Re: [Lockmeter 2.5] BKL with 51ms hold time, prove me wrong

[Andrew Morton]

Hanna Linder <[e-mail]> wrote:
>
> 
> Sure enough. I ported lockmeter to 2.5.67-mm1 and ran the same rmap-test
> and lo and behold all the ext3 issues went away. However, the one remaining
> long hold time moved to the top (unmap_vmas):

Ah, but the unmap_vmas lock was not kernel_flag, was it?  It'll be
page_table_lock.

That's OK I think.  The only time this is likely to bite anyone is if you
have a threaded application in which one thread it doing a massive munmap()
while another one is handling a pagefault, running mmap(), etc.  And given
that _establishing_ that large mapping in the first place takes tons of CPU,
the relative loss from the long hold time is small.

Famous last words.

> 
> Here is the port of lockmeter to the 2.5.67-mm1 kernel. If you would consider
> putting it in your tree that would be great and I would work on porting the
> rest of the architectures. 
> 
> http://prdownloads.sourceforge.net/lse/lockmeter1.5-2.5.67-mm1.patch?download

OK, that's pretty unintrusive.