On Tue, Jun 21, 2022 at 09:14:00AM +0800, Huang, Ying wrote: > Qian Cai <quic_qiancai@xxxxxxxxxxx> writes: > > > On Mon, Jun 20, 2022 at 10:20:07PM +0800, Muchun Song wrote: > >> The lock does not protect the read sides. So the write side should be > >> fixed by WRITTE_ONCE(). > > > > https://lwn.net/Articles/816854/ > > > > "Unmarked writes (aligned and up to word size) can be treated as if they had > > used WRITE_ONCE() by building with > > CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC=y (also selected by default). All right, CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC help us avoid KCSAN complaining. > > Experience has shown that compilers are much less likely to destructively > > optimize in-kernel writes than reads. Some developers might therefore > > choose to use READ_ONCE() but omit the corresponding WRITE_ONCE(). Other > > developers might prefer the documentation benefits and long-term peace of > > mind accruing from explicit use of WRITE_ONCE()..." > > Thanks for pointing me to this great article. So although not required > by KCSAN strictly, WRITE_ONCE() is still good for documentation, etc. > Just like we have done for swap_info_struct->highest_bit, etc. > +1 > Best Regards, > Huang, Ying >
