On 2/19/25 10:11 PM, Steven Rostedt wrote:
On Thu, 20 Feb 2025 11:40:36 +0900 Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx> wrote:Hmm, right. Since the blocked_on must be NULL before setting flag, if we can ensure the writing order so that blocked_flags is always updated before blocked_on, may it be safe? Or, (this may introduce more memory overhead) don't use union but use different blocked_on_mutex, blocked_on_rwsem, etc. Another idea is to make the owner offset same, like introducing struct common_lock { atomic_long_t owner; }; But the problem is that rt_mutex does not use atomic for storing the owner. (we can make it atomic using wrapper)Either that, or add to the task_struct: struct mutex *blocked_on_mutex; struct rwsem *blocked_on_rwsem; struct rtlock *blocked_on_rtlock; And just have each type assign to its own type. Then you only need to look at each one. But yeah, this adds even more bloat to task_struct. :-/
Another alternative is to encode the locking type into the lowest 2 bits of the address and combined them into a single atomic_long_t data item. Of course, we can only support 4 different types with this scheme.
Cheers, Longman
-- Steve
