Hi all, it seems crash shows incorrect HighMem stats with recent kernels. E.g.: crash> kmem -i [...] TOTAL HIGH 1821682 6.9 GB 93% of TOTAL MEM FREE HIGH 0 0 0% of TOTAL HIGH TOTAL LOW 132983 519.5 MB 6% of TOTAL MEM FREE LOW 1333474 5.1 GB 1002% of TOTAL LOW BTW note that total low is smaller than free low, which is obviously incorrect. I believe that this is somehow related to the Movable zone, because the code that counts free low pages checks for pages which belong to ZONE_HIGHMEM, which is initialized as: vt->ZONE_HIGHMEM = vt->nr_zones - 1; My system has 4 zones: ZONE NAME SIZE MEM_MAP START_PADDR START_MAPNR 0 DMA 4080 f4a02200 10000 0 1 Normal 221694 f4a22000 1000000 4080 2 HighMem 790002 f50e5fc0 371fe000 225774 3 Movable 0 0 0 0 And indeed, "help -v" shows: [...] dump_free_pages: dump_free_pages_zones_v2() [...] ZONE_HIGHMEM: 3 I don't know yet how to fix this, but maybe somebody can push me in the right direction.

Dne Čt 12. ledna 2012 22:59:14 Dave Anderson napsal(a): > ----- Original Message ----- > > > Hi all, > > > > it seems crash shows incorrect HighMem stats with recent kernels. > > E.g.: > > > > crash> kmem -i > > [...] > > TOTAL HIGH 1821682 6.9 GB 93% of TOTAL MEM > > FREE HIGH 0 0 0% of TOTAL HIGH > > TOTAL LOW 132983 519.5 MB 6% of TOTAL MEM > > > > FREE LOW 1333474 5.1 GB 1002% of TOTAL LOW > > > > BTW note that total low is smaller than free low, which is obviously > > incorrect. I believe that this is somehow related to the Movable zone, > > because the code that counts free low pages checks for pages which > > belong to > > > > ZONE_HIGHMEM, which is initialized as: > > vt->ZONE_HIGHMEM = vt->nr_zones - 1; > > > > My system has 4 zones: > > > > ZONE NAME SIZE MEM_MAP START_PADDR START_MAPNR > > > > 0 DMA 4080 f4a02200 10000 0 > > 1 Normal 221694 f4a22000 1000000 4080 > > 2 HighMem 790002 f50e5fc0 371fe000 225774 > > 3 Movable 0 0 0 0 > > > > And indeed, "help -v" shows: > > [...] > > > > dump_free_pages: dump_free_pages_zones_v2() > > > > [...] > > > > ZONE_HIGHMEM: 3 > > > > I don't know yet how to fix this, but maybe somebody can push me in > > the right direction. > > Seems like you're already headed down the right path with the > vt->ZONE_HIGHMEM initialization, at least w/respect to the > FREE HIGH of "0". On the other hand, I'm looking at a recent > x86 kernel with the Movable zone, and it has normal > looking statistics w/respect to total-low/free-low. Interesting. What does "kmem -n" show? How is your ZONE_HIGHMEM reported by "help -v"?

Dne Čt 12. ledna 2012 22:38:38 Petr Tesarik napsal(a): > Hi all, > > it seems crash shows incorrect HighMem stats with recent kernels. > E.g.: > > crash> kmem -i > [...] > TOTAL HIGH 1821682 6.9 GB 93% of TOTAL MEM > FREE HIGH 0 0 0% of TOTAL HIGH > TOTAL LOW 132983 519.5 MB 6% of TOTAL MEM > FREE LOW 1333474 5.1 GB 1002% of TOTAL LOW > > BTW note that total low is smaller than free low, which is obviously > incorrect. I believe that this is somehow related to the Movable zone, > because the code that counts free low pages checks for pages which belong > to ZONE_HIGHMEM, which is initialized as: > > vt->ZONE_HIGHMEM = vt->nr_zones - 1; > > My system has 4 zones: > > ZONE NAME SIZE MEM_MAP START_PADDR START_MAPNR > 0 DMA 4080 f4a02200 10000 0 > 1 Normal 221694 f4a22000 1000000 4080 > 2 HighMem 790002 f50e5fc0 371fe000 225774 > 3 Movable 0 0 0 0 > > And indeed, "help -v" shows: > [...] > dump_free_pages: dump_free_pages_zones_v2() > [...] > ZONE_HIGHMEM: 3 > > I don't know yet how to fix this, but maybe somebody can push me in > the > right direction. Maybe we should check the names in zone_names[]. AFAICS this array has existed ever since the dawn of time, and the HIGHMEM_ZONE has always been called "HighMem".

That already covers a lot of cases, but not the Movable zone, if it exists. nr_free_highpages() contains: if (zone_movable_is_highmem()) pages += zone_page_state( &pgdat->node_zones[ZONE_MOVABLE], NR_FREE_PAGES); zone_movable_is_highmem() is defined as: static inline int zone_movable_is_highmem(void) { #if defined(CONFIG_HIGHMEM) && defined(CONFIG_ARCH_POPULATES_NODE_MAP) return movable_zone == ZONE_HIGHMEM; #else return 0; #endif } Now, I think we don't have to take the preprocessor directives into account, because movable_zone is automatically initialized to zero, and ZONE_HIGHMEM can never be numerically equal to zero, so the test would be always false in that case anyway.

We should only set vt->HIGHMEM_ZONE to something special (such as -1) if CONFIG_HIGHMEM is not defined, and there is no highmem zone, because otherwise we might hit a false positive for zone_movable_is_highmem. Comments? Should I make the idea into a patch?