Based on kernel version 2.6.26. Page generated on 2008-07-16 21:13 EST.
1 Started Jan 2000 by Kanoj Sarcar <kanoj[AT]sgi[DOT]com> 2 3 Memory balancing is needed for non __GFP_WAIT as well as for non 4 __GFP_IO allocations. 5 6 There are two reasons to be requesting non __GFP_WAIT allocations: 7 the caller can not sleep (typically intr context), or does not want 8 to incur cost overheads of page stealing and possible swap io for 9 whatever reasons. 10 11 __GFP_IO allocation requests are made to prevent file system deadlocks. 12 13 In the absence of non sleepable allocation requests, it seems detrimental 14 to be doing balancing. Page reclamation can be kicked off lazily, that 15 is, only when needed (aka zone free memory is 0), instead of making it 16 a proactive process. 17 18 That being said, the kernel should try to fulfill requests for direct 19 mapped pages from the direct mapped pool, instead of falling back on 20 the dma pool, so as to keep the dma pool filled for dma requests (atomic 21 or not). A similar argument applies to highmem and direct mapped pages. 22 OTOH, if there is a lot of free dma pages, it is preferable to satisfy 23 regular memory requests by allocating one from the dma pool, instead 24 of incurring the overhead of regular zone balancing. 25 26 In 2.2, memory balancing/page reclamation would kick off only when the 27 _total_ number of free pages fell below 1/64 th of total memory. With the 28 right ratio of dma and regular memory, it is quite possible that balancing 29 would not be done even when the dma zone was completely empty. 2.2 has 30 been running production machines of varying memory sizes, and seems to be 31 doing fine even with the presence of this problem. In 2.3, due to 32 HIGHMEM, this problem is aggravated. 33 34 In 2.3, zone balancing can be done in one of two ways: depending on the 35 zone size (and possibly of the size of lower class zones), we can decide 36 at init time how many free pages we should aim for while balancing any 37 zone. The good part is, while balancing, we do not need to look at sizes 38 of lower class zones, the bad part is, we might do too frequent balancing 39 due to ignoring possibly lower usage in the lower class zones. Also, 40 with a slight change in the allocation routine, it is possible to reduce 41 the memclass() macro to be a simple equality. 42 43 Another possible solution is that we balance only when the free memory 44 of a zone _and_ all its lower class zones falls below 1/64th of the 45 total memory in the zone and its lower class zones. This fixes the 2.2 46 balancing problem, and stays as close to 2.2 behavior as possible. Also, 47 the balancing algorithm works the same way on the various architectures, 48 which have different numbers and types of zones. If we wanted to get 49 fancy, we could assign different weights to free pages in different 50 zones in the future. 51 52 Note that if the size of the regular zone is huge compared to dma zone, 53 it becomes less significant to consider the free dma pages while 54 deciding whether to balance the regular zone. The first solution 55 becomes more attractive then. 56 57 The appended patch implements the second solution. It also "fixes" two 58 problems: first, kswapd is woken up as in 2.2 on low memory conditions 59 for non-sleepable allocations. Second, the HIGHMEM zone is also balanced, 60 so as to give a fighting chance for replace_with_highmem() to get a 61 HIGHMEM page, as well as to ensure that HIGHMEM allocations do not 62 fall back into regular zone. This also makes sure that HIGHMEM pages 63 are not leaked (for example, in situations where a HIGHMEM page is in 64 the swapcache but is not being used by anyone) 65 66 kswapd also needs to know about the zones it should balance. kswapd is 67 primarily needed in a situation where balancing can not be done, 68 probably because all allocation requests are coming from intr context 69 and all process contexts are sleeping. For 2.3, kswapd does not really 70 need to balance the highmem zone, since intr context does not request 71 highmem pages. kswapd looks at the zone_wake_kswapd field in the zone 72 structure to decide whether a zone needs balancing. 73 74 Page stealing from process memory and shm is done if stealing the page would 75 alleviate memory pressure on any zone in the page's node that has fallen below 76 its watermark. 77 78 pages_min/pages_low/pages_high/low_on_memory/zone_wake_kswapd: These are 79 per-zone fields, used to determine when a zone needs to be balanced. When 80 the number of pages falls below pages_min, the hysteric field low_on_memory 81 gets set. This stays set till the number of free pages becomes pages_high. 82 When low_on_memory is set, page allocation requests will try to free some 83 pages in the zone (providing GFP_WAIT is set in the request). Orthogonal 84 to this, is the decision to poke kswapd to free some zone pages. That 85 decision is not hysteresis based, and is done when the number of free 86 pages is below pages_low; in which case zone_wake_kswapd is also set. 87 88 89 (Good) Ideas that I have heard: 90 1. Dynamic experience should influence balancing: number of failed requests 91 for a zone can be tracked and fed into the balancing scheme (jalvo[AT]mbay[DOT]net) 92 2. Implement a replace_with_highmem()-like replace_with_regular() to preserve 93 dma pages. (lkd[AT]tantalophile.demon.co[DOT]uk)