On Sun, May 06, 2018 at 09:32:15AM +0000, Huaisheng HS1 Ye wrote: > This idea is great, we can replace GFP_ZONE_TABLE and GFP_ZONE_BAD with it. > I have realized it preliminarily based on your code and tested it on a 2 sockets platform. Fortunately, we got a positive test result. Great! > I made some adjustments for __GFP_HIGHMEM, this flag is special than others, because the return result of gfp_zone has two possibilities, which depend on ___GFP_MOVABLE has been enabled or disabled. > When ___GFP_MOVABLE has been enabled, ZONE_MOVABLE shall be returned. When disabled, OPT_ZONE_HIGHMEM shall be used. > > #define __GFP_DMA ((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL) > #define __GFP_HIGHMEM ((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL) I'm not sure this is right ... Let me think about this a little. > #define __GFP_DMA32 ((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL) > #define __GFP_MOVABLE ((__force gfp_t)___GFP_MOVABLE) /* ZONE_MOVABLE allowed */ > #define GFP_ZONEMASK ((__force gfp_t)___GFP_ZONE_MASK | ___GFP_MOVABLE) > > The present situation is that, based on this change, the bits of flags, __GFP_DMA and __GFP_HIGHMEM and __GFP_DMA32, have been encoded. > That is totally different from existing code, you know in kernel scope, there are many drivers or subsystems use these flags directly to realize bit manipulations like this below, > swiotlb-xen.c (drivers\xen): flags &= ~(__GFP_DMA | __GFP_HIGHMEM); > extent_io.c (fs\btrfs): mask &= ~(__GFP_DMA32|__GFP_HIGHMEM); > > Because of these flags have been encoded, the above operations can cause problem. > I am trying to get a solution to resolve it. Any progress will be reported. These users probably want: flags &= GFP_RECLAIM_MASK;