Re: [PATCH] mm: kmem: fix a NULL pointer dereference in obj_stock_flush_required()

[Yosry Ahmed]

On Sun, Apr 23, 2023 at 11:51 PM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
> On Sun, 23 Apr 2023 at 04:26, Muchun Song <muchun.song@xxxxxxxxx> wrote:
> > > On Apr 22, 2023, at 01:40, Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote:
> > >
> > > KCSAN found an issue in obj_stock_flush_required():
> > > stock->cached_objcg can be reset between the check and dereference:
> > >
> > > ==================================================================
> > > BUG: KCSAN: data-race in drain_all_stock / drain_obj_stock
> > >
> > > write to 0xffff888237c2a2f8 of 8 bytes by task 19625 on cpu 0:
> > > drain_obj_stock+0x408/0x4e0 mm/memcontrol.c:3306
> > > refill_obj_stock+0x9c/0x1e0 mm/memcontrol.c:3340
> > > obj_cgroup_uncharge+0xe/0x10 mm/memcontrol.c:3408
> > > memcg_slab_free_hook mm/slab.h:587 [inline]
> > > __cache_free mm/slab.c:3373 [inline]
> > > __do_kmem_cache_free mm/slab.c:3577 [inline]
> > > kmem_cache_free+0x105/0x280 mm/slab.c:3602
> > > __d_free fs/dcache.c:298 [inline]
> > > dentry_free fs/dcache.c:375 [inline]
> > > __dentry_kill+0x422/0x4a0 fs/dcache.c:621
> > > dentry_kill+0x8d/0x1e0
> > > dput+0x118/0x1f0 fs/dcache.c:913
> > > __fput+0x3bf/0x570 fs/file_table.c:329
> > > ____fput+0x15/0x20 fs/file_table.c:349
> > > task_work_run+0x123/0x160 kernel/task_work.c:179
> > > resume_user_mode_work include/linux/resume_user_mode.h:49 [inline]
> > > exit_to_user_mode_loop+0xcf/0xe0 kernel/entry/common.c:171
> > > exit_to_user_mode_prepare+0x6a/0xa0 kernel/entry/common.c:203
> > > __syscall_exit_to_user_mode_work kernel/entry/common.c:285 [inline]
> > > syscall_exit_to_user_mode+0x26/0x140 kernel/entry/common.c:296
> > > do_syscall_64+0x4d/0xc0 arch/x86/entry/common.c:86
> > > entry_SYSCALL_64_after_hwframe+0x63/0xcd
> > >
> > > read to 0xffff888237c2a2f8 of 8 bytes by task 19632 on cpu 1:
> > > obj_stock_flush_required mm/memcontrol.c:3319 [inline]
> > > drain_all_stock+0x174/0x2a0 mm/memcontrol.c:2361
> > > try_charge_memcg+0x6d0/0xd10 mm/memcontrol.c:2703
> > > try_charge mm/memcontrol.c:2837 [inline]
> > > mem_cgroup_charge_skmem+0x51/0x140 mm/memcontrol.c:7290
> > > sock_reserve_memory+0xb1/0x390 net/core/sock.c:1025
> > > sk_setsockopt+0x800/0x1e70 net/core/sock.c:1525
> > > udp_lib_setsockopt+0x99/0x6c0 net/ipv4/udp.c:2692
> > > udp_setsockopt+0x73/0xa0 net/ipv4/udp.c:2817
> > > sock_common_setsockopt+0x61/0x70 net/core/sock.c:3668
> > > __sys_setsockopt+0x1c3/0x230 net/socket.c:2271
> > > __do_sys_setsockopt net/socket.c:2282 [inline]
> > > __se_sys_setsockopt net/socket.c:2279 [inline]
> > > __x64_sys_setsockopt+0x66/0x80 net/socket.c:2279
> > > do_syscall_x64 arch/x86/entry/common.c:50 [inline]
> > > do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80
> > > entry_SYSCALL_64_after_hwframe+0x63/0xcd
> > >
> > > value changed: 0xffff8881382d52c0 -> 0xffff888138893740
> > >
> > > Reported by Kernel Concurrency Sanitizer on:
> > > CPU: 1 PID: 19632 Comm: syz-executor.0 Not tainted 6.3.0-rc2-syzkaller-00387-g534293368afa #0
> > > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 03/02/2023
> > >
> > > Fix it by reading the cached_objcg with READ_ONCE().
> > >
> > > Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
> > > Reported-by: syzbot+774c29891415ab0fd29d@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Reported-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> > > Link: https://lore.kernel.org/linux-mm/CACT4Y+ZfucZhM60YPphWiCLJr6+SGFhT+jjm8k1P-a_8Kkxsjg@xxxxxxxxxxxxxx/T/#t
> > > Signed-off-by: Roman Gushchin <roman.gushchin@xxxxxxxxx>
> >
> > Acked-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> >
> > Thanks.
>
> This improves things, but strictly speaking the write side also needs
> WRITE_ONCE. Ordering is always a game of two. It's not possible to
> order things on one side, if the other side messes up the ordering.
>

It looks like most other accesses use memcg_stock.stock_lock for
synchronization. Based on the output of obj_stock_flush_required()
we call drain_local_stock(), which acquires that lock as well. Should
we refactor the code to extend the lock section to cover both
obj_stock_flush_required() and drain_local_stock()?

IIUC this may unify the synchronization handling and
READ_ONCE/WRITE_ONCE may no longer be needed. This should also avoid
any inaccuracies (e.g. unnecessary flushes) that may happen if the
cached objcg changes between obj_stock_flush_required() and
drain_local_stock().

Did I miss anything here?