Caches that shrink automatically

Hello,

Currently there is no way for modules to define dynamically sized caches that shrink upon memory pressure. We need this for implementing Extended Attribute caches on ext2, ext3, and ReiserFS. Other caches could also make use of the same mechanism (e.g., nfsd's permission cache, dcache, icache, dqache).

I propose this patch, which adds the register_cache() and unregister_cache() functions. They allow to register a callback which is invoked on memory pressure. This callback shall then try to free some memory; the parameters and semantics are similar to the other shrink functions in mm/vmscan.c.

Regards,
Andreas.

------------------------------------------------------------------
Andreas Gruenbacher SuSE Linux AG
mailto:[email protected] Deutschherrnstr. 15-19
http://www.suse.de/ D-90429 Nuernberg, Germany

Modular caches that shrink automatically

This patch adds the register_cache() and unregister_cache() functions which register a callback that is invoked on memory pressure to free some memory. The parameters and semantics to the callback are similar to the other shrink functions.

diff -Nur linux-2.5.30/include/linux/cache_def.h linux-2.5.30.patch/include/linux/cache_def.h
@@ -0,0 +1,15 @@
+/*
+ * linux/cache_def.h
+ * Modular caches that shrink automatically
+ *
+ * Copyright (C) 2002 by Andreas Gruenbacher, <[email protected]>
+ */
+
+struct cache_definition {
+	const char *name;
+	void (*shrink)(int, unsigned int);
+	struct list_head link;
+};
+
+extern void register_cache(struct cache_definition *);
+extern void unregister_cache(struct cache_definition *);

diff -Nur linux-2.5.30/kernel/ksyms.c linux-2.5.30.patch/kernel/ksyms.c
@@ -31,6 +31,7 @@
 #include <linux/genhd.h>
 #include <linux/blkpg.h>
 #include <linux/swap.h>
+#include <linux/cache_def.h>
 #include <linux/ctype.h>
 #include <linux/file.h>
 #include <linux/iobuf.h>

@@ -106,6 +107,8 @@
 EXPORT_SYMBOL(kmem_cache_alloc);
 EXPORT_SYMBOL(kmem_cache_free);
 EXPORT_SYMBOL(kmem_cache_size);
+EXPORT_SYMBOL(register_cache);
+EXPORT_SYMBOL(unregister_cache);
 EXPORT_SYMBOL(kmalloc);
 EXPORT_SYMBOL(kfree);
 EXPORT_SYMBOL(vfree);

Binary files linux-2.5.30/mm/.vmscan.c.swp and linux-2.5.30.patch/mm/.vmscan.c.swp differ

diff -Nur linux-2.5.30/mm/vmscan.c linux-2.5.30.patch/mm/vmscan.c
@@ -15,6 +15,7 @@
 #include <linux/slab.h>
 #include <linux/kernel_stat.h>
 #include <linux/swap.h>
+#include <linux/cache_def.h>
 #include <linux/smp_lock.h>
 #include <linux/pagemap.h>
 #include <linux/init.h>

@@ -61,6 +62,39 @@
 	return 0;
 }
 
+static DECLARE_MUTEX(other_caches_lock);
+static LIST_HEAD(other_caches);
+
+void register_cache(struct cache_definition *cache)
+{
+	down(&other_caches_lock);
+	list_add(&cache->link, &other_caches);
+	up(&other_caches_lock);
+}
+
+void unregister_cache(struct cache_definition *cache)
+{
+	down(&other_caches_lock);
+	list_del(&cache->link);
+	up(&other_caches_lock);
+}
+
+static void shrink_other_caches(unsigned int priority, int gfp_mask)
+{
+	struct list_head *p = other_caches.prev;
+
+	down(&other_caches_lock);
+	while (p != &other_caches) {
+		struct cache_definition *cache =
+			list_entry(p, struct cache_definition, link);
+
+		cache->shrink(priority, gfp_mask);
+		p = p->prev;
+	}
+	up(&other_caches_lock);
+}
+
+
 static int
 shrink_cache(int nr_pages, zone_t *classzone,
 		unsigned int gfp_mask, int priority, int max_scan)

@@ -395,6 +429,7 @@
 #ifdef CONFIG_QUOTA
 	shrink_dqcache_memory(DEF_PRIORITY, gfp_mask);
 #endif
+	shrink_other_caches(priority, gfp_mask);
 
 	return nr_pages;
 }

How do you ensure that caches have their (internal) aging hands pushed at a speed that is proportional to their memory usage, or is your design susceptible to all the usual complaints the unified memory manager crowd has about separate caches?

