Re: [PATCH] kmap tracking

["Randy.Dunlap" <rdunlap@xenotime.net>]

On Mon, 22 May 2006 10:58:54 -0700 Zach Brown wrote:

> 
> > I was scratching my head over this patch trying to think of any bug in
> > recent years which it would have detected.  I failed.
> 
> 2.4 nfs used to require that it be able to kmap entire RPCs for them to
> make forward progress.  Its file->write() required RPC forward progress
> before it would return.  And some callers were holding kmaps across
> file->write() calls.  So with enough concurrent callers doing that the
> system would get stuck.
> 
> We used the patch to see who the callers were when the system got into
> that state.
> 
> One of them was core dumping, believe it or not.  2.6 elf_core_dump()
> still holds a kmap across file->write(), which seems unwise, but I
> haven't gotten to seeing if it's worth worrying about or not.
> 
> So maybe these days the kmap story is less dreadful and it isn't as
> helpful, but that's what we used it for.

I was planning to add kmap_atomic* variants to the patch.
I could see it being useful for those callers, but maybe problems
with them would be more obvious anyway and wouldn't need such
a patch.

---
~Randy