1Started Jan 2000 by Kanoj Sarcar <kanoj@sgi.com> 2 3Memory balancing is needed for !__GFP_ATOMIC and !__GFP_KSWAPD_RECLAIM as 4well as for non __GFP_IO allocations. 5 6The first reason why a caller may avoid reclaim is that the caller can not 7sleep due to holding a spinlock or is in interrupt context. The second may 8be that the caller is willing to fail the allocation without incurring the 9overhead of page reclaim. This may happen for opportunistic high-order 10allocation requests that have order-0 fallback options. In such cases, 11the caller may also wish to avoid waking kswapd. 12 13__GFP_IO allocation requests are made to prevent file system deadlocks. 14 15In the absence of non sleepable allocation requests, it seems detrimental 16to be doing balancing. Page reclamation can be kicked off lazily, that 17is, only when needed (aka zone free memory is 0), instead of making it 18a proactive process. 19 20That being said, the kernel should try to fulfill requests for direct 21mapped pages from the direct mapped pool, instead of falling back on 22the dma pool, so as to keep the dma pool filled for dma requests (atomic 23or not). A similar argument applies to highmem and direct mapped pages. 24OTOH, if there is a lot of free dma pages, it is preferable to satisfy 25regular memory requests by allocating one from the dma pool, instead 26of incurring the overhead of regular zone balancing. 27 28In 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 30right ratio of dma and regular memory, it is quite possible that balancing 31would not be done even when the dma zone was completely empty. 2.2 has 32been running production machines of varying memory sizes, and seems to be 33doing fine even with the presence of this problem. In 2.3, due to 34HIGHMEM, this problem is aggravated. 35 36In 2.3, zone balancing can be done in one of two ways: depending on the 37zone size (and possibly of the size of lower class zones), we can decide 38at init time how many free pages we should aim for while balancing any 39zone. The good part is, while balancing, we do not need to look at sizes 40of lower class zones, the bad part is, we might do too frequent balancing 41due to ignoring possibly lower usage in the lower class zones. Also, 42with a slight change in the allocation routine, it is possible to reduce 43the memclass() macro to be a simple equality. 44 45Another possible solution is that we balance only when the free memory 46of a zone _and_ all its lower class zones falls below 1/64th of the 47total memory in the zone and its lower class zones. This fixes the 2.2 48balancing problem, and stays as close to 2.2 behavior as possible. Also, 49the balancing algorithm works the same way on the various architectures, 50which have different numbers and types of zones. If we wanted to get 51fancy, we could assign different weights to free pages in different 52zones in the future. 53 54Note that if the size of the regular zone is huge compared to dma zone, 55it becomes less significant to consider the free dma pages while 56deciding whether to balance the regular zone. The first solution 57becomes more attractive then. 58 59The appended patch implements the second solution. It also "fixes" two 60problems: first, kswapd is woken up as in 2.2 on low memory conditions 61for non-sleepable allocations. Second, the HIGHMEM zone is also balanced, 62so as to give a fighting chance for replace_with_highmem() to get a 63HIGHMEM page, as well as to ensure that HIGHMEM allocations do not 64fall back into regular zone. This also makes sure that HIGHMEM pages 65are not leaked (for example, in situations where a HIGHMEM page is in 66the swapcache but is not being used by anyone) 67 68kswapd also needs to know about the zones it should balance. kswapd is 69primarily needed in a situation where balancing can not be done, 70probably because all allocation requests are coming from intr context 71and all process contexts are sleeping. For 2.3, kswapd does not really 72need to balance the highmem zone, since intr context does not request 73highmem pages. kswapd looks at the zone_wake_kswapd field in the zone 74structure to decide whether a zone needs balancing. 75 76Page stealing from process memory and shm is done if stealing the page would 77alleviate memory pressure on any zone in the page's node that has fallen below 78its watermark. 79 80watemark[WMARK_MIN/WMARK_LOW/WMARK_HIGH]/low_on_memory/zone_wake_kswapd: These 81are per-zone fields, used to determine when a zone needs to be balanced. When 82the number of pages falls below watermark[WMARK_MIN], the hysteric field 83low_on_memory gets set. This stays set till the number of free pages becomes 84watermark[WMARK_HIGH]. When low_on_memory is set, page allocation requests will 85try to free some pages in the zone (providing GFP_WAIT is set in the request). 86Orthogonal to this, is the decision to poke kswapd to free some zone pages. 87That decision is not hysteresis based, and is done when the number of free 88pages is below watermark[WMARK_LOW]; in which case zone_wake_kswapd is also set. 89 90 91(Good) Ideas that I have heard: 921. Dynamic experience should influence balancing: number of failed requests 93for a zone can be tracked and fed into the balancing scheme (jalvo@mbay.net) 942. Implement a replace_with_highmem()-like replace_with_regular() to preserve 95dma pages. (lkd@tantalophile.demon.co.uk) 96