Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com> |
| 3 | * |
| 4 | * Uses a block device as cache for other block devices; optimized for SSDs. |
| 5 | * All allocation is done in buckets, which should match the erase block size |
| 6 | * of the device. |
| 7 | * |
| 8 | * Buckets containing cached data are kept on a heap sorted by priority; |
| 9 | * bucket priority is increased on cache hit, and periodically all the buckets |
| 10 | * on the heap have their priority scaled down. This currently is just used as |
| 11 | * an LRU but in the future should allow for more intelligent heuristics. |
| 12 | * |
| 13 | * Buckets have an 8 bit counter; freeing is accomplished by incrementing the |
| 14 | * counter. Garbage collection is used to remove stale pointers. |
| 15 | * |
| 16 | * Indexing is done via a btree; nodes are not necessarily fully sorted, rather |
| 17 | * as keys are inserted we only sort the pages that have not yet been written. |
| 18 | * When garbage collection is run, we resort the entire node. |
| 19 | * |
| 20 | * All configuration is done via sysfs; see Documentation/bcache.txt. |
| 21 | */ |
| 22 | |
| 23 | #include "bcache.h" |
| 24 | #include "btree.h" |
| 25 | #include "debug.h" |
Kent Overstreet | 65d4523 | 2013-12-20 17:22:05 -0800 | [diff] [blame] | 26 | #include "extents.h" |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 27 | |
| 28 | #include <linux/slab.h> |
| 29 | #include <linux/bitops.h> |
| 30 | #include <linux/hash.h> |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 31 | #include <linux/kthread.h> |
Geert Uytterhoeven | cd953ed | 2013-03-27 18:56:28 +0100 | [diff] [blame] | 32 | #include <linux/prefetch.h> |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 33 | #include <linux/random.h> |
| 34 | #include <linux/rcupdate.h> |
| 35 | #include <trace/events/bcache.h> |
| 36 | |
| 37 | /* |
| 38 | * Todo: |
| 39 | * register_bcache: Return errors out to userspace correctly |
| 40 | * |
| 41 | * Writeback: don't undirty key until after a cache flush |
| 42 | * |
| 43 | * Create an iterator for key pointers |
| 44 | * |
| 45 | * On btree write error, mark bucket such that it won't be freed from the cache |
| 46 | * |
| 47 | * Journalling: |
| 48 | * Check for bad keys in replay |
| 49 | * Propagate barriers |
| 50 | * Refcount journal entries in journal_replay |
| 51 | * |
| 52 | * Garbage collection: |
| 53 | * Finish incremental gc |
| 54 | * Gc should free old UUIDs, data for invalid UUIDs |
| 55 | * |
| 56 | * Provide a way to list backing device UUIDs we have data cached for, and |
| 57 | * probably how long it's been since we've seen them, and a way to invalidate |
| 58 | * dirty data for devices that will never be attached again |
| 59 | * |
| 60 | * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so |
| 61 | * that based on that and how much dirty data we have we can keep writeback |
| 62 | * from being starved |
| 63 | * |
| 64 | * Add a tracepoint or somesuch to watch for writeback starvation |
| 65 | * |
| 66 | * When btree depth > 1 and splitting an interior node, we have to make sure |
| 67 | * alloc_bucket() cannot fail. This should be true but is not completely |
| 68 | * obvious. |
| 69 | * |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 70 | * Plugging? |
| 71 | * |
| 72 | * If data write is less than hard sector size of ssd, round up offset in open |
| 73 | * bucket to the next whole sector |
| 74 | * |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 75 | * Superblock needs to be fleshed out for multiple cache devices |
| 76 | * |
| 77 | * Add a sysfs tunable for the number of writeback IOs in flight |
| 78 | * |
| 79 | * Add a sysfs tunable for the number of open data buckets |
| 80 | * |
| 81 | * IO tracking: Can we track when one process is doing io on behalf of another? |
| 82 | * IO tracking: Don't use just an average, weigh more recent stuff higher |
| 83 | * |
| 84 | * Test module load/unload |
| 85 | */ |
| 86 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 87 | #define MAX_NEED_GC 64 |
| 88 | #define MAX_SAVE_PRIO 72 |
| 89 | |
| 90 | #define PTR_DIRTY_BIT (((uint64_t) 1 << 36)) |
| 91 | |
| 92 | #define PTR_HASH(c, k) \ |
| 93 | (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0)) |
| 94 | |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 95 | #define insert_lock(s, b) ((b)->level <= (s)->lock) |
| 96 | |
| 97 | /* |
| 98 | * These macros are for recursing down the btree - they handle the details of |
| 99 | * locking and looking up nodes in the cache for you. They're best treated as |
| 100 | * mere syntax when reading code that uses them. |
| 101 | * |
| 102 | * op->lock determines whether we take a read or a write lock at a given depth. |
| 103 | * If you've got a read lock and find that you need a write lock (i.e. you're |
| 104 | * going to have to split), set op->lock and return -EINTR; btree_root() will |
| 105 | * call you again and you'll have the correct lock. |
| 106 | */ |
| 107 | |
| 108 | /** |
| 109 | * btree - recurse down the btree on a specified key |
| 110 | * @fn: function to call, which will be passed the child node |
| 111 | * @key: key to recurse on |
| 112 | * @b: parent btree node |
| 113 | * @op: pointer to struct btree_op |
| 114 | */ |
| 115 | #define btree(fn, key, b, op, ...) \ |
| 116 | ({ \ |
| 117 | int _r, l = (b)->level - 1; \ |
| 118 | bool _w = l <= (op)->lock; \ |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 119 | struct btree *_child = bch_btree_node_get((b)->c, op, key, l, \ |
| 120 | _w, b); \ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 121 | if (!IS_ERR(_child)) { \ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 122 | _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \ |
| 123 | rw_unlock(_w, _child); \ |
| 124 | } else \ |
| 125 | _r = PTR_ERR(_child); \ |
| 126 | _r; \ |
| 127 | }) |
| 128 | |
| 129 | /** |
| 130 | * btree_root - call a function on the root of the btree |
| 131 | * @fn: function to call, which will be passed the child node |
| 132 | * @c: cache set |
| 133 | * @op: pointer to struct btree_op |
| 134 | */ |
| 135 | #define btree_root(fn, c, op, ...) \ |
| 136 | ({ \ |
| 137 | int _r = -EINTR; \ |
| 138 | do { \ |
| 139 | struct btree *_b = (c)->root; \ |
| 140 | bool _w = insert_lock(op, _b); \ |
| 141 | rw_lock(_w, _b, _b->level); \ |
| 142 | if (_b == (c)->root && \ |
| 143 | _w == insert_lock(op, _b)) { \ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 144 | _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \ |
| 145 | } \ |
| 146 | rw_unlock(_w, _b); \ |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 147 | bch_cannibalize_unlock(c); \ |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 148 | if (_r == -EINTR) \ |
| 149 | schedule(); \ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 150 | } while (_r == -EINTR); \ |
| 151 | \ |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 152 | finish_wait(&(c)->btree_cache_wait, &(op)->wait); \ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 153 | _r; \ |
| 154 | }) |
| 155 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 156 | static inline struct bset *write_block(struct btree *b) |
| 157 | { |
| 158 | return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c); |
| 159 | } |
| 160 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 161 | static void bch_btree_init_next(struct btree *b) |
| 162 | { |
| 163 | /* If not a leaf node, always sort */ |
| 164 | if (b->level && b->keys.nsets) |
| 165 | bch_btree_sort(&b->keys, &b->c->sort); |
| 166 | else |
| 167 | bch_btree_sort_lazy(&b->keys, &b->c->sort); |
| 168 | |
| 169 | if (b->written < btree_blocks(b)) |
| 170 | bch_bset_init_next(&b->keys, write_block(b), |
| 171 | bset_magic(&b->c->sb)); |
| 172 | |
| 173 | } |
| 174 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 175 | /* Btree key manipulation */ |
| 176 | |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 177 | void bkey_put(struct cache_set *c, struct bkey *k) |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 178 | { |
| 179 | unsigned i; |
| 180 | |
| 181 | for (i = 0; i < KEY_PTRS(k); i++) |
| 182 | if (ptr_available(c, k, i)) |
| 183 | atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin); |
| 184 | } |
| 185 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 186 | /* Btree IO */ |
| 187 | |
| 188 | static uint64_t btree_csum_set(struct btree *b, struct bset *i) |
| 189 | { |
| 190 | uint64_t crc = b->key.ptr[0]; |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 191 | void *data = (void *) i + 8, *end = bset_bkey_last(i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 192 | |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 193 | crc = bch_crc64_update(crc, data, end - data); |
Kent Overstreet | c19ed23 | 2013-03-26 13:49:02 -0700 | [diff] [blame] | 194 | return crc ^ 0xffffffffffffffffULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 195 | } |
| 196 | |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 197 | void bch_btree_node_read_done(struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 198 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 199 | const char *err = "bad btree header"; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 200 | struct bset *i = btree_bset_first(b); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 201 | struct btree_iter *iter; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 202 | |
Kent Overstreet | bcf090e | 2014-05-19 08:57:55 -0700 | [diff] [blame] | 203 | iter = mempool_alloc(b->c->fill_iter, GFP_NOIO); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 204 | iter->size = b->c->sb.bucket_size / b->c->sb.block_size; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 205 | iter->used = 0; |
| 206 | |
Kent Overstreet | 280481d | 2013-10-24 16:36:03 -0700 | [diff] [blame] | 207 | #ifdef CONFIG_BCACHE_DEBUG |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 208 | iter->b = &b->keys; |
Kent Overstreet | 280481d | 2013-10-24 16:36:03 -0700 | [diff] [blame] | 209 | #endif |
| 210 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 211 | if (!i->seq) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 212 | goto err; |
| 213 | |
| 214 | for (; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 215 | b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 216 | i = write_block(b)) { |
| 217 | err = "unsupported bset version"; |
| 218 | if (i->version > BCACHE_BSET_VERSION) |
| 219 | goto err; |
| 220 | |
| 221 | err = "bad btree header"; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 222 | if (b->written + set_blocks(i, block_bytes(b->c)) > |
| 223 | btree_blocks(b)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 224 | goto err; |
| 225 | |
| 226 | err = "bad magic"; |
Kent Overstreet | 81ab419 | 2013-10-31 15:46:42 -0700 | [diff] [blame] | 227 | if (i->magic != bset_magic(&b->c->sb)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 228 | goto err; |
| 229 | |
| 230 | err = "bad checksum"; |
| 231 | switch (i->version) { |
| 232 | case 0: |
| 233 | if (i->csum != csum_set(i)) |
| 234 | goto err; |
| 235 | break; |
| 236 | case BCACHE_BSET_VERSION: |
| 237 | if (i->csum != btree_csum_set(b, i)) |
| 238 | goto err; |
| 239 | break; |
| 240 | } |
| 241 | |
| 242 | err = "empty set"; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 243 | if (i != b->keys.set[0].data && !i->keys) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 244 | goto err; |
| 245 | |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 246 | bch_btree_iter_push(iter, i->start, bset_bkey_last(i)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 247 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 248 | b->written += set_blocks(i, block_bytes(b->c)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 249 | } |
| 250 | |
| 251 | err = "corrupted btree"; |
| 252 | for (i = write_block(b); |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 253 | bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 254 | i = ((void *) i) + block_bytes(b->c)) |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 255 | if (i->seq == b->keys.set[0].data->seq) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 256 | goto err; |
| 257 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 258 | bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 259 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 260 | i = b->keys.set[0].data; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 261 | err = "short btree key"; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 262 | if (b->keys.set[0].size && |
| 263 | bkey_cmp(&b->key, &b->keys.set[0].end) < 0) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 264 | goto err; |
| 265 | |
| 266 | if (b->written < btree_blocks(b)) |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 267 | bch_bset_init_next(&b->keys, write_block(b), |
| 268 | bset_magic(&b->c->sb)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 269 | out: |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 270 | mempool_free(iter, b->c->fill_iter); |
| 271 | return; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 272 | err: |
| 273 | set_btree_node_io_error(b); |
Kent Overstreet | 88b9f8c | 2013-12-17 21:46:35 -0800 | [diff] [blame] | 274 | bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys", |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 275 | err, PTR_BUCKET_NR(b->c, &b->key, 0), |
Kent Overstreet | 88b9f8c | 2013-12-17 21:46:35 -0800 | [diff] [blame] | 276 | bset_block_offset(b, i), i->keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 277 | goto out; |
| 278 | } |
| 279 | |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 280 | static void btree_node_read_endio(struct bio *bio) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 281 | { |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 282 | struct closure *cl = bio->bi_private; |
| 283 | closure_put(cl); |
| 284 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 285 | |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 286 | static void bch_btree_node_read(struct btree *b) |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 287 | { |
| 288 | uint64_t start_time = local_clock(); |
| 289 | struct closure cl; |
| 290 | struct bio *bio; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 291 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 292 | trace_bcache_btree_read(b); |
| 293 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 294 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 295 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 296 | bio = bch_bbio_alloc(b->c); |
