On 2024-09-17 16:33, Boqun Feng wrote: [...]
The synchronization between readers and updaters is built around "hazard pointer slots": a slot readers can use to store a pointer value. Reader side protection: 1. Read the value of a pointer to the target data element. 2. Store it to a hazard pointer slot. 3. Enforce full ordering (e.g. smp_mb()). 4. Re-read the original pointer, reset the slot and retry if the value changed. 5. Otherwise, the continued existence of the target data element is guaranteed. Updater side check: 1. Unpublish the target data element (e.g. setting the pointer value to NULL). 2. Enforce full ordering. 3. Read the value from a hazard pointer slot. 4. If the value doesn't match the target data element, then this slot doesn't represent a reference to it. 5. Otherwise, updater needs to re-check (step 3).
Cool! I look forward to see where this is meant to be used. I would expect it to be a useful tool to speed up reference counting of things like the mm_struct and for TLB flush IPIs. On a related note, with a userspace port in mind, the membarrier(2) syscall can be useful to turn the smp_mb() in (3) from the reader side into a simple compiler barrier, assuming (2) from the updater is using membarrier. If IPIs are acceptable (or already required) for some kernel use-cases, this means a similar asymmetric fence scheme could be used to speed up readers. Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. https://www.efficios.com
