Pull percpu fix from Tejun Heo: "This contains one patch to fix a race condition which can lead to percpu_ref using a percpu pointer which is corrupted with a set DEAD bit. The bug was introduced while separating out the ATOMIC mode flag from the DEAD flag. The fix is pretty straight forward. I just committed the patch to the percpu tree but am sending out the pull request early as I'll be on vacation for a week. The patch should be fairly safe and while the latency will be higher I'll be checking emails" * 'for-3.18-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: percpu-ref: fix DEAD flag contamination of percpu pointer
Unified · Split
| @@ -133,7 +133,13 @@ static inline bool __ref_is_percpu(struct percpu_ref *ref, · 132 unmodified lines | ||
| 133 | 133 | /* paired with smp_store_release() in percpu_ref_reinit() */ |
| 134 | 134 | smp_read_barrier_depends(); |
| 135 | 135 | |
| 136 | − if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC)) | |
| 136 | + /* | |
| 137 | + * Theoretically, the following could test just ATOMIC; however, | |
| 138 | + * then we'd have to mask off DEAD separately as DEAD may be | |
| 139 | + * visible without ATOMIC if we race with percpu_ref_kill(). DEAD | |
| 140 | + * implies ATOMIC anyway. Test them together. | |
| 141 | + */ | |
| 142 | + if (unlikely(percpu_ptr & __PERCPU_REF_ATOMIC_DEAD)) | |
| 137 | 143 | return false; |
| 138 | 144 | |
| 139 | 145 | *percpu_countp = (unsigned long __percpu *)percpu_ptr; |