Kent Overstreet | 4f024f3 | 2013-10-11 15:44:27 -0700 | [diff] [blame] | 297 | bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9; |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 298 | bio->bi_end_io = btree_node_read_endio; |
| 299 | bio->bi_private = &cl; |
Mike Christie | ad0d9e7 | 2016-06-05 14:32:05 -0500 | [diff] [blame] | 300 | bio_set_op_attrs(bio, REQ_OP_READ, REQ_META|READ_SYNC); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 301 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 302 | bch_bio_map(bio, b->keys.set[0].data); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 303 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 304 | bch_submit_bbio(bio, b->c, &b->key, 0); |
| 305 | closure_sync(&cl); |
| 306 | |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 307 | if (bio->bi_error) |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 308 | set_btree_node_io_error(b); |
| 309 | |
| 310 | bch_bbio_free(bio, b->c); |
| 311 | |
| 312 | if (btree_node_io_error(b)) |
| 313 | goto err; |
| 314 | |
| 315 | bch_btree_node_read_done(b); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 316 | bch_time_stats_update(&b->c->btree_read_time, start_time); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 317 | |
| 318 | return; |
| 319 | err: |
Geert Uytterhoeven | 61cbd25 | 2013-09-23 23:17:30 -0700 | [diff] [blame] | 320 | bch_cache_set_error(b->c, "io error reading bucket %zu", |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 321 | PTR_BUCKET_NR(b->c, &b->key, 0)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 322 | } |
| 323 | |
| 324 | static void btree_complete_write(struct btree *b, struct btree_write *w) |
| 325 | { |
| 326 | if (w->prio_blocked && |
| 327 | !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) |
Kent Overstreet | 119ba0f | 2013-04-24 19:01:12 -0700 | [diff] [blame] | 328 | wake_up_allocators(b->c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 329 | |
| 330 | if (w->journal) { |
| 331 | atomic_dec_bug(w->journal); |
| 332 | __closure_wake_up(&b->c->journal.wait); |
| 333 | } |
| 334 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 335 | w->prio_blocked = 0; |
| 336 | w->journal = NULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 337 | } |
| 338 | |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 339 | static void btree_node_write_unlock(struct closure *cl) |
| 340 | { |
| 341 | struct btree *b = container_of(cl, struct btree, io); |
| 342 | |
| 343 | up(&b->io_mutex); |
| 344 | } |
| 345 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 346 | static void __btree_node_write_done(struct closure *cl) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 347 | { |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 348 | struct btree *b = container_of(cl, struct btree, io); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 349 | struct btree_write *w = btree_prev_write(b); |
| 350 | |
| 351 | bch_bbio_free(b->bio, b->c); |
| 352 | b->bio = NULL; |
| 353 | btree_complete_write(b, w); |
| 354 | |
| 355 | if (btree_node_dirty(b)) |
Kent Overstreet | 56b3077 | 2014-01-23 01:44:55 -0800 | [diff] [blame] | 356 | schedule_delayed_work(&b->work, 30 * HZ); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 357 | |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 358 | closure_return_with_destructor(cl, btree_node_write_unlock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 359 | } |
| 360 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 361 | static void btree_node_write_done(struct closure *cl) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 362 | { |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 363 | struct btree *b = container_of(cl, struct btree, io); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 364 | |
Guoqing Jiang | 491221f | 2016-09-22 03:10:01 -0400 | [diff] [blame] | 365 | bio_free_pages(b->bio); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 366 | __btree_node_write_done(cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 367 | } |
| 368 | |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 369 | static void btree_node_write_endio(struct bio *bio) |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 370 | { |
| 371 | struct closure *cl = bio->bi_private; |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 372 | struct btree *b = container_of(cl, struct btree, io); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 373 | |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 374 | if (bio->bi_error) |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 375 | set_btree_node_io_error(b); |
| 376 | |
Christoph Hellwig | 4246a0b | 2015-07-20 15:29:37 +0200 | [diff] [blame] | 377 | bch_bbio_count_io_errors(b->c, bio, bio->bi_error, "writing btree"); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 378 | closure_put(cl); |
| 379 | } |
| 380 | |
| 381 | static void do_btree_node_write(struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 382 | { |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 383 | struct closure *cl = &b->io; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 384 | struct bset *i = btree_bset_last(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 385 | BKEY_PADDED(key) k; |
| 386 | |
| 387 | i->version = BCACHE_BSET_VERSION; |
| 388 | i->csum = btree_csum_set(b, i); |
| 389 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 390 | BUG_ON(b->bio); |
| 391 | b->bio = bch_bbio_alloc(b->c); |
| 392 | |
| 393 | b->bio->bi_end_io = btree_node_write_endio; |
Kent Overstreet | faadf0c | 2013-11-01 18:03:08 -0700 | [diff] [blame] | 394 | b->bio->bi_private = cl; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 395 | b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c)); |
Mike Christie | ad0d9e7 | 2016-06-05 14:32:05 -0500 | [diff] [blame] | 396 | bio_set_op_attrs(b->bio, REQ_OP_WRITE, REQ_META|WRITE_SYNC|REQ_FUA); |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 397 | bch_bio_map(b->bio, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 398 | |
Kent Overstreet | e49c7c3 | 2013-06-26 17:25:38 -0700 | [diff] [blame] | 399 | /* |
| 400 | * If we're appending to a leaf node, we don't technically need FUA - |
| 401 | * this write just needs to be persisted before the next journal write, |
| 402 | * which will be marked FLUSH|FUA. |
| 403 | * |
| 404 | * Similarly if we're writing a new btree root - the pointer is going to |
| 405 | * be in the next journal entry. |
| 406 | * |
| 407 | * But if we're writing a new btree node (that isn't a root) or |
| 408 | * appending to a non leaf btree node, we need either FUA or a flush |
| 409 | * when we write the parent with the new pointer. FUA is cheaper than a |
| 410 | * flush, and writes appending to leaf nodes aren't blocking anything so |
| 411 | * just make all btree node writes FUA to keep things sane. |
| 412 | */ |
| 413 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 414 | bkey_copy(&k.key, &b->key); |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 415 | SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 416 | bset_sector_offset(&b->keys, i)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 417 | |
Kent Overstreet | 501d52a | 2014-05-19 08:55:40 -0700 | [diff] [blame] | 418 | if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 419 | int j; |
| 420 | struct bio_vec *bv; |
| 421 | void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); |
| 422 | |
Kent Overstreet | 7988613 | 2013-11-23 17:19:00 -0800 | [diff] [blame] | 423 | bio_for_each_segment_all(bv, b->bio, j) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 424 | memcpy(page_address(bv->bv_page), |
| 425 | base + j * PAGE_SIZE, PAGE_SIZE); |
| 426 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 427 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
| 428 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 429 | continue_at(cl, btree_node_write_done, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 430 | } else { |
| 431 | b->bio->bi_vcnt = 0; |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 432 | bch_bio_map(b->bio, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 433 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 434 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
| 435 | |
| 436 | closure_sync(cl); |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 437 | continue_at_nobarrier(cl, __btree_node_write_done, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 438 | } |
| 439 | } |
| 440 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 441 | void __bch_btree_node_write(struct btree *b, struct closure *parent) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 442 | { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 443 | struct bset *i = btree_bset_last(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 444 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 445 | lockdep_assert_held(&b->write_lock); |
| 446 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 447 | trace_bcache_btree_write(b); |
| 448 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 449 | BUG_ON(current->bio_list); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 450 | BUG_ON(b->written >= btree_blocks(b)); |
| 451 | BUG_ON(b->written && !i->keys); |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 452 | BUG_ON(btree_bset_first(b)->seq != i->seq); |
Kent Overstreet | dc9d98d | 2013-12-17 23:47:33 -0800 | [diff] [blame] | 453 | bch_check_keys(&b->keys, "writing"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 454 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 455 | cancel_delayed_work(&b->work); |
| 456 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 457 | /* If caller isn't waiting for write, parent refcount is cache set */ |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 458 | down(&b->io_mutex); |
| 459 | closure_init(&b->io, parent ?: &b->c->cl); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 460 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 461 | clear_bit(BTREE_NODE_dirty, &b->flags); |
| 462 | change_bit(BTREE_NODE_write_idx, &b->flags); |
| 463 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 464 | do_btree_node_write(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 465 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 466 | atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size, |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 467 | &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); |
| 468 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 469 | b->written += set_blocks(i, block_bytes(b->c)); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 470 | } |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 471 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 472 | void bch_btree_node_write(struct btree *b, struct closure *parent) |
| 473 | { |
| 474 | unsigned nsets = b->keys.nsets; |
| 475 | |
| 476 | lockdep_assert_held(&b->lock); |
| 477 | |
| 478 | __bch_btree_node_write(b, parent); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 479 | |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 480 | /* |
| 481 | * do verify if there was more than one set initially (i.e. we did a |
| 482 | * sort) and we sorted down to a single set: |
| 483 | */ |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 484 | if (nsets && !b->keys.nsets) |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 485 | bch_btree_verify(b); |
| 486 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 487 | bch_btree_init_next(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 488 | } |
| 489 | |
Kent Overstreet | f269af5 | 2013-07-23 20:48:29 -0700 | [diff] [blame] | 490 | static void bch_btree_node_write_sync(struct btree *b) |
| 491 | { |
| 492 | struct closure cl; |
| 493 | |
| 494 | closure_init_stack(&cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 495 | |
| 496 | mutex_lock(&b->write_lock); |
Kent Overstreet | f269af5 | 2013-07-23 20:48:29 -0700 | [diff] [blame] | 497 | bch_btree_node_write(b, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 498 | mutex_unlock(&b->write_lock); |
| 499 | |
Kent Overstreet | f269af5 | 2013-07-23 20:48:29 -0700 | [diff] [blame] | 500 | closure_sync(&cl); |
| 501 | } |
| 502 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 503 | static void btree_node_write_work(struct work_struct *w) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 504 | { |
| 505 | struct btree *b = container_of(to_delayed_work(w), struct btree, work); |
| 506 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 507 | mutex_lock(&b->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 508 | if (btree_node_dirty(b)) |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 509 | __bch_btree_node_write(b, NULL); |
| 510 | mutex_unlock(&b->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 511 | } |
| 512 | |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 513 | static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 514 | { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 515 | struct bset *i = btree_bset_last(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 516 | struct btree_write *w = btree_current_write(b); |
| 517 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 518 | lockdep_assert_held(&b->write_lock); |
| 519 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 520 | BUG_ON(!b->written); |
| 521 | BUG_ON(!i->keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 522 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 523 | if (!btree_node_dirty(b)) |
Kent Overstreet | 56b3077 | 2014-01-23 01:44:55 -0800 | [diff] [blame] | 524 | schedule_delayed_work(&b->work, 30 * HZ); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 525 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 526 | set_btree_node_dirty(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 527 | |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 528 | if (journal_ref) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 529 | if (w->journal && |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 530 | journal_pin_cmp(b->c, w->journal, journal_ref)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 531 | atomic_dec_bug(w->journal); |
| 532 | w->journal = NULL; |
| 533 | } |
| 534 | |
| 535 | if (!w->journal) { |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 536 | w->journal = journal_ref; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 537 | atomic_inc(w->journal); |
