On Wed, Mar 20, 2019 at 10:48:03PM +0100, Vlastimil Babka wrote: > On 3/20/2019 7:53 PM, Matthew Wilcox wrote: > > On Wed, Mar 20, 2019 at 09:48:47AM +0100, Vlastimil Babka wrote: > >> Natural alignment to size is rather well defined, no? Would anyone ever > >> assume a larger one, for what reason? > >> It's now where some make assumptions (even unknowingly) for natural > >> There are two 'odd' sizes 96 and 192, which will keep cacheline size > >> alignment, would anyone really expect more than 64 bytes? > > > > Presumably 96 will keep being aligned to 32 bytes, as aligning 96 to 64 > > just results in 128-byte allocations. > > Well, looks like that's what happens. This is with SLAB, but the alignment > calculations should be common: > > slabinfo - version: 2.1 > # name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab> : tunables <limit> <batchcount> <sharedfactor> : slabdata <active_slabs> <num_slabs> <sharedavail> > kmalloc-96 2611 4896 128 32 1 : tunables 120 60 8 : slabdata 153 153 0 > kmalloc-128 4798 5536 128 32 1 : tunables 120 60 8 : slabdata 173 173 0 Hmm. On my laptop, I see: kmalloc-96 28050 35364 96 42 1 : tunables 0 0 0 : slabdata 842 842 0 That'd take me from 842 * 4k pages to 1105 4k pages -- an extra megabyte of memory. This is running Debian's 4.19 kernel: # CONFIG_SLAB is not set CONFIG_SLUB=y # CONFIG_SLOB is not set CONFIG_SLAB_MERGE_DEFAULT=y CONFIG_SLAB_FREELIST_RANDOM=y CONFIG_SLAB_FREELIST_HARDENED=y CONFIG_SLUB_CPU_PARTIAL=y
