I have traced this bug to a missing memory barrier in the powerpc IPI handling
code. io_uring uses task_work_add() to schedule I/O worker creation, which in
turn issues an IPI, and when precise timing conditions are met the inconsistent
state between the two CPU cores can lead to corruption of
I have traced this to a regression in the Linux kernel. The issue appears to
be a type of data race that is more likely to occur on ppc64el than on other
architectures, but is also likely to affect other architectures. The issue
remains in the latest GIT version of the Linux kernel, and I am
2 matches
Mail list logo