| 538 | } |
| 539 | } |
| 540 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 541 | /* Force write if set is too big */ |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 542 | if (set_bytes(i) > PAGE_SIZE - 48 && |
| 543 | !current->bio_list) |
| 544 | bch_btree_node_write(b, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 545 | } |
| 546 | |
| 547 | /* |
| 548 | * Btree in memory cache - allocation/freeing |
| 549 | * mca -> memory cache |
| 550 | */ |
| 551 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 552 | #define mca_reserve(c) (((c->root && c->root->level) \ |
| 553 | ? c->root->level : 1) * 8 + 16) |
| 554 | #define mca_can_free(c) \ |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 555 | max_t(int, 0, c->btree_cache_used - mca_reserve(c)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 556 | |
| 557 | static void mca_data_free(struct btree *b) |
| 558 | { |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 559 | BUG_ON(b->io_mutex.count != 1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 560 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 561 | bch_btree_keys_free(&b->keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 562 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 563 | b->c->btree_cache_used--; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 564 | list_move(&b->list, &b->c->btree_cache_freed); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 565 | } |
| 566 | |
| 567 | static void mca_bucket_free(struct btree *b) |
| 568 | { |
| 569 | BUG_ON(btree_node_dirty(b)); |
| 570 | |
| 571 | b->key.ptr[0] = 0; |
| 572 | hlist_del_init_rcu(&b->hash); |
| 573 | list_move(&b->list, &b->c->btree_cache_freeable); |
| 574 | } |
| 575 | |
| 576 | static unsigned btree_order(struct bkey *k) |
| 577 | { |
| 578 | return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); |
| 579 | } |
| 580 | |
| 581 | static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) |
| 582 | { |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 583 | if (!bch_btree_keys_alloc(&b->keys, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 584 | max_t(unsigned, |
| 585 | ilog2(b->c->btree_pages), |
| 586 | btree_order(k)), |
| 587 | gfp)) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 588 | b->c->btree_cache_used++; |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 589 | list_move(&b->list, &b->c->btree_cache); |
| 590 | } else { |
| 591 | list_move(&b->list, &b->c->btree_cache_freed); |
| 592 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 593 | } |
| 594 | |
| 595 | static struct btree *mca_bucket_alloc(struct cache_set *c, |
| 596 | struct bkey *k, gfp_t gfp) |
| 597 | { |
| 598 | struct btree *b = kzalloc(sizeof(struct btree), gfp); |
| 599 | if (!b) |
| 600 | return NULL; |
| 601 | |
| 602 | init_rwsem(&b->lock); |
| 603 | lockdep_set_novalidate_class(&b->lock); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 604 | mutex_init(&b->write_lock); |
| 605 | lockdep_set_novalidate_class(&b->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 606 | INIT_LIST_HEAD(&b->list); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 607 | INIT_DELAYED_WORK(&b->work, btree_node_write_work); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 608 | b->c = c; |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 609 | sema_init(&b->io_mutex, 1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 610 | |
| 611 | mca_data_alloc(b, k, gfp); |
| 612 | return b; |
| 613 | } |
| 614 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 615 | static int mca_reap(struct btree *b, unsigned min_order, bool flush) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 616 | { |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 617 | struct closure cl; |
| 618 | |
| 619 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 620 | lockdep_assert_held(&b->c->bucket_lock); |
| 621 | |
| 622 | if (!down_write_trylock(&b->lock)) |
| 623 | return -ENOMEM; |
| 624 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 625 | BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 626 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 627 | if (b->keys.page_order < min_order) |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 628 | goto out_unlock; |
| 629 | |
| 630 | if (!flush) { |
| 631 | if (btree_node_dirty(b)) |
| 632 | goto out_unlock; |
| 633 | |
| 634 | if (down_trylock(&b->io_mutex)) |
| 635 | goto out_unlock; |
| 636 | up(&b->io_mutex); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 637 | } |
| 638 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 639 | mutex_lock(&b->write_lock); |
Kent Overstreet | f269af5 | 2013-07-23 20:48:29 -0700 | [diff] [blame] | 640 | if (btree_node_dirty(b)) |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 641 | __bch_btree_node_write(b, &cl); |
| 642 | mutex_unlock(&b->write_lock); |
| 643 | |
| 644 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 645 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 646 | /* wait for any in flight btree write */ |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 647 | down(&b->io_mutex); |
| 648 | up(&b->io_mutex); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 649 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 650 | return 0; |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 651 | out_unlock: |
| 652 | rw_unlock(true, b); |
| 653 | return -ENOMEM; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 654 | } |
| 655 | |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 656 | static unsigned long bch_mca_scan(struct shrinker *shrink, |
| 657 | struct shrink_control *sc) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 658 | { |
| 659 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); |
| 660 | struct btree *b, *t; |
| 661 | unsigned long i, nr = sc->nr_to_scan; |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 662 | unsigned long freed = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 663 | |
| 664 | if (c->shrinker_disabled) |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 665 | return SHRINK_STOP; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 666 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 667 | if (c->btree_cache_alloc_lock) |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 668 | return SHRINK_STOP; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 669 | |
| 670 | /* Return -1 if we can't do anything right now */ |
Kent Overstreet | a698e08 | 2013-09-23 23:17:34 -0700 | [diff] [blame] | 671 | if (sc->gfp_mask & __GFP_IO) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 672 | mutex_lock(&c->bucket_lock); |
| 673 | else if (!mutex_trylock(&c->bucket_lock)) |
| 674 | return -1; |
| 675 | |
Kent Overstreet | 36c9ea9 | 2013-06-03 13:04:56 -0700 | [diff] [blame] | 676 | /* |
| 677 | * It's _really_ critical that we don't free too many btree nodes - we |
| 678 | * have to always leave ourselves a reserve. The reserve is how we |
| 679 | * guarantee that allocating memory for a new btree node can always |
| 680 | * succeed, so that inserting keys into the btree can always succeed and |
| 681 | * IO can always make forward progress: |
| 682 | */ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 683 | nr /= c->btree_pages; |
| 684 | nr = min_t(unsigned long, nr, mca_can_free(c)); |
| 685 | |
| 686 | i = 0; |
| 687 | list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 688 | if (freed >= nr) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 689 | break; |
| 690 | |
| 691 | if (++i > 3 && |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 692 | !mca_reap(b, 0, false)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 693 | mca_data_free(b); |
| 694 | rw_unlock(true, b); |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 695 | freed++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 696 | } |
| 697 | } |
| 698 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 699 | for (i = 0; (nr--) && i < c->btree_cache_used; i++) { |
Kent Overstreet | b0f32a5 | 2013-12-10 13:24:26 -0800 | [diff] [blame] | 700 | if (list_empty(&c->btree_cache)) |
| 701 | goto out; |
| 702 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 703 | b = list_first_entry(&c->btree_cache, struct btree, list); |
| 704 | list_rotate_left(&c->btree_cache); |
| 705 | |
| 706 | if (!b->accessed && |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 707 | !mca_reap(b, 0, false)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 708 | mca_bucket_free(b); |
| 709 | mca_data_free(b); |
| 710 | rw_unlock(true, b); |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 711 | freed++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 712 | } else |
| 713 | b->accessed = 0; |
| 714 | } |
| 715 | out: |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 716 | mutex_unlock(&c->bucket_lock); |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 717 | return freed; |
| 718 | } |
| 719 | |
| 720 | static unsigned long bch_mca_count(struct shrinker *shrink, |
| 721 | struct shrink_control *sc) |
| 722 | { |
| 723 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); |
| 724 | |
| 725 | if (c->shrinker_disabled) |
| 726 | return 0; |
| 727 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 728 | if (c->btree_cache_alloc_lock) |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 729 | return 0; |
| 730 | |
| 731 | return mca_can_free(c) * c->btree_pages; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 732 | } |
| 733 | |
| 734 | void bch_btree_cache_free(struct cache_set *c) |
| 735 | { |
| 736 | struct btree *b; |
| 737 | struct closure cl; |
| 738 | closure_init_stack(&cl); |
| 739 | |
| 740 | if (c->shrink.list.next) |
| 741 | unregister_shrinker(&c->shrink); |
| 742 | |
| 743 | mutex_lock(&c->bucket_lock); |
| 744 | |
| 745 | #ifdef CONFIG_BCACHE_DEBUG |
| 746 | if (c->verify_data) |
| 747 | list_move(&c->verify_data->list, &c->btree_cache); |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 748 | |
| 749 | free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c))); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 750 | #endif |
| 751 | |
| 752 | list_splice(&c->btree_cache_freeable, |
| 753 | &c->btree_cache); |
| 754 | |
| 755 | while (!list_empty(&c->btree_cache)) { |
| 756 | b = list_first_entry(&c->btree_cache, struct btree, list); |
| 757 | |
| 758 | if (btree_node_dirty(b)) |
| 759 | btree_complete_write(b, btree_current_write(b)); |
| 760 | clear_bit(BTREE_NODE_dirty, &b->flags); |
| 761 | |
| 762 | mca_data_free(b); |
| 763 | } |
| 764 | |
| 765 | while (!list_empty(&c->btree_cache_freed)) { |
| 766 | b = list_first_entry(&c->btree_cache_freed, |
| 767 | struct btree, list); |
| 768 | list_del(&b->list); |
| 769 | cancel_delayed_work_sync(&b->work); |
| 770 | kfree(b); |
| 771 | } |
| 772 | |
| 773 | mutex_unlock(&c->bucket_lock); |
| 774 | } |
| 775 | |
| 776 | int bch_btree_cache_alloc(struct cache_set *c) |
| 777 | { |
| 778 | unsigned i; |
| 779 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 780 | for (i = 0; i < mca_reserve(c); i++) |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 781 | if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL)) |
| 782 | return -ENOMEM; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 783 | |
| 784 | list_splice_init(&c->btree_cache, |
| 785 | &c->btree_cache_freeable); |
| 786 | |
| 787 | #ifdef CONFIG_BCACHE_DEBUG |
| 788 | mutex_init(&c->verify_lock); |
| 789 | |
Kent Overstreet | 78b77bf | 2013-12-17 22:49:08 -0800 | [diff] [blame] | 790 | c->verify_ondisk = (void *) |
| 791 | __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c))); |
| 792 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 793 | c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); |
| 794 | |
| 795 | if (c->verify_data && |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 796 | c->verify_data->keys.set->data) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 797 | list_del_init(&c->verify_data->list); |
| 798 | else |
| 799 | c->verify_data = NULL; |
| 800 | #endif |
| 801 | |
Dave Chinner | 7dc19d5 | 2013-08-28 10:18:11 +1000 | [diff] [blame] | 802 | c->shrink.count_objects = bch_mca_count; |
| 803 | c->shrink.scan_objects = bch_mca_scan; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 804 | c->shrink.seeks = 4; |
| 805 | c->shrink.batch = c->btree_pages * 2; |
Michael Lyle | 409982c | 2017-11-24 15:14:27 -0800 | [diff] [blame^] | 806 | |
| 807 | if (register_shrinker(&c->shrink)) |
| 808 | pr_warn("bcache: %s: could not register shrinker", |
| 809 | __func__); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 810 | |
| 811 | return 0; |
| 812 | } |
| 813 | |
| 814 | /* Btree in memory cache - hash table */ |
| 815 | |
| 816 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) |
| 817 | { |
| 818 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; |
| 819 | } |
| 820 | |
| 821 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) |
| 822 | { |
| 823 | struct btree *b; |
| 824 | |
| 825 | rcu_read_lock(); |
| 826 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) |
| 827 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) |
| 828 | goto out; |
| 829 | b = NULL; |
| 830 | out: |
| 831 | rcu_read_unlock(); |
| 832 | return b; |
| 833 | } |
| 834 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 835 | static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op) |
| 836 | { |
| 837 | struct task_struct *old; |
| 838 | |
| 839 | old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current); |
| 840 | if (old && old != current) { |
| 841 | if (op) |
| 842 | prepare_to_wait(&c->btree_cache_wait, &op->wait, |
| 843 | TASK_UNINTERRUPTIBLE); |
| 844 | return -EINTR; |
| 845 | } |
| 846 | |
| 847 | return 0; |
| 848 | } |
| 849 | |
| 850 | static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op, |
| 851 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 852 | { |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 853 | struct btree *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 854 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 855 | trace_bcache_btree_cache_cannibalize(c); |
| 856 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 857 | if (mca_cannibalize_lock(c, op)) |
| 858 | return ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 859 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 860 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
| 861 | if (!mca_reap(b, btree_order(k), false)) |
