On Mon, May 3, 2021 at 8:32 AM Vlastimil Babka <vbabka@xxxxxxx> wrote: > > On 5/3/21 4:20 PM, Waiman Long wrote: > > On 5/3/21 8:22 AM, Vlastimil Babka wrote: > >> On 5/2/21 8:07 PM, Waiman Long wrote: > >>> The obj_cgroup array (memcg_data) embedded in the page structure is > >>> allocated at the first instance an accounted memory allocation happens. > >>> With the right size object, it is possible that the allocated obj_cgroup > >>> array comes from the same slab that requires memory accounting. If this > >>> happens, the slab will never become empty again as there is at least one > >>> object left (the obj_cgroup array) in the slab. > >>> > >>> With instructmentation code added to detect this situation, I got 76 > >>> hits on the kmalloc-192 slab when booting up a test kernel on a VM. > >>> So this can really happen. > >>> > >>> To avoid the creation of these unfreeable slabs, a check is added to > >>> memcg_alloc_page_obj_cgroups() to detect that and double the size > >>> of the array in case it happens to make sure that it comes from a > >>> different kmemcache. > >>> > >>> This change, however, does not completely eliminate the presence > >>> of unfreeable slabs which can still happen if a circular obj_cgroup > >>> array dependency is formed. > >> Hm this looks like only a half fix then. > >> I'm afraid the proper fix is for kmemcg to create own set of caches for the > >> arrays. It would also solve the recursive kfree() issue. > > > > Right, this is a possible solution. However, the objcg pointers array should > > need that much memory. Creating its own set of kmemcaches may seem like an > > overkill. > > Well if we go that way, there might be additional benefits: > > depending of gfp flags, kmalloc() would allocate from: > > kmalloc-* caches that never have kmemcg objects, thus can be used for the objcg > pointer arrays > kmalloc-cg-* caches that have only kmemcg unreclaimable objects > kmalloc-rcl-* and dma-kmalloc-* can stay with on-demand > memcg_alloc_page_obj_cgroups() > > This way we fully solve the issues that this patchset solves. In addition we get > better separation between kmemcg and !kmemcg thus save memory - no allocation of > the array as soon as a single object appears in slab. For "kmalloc-8" we now > have 8 bytes for the useful data and 8 bytes for the obj_cgroup pointer. > Yes this seems like a better approach.
