Messages in this thread | | | Date | Wed, 18 Mar 2009 15:35:08 +0000 | From | Mel Gorman <> | Subject | Re: [PATCH 24/27] Convert gfp_zone() to use a table of precalculated values |
| |
On Wed, Mar 18, 2009 at 10:15:26AM -0400, Christoph Lameter wrote: > On Wed, 18 Mar 2009, Mel Gorman wrote: > > > On Mon, Mar 16, 2009 at 03:12:50PM -0400, Christoph Lameter wrote: > > > On Mon, 16 Mar 2009, Mel Gorman wrote: > > > > > > > +int gfp_zone_table[GFP_ZONEMASK] __read_mostly; > > > > > > The gfp_zone_table is compile time determinable. There is no need to > > > calculate it. > > > > > > > The cost of calculating it is negligible and the code is then freed later > > in boot. Does having a const table make any difference? > > Should it not enable the compiler to determine the value at > compile time and therefore make things like gfp_zone(constant) a > constant? >
Yeah, you're right. I didn't think it would but a test program showed that code accessing const fields like this are calculated at compile time.
> > > const int gfp_zone_table[GFP_ZONEMASK] = { > > > ZONE_NORMAL, /* 00 No flags set */ > > > ZONE_DMA, /* 01 Only GFP_DMA set */ > > > ZONE_HIGHMEM, /* 02 Only GFP_HIGHMEM set */ > > > ZONE_DMA, /* 03 GFP_HIGHMEM and GFP_DMA set */ > > > ZONE_DMA32, /* 04 Only GFP_DMA32 set */ > > > ZONE_DMA, /* 05 GFP_DMA and GFP_DMA32 set */ > > > ZONE_DMA32, /* 06 GFP_DMA32 and GFP_HIGHMEM set */ > > > ZONE_DMA, /* 07 GFP_DMA, GFP_DMA32 and GFP_DMA32 set */ > > > ZONE_MOVABLE, /* 08 Only ZONE_MOVABLE set */ > > > ZONE_DMA, /* 09 MOVABLE + DMA */ > > > ZONE_MOVABLE, /* 0A MOVABLE + HIGHMEM */ > > > ZONE_DMA, /* 0B MOVABLE + DMA + HIGHMEM */ > > > ZONE_DMA32, /* 0C MOVABLE + DMA32 */ > > > ZONE_DMA, /* 0D MOVABLE + DMA + DMA32 */ > > > ZONE_DMA32, /* 0E MOVABLE + DMA32 + HIGHMEM */ > > > ZONE_DMA /* 0F MOVABLE + DMA32 + HIGHMEM + DMA > > > }; > > > > > > Hmmmm... Guess one would need to add some #ifdeffery here to setup > > > ZONE_NORMAL in cases there is no DMA, DMA32 and HIGHMEM. > > > > > > > Indeed, as I said, this is somewhat error prone which is why the patch > > calculates the table at run-time instead of compile-time trickery. > > One would need to define some macros to make it simpler I guess > > Write something like > > #ifdef CONFIG_ZONE_DMA > #define TZONE_DMA ZONE_DMA > #else > #define TZONE_DMA ZONE_NORMAL > #endif > > for each configurable item. Then just add the T to the above table. >
If you don't mind, I'd like to postpone writing such a patch until a second or third pass at improving the allocator. I don't think I'll have the time in the short-term to put together a const-initialised-table patch that will definitily be correct.
Alternatively, I can drop this patch entirely from the set.
-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
| |