Messages in this thread | | | Date | Thu, 2 Mar 2006 15:08:59 -0800 | From | Andrew Morton <> | Subject | Re: [Fwd: Re: [PATCH] reiserfs: use balance_dirty_pages_ratelimited_nr in reiserfs_file_write] |
| |
Hans Reiser <reiser@namesys.com> wrote: > > I suspect that when someone did the search and replace when creating > balance_dirty_pages_ratelimited_nr they failed to read the code and > realize this code path was already effectively ratelimited. The result > is they made it excessively infrequent (every 1MB if ratelimit is 8) in > its calling balance_dirty_pages.
?? There's been no change to balance_dirty_pages_ratelimited(). I merely widened the interface a bit: introduced the new balance_dirty_pages_ratelimited_nr() and did
static inline void balance_dirty_pages_ratelimited(struct address_space *mapping) { balance_dirty_pages_ratelimited_nr(mapping, 1); }
That being said, if reiserfs has `number of pages' in hand then yes, it makes sense to migrate over to balance_dirty_pages_ratelimited_nr().
> If anyone has ever seen this as an actual problem on a real system, I > would be curious to hear of it.
No, I wouldn't expect it to make much difference.
All that gunk is there just to avoid calling the expensive get_writeback_state() once per set_page_dirty().
Inaccuracy here will introduce the possibility that we'll transiently dirty more memory than dirty_ratio permits, but it'll only be a little bit (eight times the amount of memory which is dirtied per balance_dirty_pages_ratelimited() call).
That's a small amount of memory. But if you have 1000 filesystems mounted and they all do the same thing at the same time, things could get a bit sticky. Your patch will (greatly) reduce the possibility of even that scenario.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |