xref: /linux/Documentation/mm/balance.rst (revision 70756b49be4ea8bf36a664322df6e7e89895fa60)
1================
2Memory Balancing
3================
4
5Started Jan 2000 by Kanoj Sarcar <kanoj@sgi.com>
6
7Memory balancing is needed for !__GFP_ATOMIC and !__GFP_KSWAPD_RECLAIM as
8well as for non __GFP_IO allocations.
9
10The first reason why a caller may avoid reclaim is that the caller can not
11sleep due to holding a spinlock or is in interrupt context. The second may
12be that the caller is willing to fail the allocation without incurring the
13overhead of page reclaim. This may happen for opportunistic high-order
14allocation requests that have order-0 fallback options. In such cases,
15the caller may also wish to avoid waking kswapd.
16
17__GFP_IO allocation requests are made to prevent file system deadlocks.
18
19In the absence of non sleepable allocation requests, it seems detrimental
20to be doing balancing. Page reclamation can be kicked off lazily, that
21is, only when needed (aka zone free memory is 0), instead of making it
22a proactive process.
23
24That being said, the kernel should try to fulfill requests for direct
25mapped pages from the direct mapped pool, instead of falling back on
26the dma pool, so as to keep the dma pool filled for dma requests (atomic
27or not). A similar argument applies to highmem and direct mapped pages.
28OTOH, if there is a lot of free dma pages, it is preferable to satisfy
29regular memory requests by allocating one from the dma pool, instead
30of incurring the overhead of regular zone balancing.
31
32In 2.2, memory balancing/page reclamation would kick off only when the
33_total_ number of free pages fell below 1/64 th of total memory. With the
34right ratio of dma and regular memory, it is quite possible that balancing
35would not be done even when the dma zone was completely empty. 2.2 has
36been running production machines of varying memory sizes, and seems to be
37doing fine even with the presence of this problem. In 2.3, due to
38HIGHMEM, this problem is aggravated.
39
40In 2.3, zone balancing can be done in one of two ways: depending on the
41zone size (and possibly of the size of lower class zones), we can decide
42at init time how many free pages we should aim for while balancing any
43zone. The good part is, while balancing, we do not need to look at sizes
44of lower class zones, the bad part is, we might do too frequent balancing
45due to ignoring possibly lower usage in the lower class zones. Also,
46with a slight change in the allocation routine, it is possible to reduce
47the memclass() macro to be a simple equality.
48
49Another possible solution is that we balance only when the free memory
50of a zone _and_ all its lower class zones falls below 1/64th of the
51total memory in the zone and its lower class zones. This fixes the 2.2
52balancing problem, and stays as close to 2.2 behavior as possible. Also,
53the balancing algorithm works the same way on the various architectures,
54which have different numbers and types of zones. If we wanted to get
55fancy, we could assign different weights to free pages in different
56zones in the future.
57
58Note that if the size of the regular zone is huge compared to dma zone,
59it becomes less significant to consider the free dma pages while
60deciding whether to balance the regular zone. The first solution
61becomes more attractive then.
62
63The appended patch implements the second solution. It also "fixes" two
64problems: first, kswapd is woken up as in 2.2 on low memory conditions
65for non-sleepable allocations. Second, the HIGHMEM zone is also balanced,
66so as to give a fighting chance for replace_with_highmem() to get a
67HIGHMEM page, as well as to ensure that HIGHMEM allocations do not
68fall back into regular zone. This also makes sure that HIGHMEM pages
69are not leaked (for example, in situations where a HIGHMEM page is in
70the swapcache but is not being used by anyone)
71
72kswapd also needs to know about the zones it should balance. kswapd is
73primarily needed in a situation where balancing can not be done,
74probably because all allocation requests are coming from intr context
75and all process contexts are sleeping. For 2.3, kswapd does not really
76need to balance the highmem zone, since intr context does not request
77highmem pages. kswapd looks at the zone_wake_kswapd field in the zone
78structure to decide whether a zone needs balancing.
79
80Page stealing from process memory and shm is done if stealing the page would
81alleviate memory pressure on any zone in the page's node that has fallen below
82its watermark.
83
84watemark[WMARK_MIN/WMARK_LOW/WMARK_HIGH]/low_on_memory/zone_wake_kswapd: These
85are per-zone fields, used to determine when a zone needs to be balanced. When
86the number of pages falls below watermark[WMARK_MIN], the hysteric field
87low_on_memory gets set. This stays set till the number of free pages becomes
88watermark[WMARK_HIGH]. When low_on_memory is set, page allocation requests will
89try to free some pages in the zone (providing GFP_WAIT is set in the request).
90Orthogonal to this, is the decision to poke kswapd to free some zone pages.
91That decision is not hysteresis based, and is done when the number of free
92pages is below watermark[WMARK_LOW]; in which case zone_wake_kswapd is also set.
93
94
95(Good) Ideas that I have heard:
96
971. Dynamic experience should influence balancing: number of failed requests
98   for a zone can be tracked and fed into the balancing scheme (jalvo@mbay.net)
992. Implement a replace_with_highmem()-like replace_with_regular() to preserve
100   dma pages. (lkd@tantalophile.demon.co.uk)
101