GC starves write's in 2.6.20

[Joakim Tjernlund]

On Fri, 2007-06-01 at 13:11 +0100, David Woodhouse wrote:
> On Thu, 2007-05-31 at 16:41 +0200, Joakim Tjernlund wrote:
> > We have noticed that deleting some 50 files, some a few MB big, in our
> > JFFS2 FS and then directly start downloading a large rpm file sometimes
> > stalls a write(2) call 10-15 secs, possibly more. The GC process is
> > running and it appears to be blocking the write(2). However the FS is
> > just 50-60% full so I don't think stalling writes for that long time is
> > justified. These long delays causes a timeout in the tftp demon and
> > the transmission is aborted. Any ideas? 
> 
> Hm. What's the GC process actually doing? If you kill it, does the file
> system still run GC in the context of the processes which are writing?
> 

Sorry, that was bit premature. It was pdflush that was running most of
the time. GC was running also but only for a short while. Then I
remebered Radoslaw's patch and tried it and the problem was gone.

 Jocke