On 05/07, Eric W. Biederman wrote: > > Oleg Nesterov <o...@redhat.com> writes: > > > before your patch get_mem_cgroup_from_mm() looks at mm->owner == current > > (in this case) and mem_cgroup_from_task() should return the correct memcg > > even if execing task migrates after bprm_mm_init(). At least in the common > > case when the old mm is not shared. > > > > After your patch the memory allocations in copy_strings() won't be accounted > > correctly, bprm->mm->memcg is wrong if this task migrates. And iiuc your > > recent > > "[PATCH 2/2] memcg: Close the race between migration and installing > > bprm->mm as mm" > > doesn't fix the problem. > > > > No? > > The patch does solve the issue. There should be nothing a userspace > process can observe that should tell it where in the middle of exec > such a migration happend so placing the migration at what from the > kernel's perspective might be technically later should not be a problem. > > If it is a problem the issue is that there is a way to observe the > difference.
So. The task migrates from some MEMCG right after bprm_mm_init(). copy_strings() triggers OOM in MEMCG. This is quite possible, it can use a lot of memory and that is why we have acct_arg_size() to make these allocations visible to oom killer. task_in_mem_cgroup(MEMCG) returns false and oom killer has to kill another innocent process in MEMCG. Does this look like a way to observe the difference? > > Perhaps we can change get_mem_cgroup_from_mm() to use > > mem_cgroup_from_css(current, memory_cgrp_id) if mm->memcg == NULL? > > Please God no. Having any unnecessary special case is just going to > confuse people and cause bugs. To me the unnecessary special case is the new_mm->memcg which is used for accounting but doesn't follow migration till exec_mmap(). But I won't argue. Oleg.
