1============================= 2BTT - Block Translation Table 3============================= 4 5 61. Introduction 7=============== 8 9Persistent memory based storage is able to perform IO at byte (or more 10accurately, cache line) granularity. However, we often want to expose such 11storage as traditional block devices. The block drivers for persistent memory 12will do exactly this. However, they do not provide any atomicity guarantees. 13Traditional SSDs typically provide protection against torn sectors in hardware, 14using stored energy in capacitors to complete in-flight block writes, or perhaps 15in firmware. We don't have this luxury with persistent memory - if a write is in 16progress, and we experience a power failure, the block will contain a mix of old 17and new data. Applications may not be prepared to handle such a scenario. 18 19The Block Translation Table (BTT) provides atomic sector update semantics for 20persistent memory devices, so that applications that rely on sector writes not 21being torn can continue to do so. The BTT manifests itself as a stacked block 22device, and reserves a portion of the underlying storage for its metadata. At 23the heart of it, is an indirection table that re-maps all the blocks on the 24volume. It can be thought of as an extremely simple file system that only 25provides atomic sector updates. 26 27 282. Static Layout 29================ 30 31The underlying storage on which a BTT can be laid out is not limited in any way. 32The BTT, however, splits the available space into chunks of up to 512 GiB, 33called "Arenas". 34 35Each arena follows the same layout for its metadata, and all references in an 36arena are internal to it (with the exception of one field that points to the 37next arena). The following depicts the "On-disk" metadata layout:: 38 39 40 Backing Store +-------> Arena 41 +---------------+ | +------------------+ 42 | | | | Arena info block | 43 | Arena 0 +---+ | 4K | 44 | 512G | +------------------+ 45 | | | | 46 +---------------+ | | 47 | | | | 48 | Arena 1 | | Data Blocks | 49 | 512G | | | 50 | | | | 51 +---------------+ | | 52 | . | | | 53 | . | | | 54 | . | | | 55 | | | | 56 | | | | 57 +---------------+ +------------------+ 58 | | 59 | BTT Map | 60 | | 61 | | 62 +------------------+ 63 | | 64 | BTT Flog | 65 | | 66 +------------------+ 67 | Info block copy | 68 | 4K | 69 +------------------+ 70 71 723. Theory of Operation 73====================== 74 75 76a. The BTT Map 77-------------- 78 79The map is a simple lookup/indirection table that maps an LBA to an internal 80block. Each map entry is 32 bits. The two most significant bits are special 81flags, and the remaining form the internal block number. 82 83======== ============================================================= 84Bit Description 85======== ============================================================= 8631 - 30 Error and Zero flags - Used in the following way:: 87 88 == == ==================================================== 89 31 30 Description 90 == == ==================================================== 91 0 0 Initial state. Reads return zeroes; Premap = Postmap 92 0 1 Zero state: Reads return zeroes 93 1 0 Error state: Reads fail; Writes clear 'E' bit 94 1 1 Normal Block – has valid postmap 95 == == ==================================================== 96 9729 - 0 Mappings to internal 'postmap' blocks 98======== ============================================================= 99 100 101Some of the terminology that will be subsequently used: 102 103============ ================================================================ 104External LBA LBA as made visible to upper layers. 105ABA Arena Block Address - Block offset/number within an arena 106Premap ABA The block offset into an arena, which was decided upon by range 107 checking the External LBA 108Postmap ABA The block number in the "Data Blocks" area obtained after 109 indirection from the map 110nfree The number of free blocks that are maintained at any given time. 111 This is the number of concurrent writes that can happen to the 112 arena. 113============ ================================================================ 114 115 116For example, after adding a BTT, we surface a disk of 1024G. We get a read for 117the external LBA at 768G. This falls into the second arena, and of the 512G 118worth of blocks that this arena contributes, this block is at 256G. Thus, the 119premap ABA is 256G. We now refer to the map, and find out the mapping for block 120'X' (256G) points to block 'Y', say '64'. Thus the postmap ABA is 64. 121 122 123b. The BTT Flog 124--------------- 125 126The BTT provides sector atomicity by making every write an "allocating write", 127i.e. Every write goes to a "free" block. A running list of free blocks is 128maintained in the form of the BTT flog. 'Flog' is a combination of the words 129"free list" and "log". The flog contains 'nfree' entries, and an entry contains: 130 131======== ===================================================================== 132lba The premap ABA that is being written to 133old_map The old postmap ABA - after 'this' write completes, this will be a 134 free block. 135new_map The new postmap ABA. The map will up updated to reflect this 136 lba->postmap_aba mapping, but we log it here in case we have to 137 recover. 138seq Sequence number to mark which of the 2 sections of this flog entry is 139 valid/newest. It cycles between 01->10->11->01 (binary) under normal 140 operation, with 00 indicating an uninitialized state. 141lba' alternate lba entry 142old_map' alternate old postmap entry 143new_map' alternate new postmap entry 144seq' alternate sequence number. 