On 4/13/21 11:46 PM, Coly Li wrote: > Hi Jens, > > This is the 2nd wave of bcache patches for Linux v5.13. This series are > patches to use NVDIMM to store bcache journal, which is the first effort > to support NVDIMM for bcache [EXPERIMENTAL]. > > All concerns from Linux v5.12 merge window are fixed, especially the > data type defined in include/uapi/linux/bcache-nvm.h. And in this > series, all the lists defined in bcache-nvm.h uapi file are stored and > accessed directly on NVDIMM as memory objects. > > Intel developers Jianpeng Ma and Qiaowei Ren compose the initial code of > nvm-pages, the related patches are, > - bcache: initialize the nvm pages allocator > - bcache: initialization of the buddy > - bcache: bch_nvm_alloc_pages() of the buddy > - bcache: bch_nvm_free_pages() of the buddy > - bcache: get allocated pages from specific owner > All the code depends on Linux libnvdimm and dax drivers, the bcache nvm- > pages allocator can be treated as user of these two drivers. > > The nvm-pages allocator is a buddy-like allocator, which allocates size > in power-of-2 pages from the NVDIMM namespace. User space tool 'bcache' > has a new added '-M' option to format a NVDIMM namespace and register it > via sysfs interface as a bcache meta device. The nvm-pages kernel code > does a DAX mapping to map the whole namespace into system's memory > address range, and allocating the pages to requestion like typical buddy > allocator does. The major difference is nvm-pages allocator maintains > the pages allocated to each requester by a owner list which stored on > NVDIMM too. Owner list of different requester is tracked by a pre- > defined UUID, all the pages tracked in all owner lists are treated as > allocated busy pages and won't be initialized into buddy system after > the system reboot. > > I modify the bcache code to recognize the nvm meta device feature, > initialize journal on NVDIMM, and do journal I/Os on NVDIMM in the > following patches, > - bcache: add initial data structures for nvm pages > - bcache: use bucket index to set GC_MARK_METADATA for journal buckets > in bch_btree_gc_finish() > - bcache: add BCH_FEATURE_INCOMPAT_NVDIMM_META into incompat feature set > - bcache: initialize bcache journal for NVDIMM meta device > - bcache: support storing bcache journal into NVDIMM meta device > - bcache: read jset from NVDIMM pages for journal replay > - bcache: add sysfs interface register_nvdimm_meta to register NVDIMM > meta device > - bcache: use div_u64() in init_owner_info() > > The bcache journal code may request a block of power-of-2 size pages > from the nvm-pages allocator, normally it is a range of 256MB or 512MB > continuous pages range. During meta data journaling, the in-memory jsets > go into the calculated nvdimm pages location by kernel memcpy routine. > So the journaling I/Os won't go into block device (e.g. SSD) anymore, > the write and read for journal jsets happen on NVDIMM. > > The whole series is testing for a while and all addressed issues are > verified to be fixed. Now it is time to consider this series as an > initial code base of a commnity cooperation and have them in bcache > upstream for future development. > > Thanks in advance for taking this. Applied, with 13/13 folded in. -- Jens Axboe
