On Thu 30-12-21 11:06:14, Shakeel Butt wrote: > Hi Michal, > > On Mon, Dec 27, 2021 at 2:48 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > [...] > > > atomic_long_add(area->nr_pages, &nr_vmalloc_pages); > > > + mod_memcg_page_state(area->pages[0], MEMCG_VMALLOC, area->nr_pages); > > > > > > /* > > > * If not enough pages were obtained to accomplish an > > > > Is it safe to assume that the whole area is always charged to the same > > memcg? I am not really deeply familiar with vmalloc internals but is it > > possible that an area could get resized/partially reused with a > > different charging context? > > From what I understand, vmalloc areas are not resized or partially > reused at the moment. There is some ongoing discussion on caching it > but caching would also require updating the accounting part as well. OK. > Regarding the whole area charged to the same memcg, the only way it > may get charged to different memcgs is if the process in which the > allocations are happening is migrated to a different memcg. We can > resolve this by traversing the pages in area->pages array (and use > lruvec based stats instead). I haven't even thought of a task migration. I expect that this is not a very likely scenario but it would lead to weird numbers and I guess we would like to prevent from that. A loop over all pages in the area and accounting them each separately should be good enough to cover that as well as the existing problem that has already been observed by syzbot. > I did contemplate on making this a lruvec stat but decided to start > simple and if we ever need per-node stat then we can easily move to > lruvec based stats. Let me know what you think. I am not really sure here. For now I would go with page by page stats gathering. Thanks! -- Michal Hocko SUSE Labs