145======== ===================================================================== 146 147Each of the above fields is 32-bit, making one entry 32 bytes. Entries are also 148padded to 64 bytes to avoid cache line sharing or aliasing. Flog updates are 149done such that for any entry being written, it: 150a. overwrites the 'old' section in the entry based on sequence numbers 151b. writes the 'new' section such that the sequence number is written last. 152 153 154c. The concept of lanes 155----------------------- 156 157While 'nfree' describes the number of concurrent IOs an arena can process 158concurrently, 'nlanes' is the number of IOs the BTT device as a whole can 159process:: 160 161 nlanes = min(nfree, num_cpus) 162 163A lane number is obtained at the start of any IO, and is used for indexing into 164all the on-disk and in-memory data structures for the duration of the IO. If 165there are more CPUs than the max number of available lanes, than lanes are 166protected by spinlocks. 167 168 169d. In-memory data structure: Read Tracking Table (RTT) 170------------------------------------------------------ 171 172Consider a case where we have two threads, one doing reads and the other, 173writes. We can hit a condition where the writer thread grabs a free block to do 174a new IO, but the (slow) reader thread is still reading from it. In other words, 175the reader consulted a map entry, and started reading the corresponding block. A 176writer started writing to the same external LBA, and finished the write updating 177the map for that external LBA to point to its new postmap ABA. At this point the 178internal, postmap block that the reader is (still) reading has been inserted 179into the list of free blocks. If another write comes in for the same LBA, it can 180grab this free block, and start writing to it, causing the reader to read 181incorrect data. To prevent this, we introduce the RTT. 182 183The RTT is a simple, per arena table with 'nfree' entries. Every reader inserts 184into rtt[lane_number], the postmap ABA it is reading, and clears it after the 185read is complete. Every writer thread, after grabbing a free block, checks the 186RTT for its presence. If the postmap free block is in the RTT, it waits till the 187reader clears the RTT entry, and only then starts writing to it. 188 189 190e. In-memory data structure: map locks 191-------------------------------------- 192 193Consider a case where two writer threads are writing to the same LBA. There can 194be a race in the following sequence of steps:: 195 196 free[lane] = map[premap_aba] 197 map[premap_aba] = postmap_aba 198 199Both threads can update their respective free[lane] with the same old, freed 200postmap_aba. This has made the layout inconsistent by losing a free entry, and 201at the same time, duplicating another free entry for two lanes. 202 203To solve this, we could have a single map lock (per arena) that has to be taken 204before performing the above sequence, but we feel that could be too contentious. 205Instead we use an array of (nfree) map_locks that is indexed by 206(premap_aba modulo nfree). 207 208 209f. Reconstruction from the Flog 210------------------------------- 211 212On startup, we analyze the BTT flog to create our list of free blocks. We walk 213through all the entries, and for each lane, of the set of two possible 214'sections', we always look at the most recent one only (based on the sequence 215number). The reconstruction rules/steps are simple: 216 217- Read map[log_entry.lba]. 218- If log_entry.new matches the map entry, then log_entry.old is free. 219- If log_entry.new does not match the map entry, then log_entry.new is free. 220 (This case can only be caused by power-fails/unsafe shutdowns) 221 222 223g. Summarizing - Read and Write flows 224------------------------------------- 225 226Read: 227 2281. Convert external LBA to arena number + pre-map ABA 2292. Get a lane (and take lane_lock) 2303. Read map to get the entry for this pre-map ABA 2314. Enter post-map ABA into RTT[lane] 2325. If TRIM flag set in map, return zeroes, and end IO (go to step 8) 2336. If ERROR flag set in map, end IO with EIO (go to step 8) 2347. Read data from this block 2358. Remove post-map ABA entry from RTT[lane] 2369. Release lane (and lane_lock) 237 238Write: 239 2401. Convert external LBA to Arena number + pre-map ABA 2412. Get a lane (and take lane_lock) 2423. Use lane to index into in-memory free list and obtain a new block, next flog 243 index, next sequence number 2444. Scan the RTT to check if free block is present, and spin/wait if it is. 2455. Write data to this free block 2466. Read map to get the existing post-map ABA entry for this pre-map ABA 2477. Write flog entry: [premap_aba / old postmap_aba / new postmap_aba / seq_num] 2488. Write new post-map ABA into map. 2499. Write old post-map entry into the free list 25010. Calculate next sequence number and write into the free list entry 25111. Release lane (and lane_lock) 252 253 2544. Error Handling 255================= 256 257An arena would be in an error state if any of the metadata is corrupted 258irrecoverably, either due to a bug or a media error. The following conditions 259indicate an error: 260 261- Info block checksum does not match (and recovering from the copy also fails) 262- All internal available blocks are not uniquely and entirely addressed by the 263 sum of mapped blocks and free blocks (from the BTT flog). 264- Rebuilding free list from the flog reveals missing/duplicate/impossible 265 entries 266- A map entry is out of bounds 267 268If any of these error conditions are encountered, the arena is put into a read 269only state using a flag in the info block. 270 271 2725. Usage 273======== 274 275The BTT can be set up on any disk (namespace) exposed by the libnvdimm subsystem 276(pmem, or blk mode). The easiest way to set up such a namespace is using the 277'ndctl' utility [1]: 278 279For example, the ndctl command line to setup a btt with a 4k sector size is:: 280 281 ndctl create-namespace -f -e namespace0.0 -m sector -l 4k 282 283See ndctl create-namespace --help for more options. 284 285[1]: https://github.com/pmem/ndctl 286