On OpenRISC I noticed our jump_label implementation based on arm64 uses kick_all_cpus_sync() to force remote cores to pick up static_key updates. The call to kick_all_cpus_sync() on OpenRISC does run IPIs on remote cores, but it does not have any guarantees about instruction cache invalidation.
The jump_label code seems to work fine and passes tests, but I had suspicions that this was incorrect. My suspicion was correct, and I was able to reproduce issues where static_keys on remote cores were left stale after jump_label updates using a custom test [0]. This series fixes the issues by triggering remote icache invalidation after jump_label text patching is done. [0] https://github.com/stffrdhrn/or1k-utils/tree/master/tests/smp_static_key_test Stafford Horne (3): openrisc: Cache invalidation cleanup openrisc: Add full instruction cache invalidate functions openrisc: Fix jump_label smp syncing arch/openrisc/include/asm/cacheflush.h | 4 ++++ arch/openrisc/kernel/head.S | 10 ---------- arch/openrisc/kernel/jump_label.c | 2 +- arch/openrisc/kernel/patching.c | 3 +++ arch/openrisc/kernel/smp.c | 21 +++++++++++++++++++++ arch/openrisc/mm/cache.c | 16 ++++++++++++++++ 6 files changed, 45 insertions(+), 11 deletions(-) -- 2.53.0
