| blob | b91443f577dcff68aa8961a73d77141ba8a73cd4 |
1 BTT - Block Translation Table
2 =============================
5 1. Introduction
6 ---------------
8 Persistent memory based storage is able to perform IO at byte (or more
9 accurately, cache line) granularity. However, we often want to expose such
10 storage as traditional block devices. The block drivers for persistent memory
11 will do exactly this. However, they do not provide any atomicity guarantees.
12 Traditional SSDs typically provide protection against torn sectors in hardware,
13 using stored energy in capacitors to complete in-flight block writes, or perhaps
14 in firmware. We don't have this luxury with persistent memory - if a write is in
15 progress, and we experience a power failure, the block will contain a mix of old
16 and new data. Applications may not be prepared to handle such a scenario.
18 The Block Translation Table (BTT) provides atomic sector update semantics for
19 persistent memory devices, so that applications that rely on sector writes not
20 being torn can continue to do so. The BTT manifests itself as a stacked block
21 device, and reserves a portion of the underlying storage for its metadata. At
22 the heart of it, is an indirection table that re-maps all the blocks on the
23 volume. It can be thought of as an extremely simple file system that only
24 provides atomic sector updates.
27 2. Static Layout
28 ----------------
30 The underlying storage on which a BTT can be laid out is not limited in any way.
31 The BTT, however, splits the available space into chunks of up to 512 GiB,
32 called "Arenas".
34 Each arena follows the same layout for its metadata, and all references in an
35 arena are internal to it (with the exception of one field that points to the
36 next arena). The following depicts the "On-disk" metadata layout:
39 Backing Store +-------> Arena
40 +---------------+ | +------------------+
41 | | | | Arena info block |
42 | Arena 0 +---+ | 4K |
43 | 512G | +------------------+
44 | | | |
45 +---------------+ | |
46 | | | |
47 | Arena 1 | | Data Blocks |
48 | 512G | | |
49 | | | |
50 +---------------+ | |
51 | . | | |
52 | . | | |
53 | . | | |
54 | | | |
55 | | | |
56 +---------------+ +------------------+
57 | |
58 | BTT Map |
59 | |
60 | |
61 +------------------+
62 | |
63 | BTT Flog |
64 | |
65 +------------------+
66 | Info block copy |
67 | 4K |
68 +------------------+
71 3. Theory of Operation
72 ----------------------
75 a. The BTT Map
76 --------------
78 The map is a simple lookup/indirection table that maps an LBA to an internal
79 block. Each map entry is 32 bits. The two most significant bits are special
80 flags, and the remaining form the internal block number.
82 Bit Description
83 31 - 30 : Error and Zero flags - Used in the following way:
84 Bit Description
85 31 30
86 -----------------------------------------------------------------------
87 00 Initial state. Reads return zeroes; Premap = Postmap
88 01 Zero state: Reads return zeroes
89 10 Error state: Reads fail; Writes clear 'E' bit
90 11 Normal Block – has valid postmap
93 29 - 0 : Mappings to internal 'postmap' blocks
96 Some of the terminology that will be subsequently used:
98 External LBA : LBA as made visible to upper layers.
99 ABA : Arena Block Address - Block offset/number within an arena
100 Premap ABA : The block offset into an arena, which was decided upon by range
101 checking the External LBA
102 Postmap ABA : The block number in the "Data Blocks" area obtained after
103 indirection from the map
104 nfree : The number of free blocks that are maintained at any given time.
105 This is the number of concurrent writes that can happen to the
106 arena.
109 For example, after adding a BTT, we surface a disk of 1024G. We get a read for
110 the external LBA at 768G. This falls into the second arena, and of the 512G
111 worth of blocks that this arena contributes, this block is at 256G. Thus, the
112 premap ABA is 256G. We now refer to the map, and find out the mapping for block
113 'X' (256G) points to block 'Y', say '64'. Thus the postmap ABA is 64.
116 b. The BTT Flog
117 ---------------
119 The BTT provides sector atomicity by making every write an "allocating write",
120 i.e. Every write goes to a "free" block. A running list of free blocks is
121 maintained in the form of the BTT flog. 'Flog' is a combination of the words
122 "free list" and "log". The flog contains 'nfree' entries, and an entry contains:
124 lba : The premap ABA that is being written to
125 old_map : The old postmap ABA - after 'this' write completes, this will be a
126 free block.
127 new_map : The new postmap ABA. The map will up updated to reflect this
128 lba->postmap_aba mapping, but we log it here in case we have to
129 recover.
130 seq : Sequence number to mark which of the 2 sections of this flog entry is
131 valid/newest. It cycles between 01->10->11->01 (binary) under normal
132 operation, with 00 indicating an uninitialized state.
133 lba' : alternate lba entry
134 old_map': alternate old postmap entry
135 new_map': alternate new postmap entry
136 seq' : alternate sequence number.
138 Each of the above fields is 32-bit, making one entry 32 bytes. Entries are also
139 padded to 64 bytes to avoid cache line sharing or aliasing. Flog updates are
140 done such that for any entry being written, it:
141 a. overwrites the 'old' section in the entry based on sequence numbers
142 b. writes the 'new' section such that the sequence number is written last.
145 c. The concept of lanes
146 -----------------------
148 While 'nfree' describes the number of concurrent IOs an arena can process
149 concurrently, 'nlanes' is the number of IOs the BTT device as a whole can
150 process.
