As you can see, if we're doing like the below.
sbi->compr_written_block += blocks;
Let's assume the initial value as 0.
<thread A> <thread B>
sbi->compr_written_block = 0;
sbi->compr_written_block = 0;
+blocks(3);
+ blocks(2);
sbi->compr_written_block = 3;
sbi->compr_written_block = 2;
Finally, we end up with 2, not 5.
As more threads are participating it, we might miss more counting.
2021년 3월 12일 (금) 오후 11:04, Greg KH <gre...@linuxfoundation.org>님이 작성:
>
> On Fri, Mar 12, 2021 at 10:56:13PM +0900, Daeho Jeong wrote:
> > Thanks for suggesting me sysfs_emit().
> >
> > For atomic values, actually, those are needed for writer part, not reader.
> >
> > +#define add_compr_block_stat(inode, blocks) \
> > + do { \
> > + struct f2fs_sb_info *sbi = F2FS_I_SB(inode); \
> > + int diff = F2FS_I(inode)->i_cluster_size - blocks; \
> > + atomic64_add(blocks, &sbi->compr_written_block); \
> > + atomic64_add(diff, &sbi->compr_saved_block); \
> > + } while (0)
> >
> > I needed a protection here, because they might be updated in the race
> > condition.
>
> Why? What are you trying to protect from "racing" here?
>
> thanks,
>
> greg k-h
