On 10/24/2014 05:08 AM, Thomas Gleixner wrote: > On Sun, 12 Oct 2014, Qiaowei Ren wrote: >> + /* >> + * Go poke the address of the new bounds table in to the >> + * bounds directory entry out in userspace memory. Note: >> + * we may race with another CPU instantiating the same table. >> + * In that case the cmpxchg will see an unexpected >> + * 'actual_old_val'. >> + */ >> + ret = user_atomic_cmpxchg_inatomic(&actual_old_val, bd_entry, >> + expected_old_val, bt_addr); > > This is fully preemptible non-atomic context, right? > > So this wants a proper comment, why using > user_atomic_cmpxchg_inatomic() is the right thing to do here. Hey Thomas, How's this for a new comment? Does this cover the points you think need clarified? ==== The kernel has allocated a bounds table and needs to point the (userspace-allocated) directory to it. The directory entry is the *only* place we track that this table was allocated, so we essentially use it instead of an kernel data structure for synchronization. A copy_to_user()-style function would not give us the atomicity that we need. If two threads race to instantiate a table, the cmpxchg ensures we know which one lost the race and that the loser frees the table that they just allocated.
