Excerpts from Christoph Hellwig's message of April 14, 2020 5:23 pm: > On Mon, Apr 13, 2020 at 10:53:03PM +1000, Nicholas Piggin wrote: >> For platforms that define HAVE_ARCH_HUGE_VMAP and support PMD vmap mappings, >> have vmalloc attempt to allocate PMD-sized pages first, before falling back >> to small pages. Allocations which use something other than PAGE_KERNEL >> protections are not permitted to use huge pages yet, not all callers expect >> this (e.g., module allocations vs strict module rwx). >> >> This gives a 6x reduction in dTLB misses for a `git diff` (of linux), from >> 45600 to 6500 and a 2.2% reduction in cycles on a 2-node POWER9. >> >> This can result in more internal fragmentation and memory overhead for a >> given allocation. It can also cause greater NUMA unbalance on hashdist >> allocations. >> >> There may be other callers that expect small pages under vmalloc but use >> PAGE_KERNEL, I'm not sure if it's feasible to catch them all. An >> alternative would be a new function or flag which enables large mappings, >> and use that in callers. > > Why do we even use vmalloc in this case rather than just doing a huge > page allocation? Which case? Usually the answer would be because you don't want to use contiguous physical memory and/or you don't want to use the linear mapping. > What callers are you intersted in? The dentry and inode caches for this test, obviously. Lots of other things could possibly benefit though, other system hashes like networking, but lot of other vmalloc callers that might benefit right away, some others could use some work to batch up allocation sizes to benefit. Thanks, Nick
