>> Platforms supporting NVDIMMs are now required to provide persistence >> guarantees once pmem stores are accepted by the memory subsystem. > >Can you point us to a precise definition of what exactly constitutes >stores being "accepted by the memory subsystem"? Back when pcommit >was a thing (hah!), it was precisely documented in the SDM.
The Instruction Set Extensions document that describes PCOMMIT says it applies to stores that were accepted by the memory subsystem but were still potentially queued there. So Dan’s description was building on that same language. A concrete example might help: MOV to location X CLWB X SFENCE is the sequence for flushing a store to pmem. Before PCOMMIT was deprecated, software needed to understand if the platform had ADR and add: PCOMMIT ; the old way SFENCE ; the old way to the above sequence if ADR wasn’t present on the platform. Now that ADR is required for all persistent memory, as it always has been for NVDIMM-N devices available today, the simpler programming model already in use for NVDIMM-N is used for all pmem. >> is usually achieved by a platform-level feature known as ADR >> (Asynchronous DRAM Refresh) that flushes any memory subsystem write >> pending queues on power loss/shutdown. > >Blech. Does this mean that we need some NMI handler or similar that >does wbinvd? Will CPU caches automatically write themselves back on >power loss? Will we lose data if something goes wrong with an SMI >handler? No, ADR is a platform-level feature that flushes any buffers in the memory subsystem on power loss. No software is involved and the CPU caches are not touched by ADR. So any stores flushed by CLFLUSH, CLFLUSHOPT, CLWB, WBINVD, or non-temporal stores that go around the CPU caches can now be considered persistent once those operations are complete. >> The 'pcommit' instruction (which has not shipped on any product) is no >> longer needed and is deprecated. > >Does this mean it will never ship? Yes. The new instructions like CLWB and CLFLUSHOPT will be rolled into the SDM but PCOMMIT will be removed from the Extensions doc and not rolled into the SDM. -andy
