On Tue, Jun 26, 2018 at 11:05 PM, Shakeel Butt <shakeelb@xxxxxxxxxx> wrote: > On Tue, Jun 26, 2018 at 12:03 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: >> >> On Mon, Jun 25, 2018 at 04:06:58PM -0700, Shakeel Butt wrote: >> > @@ -140,8 +141,9 @@ struct fanotify_event_info *fanotify_alloc_event(struct fsnotify_group *group, >> > struct inode *inode, u32 mask, >> > const struct path *path) >> > { >> > - struct fanotify_event_info *event; >> > + struct fanotify_event_info *event = NULL; >> > gfp_t gfp = GFP_KERNEL; >> > + struct mem_cgroup *old_memcg = NULL; >> > >> > /* >> > * For queues with unlimited length lost events are not expected and >> > @@ -151,19 +153,25 @@ struct fanotify_event_info *fanotify_alloc_event(struct fsnotify_group *group, >> > if (group->max_events == UINT_MAX) >> > gfp |= __GFP_NOFAIL; >> > >> > + /* Whoever is interested in the event, pays for the allocation. */ >> > + if (group->memcg) { >> > + gfp |= __GFP_ACCOUNT; >> > + old_memcg = memalloc_use_memcg(group->memcg); >> > + } >> >> group->memcg is only NULL when memcg is disabled or there is some >> offlining race. Can you make memalloc_use_memcg(NULL) mean that it >> should charge root_mem_cgroup instead of current->mm->memcg? That way >> we can make this site unconditional while retaining the behavior: >> >> gfp_t gfp = GFP_KERNEL | __GFP_ACCOUNT; >> >> memalloc_use_memcg(group->memcg); >> kmem_cache_alloc(..., gfp); >> out: >> memalloc_unuse_memcg(); >> >> (dropping old_memcg and the unuse parameter as per the other mail) >> > > group->memcg is only NULL when memcg is disabled (i.e. > get_mem_cgroup_from_mm() returns root_mem_cgroup for offlined > mm->memcg). Though group->memcg can point to an offlined memcg. > > If I understand you correctly this is what we want: > > 1. If group->memcg is NULL then __GFP_ACCOUNT is a noop i.e. memcg is disabled. > 2. If group->memcg is root_mem_cgroup, then __GFP_ACCOUNT again is a > kind of noop (charges to root_mem_cgroups are bypassed). > 3. If group->memcg is offlined memcg, then make __GFP_ACCOUNT noop by > returning root_mem_cgroup from get_mem_cgroup_from_current(). > 4. Else charge group->memcg. > > This seems reasonable. After your Ack and Amir's or Jan's answer to > the nesting query, I will resend the next version of this patch > series. > > In future if we find any use-cases of memalloc_use_memcg nesting then > we can make it work for nesting. > For the fsnotify use case memalloc_use_memcg() certainly doesn't need to nest, but I wonder, if that facility becomes popular among different subsystems, how exactly do you intend to monitor that it doesn't grow nested use cases? I would suggest that you at least leave a WARN_ON_ONCE if memalloc_use_memcg() is called and active_memcg is already set. Thanks, Amir.