| 862 | return b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 863 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 864 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
| 865 | if (!mca_reap(b, btree_order(k), true)) |
| 866 | return b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 867 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 868 | WARN(1, "btree cache cannibalize failed\n"); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 869 | return ERR_PTR(-ENOMEM); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 870 | } |
| 871 | |
| 872 | /* |
| 873 | * We can only have one thread cannibalizing other cached btree nodes at a time, |
| 874 | * or we'll deadlock. We use an open coded mutex to ensure that, which a |
| 875 | * cannibalize_bucket() will take. This means every time we unlock the root of |
| 876 | * the btree, we need to release this lock if we have it held. |
| 877 | */ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 878 | static void bch_cannibalize_unlock(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 879 | { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 880 | if (c->btree_cache_alloc_lock == current) { |
| 881 | c->btree_cache_alloc_lock = NULL; |
| 882 | wake_up(&c->btree_cache_wait); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 883 | } |
| 884 | } |
| 885 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 886 | static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op, |
| 887 | struct bkey *k, int level) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 888 | { |
| 889 | struct btree *b; |
| 890 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 891 | BUG_ON(current->bio_list); |
| 892 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 893 | lockdep_assert_held(&c->bucket_lock); |
| 894 | |
| 895 | if (mca_find(c, k)) |
| 896 | return NULL; |
| 897 | |
| 898 | /* btree_free() doesn't free memory; it sticks the node on the end of |
| 899 | * the list. Check if there's any freed nodes there: |
| 900 | */ |
| 901 | list_for_each_entry(b, &c->btree_cache_freeable, list) |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 902 | if (!mca_reap(b, btree_order(k), false)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 903 | goto out; |
| 904 | |
| 905 | /* We never free struct btree itself, just the memory that holds the on |
| 906 | * disk node. Check the freed list before allocating a new one: |
| 907 | */ |
| 908 | list_for_each_entry(b, &c->btree_cache_freed, list) |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 909 | if (!mca_reap(b, 0, false)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 910 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 911 | if (!b->keys.set[0].data) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 912 | goto err; |
| 913 | else |
| 914 | goto out; |
| 915 | } |
| 916 | |
| 917 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); |
| 918 | if (!b) |
| 919 | goto err; |
| 920 | |
| 921 | BUG_ON(!down_write_trylock(&b->lock)); |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 922 | if (!b->keys.set->data) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 923 | goto err; |
| 924 | out: |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 925 | BUG_ON(b->io_mutex.count != 1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 926 | |
| 927 | bkey_copy(&b->key, k); |
| 928 | list_move(&b->list, &c->btree_cache); |
| 929 | hlist_del_init_rcu(&b->hash); |
| 930 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); |
| 931 | |
| 932 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 933 | b->parent = (void *) ~0UL; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 934 | b->flags = 0; |
| 935 | b->written = 0; |
| 936 | b->level = level; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 937 | |
Kent Overstreet | 65d4523 | 2013-12-20 17:22:05 -0800 | [diff] [blame] | 938 | if (!b->level) |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 939 | bch_btree_keys_init(&b->keys, &bch_extent_keys_ops, |
| 940 | &b->c->expensive_debug_checks); |
Kent Overstreet | 65d4523 | 2013-12-20 17:22:05 -0800 | [diff] [blame] | 941 | else |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 942 | bch_btree_keys_init(&b->keys, &bch_btree_keys_ops, |
| 943 | &b->c->expensive_debug_checks); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 944 | |
| 945 | return b; |
| 946 | err: |
| 947 | if (b) |
| 948 | rw_unlock(true, b); |
| 949 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 950 | b = mca_cannibalize(c, op, k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 951 | if (!IS_ERR(b)) |
| 952 | goto out; |
| 953 | |
| 954 | return b; |
| 955 | } |
| 956 | |
| 957 | /** |
| 958 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it |
| 959 | * in from disk if necessary. |
| 960 | * |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 961 | * If IO is necessary and running under generic_make_request, returns -EAGAIN. |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 962 | * |
| 963 | * The btree node will have either a read or a write lock held, depending on |
| 964 | * level and op->lock. |
| 965 | */ |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 966 | struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op, |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 967 | struct bkey *k, int level, bool write, |
| 968 | struct btree *parent) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 969 | { |
| 970 | int i = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 971 | struct btree *b; |
| 972 | |
| 973 | BUG_ON(level < 0); |
| 974 | retry: |
| 975 | b = mca_find(c, k); |
| 976 | |
| 977 | if (!b) { |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 978 | if (current->bio_list) |
| 979 | return ERR_PTR(-EAGAIN); |
| 980 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 981 | mutex_lock(&c->bucket_lock); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 982 | b = mca_alloc(c, op, k, level); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 983 | mutex_unlock(&c->bucket_lock); |
| 984 | |
| 985 | if (!b) |
| 986 | goto retry; |
| 987 | if (IS_ERR(b)) |
| 988 | return b; |
| 989 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 990 | bch_btree_node_read(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 991 | |
| 992 | if (!write) |
| 993 | downgrade_write(&b->lock); |
| 994 | } else { |
| 995 | rw_lock(write, b, level); |
| 996 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { |
| 997 | rw_unlock(write, b); |
| 998 | goto retry; |
| 999 | } |
| 1000 | BUG_ON(b->level != level); |
| 1001 | } |
| 1002 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1003 | b->parent = parent; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1004 | b->accessed = 1; |
| 1005 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1006 | for (; i <= b->keys.nsets && b->keys.set[i].size; i++) { |
| 1007 | prefetch(b->keys.set[i].tree); |
| 1008 | prefetch(b->keys.set[i].data); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1009 | } |
| 1010 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1011 | for (; i <= b->keys.nsets; i++) |
| 1012 | prefetch(b->keys.set[i].data); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1013 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1014 | if (btree_node_io_error(b)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1015 | rw_unlock(write, b); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1016 | return ERR_PTR(-EIO); |
| 1017 | } |
| 1018 | |
| 1019 | BUG_ON(!b->written); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1020 | |
| 1021 | return b; |
| 1022 | } |
| 1023 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1024 | static void btree_node_prefetch(struct btree *parent, struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1025 | { |
| 1026 | struct btree *b; |
| 1027 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1028 | mutex_lock(&parent->c->bucket_lock); |
| 1029 | b = mca_alloc(parent->c, NULL, k, parent->level - 1); |
| 1030 | mutex_unlock(&parent->c->bucket_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1031 | |
| 1032 | if (!IS_ERR_OR_NULL(b)) { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1033 | b->parent = parent; |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1034 | bch_btree_node_read(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1035 | rw_unlock(true, b); |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | /* Btree alloc */ |
| 1040 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 1041 | static void btree_node_free(struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1042 | { |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1043 | trace_bcache_btree_node_free(b); |
| 1044 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1045 | BUG_ON(b == b->c->root); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1046 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1047 | mutex_lock(&b->write_lock); |
| 1048 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1049 | if (btree_node_dirty(b)) |
| 1050 | btree_complete_write(b, btree_current_write(b)); |
| 1051 | clear_bit(BTREE_NODE_dirty, &b->flags); |
| 1052 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1053 | mutex_unlock(&b->write_lock); |
| 1054 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1055 | cancel_delayed_work(&b->work); |
| 1056 | |
| 1057 | mutex_lock(&b->c->bucket_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1058 | bch_bucket_free(b->c, &b->key); |
| 1059 | mca_bucket_free(b); |
| 1060 | mutex_unlock(&b->c->bucket_lock); |
| 1061 | } |
| 1062 | |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1063 | struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op, |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1064 | int level, bool wait, |
| 1065 | struct btree *parent) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1066 | { |
| 1067 | BKEY_PADDED(key) k; |
| 1068 | struct btree *b = ERR_PTR(-EAGAIN); |
| 1069 | |
| 1070 | mutex_lock(&c->bucket_lock); |
| 1071 | retry: |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1072 | if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1073 | goto err; |
| 1074 | |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 1075 | bkey_put(c, &k.key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1076 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); |
| 1077 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1078 | b = mca_alloc(c, op, &k.key, level); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1079 | if (IS_ERR(b)) |
| 1080 | goto err_free; |
| 1081 | |
| 1082 | if (!b) { |
Kent Overstreet | b1a67b0 | 2013-03-25 11:46:44 -0700 | [diff] [blame] | 1083 | cache_bug(c, |
| 1084 | "Tried to allocate bucket that was in btree cache"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1085 | goto retry; |
| 1086 | } |
| 1087 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1088 | b->accessed = 1; |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1089 | b->parent = parent; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1090 | bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1091 | |
| 1092 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1093 | |
| 1094 | trace_bcache_btree_node_alloc(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1095 | return b; |
| 1096 | err_free: |
| 1097 | bch_bucket_free(c, &k.key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1098 | err: |
| 1099 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1100 | |
Slava Pestov | 913dc33 | 2014-05-23 11:18:35 -0700 | [diff] [blame] | 1101 | trace_bcache_btree_node_alloc_fail(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1102 | return b; |
| 1103 | } |
| 1104 | |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1105 | static struct btree *bch_btree_node_alloc(struct cache_set *c, |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1106 | struct btree_op *op, int level, |
| 1107 | struct btree *parent) |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1108 | { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1109 | return __bch_btree_node_alloc(c, op, level, op != NULL, parent); |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1110 | } |
| 1111 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1112 | static struct btree *btree_node_alloc_replacement(struct btree *b, |
| 1113 | struct btree_op *op) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1114 | { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1115 | struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent); |
Kent Overstreet | 67539e8 | 2013-09-10 22:53:34 -0700 | [diff] [blame] | 1116 | if (!IS_ERR_OR_NULL(n)) { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1117 | mutex_lock(&n->write_lock); |
Kent Overstreet | 89ebb4a | 2013-11-11 18:38:51 -0800 | [diff] [blame] | 1118 | bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort); |
Kent Overstreet | 67539e8 | 2013-09-10 22:53:34 -0700 | [diff] [blame] | 1119 | bkey_copy_key(&n->key, &b->key); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1120 | mutex_unlock(&n->write_lock); |
Kent Overstreet | 67539e8 | 2013-09-10 22:53:34 -0700 | [diff] [blame] | 1121 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1122 | |
| 1123 | return n; |
| 1124 | } |
| 1125 | |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1126 | static void make_btree_freeing_key(struct btree *b, struct bkey *k) |
| 1127 | { |
| 1128 | unsigned i; |
| 1129 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1130 | mutex_lock(&b->c->bucket_lock); |
| 1131 | |
| 1132 | atomic_inc(&b->c->prio_blocked); |
| 1133 | |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1134 | bkey_copy(k, &b->key); |
| 1135 | bkey_copy_key(k, &ZERO_KEY); |
| 1136 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1137 | for (i = 0; i < KEY_PTRS(k); i++) |
| 1138 | SET_PTR_GEN(k, i, |
| 1139 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), |
| 1140 | PTR_BUCKET(b->c, &b->key, i))); |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1141 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1142 | mutex_unlock(&b->c->bucket_lock); |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1143 | } |
| 1144 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1145 | static int btree_check_reserve(struct btree *b, struct btree_op *op) |
| 1146 | { |
| 1147 | struct cache_set *c = b->c; |
| 1148 | struct cache *ca; |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1149 | unsigned i, reserve = (c->root->level - b->level) * 2 + 1; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1150 | |