151 nlanes = min(nfree, num_cpus)
152 A lane number is obtained at the start of any IO, and is used for indexing into
153 all the on-disk and in-memory data structures for the duration of the IO. If
154 there are more CPUs than the max number of available lanes, than lanes are
155 protected by spinlocks.
158 d. In-memory data structure: Read Tracking Table (RTT)
159 ------------------------------------------------------
161 Consider a case where we have two threads, one doing reads and the other,
162 writes. We can hit a condition where the writer thread grabs a free block to do
163 a new IO, but the (slow) reader thread is still reading from it. In other words,
164 the reader consulted a map entry, and started reading the corresponding block. A
165 writer started writing to the same external LBA, and finished the write updating
166 the map for that external LBA to point to its new postmap ABA. At this point the
167 internal, postmap block that the reader is (still) reading has been inserted
168 into the list of free blocks. If another write comes in for the same LBA, it can
169 grab this free block, and start writing to it, causing the reader to read
170 incorrect data. To prevent this, we introduce the RTT.
172 The RTT is a simple, per arena table with 'nfree' entries. Every reader inserts
173 into rtt[lane_number], the postmap ABA it is reading, and clears it after the
174 read is complete. Every writer thread, after grabbing a free block, checks the
175 RTT for its presence. If the postmap free block is in the RTT, it waits till the
176 reader clears the RTT entry, and only then starts writing to it.
179 e. In-memory data structure: map locks
180 --------------------------------------
182 Consider a case where two writer threads are writing to the same LBA. There can
183 be a race in the following sequence of steps:
185 free[lane] = map[premap_aba]
186 map[premap_aba] = postmap_aba
188 Both threads can update their respective free[lane] with the same old, freed
189 postmap_aba. This has made the layout inconsistent by losing a free entry, and
190 at the same time, duplicating another free entry for two lanes.
192 To solve this, we could have a single map lock (per arena) that has to be taken
193 before performing the above sequence, but we feel that could be too contentious.
194 Instead we use an array of (nfree) map_locks that is indexed by
195 (premap_aba modulo nfree).
198 f. Reconstruction from the Flog
199 -------------------------------
201 On startup, we analyze the BTT flog to create our list of free blocks. We walk
202 through all the entries, and for each lane, of the set of two possible
203 'sections', we always look at the most recent one only (based on the sequence
204 number). The reconstruction rules/steps are simple:
205 - Read map[log_entry.lba].
206 - If log_entry.new matches the map entry, then log_entry.old is free.
207 - If log_entry.new does not match the map entry, then log_entry.new is free.
208 (This case can only be caused by power-fails/unsafe shutdowns)
211 g. Summarizing - Read and Write flows
212 -------------------------------------
214 Read:
216 1. Convert external LBA to arena number + pre-map ABA
217 2. Get a lane (and take lane_lock)
218 3. Read map to get the entry for this pre-map ABA
219 4. Enter post-map ABA into RTT[lane]
220 5. If TRIM flag set in map, return zeroes, and end IO (go to step 8)
221 6. If ERROR flag set in map, end IO with EIO (go to step 8)
222 7. Read data from this block
223 8. Remove post-map ABA entry from RTT[lane]
224 9. Release lane (and lane_lock)
226 Write:
228 1. Convert external LBA to Arena number + pre-map ABA
229 2. Get a lane (and take lane_lock)
230 3. Use lane to index into in-memory free list and obtain a new block, next flog
231 index, next sequence number
232 4. Scan the RTT to check if free block is present, and spin/wait if it is.
233 5. Write data to this free block
234 6. Read map to get the existing post-map ABA entry for this pre-map ABA
235 7. Write flog entry: [premap_aba / old postmap_aba / new postmap_aba / seq_num]
236 8. Write new post-map ABA into map.
237 9. Write old post-map entry into the free list
238 10. Calculate next sequence number and write into the free list entry
239 11. Release lane (and lane_lock)
242 4. Error Handling
243 =================
245 An arena would be in an error state if any of the metadata is corrupted
246 irrecoverably, either due to a bug or a media error. The following conditions
247 indicate an error:
248 - Info block checksum does not match (and recovering from the copy also fails)
249 - All internal available blocks are not uniquely and entirely addressed by the
250 sum of mapped blocks and free blocks (from the BTT flog).
251 - Rebuilding free list from the flog reveals missing/duplicate/impossible
252 entries
253 - A map entry is out of bounds
255 If any of these error conditions are encountered, the arena is put into a read
256 only state using a flag in the info block.
259 5. In-kernel usage
260 ==================
262 Any block driver that supports byte granularity IO to the storage may register
263 with the BTT. It will have to provide the rw_bytes interface in its
264 block_device_operations struct:
266 int (*rw_bytes)(struct gendisk *, void *, size_t, off_t, int rw);
268 It may register with the BTT after it adds its own gendisk, using btt_init:
270 struct btt *btt_init(struct gendisk *disk, unsigned long long rawsize,
271 u32 lbasize, u8 uuid[], int maxlane);
273 note that maxlane is the maximum amount of concurrency the driver wishes to
274 allow the BTT to use.
276 The BTT 'disk' appears as a stacked block device that grabs the underlying block
277 device in the O_EXCL mode.
279 When the driver wishes to remove the backing disk, it should similarly call
280 btt_fini using the same struct btt* handle that was provided to it by btt_init.
282 void btt_fini(struct btt *btt);