note that glibc implementation of
readdir() (on the top of getdents64()) does call lseek on the
directory. So does seekdir(3).

 > 
 > 2) procps is forced to parse every thread in /proc to build up accurate
 >    'process CPU usage' counters. The parsing and accessing of every

Nikita.