balance.rst 5.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102
  1. .. _balance:
  2. ================
  3. Memory Balancing
  4. ================
  5. Started Jan 2000 by Kanoj Sarcar <kanoj@sgi.com>
  6. Memory balancing is needed for !__GFP_ATOMIC and !__GFP_KSWAPD_RECLAIM as
  7. well as for non __GFP_IO allocations.
  8. The first reason why a caller may avoid reclaim is that the caller can not
  9. sleep due to holding a spinlock or is in interrupt context. The second may
  10. be that the caller is willing to fail the allocation without incurring the
  11. overhead of page reclaim. This may happen for opportunistic high-order
  12. allocation requests that have order-0 fallback options. In such cases,
  13. the caller may also wish to avoid waking kswapd.
  14. __GFP_IO allocation requests are made to prevent file system deadlocks.
  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. That being said, the kernel should try to fulfill requests for direct
  20. mapped pages from the direct mapped pool, instead of falling back on
  21. the dma pool, so as to keep the dma pool filled for dma requests (atomic
  22. or not). A similar argument applies to highmem and direct mapped pages.
  23. OTOH, if there is a lot of free dma pages, it is preferable to satisfy
  24. regular memory requests by allocating one from the dma pool, instead
  25. of incurring the overhead of regular zone balancing.
  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. In 2.3, zone balancing can be done in one of two ways: depending on the
  34. zone size (and possibly of the size of lower class zones), we can decide
  35. at init time how many free pages we should aim for while balancing any
  36. zone. The good part is, while balancing, we do not need to look at sizes
  37. of lower class zones, the bad part is, we might do too frequent balancing
  38. due to ignoring possibly lower usage in the lower class zones. Also,
  39. with a slight change in the allocation routine, it is possible to reduce
  40. the memclass() macro to be a simple equality.
  41. Another possible solution is that we balance only when the free memory
  42. of a zone _and_ all its lower class zones falls below 1/64th of the
  43. total memory in the zone and its lower class zones. This fixes the 2.2
  44. balancing problem, and stays as close to 2.2 behavior as possible. Also,
  45. the balancing algorithm works the same way on the various architectures,
  46. which have different numbers and types of zones. If we wanted to get
  47. fancy, we could assign different weights to free pages in different
  48. zones in the future.
  49. Note that if the size of the regular zone is huge compared to dma zone,
  50. it becomes less significant to consider the free dma pages while
  51. deciding whether to balance the regular zone. The first solution
  52. becomes more attractive then.
  53. The appended patch implements the second solution. It also "fixes" two
  54. problems: first, kswapd is woken up as in 2.2 on low memory conditions
  55. for non-sleepable allocations. Second, the HIGHMEM zone is also balanced,
  56. so as to give a fighting chance for replace_with_highmem() to get a
  57. HIGHMEM page, as well as to ensure that HIGHMEM allocations do not
  58. fall back into regular zone. This also makes sure that HIGHMEM pages
  59. are not leaked (for example, in situations where a HIGHMEM page is in
  60. the swapcache but is not being used by anyone)
  61. kswapd also needs to know about the zones it should balance. kswapd is
  62. primarily needed in a situation where balancing can not be done,
  63. probably because all allocation requests are coming from intr context
  64. and all process contexts are sleeping. For 2.3, kswapd does not really
  65. need to balance the highmem zone, since intr context does not request
  66. highmem pages. kswapd looks at the zone_wake_kswapd field in the zone
  67. structure to decide whether a zone needs balancing.
  68. Page stealing from process memory and shm is done if stealing the page would
  69. alleviate memory pressure on any zone in the page's node that has fallen below
  70. its watermark.
  71. watemark[WMARK_MIN/WMARK_LOW/WMARK_HIGH]/low_on_memory/zone_wake_kswapd: These
  72. are per-zone fields, used to determine when a zone needs to be balanced. When
  73. the number of pages falls below watermark[WMARK_MIN], the hysteric field
  74. low_on_memory gets set. This stays set till the number of free pages becomes
  75. watermark[WMARK_HIGH]. When low_on_memory is set, page allocation requests will
  76. try to free some pages in the zone (providing GFP_WAIT is set in the request).
  77. Orthogonal to this, is the decision to poke kswapd to free some zone pages.
  78. That decision is not hysteresis based, and is done when the number of free
  79. pages is below watermark[WMARK_LOW]; in which case zone_wake_kswapd is also set.
  80. (Good) Ideas that I have heard:
  81. 1. Dynamic experience should influence balancing: number of failed requests
  82. for a zone can be tracked and fed into the balancing scheme (jalvo@mbay.net)
  83. 2. Implement a replace_with_highmem()-like replace_with_regular() to preserve
  84. dma pages. (lkd@tantalophile.demon.co.uk)