On Mon, Nov 25, 2024 at 05:24:05AM +0000, Ruan Bonan wrote: > From the discussion, it appears that the root cause might involve > specific printk or BPF operations in the given context. To clarify and > possibly avoid similar issues in the future, are there guidelines or > best practices for writing BPF programs/hooks that interact with > tracepoints, especially those related to scheduler events, to prevent > such deadlocks? The general guideline and recommendation for all tracepoints is to be wait-free. Typically all tracer code should be. Now, BPF (users) (ab)uses tracepoints to do all sorts and takes certain liberties with them, but it is very much at the discretion of the BPF user. Slightly relaxed guideline would perhaps be to consider the context of the tracepoint, notably one of: NMI, IRQ, SoftIRQ or Task context -- and to not exceed the bounds of the given context. More specifically, when the tracepoint is inside critical sections of any sort (as is the case here) then it very much is on the BPF user to not cause inversions. At this point there really is no substitute for knowing what you're doing. Knowledge is key. In short; tracepoints should be wait-free, if you know what you're doing you can perhaps get away with a little more.
