Re: [oom]: [1/4] add __GFP_WIRED to pinned allocations

[William Lee Irwin III <wli@holomorphy.com>]

William Lee Irwin III <wli@holomorphy.com> wrote:
>> +#define __GFP_WIRED	0x8000	/* pinned */

On Wed, Jun 23, 2004 at 03:05:46PM -0700, Andrew Morton wrote:
> This would be a nice thing to keep track of.
> Isn't it the case that reclaimable slab pages (dentry, inode, mbcache,
> dquot) should not be accounted as wired memory?  Could perhaps use
> SLAB_RECLAIM_ACCOUNT for that.
> It would need to be overridden for, say, sysfs inodes and dentries, but
> they're about to become reclaimable anyway so no prob.
> It would need to be overridden for, say, ramfs dentries and inodes though.
> rd.c's blockdev pagecache pages are wired.

It's difficult to come up with comprehensible semantic refinements.
The trick with slab is whole slabs are frequently pinned by active
references, so there's quite a bit of squishiness there even after
SLAB_RECLAIM_ACCOUNT is figured out. A counter in the slab header for
references is plausibly precise, but appears to too invasive to ever
put into place. A less invasive refinement may be to clear_page_wired()
for empty slab pages.

Thanks for spotting rd.c's blkdev pagecache; that's a bogon in my
patches.

Also, I should mention this concept is based on code seen in RHEL3,
which uses a __GFP_WIRED flag, albeit for a different purpose (it
appears to be centered around removing dirty ramfs pagecache from
the LRU lists(s) as opposed to accounting for OOM-related reasons).


-- wli

Index: linux-2.6.7/drivers/block/rd.c
===================================================================
--- linux-2.6.7.orig/drivers/block/rd.c	2004-06-16 05:19:37.000000000 +0000
+++ linux-2.6.7/drivers/block/rd.c	2004-06-23 22:16:45.000000000 +0000
@@ -378,7 +378,7 @@
 		 */
 		gfp_mask = mapping_gfp_mask(mapping);
 		gfp_mask &= ~(__GFP_FS|__GFP_IO);
-		gfp_mask |= __GFP_HIGH;
+		gfp_mask |= __GFP_HIGH|__GFP_WIRED;
 		mapping_set_gfp_mask(mapping, gfp_mask);
 	}