| 1151 | mutex_lock(&c->bucket_lock); |
| 1152 | |
| 1153 | for_each_cache(ca, c, i) |
| 1154 | if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) { |
| 1155 | if (op) |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1156 | prepare_to_wait(&c->btree_cache_wait, &op->wait, |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1157 | TASK_UNINTERRUPTIBLE); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1158 | mutex_unlock(&c->bucket_lock); |
| 1159 | return -EINTR; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1160 | } |
| 1161 | |
| 1162 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1163 | |
| 1164 | return mca_cannibalize_lock(b->c, op); |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1165 | } |
| 1166 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1167 | /* Garbage collection */ |
| 1168 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1169 | static uint8_t __bch_btree_mark_key(struct cache_set *c, int level, |
| 1170 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1171 | { |
| 1172 | uint8_t stale = 0; |
| 1173 | unsigned i; |
| 1174 | struct bucket *g; |
| 1175 | |
| 1176 | /* |
| 1177 | * ptr_invalid() can't return true for the keys that mark btree nodes as |
| 1178 | * freed, but since ptr_bad() returns true we'll never actually use them |
| 1179 | * for anything and thus we don't want mark their pointers here |
| 1180 | */ |
| 1181 | if (!bkey_cmp(k, &ZERO_KEY)) |
| 1182 | return stale; |
| 1183 | |
| 1184 | for (i = 0; i < KEY_PTRS(k); i++) { |
| 1185 | if (!ptr_available(c, k, i)) |
| 1186 | continue; |
| 1187 | |
| 1188 | g = PTR_BUCKET(c, k, i); |
| 1189 | |
Kent Overstreet | 3a2fd9d | 2014-02-27 17:51:12 -0800 | [diff] [blame] | 1190 | if (gen_after(g->last_gc, PTR_GEN(k, i))) |
| 1191 | g->last_gc = PTR_GEN(k, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1192 | |
| 1193 | if (ptr_stale(c, k, i)) { |
| 1194 | stale = max(stale, ptr_stale(c, k, i)); |
| 1195 | continue; |
| 1196 | } |
| 1197 | |
| 1198 | cache_bug_on(GC_MARK(g) && |
| 1199 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), |
| 1200 | c, "inconsistent ptrs: mark = %llu, level = %i", |
| 1201 | GC_MARK(g), level); |
| 1202 | |
| 1203 | if (level) |
| 1204 | SET_GC_MARK(g, GC_MARK_METADATA); |
| 1205 | else if (KEY_DIRTY(k)) |
| 1206 | SET_GC_MARK(g, GC_MARK_DIRTY); |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1207 | else if (!GC_MARK(g)) |
| 1208 | SET_GC_MARK(g, GC_MARK_RECLAIMABLE); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1209 | |
| 1210 | /* guard against overflow */ |
| 1211 | SET_GC_SECTORS_USED(g, min_t(unsigned, |
| 1212 | GC_SECTORS_USED(g) + KEY_SIZE(k), |
Darrick J. Wong | 9471744 | 2014-01-28 16:57:39 -0800 | [diff] [blame] | 1213 | MAX_GC_SECTORS_USED)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1214 | |
| 1215 | BUG_ON(!GC_SECTORS_USED(g)); |
| 1216 | } |
| 1217 | |
| 1218 | return stale; |
| 1219 | } |
| 1220 | |
| 1221 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) |
| 1222 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1223 | void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k) |
| 1224 | { |
| 1225 | unsigned i; |
| 1226 | |
| 1227 | for (i = 0; i < KEY_PTRS(k); i++) |
| 1228 | if (ptr_available(c, k, i) && |
| 1229 | !ptr_stale(c, k, i)) { |
| 1230 | struct bucket *b = PTR_BUCKET(c, k, i); |
| 1231 | |
| 1232 | b->gen = PTR_GEN(k, i); |
| 1233 | |
| 1234 | if (level && bkey_cmp(k, &ZERO_KEY)) |
| 1235 | b->prio = BTREE_PRIO; |
| 1236 | else if (!level && b->prio == BTREE_PRIO) |
| 1237 | b->prio = INITIAL_PRIO; |
| 1238 | } |
| 1239 | |
| 1240 | __bch_btree_mark_key(c, level, k); |
| 1241 | } |
| 1242 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1243 | static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1244 | { |
| 1245 | uint8_t stale = 0; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1246 | unsigned keys = 0, good_keys = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1247 | struct bkey *k; |
| 1248 | struct btree_iter iter; |
| 1249 | struct bset_tree *t; |
| 1250 | |
| 1251 | gc->nodes++; |
| 1252 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1253 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1254 | stale = max(stale, btree_mark_key(b, k)); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1255 | keys++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1256 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1257 | if (bch_ptr_bad(&b->keys, k)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1258 | continue; |
| 1259 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1260 | gc->key_bytes += bkey_u64s(k); |
| 1261 | gc->nkeys++; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1262 | good_keys++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1263 | |
| 1264 | gc->data += KEY_SIZE(k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1265 | } |
| 1266 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1267 | for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1268 | btree_bug_on(t->size && |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1269 | bset_written(&b->keys, t) && |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1270 | bkey_cmp(&b->key, &t->end) < 0, |
| 1271 | b, "found short btree key in gc"); |
| 1272 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1273 | if (b->c->gc_always_rewrite) |
| 1274 | return true; |
| 1275 | |
| 1276 | if (stale > 10) |
| 1277 | return true; |
| 1278 | |
| 1279 | if ((keys - good_keys) * 2 > keys) |
| 1280 | return true; |
| 1281 | |
| 1282 | return false; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1283 | } |
| 1284 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1285 | #define GC_MERGE_NODES 4U |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1286 | |
| 1287 | struct gc_merge_info { |
| 1288 | struct btree *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1289 | unsigned keys; |
| 1290 | }; |
| 1291 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1292 | static int bch_btree_insert_node(struct btree *, struct btree_op *, |
| 1293 | struct keylist *, atomic_t *, struct bkey *); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1294 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1295 | static int btree_gc_coalesce(struct btree *b, struct btree_op *op, |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1296 | struct gc_stat *gc, struct gc_merge_info *r) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1297 | { |
| 1298 | unsigned i, nodes = 0, keys = 0, blocks; |
| 1299 | struct btree *new_nodes[GC_MERGE_NODES]; |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1300 | struct keylist keylist; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1301 | struct closure cl; |
| 1302 | struct bkey *k; |
| 1303 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1304 | bch_keylist_init(&keylist); |
| 1305 | |
| 1306 | if (btree_check_reserve(b, NULL)) |
| 1307 | return 0; |
| 1308 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1309 | memset(new_nodes, 0, sizeof(new_nodes)); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1310 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1311 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1312 | while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1313 | keys += r[nodes++].keys; |
| 1314 | |
| 1315 | blocks = btree_default_blocks(b->c) * 2 / 3; |
| 1316 | |
| 1317 | if (nodes < 2 || |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1318 | __set_blocks(b->keys.set[0].data, keys, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1319 | block_bytes(b->c)) > blocks * (nodes - 1)) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1320 | return 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1321 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1322 | for (i = 0; i < nodes; i++) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1323 | new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1324 | if (IS_ERR_OR_NULL(new_nodes[i])) |
| 1325 | goto out_nocoalesce; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1326 | } |
| 1327 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1328 | /* |
| 1329 | * We have to check the reserve here, after we've allocated our new |
| 1330 | * nodes, to make sure the insert below will succeed - we also check |
| 1331 | * before as an optimization to potentially avoid a bunch of expensive |
| 1332 | * allocs/sorts |
| 1333 | */ |
| 1334 | if (btree_check_reserve(b, NULL)) |
| 1335 | goto out_nocoalesce; |
| 1336 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1337 | for (i = 0; i < nodes; i++) |
| 1338 | mutex_lock(&new_nodes[i]->write_lock); |
| 1339 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1340 | for (i = nodes - 1; i > 0; --i) { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1341 | struct bset *n1 = btree_bset_first(new_nodes[i]); |
| 1342 | struct bset *n2 = btree_bset_first(new_nodes[i - 1]); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1343 | struct bkey *k, *last = NULL; |
| 1344 | |
| 1345 | keys = 0; |
| 1346 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1347 | if (i > 1) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1348 | for (k = n2->start; |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1349 | k < bset_bkey_last(n2); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1350 | k = bkey_next(k)) { |
| 1351 | if (__set_blocks(n1, n1->keys + keys + |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1352 | bkey_u64s(k), |
| 1353 | block_bytes(b->c)) > blocks) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1354 | break; |
| 1355 | |
| 1356 | last = k; |
| 1357 | keys += bkey_u64s(k); |
| 1358 | } |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1359 | } else { |
| 1360 | /* |
| 1361 | * Last node we're not getting rid of - we're getting |
| 1362 | * rid of the node at r[0]. Have to try and fit all of |
| 1363 | * the remaining keys into this node; we can't ensure |
| 1364 | * they will always fit due to rounding and variable |
| 1365 | * length keys (shouldn't be possible in practice, |
| 1366 | * though) |
| 1367 | */ |
| 1368 | if (__set_blocks(n1, n1->keys + n2->keys, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1369 | block_bytes(b->c)) > |
| 1370 | btree_blocks(new_nodes[i])) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1371 | goto out_nocoalesce; |
| 1372 | |
| 1373 | keys = n2->keys; |
| 1374 | /* Take the key of the node we're getting rid of */ |
| 1375 | last = &r->b->key; |
| 1376 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1377 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1378 | BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) > |
| 1379 | btree_blocks(new_nodes[i])); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1380 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1381 | if (last) |
| 1382 | bkey_copy_key(&new_nodes[i]->key, last); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1383 | |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1384 | memcpy(bset_bkey_last(n1), |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1385 | n2->start, |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1386 | (void *) bset_bkey_idx(n2, keys) - (void *) n2->start); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1387 | |
| 1388 | n1->keys += keys; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1389 | r[i].keys = n1->keys; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1390 | |
| 1391 | memmove(n2->start, |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1392 | bset_bkey_idx(n2, keys), |
| 1393 | (void *) bset_bkey_last(n2) - |
| 1394 | (void *) bset_bkey_idx(n2, keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1395 | |
| 1396 | n2->keys -= keys; |
| 1397 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1398 | if (__bch_keylist_realloc(&keylist, |
Kent Overstreet | 085d2a3 | 2013-11-11 18:20:51 -0800 | [diff] [blame] | 1399 | bkey_u64s(&new_nodes[i]->key))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1400 | goto out_nocoalesce; |
| 1401 | |
| 1402 | bch_btree_node_write(new_nodes[i], &cl); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1403 | bch_keylist_add(&keylist, &new_nodes[i]->key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1404 | } |
| 1405 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1406 | for (i = 0; i < nodes; i++) |
| 1407 | mutex_unlock(&new_nodes[i]->write_lock); |
| 1408 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1409 | closure_sync(&cl); |
| 1410 | |
| 1411 | /* We emptied out this node */ |
| 1412 | BUG_ON(btree_bset_first(new_nodes[0])->keys); |
| 1413 | btree_node_free(new_nodes[0]); |
| 1414 | rw_unlock(true, new_nodes[0]); |
Slava Pestov | 400ffaa | 2014-07-12 21:53:11 -0700 | [diff] [blame] | 1415 | new_nodes[0] = NULL; |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1416 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1417 | for (i = 0; i < nodes; i++) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1418 | if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1419 | goto out_nocoalesce; |
| 1420 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1421 | make_btree_freeing_key(r[i].b, keylist.top); |
| 1422 | bch_keylist_push(&keylist); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1423 | } |
| 1424 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1425 | bch_btree_insert_node(b, op, &keylist, NULL, NULL); |
| 1426 | BUG_ON(!bch_keylist_empty(&keylist)); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1427 | |
| 1428 | for (i = 0; i < nodes; i++) { |
| 1429 | btree_node_free(r[i].b); |
| 1430 | rw_unlock(true, r[i].b); |
| 1431 | |
| 1432 | r[i].b = new_nodes[i]; |
| 1433 | } |
| 1434 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1435 | memmove(r, r + 1, sizeof(r[0]) * (nodes - 1)); |
| 1436 | r[nodes - 1].b = ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1437 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1438 | trace_bcache_btree_gc_coalesce(nodes); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1439 | gc->nodes--; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1440 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1441 | bch_keylist_free(&keylist); |
| 1442 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1443 | /* Invalidated our iterator */ |
| 1444 | return -EINTR; |
| 1445 | |
| 1446 | out_nocoalesce: |
