Messages in this thread | | | Date | Tue, 28 Apr 2009 16:42:48 -0700 | From | Andrew Morton <> | Subject | Re: [PATCH 5/5] proc: export more page flags in /proc/kpageflags |
| |
On Tue, 28 Apr 2009 18:31:09 -0500 Matt Mackall <mpm@selenic.com> wrote:
> On Tue, 2009-04-28 at 16:02 -0700, Andrew Morton wrote: > > On Tue, 28 Apr 2009 17:46:34 -0500 > > Matt Mackall <mpm@selenic.com> wrote: > > > > > > > +/* a helper function _not_ intended for more general uses */ > > > > > +static inline int page_cap_writeback_dirty(struct page *page) > > > > > +{ > > > > > + struct address_space *mapping; > > > > > + > > > > > + if (!PageSlab(page)) > > > > > + mapping = page_mapping(page); > > > > > + else > > > > > + mapping = NULL; > > > > > + > > > > > + return mapping && mapping_cap_writeback_dirty(mapping); > > > > > +} > > > > > > > > If the page isn't locked then page->mapping can be concurrently removed > > > > and freed. This actually happened to me in real-life testing several > > > > years ago. > > > > > > We certainly don't want to be taking locks per page to build the flags > > > data here. As we don't have any pretense of being atomic, it's ok if we > > > can find a way to do the test that's inaccurate when a race occurs, so > > > long as it doesn't dereference null. > > > > > > But if there's not an obvious way to do that, we should probably just > > > drop this flag bit for this iteration. > > > > trylock_page() could be used here, perhaps. > > > > Then again, why _not_ just do lock_page()? After all, few pages are > > ever locked. There will be latency if the caller stumbles across a > > page which is under read I/O, but so be it? > > As I mentioned just a bit ago, it's really not an unreasonable use case > to want to do this on every page in the system back to back. So per page > overhead matters. And the odds of stalling on a locked page when > visiting 1M pages while under load are probably not negligible.
The chances of stalling on a locked page are pretty good, and the duration of the stall might be long indeed. Perhaps a trylock is a decent compromise - it depends on the value of this metric, and I've forgotten what we're talking about ;)
umm, seems that this flag is needed to enable PG_error, PG_dirty, PG_uptodate and PG_writeback reporting. So simply removing this code would put a huge hole in the patchset, no?
> Our lock primitives are pretty low overhead in the fast path, but every > cycle counts. The new tests and branches this code already adds are a > bit worrisome, but on balance probably worth it.
That should be easy to quantify (hint).
| |