+Marco On Tue, Oct 29, 2019 at 2:03 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Mon 28-10-19 17:54:05, Shakeel Butt wrote: > > Syzbot reported the following bug: > > > > BUG: KCSAN: data-race in mem_cgroup_select_victim_node / mem_cgroup_select_victim_node > > > > write to 0xffff88809fade9b0 of 4 bytes by task 8603 on cpu 0: > > mem_cgroup_select_victim_node+0xb5/0x3d0 mm/memcontrol.c:1686 > > try_to_free_mem_cgroup_pages+0x175/0x4c0 mm/vmscan.c:3376 > > reclaim_high.constprop.0+0xf7/0x140 mm/memcontrol.c:2349 > > mem_cgroup_handle_over_high+0x96/0x180 mm/memcontrol.c:2430 > > tracehook_notify_resume include/linux/tracehook.h:197 [inline] > > exit_to_usermode_loop+0x20c/0x2c0 arch/x86/entry/common.c:163 > > prepare_exit_to_usermode+0x180/0x1a0 arch/x86/entry/common.c:194 > > swapgs_restore_regs_and_return_to_usermode+0x0/0x40 > > > > read to 0xffff88809fade9b0 of 4 bytes by task 7290 on cpu 1: > > mem_cgroup_select_victim_node+0x92/0x3d0 mm/memcontrol.c:1675 > > try_to_free_mem_cgroup_pages+0x175/0x4c0 mm/vmscan.c:3376 > > reclaim_high.constprop.0+0xf7/0x140 mm/memcontrol.c:2349 > > mem_cgroup_handle_over_high+0x96/0x180 mm/memcontrol.c:2430 > > tracehook_notify_resume include/linux/tracehook.h:197 [inline] > > exit_to_usermode_loop+0x20c/0x2c0 arch/x86/entry/common.c:163 > > prepare_exit_to_usermode+0x180/0x1a0 arch/x86/entry/common.c:194 > > swapgs_restore_regs_and_return_to_usermode+0x0/0x40 > > > > mem_cgroup_select_victim_node() can be called concurrently which reads > > and modifies memcg->last_scanned_node without any synchrnonization. So, > > read and modify memcg->last_scanned_node with READ_ONCE()/WRITE_ONCE() > > to stop potential reordering. > > I am sorry but I do not understand the problem and the fix. Why does the > race happen and why does _ONCE fixes it? There is still no > synchronization. Do you want to prevent from memcg->last_scanned_node > reloading? > The problem is memcg->last_scanned_node can read and modified concurrently. Though to me it seems like a tolerable race and not worth to add an explicit lock. My aim was to make KCSAN happy here to look elsewhere for the concurrency bugs. However I see that it might complain next on memcg->scan_nodes. Now taking a step back, I am questioning the whole motivation behind mem_cgroup_select_victim_node(). Since we pass ZONELIST_FALLBACK zonelist to the reclaimer, the shrink_node will be called for all potential nodes. Also we don't short circuit the traversal of shrink_node for all nodes on nr_reclaimed and we scan (size_on_node >> priority) for all nodes, I don't see the reason behind having round robin order of node traversal. I am thinking of removing the whole mem_cgroup_select_victim_node() heuristic. Please let me know if there are any objections. thanks, Shakeel
