On 16.04.21 13:06, Vlastimil Babka wrote:
On 4/14/21 4:38 AM, lipeifeng@xxxxxxxx wrote:From: lipeifeng <lipeifeng@xxxxxxxx> This patch would "sort" the free-pages in buddy by pages-PFN to concentrate low-order-pages allocation in the front area of memory and high-order-pages allcation on the contrary so that few memory-pollution in the back area of memory and the probablity of high-order-pages allocation would be increased significantly. ----------------------------------------------------------------------- 1) Divide memory into several segments by pages-PFN "Multi_freearea" would divide memory into FREE_AREA_COUNTS segments by pages-PFN,each memory-segment corresponds to a free_area. Example: machine(4G of physical memery) and FREE_AREA_COUNTS(4): page-PFN:0x0 0x40000(1G) 0x80000(2G) 0xc0000(3G) 0xFFFFF(4G) |------------|--------------|--------------|-------------| free_area: [0][] [1][] [2][] [3][] NOTE: Selecting the corresponding freearea when pages are freed back to buddy: - pages-PFN[0, free_area_segment[0].max_pfn] -> free_area[0][] - pages-PFN[free_area_segment[flc - 1].max_pfn, free_area_segment[flc].max_pfn] -> free_area[flc][] (flc > 0) By this way, all pages in the same segment/free_area is within a certain range of pages-PFN. 2) Select the corresponding freearea to alloc-pages "Multi_freearea" would select the corresponding free_area by the allocation-order when alloc-pages. - order < HIGH_ORDER_TO_FLC: free_area[0] -> ... -> free_area[FREE_AREA_COUNTS - 1] - order >= HIGH_ORDER_TO_FLC: free_area[FREE_AREA_COUNTS - 1] -> ... -> free_area[0] Example: The machine(4G of physical memery) and FREE_AREA_COUNTS(4) and HIGH_ORDER_TO_FLC(3). If user allocs page(order = 0),it would take page from free_area[0][] first, if that fails,try free_area[1][] and so on. If user allocs page(order = 4),it would take page from free_area[3][] first, if that fails,try free_area[2][] and so on. By this way,low-order pages will be concentrated in the front area of memory.Because of few memory-pollution in the back area of memory, the sussessful probablity of high-order allocation would be improved. 3) Adjust the location of free-pages in the free_list "Multi_freearea" would place free-pages in the head of free_list if pages-PFN is smaller than free_area_segment[flc]->median_pfn and in the tail of free_list on the contrary. Example: page-PFN: free_area_segment[flc]->median_pfn | free_list: page->page->page->...|...page->page->page pages-PFN:| < median_pfn | >= median_pfn | Because it would take pages from the head of the freelist first in buddy system,the free-pages in the tail are more likely to keep in the buddy system.The closer the PFN of pages kept in buddy system, the greater the probablity of merging that into high-order pages.I think this part 3) would be worth to be tried separately first, as it's not a big change compared to the other ones.
Let's consider part 3 only and ignore the 1) multi freearea (which might be problematic with sparcity) and 2) the modified allocation scheme (which doesn't yet quite sense to me yet, e.g., because we group by mobility and have compaction in place; I assume this really only helps in some special cases -- like the test case you are giving; I might be wrong)
Right now, we decide whether to but to head or tail based on how likely it is that we might merge to a higher-order page (buddy_merge_likely()) in the future. So we only consider the current "neighborhood" of the page we're freeing. As we restrict our neighborhood to MAX_ORDER - 1 pages (what we can actually merge). Of course, we can easily be wrong here. Grouping by movability and compaction only helps to some degree I guess.
AFAIK, what you propose is basing the decisions where to place a page (in addition?) on a median_pfn. Without 1) and 2) I cannot completely understand if 3) itself would help at all (and how to set the median_pfn). But it would certainly be interesting if we can tweak the current logic to better identify merge targets simply by tweaking buddy_merge_likely() or the assumptions it is making.
-- Thanks, David / dhildenb