Hans

Hans

That's a policy/optimization issue; it's not even desirable to shrink the caches with priorities proportional to their size---they would all tend to become equally large.

The patch shrinks all the caches equally often, with the same priorities. The caches can then decide themselves how they will react, depending on their cache size and entry size, replacement strategy, taking care of page entry clustering or not, etc.

The icache, dcache, and dqcache are shrunk using the same strategy (except the priority is a constant for some of the caches, which could be coded in the shrink function as well). This scheme has worked out pretty well so far, right?

For Extended Attributes we are currently using a very simple cache with LRU entry replacement, and small entries. The cache doesn't grow very big, either.

Regards,
Andreas.

------------------------------------------------------------------
Andreas Gruenbacher SuSE Linux AG
mailto:[email protected] Deutschherrnstr. 15-19
http://www.suse.de/ D-90429 Nuernberg, Germany

Nope, the idea is to push all caches according to size, but often-used caches should shrink less than caches that are hardly ever used.

Not quite, we still have some bad problems balancing the size of these caches versus the size of the other VM occupants.

However, your shrinking function is good enough for now and can be used with something like Ed Tomlinson's approach later on to make reclaiming better balanced.

regards,

Rik

Bravely reimplemented by the knights who say "NIH".

http://www.surriel.com/		http://distro.conectiva.com/

Do you let the subcache decide how to move the aging hand and track it? Have I convinced you of that one yet? Or is it still page based?

Hans

Linus has indicated that he would like to have it page based, but implementation issues point towards letting the subcache manage its objects by itself ;)

Rik

Bravely reimplemented by the knights who say "NIH".

http://www.surriel.com/		http://distro.conectiva.com/

The two are not mutually incompatible.

I think we've all seen that non-global shrinking just doesn't work very well, because one cache that grows too large will end up asking everybody else to shrink, even if a global shrinking policy would have realized that the memory pressure is due to that one overly large cache. The resulting balancing problems are "interesting".

Being purely page-based, together with support for the sub-caches knowing about the page-based aging, should be fine.

In particular, it is useless for the sub-caches to try to maintain their own LRU lists and their own accessed bits. But that doesn't mean that they can _act_ as if they updated their own accessed bits, while really just telling the page-based thing that that page is active.

This is what the buffer cache has been doing for the last two years, ie "touch_buffer()" actually ends up touching the page. Which seems to be working quite well.

Linus

I'm not sure I agree with this. For eg. the dcache you will want to reclaim the less used entries on a page even if there are a few very intensely used entries on that page.

This is because then new dcache allocations can come from the empty space on this page and the dcache doesn't have to grow in order to allocate new entries.

If we were to go fully page-based, it'd become impossible to evict dcache entries from any page with at least one heavily used entry and the dcache will use much more space than otherwise required. In reality it'd be even worse because we cannot evict any dcache entry which is pinned by eg. inodes or directory entries.

Please tell me if I've overlooked something, it would be nice if the page based scheme could work out after all ;)

regards,

Rik

Bravely reimplemented by the knights who say "NIH".

http://www.surriel.com/		http://distro.conectiva.com/

True in theory, but I doubt you will see it very much in practice.

Most of the time when you want to free dentries, it is because you have a _ton_ of them.

The fact that some will look cold even if they aren't should not matter that much statistically.

Yah, it's a guess. We can test it.