| 1447 | closure_sync(&cl); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1448 | bch_keylist_free(&keylist); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1449 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1450 | while ((k = bch_keylist_pop(&keylist))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1451 | if (!bkey_cmp(k, &ZERO_KEY)) |
| 1452 | atomic_dec(&b->c->prio_blocked); |
| 1453 | |
| 1454 | for (i = 0; i < nodes; i++) |
| 1455 | if (!IS_ERR_OR_NULL(new_nodes[i])) { |
| 1456 | btree_node_free(new_nodes[i]); |
| 1457 | rw_unlock(true, new_nodes[i]); |
| 1458 | } |
| 1459 | return 0; |
| 1460 | } |
| 1461 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1462 | static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op, |
| 1463 | struct btree *replace) |
| 1464 | { |
| 1465 | struct keylist keys; |
| 1466 | struct btree *n; |
| 1467 | |
| 1468 | if (btree_check_reserve(b, NULL)) |
| 1469 | return 0; |
| 1470 | |
| 1471 | n = btree_node_alloc_replacement(replace, NULL); |
| 1472 | |
| 1473 | /* recheck reserve after allocating replacement node */ |
| 1474 | if (btree_check_reserve(b, NULL)) { |
| 1475 | btree_node_free(n); |
| 1476 | rw_unlock(true, n); |
| 1477 | return 0; |
| 1478 | } |
| 1479 | |
| 1480 | bch_btree_node_write_sync(n); |
| 1481 | |
| 1482 | bch_keylist_init(&keys); |
| 1483 | bch_keylist_add(&keys, &n->key); |
| 1484 | |
| 1485 | make_btree_freeing_key(replace, keys.top); |
| 1486 | bch_keylist_push(&keys); |
| 1487 | |
| 1488 | bch_btree_insert_node(b, op, &keys, NULL, NULL); |
| 1489 | BUG_ON(!bch_keylist_empty(&keys)); |
| 1490 | |
| 1491 | btree_node_free(replace); |
| 1492 | rw_unlock(true, n); |
| 1493 | |
| 1494 | /* Invalidated our iterator */ |
| 1495 | return -EINTR; |
| 1496 | } |
| 1497 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1498 | static unsigned btree_gc_count_keys(struct btree *b) |
| 1499 | { |
| 1500 | struct bkey *k; |
| 1501 | struct btree_iter iter; |
| 1502 | unsigned ret = 0; |
| 1503 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1504 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1505 | ret += bkey_u64s(k); |
| 1506 | |
| 1507 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1508 | } |
| 1509 | |
| 1510 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, |
| 1511 | struct closure *writes, struct gc_stat *gc) |
| 1512 | { |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1513 | int ret = 0; |
| 1514 | bool should_rewrite; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1515 | struct bkey *k; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1516 | struct btree_iter iter; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1517 | struct gc_merge_info r[GC_MERGE_NODES]; |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1518 | struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1519 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1520 | bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1521 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1522 | for (i = r; i < r + ARRAY_SIZE(r); i++) |
| 1523 | i->b = ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1524 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1525 | while (1) { |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1526 | k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1527 | if (k) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1528 | r->b = bch_btree_node_get(b->c, op, k, b->level - 1, |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1529 | true, b); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1530 | if (IS_ERR(r->b)) { |
| 1531 | ret = PTR_ERR(r->b); |
| 1532 | break; |
| 1533 | } |
| 1534 | |
| 1535 | r->keys = btree_gc_count_keys(r->b); |
| 1536 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1537 | ret = btree_gc_coalesce(b, op, gc, r); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1538 | if (ret) |
| 1539 | break; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1540 | } |
| 1541 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1542 | if (!last->b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1543 | break; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1544 | |
| 1545 | if (!IS_ERR(last->b)) { |
| 1546 | should_rewrite = btree_gc_mark_node(last->b, gc); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1547 | if (should_rewrite) { |
| 1548 | ret = btree_gc_rewrite_node(b, op, last->b); |
| 1549 | if (ret) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1550 | break; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1551 | } |
| 1552 | |
| 1553 | if (last->b->level) { |
| 1554 | ret = btree_gc_recurse(last->b, op, writes, gc); |
| 1555 | if (ret) |
| 1556 | break; |
| 1557 | } |
| 1558 | |
| 1559 | bkey_copy_key(&b->c->gc_done, &last->b->key); |
| 1560 | |
| 1561 | /* |
| 1562 | * Must flush leaf nodes before gc ends, since replace |
| 1563 | * operations aren't journalled |
| 1564 | */ |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1565 | mutex_lock(&last->b->write_lock); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1566 | if (btree_node_dirty(last->b)) |
| 1567 | bch_btree_node_write(last->b, writes); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1568 | mutex_unlock(&last->b->write_lock); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1569 | rw_unlock(true, last->b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1570 | } |
| 1571 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1572 | memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1)); |
| 1573 | r->b = NULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1574 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1575 | if (need_resched()) { |
| 1576 | ret = -EAGAIN; |
| 1577 | break; |
| 1578 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1579 | } |
| 1580 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1581 | for (i = r; i < r + ARRAY_SIZE(r); i++) |
| 1582 | if (!IS_ERR_OR_NULL(i->b)) { |
| 1583 | mutex_lock(&i->b->write_lock); |
| 1584 | if (btree_node_dirty(i->b)) |
| 1585 | bch_btree_node_write(i->b, writes); |
| 1586 | mutex_unlock(&i->b->write_lock); |
| 1587 | rw_unlock(true, i->b); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1588 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1589 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1590 | return ret; |
| 1591 | } |
| 1592 | |
| 1593 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, |
| 1594 | struct closure *writes, struct gc_stat *gc) |
| 1595 | { |
| 1596 | struct btree *n = NULL; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1597 | int ret = 0; |
| 1598 | bool should_rewrite; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1599 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1600 | should_rewrite = btree_gc_mark_node(b, gc); |
| 1601 | if (should_rewrite) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1602 | n = btree_node_alloc_replacement(b, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1603 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1604 | if (!IS_ERR_OR_NULL(n)) { |
| 1605 | bch_btree_node_write_sync(n); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1606 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1607 | bch_btree_set_root(n); |
| 1608 | btree_node_free(b); |
| 1609 | rw_unlock(true, n); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1610 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1611 | return -EINTR; |
| 1612 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1613 | } |
| 1614 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1615 | __bch_btree_mark_key(b->c, b->level + 1, &b->key); |
| 1616 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1617 | if (b->level) { |
| 1618 | ret = btree_gc_recurse(b, op, writes, gc); |
| 1619 | if (ret) |
| 1620 | return ret; |
| 1621 | } |
| 1622 | |
| 1623 | bkey_copy_key(&b->c->gc_done, &b->key); |
| 1624 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1625 | return ret; |
| 1626 | } |
| 1627 | |
| 1628 | static void btree_gc_start(struct cache_set *c) |
| 1629 | { |
| 1630 | struct cache *ca; |
| 1631 | struct bucket *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1632 | unsigned i; |
| 1633 | |
| 1634 | if (!c->gc_mark_valid) |
| 1635 | return; |
| 1636 | |
| 1637 | mutex_lock(&c->bucket_lock); |
| 1638 | |
| 1639 | c->gc_mark_valid = 0; |
| 1640 | c->gc_done = ZERO_KEY; |
| 1641 | |
| 1642 | for_each_cache(ca, c, i) |
| 1643 | for_each_bucket(b, ca) { |
Kent Overstreet | 3a2fd9d | 2014-02-27 17:51:12 -0800 | [diff] [blame] | 1644 | b->last_gc = b->gen; |
Kent Overstreet | 29ebf46 | 2013-07-11 19:43:21 -0700 | [diff] [blame] | 1645 | if (!atomic_read(&b->pin)) { |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1646 | SET_GC_MARK(b, 0); |
Kent Overstreet | 29ebf46 | 2013-07-11 19:43:21 -0700 | [diff] [blame] | 1647 | SET_GC_SECTORS_USED(b, 0); |
| 1648 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1649 | } |
| 1650 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1651 | mutex_unlock(&c->bucket_lock); |
| 1652 | } |
| 1653 | |
Kent Overstreet | 2531d9ee | 2014-03-17 16:55:55 -0700 | [diff] [blame] | 1654 | static size_t bch_btree_gc_finish(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1655 | { |
| 1656 | size_t available = 0; |
| 1657 | struct bucket *b; |
| 1658 | struct cache *ca; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1659 | unsigned i; |
| 1660 | |
| 1661 | mutex_lock(&c->bucket_lock); |
| 1662 | |
| 1663 | set_gc_sectors(c); |
| 1664 | c->gc_mark_valid = 1; |
| 1665 | c->need_gc = 0; |
| 1666 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1667 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) |
| 1668 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), |
| 1669 | GC_MARK_METADATA); |
| 1670 | |
Nicholas Swenson | bf0a628 | 2013-11-26 19:14:23 -0800 | [diff] [blame] | 1671 | /* don't reclaim buckets to which writeback keys point */ |
| 1672 | rcu_read_lock(); |
| 1673 | for (i = 0; i < c->nr_uuids; i++) { |
| 1674 | struct bcache_device *d = c->devices[i]; |
| 1675 | struct cached_dev *dc; |
| 1676 | struct keybuf_key *w, *n; |
| 1677 | unsigned j; |
| 1678 | |
| 1679 | if (!d || UUID_FLASH_ONLY(&c->uuids[i])) |
| 1680 | continue; |
| 1681 | dc = container_of(d, struct cached_dev, disk); |
| 1682 | |
| 1683 | spin_lock(&dc->writeback_keys.lock); |
| 1684 | rbtree_postorder_for_each_entry_safe(w, n, |
| 1685 | &dc->writeback_keys.keys, node) |
| 1686 | for (j = 0; j < KEY_PTRS(&w->key); j++) |
| 1687 | SET_GC_MARK(PTR_BUCKET(c, &w->key, j), |
| 1688 | GC_MARK_DIRTY); |
| 1689 | spin_unlock(&dc->writeback_keys.lock); |
| 1690 | } |
| 1691 | rcu_read_unlock(); |
| 1692 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1693 | for_each_cache(ca, c, i) { |
| 1694 | uint64_t *i; |
| 1695 | |
| 1696 | ca->invalidate_needs_gc = 0; |
| 1697 | |
| 1698 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) |
| 1699 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); |
| 1700 | |
| 1701 | for (i = ca->prio_buckets; |
| 1702 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) |
| 1703 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); |
| 1704 | |
| 1705 | for_each_bucket(b, ca) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1706 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); |
| 1707 | |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1708 | if (atomic_read(&b->pin)) |
| 1709 | continue; |
| 1710 | |
| 1711 | BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b)); |
| 1712 | |
| 1713 | if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1714 | available++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1715 | } |
| 1716 | } |
| 1717 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1718 | mutex_unlock(&c->bucket_lock); |
| 1719 | return available; |
| 1720 | } |
| 1721 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1722 | static void bch_btree_gc(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1723 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1724 | int ret; |
| 1725 | unsigned long available; |
| 1726 | struct gc_stat stats; |
| 1727 | struct closure writes; |
| 1728 | struct btree_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1729 | uint64_t start_time = local_clock(); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1730 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1731 | trace_bcache_gc_start(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1732 | |
| 1733 | memset(&stats, 0, sizeof(struct gc_stat)); |
| 1734 | closure_init_stack(&writes); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1735 | bch_btree_op_init(&op, SHRT_MAX); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1736 | |
| 1737 | btree_gc_start(c); |
| 1738 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1739 | do { |
| 1740 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
| 1741 | closure_sync(&writes); |
Kent Overstreet | c5f1e5a | 2015-11-29 17:18:33 -0800 | [diff] [blame] | 1742 | cond_resched(); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1743 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1744 | if (ret && ret != -EAGAIN) |
| 1745 | pr_warn("gc failed!"); |
| 1746 | } while (ret); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1747 | |
| 1748 | available = bch_btree_gc_finish(c); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1749 | wake_up_allocators(c); |
| 1750 | |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 1751 | bch_time_stats_update(&c->btree_gc_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1752 | |
| 1753 | stats.key_bytes *= sizeof(uint64_t); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1754 | stats.data <<= 9; |
| 1755 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; |
| 1756 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1757 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1758 | trace_bcache_gc_end(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1759 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1760 | bch_moving_gc(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1761 | } |
| 1762 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1763 | static bool gc_should_run(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1764 | { |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1765 | struct cache *ca; |
| 1766 | unsigned i; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1767 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1768 | for_each_cache(ca, c, i) |
| 1769 | if (ca->invalidate_needs_gc) |
