On Mon, Oct 18, 2021 at 11:22:46AM +0200, Vlastimil Babka wrote: > On 10/18/21 05:38, Rustam Kovhaev wrote: > > Let's prepend all allocations of (PAGE_SIZE - align_offset) and less > > with the size header. This way kmem_cache_alloc() memory can be freed > > with kfree() and the other way around, as long as they are less than > > (PAGE_SIZE - align_offset). > > This size limitation seems like an unnecessary gotcha. Couldn't we make > these large allocations in slob_alloc_node() (that use slob_new_pages() > directly) similar enough to large kmalloc() ones, so that kfree() can > recognize them and free properly? AFAICS it might mean just adding > __GFP_COMP to make sure there's a compound order stored, as these already > don't seem to set PageSlab. Thanks for the pointers, I'll send a new version. > > The main reason for this change is to simplify SLOB a little bit, make > > it a bit easier to debug whenever something goes wrong. > > I would say the main reason is to simplify the slab API and guarantee that > both kmem_cache_alloc() and kmalloc() can be freed by kfree(). > > We should also update the comments at top of slob.c to reflect the change. > And Documentation/core-api/memory-allocation.rst (the last paragraph). OK, thank you! > > meminfo right after the system boot, without the patch: > > Slab: 35500 kB > > > > the same, with the patch: > > Slab: 36396 kB > > 2.5% increase, hopefully acceptable. > > Thanks!
