Re: [PATCH -next v2] mm/kmemleak: annotate various data races obj->ptr

[Catalin Marinas]

On Thu, Feb 13, 2020 at 12:36:30PM -0500, Qian Cai wrote:
> The value of object->pointer could be accessed concurrently as noticed
> by KCSAN,
> 
>  write to 0xffffb0ea683a7d50 of 4 bytes by task 23575 on cpu 12:
>   do_raw_spin_lock+0x114/0x200
>   debug_spin_lock_after at kernel/locking/spinlock_debug.c:91
>   (inlined by) do_raw_spin_lock at kernel/locking/spinlock_debug.c:115
>   _raw_spin_lock+0x40/0x50
>   __handle_mm_fault+0xa9e/0xd00
>   handle_mm_fault+0xfc/0x2f0
>   do_page_fault+0x263/0x6f9
>   page_fault+0x34/0x40
> 
>  read to 0xffffb0ea683a7d50 of 4 bytes by task 839 on cpu 60:
>   crc32_le_base+0x67/0x350
>   crc32_le_base+0x67/0x350:
>   crc32_body at lib/crc32.c:106
>   (inlined by) crc32_le_generic at lib/crc32.c:179
>   (inlined by) crc32_le at lib/crc32.c:197
>   kmemleak_scan+0x528/0xd90
>   update_checksum at mm/kmemleak.c:1172
>   (inlined by) kmemleak_scan at mm/kmemleak.c:1497
>   kmemleak_scan_thread+0xcc/0xfa
>   kthread+0x1e0/0x200
>   ret_from_fork+0x27/0x50
> 
>  write to 0xffff939bf07b95b8 of 4 bytes by interrupt on cpu 119:
>   __free_object+0x884/0xcb0
>   __free_object at lib/debugobjects.c:359
>   __debug_check_no_obj_freed+0x19d/0x370
>   debug_check_no_obj_freed+0x41/0x4b
>   slab_free_freelist_hook+0xfb/0x1c0
>   kmem_cache_free+0x10c/0x3a0
>   free_object_rcu+0x1ca/0x260
>   rcu_core+0x677/0xcc0
>   rcu_core_si+0x17/0x20
>   __do_softirq+0xd9/0x57c
>   run_ksoftirqd+0x29/0x50
>   smpboot_thread_fn+0x222/0x3f0
>   kthread+0x1e0/0x200
>   ret_from_fork+0x27/0x50
> 
>  read to 0xffff939bf07b95b8 of 8 bytes by task 838 on cpu 109:
>   scan_block+0x69/0x190
>   scan_block at mm/kmemleak.c:1250
>   kmemleak_scan+0x249/0xd90
>   scan_large_block at mm/kmemleak.c:1309
>   (inlined by) kmemleak_scan at mm/kmemleak.c:1434
>   kmemleak_scan_thread+0xcc/0xfa
>   kthread+0x1e0/0x200
>   ret_from_fork+0x27/0x50
> 
> crc32() will dereference object->pointer. If a shattered value was
> returned due to a data race, it will be corrected in the next scan.
> scan_block() will dereference a range of addresses (e.g., percpu
> sections) to search for valid pointers. Even if a data race heppens, it
> will cause no issue because the code here does not care about the exact
> value of a non-pointer. Thus, mark them as intentional data races using
> the data_race() macro.
> 
> Signed-off-by: Qian Cai <cai@xxxxxx>

Acked-by: Catalin Marinas <catalin.marinas@xxxxxxx>