| 1770 | return true; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1771 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1772 | if (atomic_read(&c->sectors_to_gc) < 0) |
| 1773 | return true; |
| 1774 | |
| 1775 | return false; |
| 1776 | } |
| 1777 | |
| 1778 | static int bch_gc_thread(void *arg) |
| 1779 | { |
| 1780 | struct cache_set *c = arg; |
| 1781 | |
| 1782 | while (1) { |
| 1783 | wait_event_interruptible(c->gc_wait, |
| 1784 | kthread_should_stop() || gc_should_run(c)); |
| 1785 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1786 | if (kthread_should_stop()) |
| 1787 | break; |
| 1788 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1789 | set_gc_sectors(c); |
| 1790 | bch_btree_gc(c); |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1791 | } |
| 1792 | |
| 1793 | return 0; |
| 1794 | } |
| 1795 | |
| 1796 | int bch_gc_thread_start(struct cache_set *c) |
| 1797 | { |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1798 | c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc"); |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1799 | if (IS_ERR(c->gc_thread)) |
| 1800 | return PTR_ERR(c->gc_thread); |
| 1801 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1802 | return 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1803 | } |
| 1804 | |
| 1805 | /* Initial partial gc */ |
| 1806 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1807 | static int bch_btree_check_recurse(struct btree *b, struct btree_op *op) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1808 | { |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1809 | int ret = 0; |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1810 | struct bkey *k, *p = NULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1811 | struct btree_iter iter; |
| 1812 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1813 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) |
| 1814 | bch_initial_mark_key(b->c, b->level, k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1815 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1816 | bch_initial_mark_key(b->c, b->level + 1, &b->key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1817 | |
| 1818 | if (b->level) { |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1819 | bch_btree_iter_init(&b->keys, &iter, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1820 | |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1821 | do { |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1822 | k = bch_btree_iter_next_filter(&iter, &b->keys, |
| 1823 | bch_ptr_bad); |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1824 | if (k) |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1825 | btree_node_prefetch(b, k); |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1826 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1827 | if (p) |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1828 | ret = btree(check_recurse, p, b, op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1829 | |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1830 | p = k; |
| 1831 | } while (p && !ret); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1832 | } |
| 1833 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1834 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1835 | } |
| 1836 | |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 1837 | int bch_btree_check(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1838 | { |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 1839 | struct btree_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1840 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1841 | bch_btree_op_init(&op, SHRT_MAX); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1842 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1843 | return btree_root(check_recurse, c, &op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1844 | } |
| 1845 | |
Kent Overstreet | 2531d9ee | 2014-03-17 16:55:55 -0700 | [diff] [blame] | 1846 | void bch_initial_gc_finish(struct cache_set *c) |
| 1847 | { |
| 1848 | struct cache *ca; |
| 1849 | struct bucket *b; |
| 1850 | unsigned i; |
| 1851 | |
| 1852 | bch_btree_gc_finish(c); |
| 1853 | |
| 1854 | mutex_lock(&c->bucket_lock); |
| 1855 | |
| 1856 | /* |
| 1857 | * We need to put some unused buckets directly on the prio freelist in |
| 1858 | * order to get the allocator thread started - it needs freed buckets in |
| 1859 | * order to rewrite the prios and gens, and it needs to rewrite prios |
| 1860 | * and gens in order to free buckets. |
| 1861 | * |
| 1862 | * This is only safe for buckets that have no live data in them, which |
| 1863 | * there should always be some of. |
| 1864 | */ |
| 1865 | for_each_cache(ca, c, i) { |
| 1866 | for_each_bucket(b, ca) { |
| 1867 | if (fifo_full(&ca->free[RESERVE_PRIO])) |
| 1868 | break; |
| 1869 | |
| 1870 | if (bch_can_invalidate_bucket(ca, b) && |
| 1871 | !GC_MARK(b)) { |
| 1872 | __bch_invalidate_one_bucket(ca, b); |
| 1873 | fifo_push(&ca->free[RESERVE_PRIO], |
| 1874 | b - ca->buckets); |
| 1875 | } |
| 1876 | } |
| 1877 | } |
| 1878 | |
| 1879 | mutex_unlock(&c->bucket_lock); |
| 1880 | } |
| 1881 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1882 | /* Btree insertion */ |
| 1883 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1884 | static bool btree_insert_key(struct btree *b, struct bkey *k, |
| 1885 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1886 | { |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1887 | unsigned status; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1888 | |
| 1889 | BUG_ON(bkey_cmp(k, &b->key) > 0); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1890 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1891 | status = bch_btree_insert_key(&b->keys, k, replace_key); |
| 1892 | if (status != BTREE_INSERT_STATUS_NO_INSERT) { |
| 1893 | bch_check_keys(&b->keys, "%u for %s", status, |
| 1894 | replace_key ? "replace" : "insert"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1895 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1896 | trace_bcache_btree_insert_key(b, k, replace_key != NULL, |
| 1897 | status); |
| 1898 | return true; |
| 1899 | } else |
| 1900 | return false; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1901 | } |
| 1902 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1903 | static size_t insert_u64s_remaining(struct btree *b) |
| 1904 | { |
Kent Overstreet | 3572324 | 2014-01-10 18:53:02 -0800 | [diff] [blame] | 1905 | long ret = bch_btree_keys_u64s_remaining(&b->keys); |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1906 | |
| 1907 | /* |
| 1908 | * Might land in the middle of an existing extent and have to split it |
| 1909 | */ |
| 1910 | if (b->keys.ops->is_extents) |
| 1911 | ret -= KEY_MAX_U64S; |
| 1912 | |
| 1913 | return max(ret, 0L); |
| 1914 | } |
| 1915 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1916 | static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 1917 | struct keylist *insert_keys, |
| 1918 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1919 | { |
| 1920 | bool ret = false; |
Kent Overstreet | dc9d98d | 2013-12-17 23:47:33 -0800 | [diff] [blame] | 1921 | int oldsize = bch_count_data(&b->keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1922 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1923 | while (!bch_keylist_empty(insert_keys)) { |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1924 | struct bkey *k = insert_keys->keys; |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1925 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1926 | if (bkey_u64s(k) > insert_u64s_remaining(b)) |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 1927 | break; |
| 1928 | |
| 1929 | if (bkey_cmp(k, &b->key) <= 0) { |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 1930 | if (!b->level) |
| 1931 | bkey_put(b->c, k); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1932 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1933 | ret |= btree_insert_key(b, k, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1934 | bch_keylist_pop_front(insert_keys); |
| 1935 | } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1936 | BKEY_PADDED(key) temp; |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1937 | bkey_copy(&temp.key, insert_keys->keys); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1938 | |
| 1939 | bch_cut_back(&b->key, &temp.key); |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1940 | bch_cut_front(&b->key, insert_keys->keys); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1941 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1942 | ret |= btree_insert_key(b, &temp.key, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1943 | break; |
| 1944 | } else { |
| 1945 | break; |
| 1946 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1947 | } |
| 1948 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1949 | if (!ret) |
| 1950 | op->insert_collision = true; |
| 1951 | |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 1952 | BUG_ON(!bch_keylist_empty(insert_keys) && b->level); |
| 1953 | |
Kent Overstreet | dc9d98d | 2013-12-17 23:47:33 -0800 | [diff] [blame] | 1954 | BUG_ON(bch_count_data(&b->keys) < oldsize); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1955 | return ret; |
| 1956 | } |
| 1957 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1958 | static int btree_split(struct btree *b, struct btree_op *op, |
| 1959 | struct keylist *insert_keys, |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 1960 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1961 | { |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 1962 | bool split; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1963 | struct btree *n1, *n2 = NULL, *n3 = NULL; |
| 1964 | uint64_t start_time = local_clock(); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1965 | struct closure cl; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 1966 | struct keylist parent_keys; |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1967 | |
| 1968 | closure_init_stack(&cl); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 1969 | bch_keylist_init(&parent_keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1970 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1971 | if (btree_check_reserve(b, op)) { |
| 1972 | if (!b->level) |
| 1973 | return -EINTR; |
| 1974 | else |
| 1975 | WARN(1, "insufficient reserve for split\n"); |
| 1976 | } |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1977 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1978 | n1 = btree_node_alloc_replacement(b, op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1979 | if (IS_ERR(n1)) |
| 1980 | goto err; |
| 1981 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1982 | split = set_blocks(btree_bset_first(n1), |
| 1983 | block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1984 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1985 | if (split) { |
| 1986 | unsigned keys = 0; |
| 1987 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1988 | trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1989 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1990 | n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1991 | if (IS_ERR(n2)) |
| 1992 | goto err_free1; |
| 1993 | |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 1994 | if (!b->parent) { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1995 | n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1996 | if (IS_ERR(n3)) |
| 1997 | goto err_free2; |
| 1998 | } |
| 1999 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2000 | mutex_lock(&n1->write_lock); |
| 2001 | mutex_lock(&n2->write_lock); |
| 2002 | |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2003 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2004 | |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2005 | /* |
| 2006 | * Has to be a linear search because we don't have an auxiliary |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2007 | * search tree yet |
| 2008 | */ |
| 2009 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2010 | while (keys < (btree_bset_first(n1)->keys * 3) / 5) |
| 2011 | keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 2012 | keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2013 | |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 2014 | bkey_copy_key(&n1->key, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2015 | bset_bkey_idx(btree_bset_first(n1), keys)); |
| 2016 | keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2017 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2018 | btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys; |
| 2019 | btree_bset_first(n1)->keys = keys; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2020 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2021 | memcpy(btree_bset_first(n2)->start, |
| 2022 | bset_bkey_last(btree_bset_first(n1)), |
| 2023 | btree_bset_first(n2)->keys * sizeof(uint64_t)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2024 | |
| 2025 | bkey_copy_key(&n2->key, &b->key); |
| 2026 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2027 | bch_keylist_add(&parent_keys, &n2->key); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2028 | bch_btree_node_write(n2, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2029 | mutex_unlock(&n2->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2030 | rw_unlock(true, n2); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2031 | } else { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2032 | trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2033 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2034 | mutex_lock(&n1->write_lock); |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2035 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2036 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2037 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2038 | bch_keylist_add(&parent_keys, &n1->key); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2039 | bch_btree_node_write(n1, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2040 | mutex_unlock(&n1->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2041 | |
| 2042 | if (n3) { |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2043 | /* Depth increases, make a new root */ |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2044 | mutex_lock(&n3->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2045 | bkey_copy_key(&n3->key, &MAX_KEY); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2046 | bch_btree_insert_keys(n3, op, &parent_keys, NULL); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2047 | bch_btree_node_write(n3, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2048 | mutex_unlock(&n3->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2049 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2050 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2051 | bch_btree_set_root(n3); |
