On Fri, 3 Nov 2006, Zou, Nanhai wrote:
Hi, This patch should fix the issue.
It would appear to fix the issue for IA64 but you are blotting over the issue that the map is reporting a one page hole. On arches with really adjacent regions that are getting merged, the regions will appear to overlap by one page. What can happen is something like this
PFN ranges for nodes Node 1: 0 -> 1000 Node 0: 1000 -> 2000If your fix was applied, the end effect would be that the nodes appear to overlap. The last page on node 1 would end up on the free lists of node 0 because the memmap for that page will be incorrectly initialised twice. While it wouldn't be very noticeable, it's still buggy.
The right fix is to fix the map. Alternatively, change add_active_range(nid, start_pfn, end_pfn) to add_active-range(nid, start_pfn, end_pfn+1) in the arch-specific code if you know it's the right thing to do.
Thanks Zou Nan hai-----Original Message----- From: Mel Gorman [mailto:mel@xxxxxxxxx] Sent: 2006Äê11ÔÂ3ÈÕ 0:27 To: Horms Cc: Andy Whitcroft; Zou, Nanhai; linux-ia64@xxxxxxxxxxxxxxx; Bob Picco; Andrew Morton; Dave Hansen; Andi Kleen; Benjamin Herrenschmidt; Paul Mackerras; Keith Mannthey; Luck, Tony; KAMEZAWA Hiroyuki; Yasunori Goto; Khalid Aziz Subject: Re: 05e0caad3b7bd0d0fbeff980bca22f186241a501 breaks ia64 kdump On Thu, 2 Nov 2006, Horms wrote:On Wed, Nov 01, 2006 at 09:38:38AM +0000, Mel Gorman wrote:Plan C Call add_active_ranges() for the the PFN ranges initmem reside in. Look at free_initmem() to see how to get the virtual address of the range is found and make the appropriate conversion. A variation is to call add_active_ranges() for the PFN range the whole kernel resides in if you have that information for the crash kernel.I've made a version of this approach that seems to work. I had to update add_active_range so that it merges ranges that are not overlapping, but do create a continous range. Is this a correct thing to do?It's a bit unlikely it's the correct fix. As you can see, the add_active_range() is already able to merge contiguous ranges. If the calls imply there is a one page hole, then there is a one page hole. If the caller is wrong and there isn't a one-page hole, then fix the caller walking the map. Power for example currently merges large numbers of adjacent regions that are registered separetly.-- Horms H: http://www.vergenet.net/~horms/ W: http://www.valinux.co.jp/en/ Index: linux-2.6/mm/page_alloc.c =================================================================== --- linux-2.6.orig/mm/page_alloc.c 2006-11-02 16:17:17.000000000 +0900 +++ linux-2.6/mm/page_alloc.c 2006-11-02 16:17:22.000000000 +0900 @@ -2407,7 +2407,7 @@ return; /* Merge forward if suitable */ - if (start_pfn <= early_node_map[i].end_pfn && + if (start_pfn <= early_node_map[i].end_pfn + 1 && end_pfn > early_node_map[i].end_pfn) { early_node_map[i].end_pfn = end_pfn; return; @@ -2415,7 +2415,7 @@ /* Merge backward if suitable */ if (start_pfn < early_node_map[i].end_pfn && - end_pfn >= early_node_map[i].start_pfn) { + end_pfn + 1 >= early_node_map[i].start_pfn) { early_node_map[i].start_pfn = start_pfn; return; } Index: linux-2.6/arch/ia64/mm/discontig.c =================================================================== --- linux-2.6.orig/arch/ia64/mm/discontig.c 2006-11-0216:17:59.000000000 +0900+++ linux-2.6/arch/ia64/mm/discontig.c 2006-11-02 16:18:03.000000000+0900@@ -689,6 +689,8 @@ arch_sparse_init(); efi_memmap_walk(filter_rsvd_memory, count_node_pages); + add_active_range(0, ia64_tpa(_text) >> PAGE_SHIFT, + (ia64_tpa(_end) -1 ) >> PAGE_SHIFT); #ifdef CONFIG_VIRTUAL_MEM_MAP vmalloc_end -= PAGE_ALIGN(ALIGN(max_low_pfn, MAX_ORDER_NR_PAGES) *-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
|