Linus

Josh tested it. He posted on it. I'll have him find his original post and repost tomorrow, but summarized in brief, the current dcache shrinking/management approach was quite inefficient. Each active dcache entry kept a whole lot of dead ones around.

Hans

It seems the rmap+slab-LRU patches work to address this problem, but even with those patches applied Ed Tomlinson's test shows a ~50% utilization of dcache slab pages under memory pressure. My original test on 2.4.18 showed a ~25% utilization under memory pressure, so the slab-LRU patches definetly show an improvement. In practice and theory, we see that the slab caches are inefficient at using memory under pressure, so probably there is more work to be done. Ed's posted results:

http://groups.google.com/groups?hl=en&lr=&ie=UTF-8&selm=01041309360100.32757%40oscar

Here is the text of the message that Hans refers to:

http://groups.google.com/groups?q=g:thl2323088796d&dq=&hl=en&lr=&ie=UTF-8&selm=20020128091338.D6578%40helen.CS.Berkeley.EDU

--

When memory pressure becomes high, the Linux kswapd begins calling shrink_caches() from try_to_free_pages() with an integer priority from 6 (the default, lowest priority) to 1 (high priority). Looking specifically at the dcache, this results in a calls to shrink_dcache_memory() that attempt to free a fraction (1/priority) of the inactive dcache entries. This ultimately leads to prune_dcache() scanning the dcache in least-recently-used order attempting to call kmem_cache_free() on some number of dcache entries.

Dcache entries are allocated from the kmem_slab_cache, which manages objects in page-size "slabs", but the kmem_slab_cache cannot free a page until every object in a slab becomes unused. The problem is that freeing dcache entries in LRU-order is effectively freeing entries from randomly-selected slabs, and therefore applying shrink_caches() pressure to the dcache has an undesired result. In the attempt to reduce its size, the dcache must free objects from random slabs in order to actually release full pages. The result is that under high memory pressure the dcache utilization drops dramatically. The prune_dcache() mechanism doesn't just reduce the page utilization as desired, it reduces the intra-page utilization, which is bad.

In order to measure this effect (via /proc/slabinfo) I first populated a large dcache and then ran a memory-hog to force swapping to occur. The dcache utilization drops to between 20-35%. For example, before running the memory-hog my dcache reports:

dentry_cache 10170 10170 128 339 339 1 : 252 126

(i.e., 10170 active dentry objects, 10170 available dentry objects @ 128 bytes each, 339 pages with at least one object, and 339 allocated pages, an approximately 1.4MB dcache)

While running the memory-hog program to initiate swapping, the dcache stands at:

dentry_cache 693 3150 128 105 105 1 : 252 126

Meaning, the randomly-applied cache pressure was successful at freeing 234 (= 339-105) pages, leaving a 430KB dcache, but at the same time it reduced the cache utilization to 22%, meaning that although it was able to free nearly 1MB of space, 335KB are now wasted as a result of the high memory-pressure condition.

So, it would seem that the dcache and kmem_slab_cache memory allocator could benefit from a way to shrink the dcache in a less random way. Any thoughts?

Another angle worth taking into account is that for any data that is purely cached, (that is not currently pinned by a user), it is possible to compact the data. With rmap is property even applies to pinned page cache pages.

For cases where internal fragmentation is likely it may be worth taking advantage of this property.

Eric

That is not what I said, I said move the aging hand at a speed proportional to size, what affect the aging has depends on the usage of the objects in the cache.

This is wrong, because frequently used objects should not be shrunk out of their caches.

How can they decide this without a pressure signal from the master memory manager that is proportional to their size? Or is the idea that they figure out their own size? I suppose that is equivalent....

What does priority

Our cache management has deep algorithm flaws which result in things like one active dcache entry keeping an entire page of unused dcache entries from expiring from the cache. Did you see Joshua MacDonald's post on that topic?

You don't support eas that are larger than a page? Is it LRU on each ea, or is it LRU on the page containing the ea?

Hans

This is what the slablru patches should do.. Any slab user should be able to get a callback when the LRU decides that the page seems to be unused.

Linus