Based on kernel version 4.7.2. Page generated on 2016-08-22 22:48 EST.
1 Started Jan 2000 by Kanoj Sarcar <firstname.lastname@example.org> 2 3 Memory balancing is needed for !__GFP_ATOMIC and !__GFP_KSWAPD_RECLAIM as 4 well as for non __GFP_IO allocations. 5 6 The first reason why a caller may avoid reclaim is that the caller can not 7 sleep due to holding a spinlock or is in interrupt context. The second may 8 be that the caller is willing to fail the allocation without incurring the 9 overhead of page reclaim. This may happen for opportunistic high-order 10 allocation requests that have order-0 fallback options. In such cases, 11 the caller may also wish to avoid waking kswapd. 12 13 __GFP_IO allocation requests are made to prevent file system deadlocks. 14 15 In the absence of non sleepable allocation requests, it seems detrimental 16 to be doing balancing. Page reclamation can be kicked off lazily, that 17 is, only when needed (aka zone free memory is 0), instead of making it 18 a proactive process. 19 20 That being said, the kernel should try to fulfill requests for direct 21 mapped pages from the direct mapped pool, instead of falling back on 22 the dma pool, so as to keep the dma pool filled for dma requests (atomic 23 or not). A similar argument applies to highmem and direct mapped pages. 24 OTOH, if there is a lot of free dma pages, it is preferable to satisfy 25 regular memory requests by allocating one from the dma pool, instead 26 of incurring the overhead of regular zone balancing. 27 28 In 2.2, memory balancing/page reclamation would kick off only when the 29 _total_ number of free pages fell below 1/64 th of total memory. With the 30 right ratio of dma and regular memory, it is quite possible that balancing 31 would not be done even when the dma zone was completely empty. 2.2 has 32 been running production machines of varying memory sizes, and seems to be 33 doing fine even with the presence of this problem. In 2.3, due to 34 HIGHMEM, this problem is aggravated. 35 36 In 2.3, zone balancing can be done in one of two ways: depending on the 37 zone size (and possibly of the size of lower class zones), we can decide 38 at init time how many free pages we should aim for while balancing any 39 zone. The good part is, while balancing, we do not need to look at sizes 40 of lower class zones, the bad part is, we might do too frequent balancing 41 due to ignoring possibly lower usage in the lower class zones. Also, 42 with a slight change in the allocation routine, it is possible to reduce 43 the memclass() macro to be a simple equality. 44 45 Another possible solution is that we balance only when the free memory 46 of a zone _and_ all its lower class zones falls below 1/64th of the 47 total memory in the zone and its lower class zones. This fixes the 2.2 48 balancing problem, and stays as close to 2.2 behavior as possible. Also, 49 the balancing algorithm works the same way on the various architectures, 50 which have different numbers and types of zones. If we wanted to get 51 fancy, we could assign different weights to free pages in different 52 zones in the future. 53 54 Note that if the size of the regular zone is huge compared to dma zone, 55 it becomes less significant to consider the free dma pages while 56 deciding whether to balance the regular zone. The first solution 57 becomes more attractive then. 58 59 The appended patch implements the second solution. It also "fixes" two 60 problems: first, kswapd is woken up as in 2.2 on low memory conditions 61 for non-sleepable allocations. Second, the HIGHMEM zone is also balanced, 62 so as to give a fighting chance for replace_with_highmem() to get a 63 HIGHMEM page, as well as to ensure that HIGHMEM allocations do not 64 fall back into regular zone. This also makes sure that HIGHMEM pages 65 are not leaked (for example, in situations where a HIGHMEM page is in 66 the swapcache but is not being used by anyone) 67 68 kswapd also needs to know about the zones it should balance. kswapd is 69 primarily needed in a situation where balancing can not be done, 70 probably because all allocation requests are coming from intr context 71 and all process contexts are sleeping. For 2.3, kswapd does not really 72 need to balance the highmem zone, since intr context does not request 73 highmem pages. kswapd looks at the zone_wake_kswapd field in the zone 74 structure to decide whether a zone needs balancing. 75 76 Page stealing from process memory and shm is done if stealing the page would 77 alleviate memory pressure on any zone in the page's node that has fallen below 78 its watermark. 79 80 watemark[WMARK_MIN/WMARK_LOW/WMARK_HIGH]/low_on_memory/zone_wake_kswapd: These 81 are per-zone fields, used to determine when a zone needs to be balanced. When 82 the number of pages falls below watermark[WMARK_MIN], the hysteric field 83 low_on_memory gets set. This stays set till the number of free pages becomes 84 watermark[WMARK_HIGH]. When low_on_memory is set, page allocation requests will 85 try to free some pages in the zone (providing GFP_WAIT is set in the request). 86 Orthogonal to this, is the decision to poke kswapd to free some zone pages. 87 That decision is not hysteresis based, and is done when the number of free 88 pages is below watermark[WMARK_LOW]; in which case zone_wake_kswapd is also set. 89 90 91 (Good) Ideas that I have heard: 92 1. Dynamic experience should influence balancing: number of failed requests 93 for a zone can be tracked and fed into the balancing scheme (email@example.com) 94 2. Implement a replace_with_highmem()-like replace_with_regular() to preserve 95 dma pages. (firstname.lastname@example.org)