On Thu, Sep 03, 2009 at 10:27:53PM +0800, Wu Fei wrote: > This patch tries to shrink the size of the 64bit tlb handler and also fix > an vmalloc bug at the same time. > Forgot to say, it's only tested on 2.6.27, not the master, I can't get the environment to run the latest kernel currently. Thanks, Wufei. > By combining the swapper_pg_dir and module_pg_dir, several checks in tlb > handler, particularly build_get_pgd_vmalloc64, are not necessary. The reason > they can be combined is that, the effective virtual address of vmalloc returned > is at the bottom, and of module_alloc returned is at the top. > > In the normal case of 4KB page size: > VMALLOC_START, VMALLOC_END 0xc0000000 00000000 - 0xc0000100 00000000 > MODULE_START, MODULE_END 0xffffffff c0000000 - 0xffffffff +xxxxxxx > Change it to: > VMALLOC_START, VMALLOC_END 0xc0000000 00000000 - 0xc00000ff 00000000 > MODULE_START, MODULE_END 0xffffffff c0000000 - 0xffffffff +xxxxxxx > We use the least 40 bits to traverse the page table, the change makes it still > one-to-one mapping without more checking. "+" is in the range of [c,d,e,f], > so there even are big holes bewteen them. > > With this patch, the tlb refill handler only contains about 28 instructions, > instead of the original 38. > > > And this patch also fix a bug in vmalloc, which happens when its returned > address is not covered by the first pgd. e.g. if we do two vmallocs, the first > returned address is 0xc0000000 00000000, and the 2nd is 0xc0000000 40000000, > > vmalloc -> __vmalloc_node -> __vmalloc_area_node -> __vmalloc_area_node > -> map_vm_area -> pgd_offset_k > > pgd_offset_k doesn't use the address to index the pgd, just return the first one: > > #define pgd_offset_k(address) \ > ((address) >= MODULE_START ? module_pg_dir : pgd_offset(&init_mm, 0UL)) > > This is wrong, then the 2 addresses are mapped to the same pte. This bug doesn't > happen because even in the 4KB page case, one pgd can cover 1GB size, and it looks > like the system won't vmalloc so much area.
