1Reference counting in pnfs: 2========================== 3 4The are several inter-related caches. We have layouts which can 5reference multiple devices, each of which can reference multiple data servers. 6Each data server can be referenced by multiple devices. Each device 7can be referenced by multiple layouts. To keep all of this straight, 8we need to reference count. 9 10 11struct pnfs_layout_hdr 12---------------------- 13The on-the-wire command LAYOUTGET corresponds to struct 14pnfs_layout_segment, usually referred to by the variable name lseg. 15Each nfs_inode may hold a pointer to a cache of these layout 16segments in nfsi->layout, of type struct pnfs_layout_hdr. 17 18We reference the header for the inode pointing to it, across each 19outstanding RPC call that references it (LAYOUTGET, LAYOUTRETURN, 20LAYOUTCOMMIT), and for each lseg held within. 21 22Each header is also (when non-empty) put on a list associated with 23struct nfs_client (cl_layouts). Being put on this list does not bump 24the reference count, as the layout is kept around by the lseg that 25keeps it in the list. 26 27deviceid_cache 28-------------- 29lsegs reference device ids, which are resolved per nfs_client and 30layout driver type. The device ids are held in a RCU cache (struct 31nfs4_deviceid_cache). The cache itself is referenced across each 32mount. The entries (struct nfs4_deviceid) themselves are held across 33the lifetime of each lseg referencing them. 34 35RCU is used because the deviceid is basically a write once, read many 36data structure. The hlist size of 32 buckets needs better 37justification, but seems reasonable given that we can have multiple 38deviceid's per filesystem, and multiple filesystems per nfs_client. 39 40The hash code is copied from the nfsd code base. A discussion of 41hashing and variations of this algorithm can be found at: 42http://groups.google.com/group/comp.lang.c/browse_thread/thread/9522965e2b8d3809 43 44data server cache 45----------------- 46file driver devices refer to data servers, which are kept in a module 47level cache. Its reference is held over the lifetime of the deviceid 48pointing to it. 49 50lseg 51---- 52lseg maintains an extra reference corresponding to the NFS_LSEG_VALID 53bit which holds it in the pnfs_layout_hdr's list. When the final lseg 54is removed from the pnfs_layout_hdr's list, the NFS_LAYOUT_DESTROYED 55bit is set, preventing any new lsegs from being added. 56 57layout drivers 58-------------- 59 60PNFS utilizes what is called layout drivers. The STD defines 4 basic 61layout types: "files", "objects", "blocks", and "flexfiles". For each 62of these types there is a layout-driver with a common function-vectors 63table which are called by the nfs-client pnfs-core to implement the 64different layout types. 65 66Files-layout-driver code is in: fs/nfs/filelayout/.. directory 67Objects-layout-deriver code is in: fs/nfs/objlayout/.. directory 68Blocks-layout-deriver code is in: fs/nfs/blocklayout/.. directory 69Flexfiles-layout-driver code is in: fs/nfs/flexfilelayout/.. directory 70 71objects-layout setup 72-------------------- 73 74As part of the full STD implementation the objlayoutdriver.ko needs, at times, 75to automatically login to yet undiscovered iscsi/osd devices. For this the 76driver makes up-calles to a user-mode script called *osd_login* 77 78The path_name of the script to use is by default: 79 /sbin/osd_login. 80This name can be overridden by the Kernel module parameter: 81 objlayoutdriver.osd_login_prog 82 83If Kernel does not find the osd_login_prog path it will zero it out 84and will not attempt farther logins. An admin can then write new value 85to the objlayoutdriver.osd_login_prog Kernel parameter to re-enable it. 86 87The /sbin/osd_login is part of the nfs-utils package, and should usually 88be installed on distributions that support this Kernel version. 89 90The API to the login script is as follows: 91 Usage: $0 -u <URI> -o <OSDNAME> -s <SYSTEMID> 92 Options: 93 -u target uri e.g. iscsi://<ip>:<port> 94 (allways exists) 95 (More protocols can be defined in the future. 96 The client does not interpret this string it is 97 passed unchanged as received from the Server) 98 -o osdname of the requested target OSD 99 (Might be empty) 100 (A string which denotes the OSD name, there is a 101 limit of 64 chars on this string) 102 -s systemid of the requested target OSD 103 (Might be empty) 104 (This string, if not empty is always an hex 105 representation of the 20 bytes osd_system_id) 106 107blocks-layout setup 108------------------- 109 110TODO: Document the setup needs of the blocks layout driver 111