aboutsummaryrefslogtreecommitdiff
path: root/Documentation/vm/balance
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 15:20:36 -0700
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 15:20:36 -0700
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /Documentation/vm/balance
Linux-2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'Documentation/vm/balance')
-rw-r--r--Documentation/vm/balance93
1 files changed, 93 insertions, 0 deletions
diff --git a/Documentation/vm/balance b/Documentation/vm/balance
new file mode 100644
index 00000000000..bd3d31bc491
--- /dev/null
+++ b/Documentation/vm/balance
@@ -0,0 +1,93 @@
+Started Jan 2000 by Kanoj Sarcar <kanoj@sgi.com>
+
+Memory balancing is needed for non __GFP_WAIT as well as for non
+__GFP_IO allocations.
+
+There are two reasons to be requesting non __GFP_WAIT allocations:
+the caller can not sleep (typically intr context), or does not want
+to incur cost overheads of page stealing and possible swap io for
+whatever reasons.
+
+__GFP_IO allocation requests are made to prevent file system deadlocks.
+
+In the absence of non sleepable allocation requests, it seems detrimental
+to be doing balancing. Page reclamation can be kicked off lazily, that
+is, only when needed (aka zone free memory is 0), instead of making it
+a proactive process.
+
+That being said, the kernel should try to fulfill requests for direct
+mapped pages from the direct mapped pool, instead of falling back on
+the dma pool, so as to keep the dma pool filled for dma requests (atomic
+or not). A similar argument applies to highmem and direct mapped pages.
+OTOH, if there is a lot of free dma pages, it is preferable to satisfy
+regular memory requests by allocating one from the dma pool, instead
+of incurring the overhead of regular zone balancing.
+
+In 2.2, memory balancing/page reclamation would kick off only when the
+_total_ number of free pages fell below 1/64 th of total memory. With the
+right ratio of dma and regular memory, it is quite possible that balancing
+would not be done even when the dma zone was completely empty. 2.2 has
+been running production machines of varying memory sizes, and seems to be
+doing fine even with the presence of this problem. In 2.3, due to
+HIGHMEM, this problem is aggravated.
+
+In 2.3, zone balancing can be done in one of two ways: depending on the
+zone size (and possibly of the size of lower class zones), we can decide
+at init time how many free pages we should aim for while balancing any
+zone. The good part is, while balancing, we do not need to look at sizes
+of lower class zones, the bad part is, we might do too frequent balancing
+due to ignoring possibly lower usage in the lower class zones. Also,
+with a slight change in the allocation routine, it is possible to reduce
+the memclass() macro to be a simple equality.
+
+Another possible solution is that we balance only when the free memory
+of a zone _and_ all its lower class zones falls below 1/64th of the
+total memory in the zone and its lower class zones. This fixes the 2.2
+balancing problem, and stays as close to 2.2 behavior as possible. Also,
+the balancing algorithm works the same way on the various architectures,
+which have different numbers and types of zones. If we wanted to get
+fancy, we could assign different weights to free pages in different
+zones in the future.
+
+Note that if the size of the regular zone is huge compared to dma zone,
+it becomes less significant to consider the free dma pages while
+deciding whether to balance the regular zone. The first solution
+becomes more attractive then.
+
+The appended patch implements the second solution. It also "fixes" two
+problems: first, kswapd is woken up as in 2.2 on low memory conditions
+for non-sleepable allocations. Second, the HIGHMEM zone is also balanced,
+so as to give a fighting chance for replace_with_highmem() to get a
+HIGHMEM page, as well as to ensure that HIGHMEM allocations do not
+fall back into regular zone. This also makes sure that HIGHMEM pages
+are not leaked (for example, in situations where a HIGHMEM page is in
+the swapcache but is not being used by anyone)
+
+kswapd also needs to know about the zones it should balance. kswapd is
+primarily needed in a situation where balancing can not be done,
+probably because all allocation requests are coming from intr context
+and all process contexts are sleeping. For 2.3, kswapd does not really
+need to balance the highmem zone, since intr context does not request
+highmem pages. kswapd looks at the zone_wake_kswapd field in the zone
+structure to decide whether a zone needs balancing.
+
+Page stealing from process memory and shm is done if stealing the page would
+alleviate memory pressure on any zone in the page's node that has fallen below
+its watermark.
+
+pages_min/pages_low/pages_high/low_on_memory/zone_wake_kswapd: These are
+per-zone fields, used to determine when a zone needs to be balanced. When
+the number of pages falls below pages_min, the hysteric field low_on_memory
+gets set. This stays set till the number of free pages becomes pages_high.
+When low_on_memory is set, page allocation requests will try to free some
+pages in the zone (providing GFP_WAIT is set in the request). Orthogonal
+to this, is the decision to poke kswapd to free some zone pages. That
+decision is not hysteresis based, and is done when the number of free
+pages is below pages_low; in which case zone_wake_kswapd is also set.
+
+
+(Good) Ideas that I have heard:
+1. Dynamic experience should influence balancing: number of failed requests
+for a zone can be tracked and fed into the balancing scheme (jalvo@mbay.net)
+2. Implement a replace_with_highmem()-like replace_with_regular() to preserve
+dma pages. (lkd@tantalophile.demon.co.uk)