On Thu, Jun 02, 2022 at 10:13:25PM -0700, Bart Van Assche wrote: > On 6/1/22 10:30, Jason Gunthorpe wrote: > > That is just the keys, not the graph arcs. lockdep records an arc > > between every key that establishes a locking relationship and > > minimizing the number of keys also de-duplicates those arcs. > > Do you agree that this overhead is acceptable since lockdep is only > used to debug kernel code? I don't know lockdep well enough to know - if you bloat the graph too much it may become excessively slow when searching for cycles/edges. Debug workloads with CM may trigger creating of 1000's of keys and we still want the testing to be workable. IMHO it was not designed to have one key per instance. Jason