| 2052 | rw_unlock(true, n3); |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2053 | } else if (!b->parent) { |
| 2054 | /* Root filled up but didn't need to be split */ |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2055 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2056 | bch_btree_set_root(n1); |
| 2057 | } else { |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2058 | /* Split a non root node */ |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2059 | closure_sync(&cl); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2060 | make_btree_freeing_key(b, parent_keys.top); |
| 2061 | bch_keylist_push(&parent_keys); |
| 2062 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2063 | bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL); |
| 2064 | BUG_ON(!bch_keylist_empty(&parent_keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2065 | } |
| 2066 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 2067 | btree_node_free(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2068 | rw_unlock(true, n1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2069 | |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 2070 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2071 | |
| 2072 | return 0; |
| 2073 | err_free2: |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2074 | bkey_put(b->c, &n2->key); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 2075 | btree_node_free(n2); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2076 | rw_unlock(true, n2); |
| 2077 | err_free1: |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2078 | bkey_put(b->c, &n1->key); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 2079 | btree_node_free(n1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2080 | rw_unlock(true, n1); |
| 2081 | err: |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 2082 | WARN(1, "bcache: btree split failed (level %u)", b->level); |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2083 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2084 | if (n3 == ERR_PTR(-EAGAIN) || |
| 2085 | n2 == ERR_PTR(-EAGAIN) || |
| 2086 | n1 == ERR_PTR(-EAGAIN)) |
| 2087 | return -EAGAIN; |
| 2088 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2089 | return -ENOMEM; |
| 2090 | } |
| 2091 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2092 | static int bch_btree_insert_node(struct btree *b, struct btree_op *op, |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 2093 | struct keylist *insert_keys, |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2094 | atomic_t *journal_ref, |
| 2095 | struct bkey *replace_key) |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2096 | { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2097 | struct closure cl; |
| 2098 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2099 | BUG_ON(b->level && replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2100 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2101 | closure_init_stack(&cl); |
| 2102 | |
| 2103 | mutex_lock(&b->write_lock); |
| 2104 | |
| 2105 | if (write_block(b) != btree_bset_last(b) && |
| 2106 | b->keys.last_set_unwritten) |
| 2107 | bch_btree_init_next(b); /* just wrote a set */ |
| 2108 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 2109 | if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2110 | mutex_unlock(&b->write_lock); |
| 2111 | goto split; |
| 2112 | } |
Kent Overstreet | 3b3e9e5 | 2013-12-07 03:57:58 -0800 | [diff] [blame] | 2113 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2114 | BUG_ON(write_block(b) != btree_bset_last(b)); |
| 2115 | |
| 2116 | if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) { |
| 2117 | if (!b->level) |
| 2118 | bch_btree_leaf_dirty(b, journal_ref); |
| 2119 | else |
| 2120 | bch_btree_node_write(b, &cl); |
| 2121 | } |
| 2122 | |
| 2123 | mutex_unlock(&b->write_lock); |
| 2124 | |
| 2125 | /* wait for btree node write if necessary, after unlock */ |
| 2126 | closure_sync(&cl); |
| 2127 | |
| 2128 | return 0; |
| 2129 | split: |
| 2130 | if (current->bio_list) { |
| 2131 | op->lock = b->c->root->level + 1; |
| 2132 | return -EAGAIN; |
| 2133 | } else if (op->lock <= b->c->root->level) { |
| 2134 | op->lock = b->c->root->level + 1; |
| 2135 | return -EINTR; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2136 | } else { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2137 | /* Invalidated all iterators */ |
| 2138 | int ret = btree_split(b, op, insert_keys, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2139 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2140 | if (bch_keylist_empty(insert_keys)) |
| 2141 | return 0; |
| 2142 | else if (!ret) |
| 2143 | return -EINTR; |
| 2144 | return ret; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2145 | } |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2146 | } |
| 2147 | |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2148 | int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, |
| 2149 | struct bkey *check_key) |
| 2150 | { |
| 2151 | int ret = -EINTR; |
| 2152 | uint64_t btree_ptr = b->key.ptr[0]; |
| 2153 | unsigned long seq = b->seq; |
| 2154 | struct keylist insert; |
| 2155 | bool upgrade = op->lock == -1; |
| 2156 | |
| 2157 | bch_keylist_init(&insert); |
| 2158 | |
| 2159 | if (upgrade) { |
| 2160 | rw_unlock(false, b); |
| 2161 | rw_lock(true, b, b->level); |
| 2162 | |
| 2163 | if (b->key.ptr[0] != btree_ptr || |
Zheng Liu | 2ef9ccb | 2015-11-29 17:17:05 -0800 | [diff] [blame] | 2164 | b->seq != seq + 1) { |
| 2165 | op->lock = b->level; |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2166 | goto out; |
Zheng Liu | 2ef9ccb | 2015-11-29 17:17:05 -0800 | [diff] [blame] | 2167 | } |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2168 | } |
| 2169 | |
| 2170 | SET_KEY_PTRS(check_key, 1); |
| 2171 | get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); |
| 2172 | |
| 2173 | SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); |
| 2174 | |
| 2175 | bch_keylist_add(&insert, check_key); |
| 2176 | |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2177 | ret = bch_btree_insert_node(b, op, &insert, NULL, NULL); |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2178 | |
| 2179 | BUG_ON(!ret && !bch_keylist_empty(&insert)); |
| 2180 | out: |
| 2181 | if (upgrade) |
| 2182 | downgrade_write(&b->lock); |
| 2183 | return ret; |
| 2184 | } |
| 2185 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2186 | struct btree_insert_op { |
| 2187 | struct btree_op op; |
| 2188 | struct keylist *keys; |
| 2189 | atomic_t *journal_ref; |
| 2190 | struct bkey *replace_key; |
| 2191 | }; |
| 2192 | |
Wei Yongjun | 08239ca | 2013-11-28 10:31:35 +0800 | [diff] [blame] | 2193 | static int btree_insert_fn(struct btree_op *b_op, struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2194 | { |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2195 | struct btree_insert_op *op = container_of(b_op, |
| 2196 | struct btree_insert_op, op); |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 2197 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2198 | int ret = bch_btree_insert_node(b, &op->op, op->keys, |
| 2199 | op->journal_ref, op->replace_key); |
| 2200 | if (ret && !bch_keylist_empty(op->keys)) |
| 2201 | return ret; |
| 2202 | else |
| 2203 | return MAP_DONE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2204 | } |
| 2205 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2206 | int bch_btree_insert(struct cache_set *c, struct keylist *keys, |
| 2207 | atomic_t *journal_ref, struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2208 | { |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2209 | struct btree_insert_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2210 | int ret = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2211 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2212 | BUG_ON(current->bio_list); |
Kent Overstreet | 4f3d401 | 2013-09-10 18:46:36 -0700 | [diff] [blame] | 2213 | BUG_ON(bch_keylist_empty(keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2214 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2215 | bch_btree_op_init(&op.op, 0); |
| 2216 | op.keys = keys; |
| 2217 | op.journal_ref = journal_ref; |
| 2218 | op.replace_key = replace_key; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2219 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2220 | while (!ret && !bch_keylist_empty(keys)) { |
| 2221 | op.op.lock = 0; |
| 2222 | ret = bch_btree_map_leaf_nodes(&op.op, c, |
| 2223 | &START_KEY(keys->keys), |
| 2224 | btree_insert_fn); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2225 | } |
| 2226 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2227 | if (ret) { |
| 2228 | struct bkey *k; |
| 2229 | |
| 2230 | pr_err("error %i", ret); |
| 2231 | |
| 2232 | while ((k = bch_keylist_pop(keys))) |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 2233 | bkey_put(c, k); |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2234 | } else if (op.op.insert_collision) |
| 2235 | ret = -ESRCH; |
Kent Overstreet | 6054c6d | 2013-07-24 18:06:22 -0700 | [diff] [blame] | 2236 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2237 | return ret; |
| 2238 | } |
| 2239 | |
| 2240 | void bch_btree_set_root(struct btree *b) |
| 2241 | { |
| 2242 | unsigned i; |
Kent Overstreet | e49c7c3 | 2013-06-26 17:25:38 -0700 | [diff] [blame] | 2243 | struct closure cl; |
| 2244 | |
| 2245 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2246 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2247 | trace_bcache_btree_set_root(b); |
| 2248 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2249 | BUG_ON(!b->written); |
| 2250 | |
| 2251 | for (i = 0; i < KEY_PTRS(&b->key); i++) |
| 2252 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); |
| 2253 | |
| 2254 | mutex_lock(&b->c->bucket_lock); |
| 2255 | list_del_init(&b->list); |
| 2256 | mutex_unlock(&b->c->bucket_lock); |
| 2257 | |
| 2258 | b->c->root = b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2259 | |
Kent Overstreet | e49c7c3 | 2013-06-26 17:25:38 -0700 | [diff] [blame] | 2260 | bch_journal_meta(b->c, &cl); |
| 2261 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2262 | } |
| 2263 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2264 | /* Map across nodes or keys */ |
| 2265 | |
| 2266 | static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, |
| 2267 | struct bkey *from, |
| 2268 | btree_map_nodes_fn *fn, int flags) |
| 2269 | { |
| 2270 | int ret = MAP_CONTINUE; |
| 2271 | |
| 2272 | if (b->level) { |
| 2273 | struct bkey *k; |
| 2274 | struct btree_iter iter; |
| 2275 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 2276 | bch_btree_iter_init(&b->keys, &iter, from); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2277 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 2278 | while ((k = bch_btree_iter_next_filter(&iter, &b->keys, |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2279 | bch_ptr_bad))) { |
| 2280 | ret = btree(map_nodes_recurse, k, b, |
| 2281 | op, from, fn, flags); |
| 2282 | from = NULL; |
| 2283 | |
| 2284 | if (ret != MAP_CONTINUE) |
| 2285 | return ret; |
| 2286 | } |
| 2287 | } |
| 2288 | |
| 2289 | if (!b->level || flags == MAP_ALL_NODES) |
| 2290 | ret = fn(op, b); |
| 2291 | |
| 2292 | return ret; |
| 2293 | } |
| 2294 | |
| 2295 | int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, |
| 2296 | struct bkey *from, btree_map_nodes_fn *fn, int flags) |
| 2297 | { |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2298 | return btree_root(map_nodes_recurse, c, op, from, fn, flags); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2299 | } |
| 2300 | |
| 2301 | static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, |
| 2302 | struct bkey *from, btree_map_keys_fn *fn, |
| 2303 | int flags) |
| 2304 | { |
| 2305 | int ret = MAP_CONTINUE; |
| 2306 | struct bkey *k; |
| 2307 | struct btree_iter iter; |
| 2308 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 2309 | bch_btree_iter_init(&b->keys, &iter, from); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2310 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 2311 | while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) { |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2312 | ret = !b->level |
| 2313 | ? fn(op, b, k) |
| 2314 | : btree(map_keys_recurse, k, b, op, from, fn, flags); |
| 2315 | from = NULL; |
| 2316 | |
| 2317 | if (ret != MAP_CONTINUE) |
| 2318 | return ret; |
| 2319 | } |
| 2320 | |
| 2321 | if (!b->level && (flags & MAP_END_KEY)) |
| 2322 | ret = fn(op, b, &KEY(KEY_INODE(&b->key), |
| 2323 | KEY_OFFSET(&b->key), 0)); |
| 2324 | |
| 2325 | return ret; |
| 2326 | } |
| 2327 | |
| 2328 | int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, |
| 2329 | struct bkey *from, btree_map_keys_fn *fn, int flags) |
| 2330 | { |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2331 | return btree_root(map_keys_recurse, c, op, from, fn, flags); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2332 | } |
| 2333 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2334 | /* Keybuf code */ |
| 2335 | |
| 2336 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) |
| 2337 | { |
| 2338 | /* Overlapping keys compare equal */ |
| 2339 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) |
| 2340 | return -1; |
| 2341 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) |
| 2342 | return 1; |
| 2343 | return 0; |
| 2344 | } |
| 2345 | |
| 2346 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, |
| 2347 | struct keybuf_key *r) |
| 2348 | { |
| 2349 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); |
| 2350 | } |
| 2351 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2352 | struct refill { |
| 2353 | struct btree_op op; |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2354 | unsigned nr_found; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2355 | struct keybuf *buf; |
| 2356 | struct bkey *end; |
| 2357 | keybuf_pred_fn *pred; |
| 2358 | }; |
| 2359 | |
| 2360 | static int refill_keybuf_fn(struct btree_op *op, struct btree *b, |
| 2361 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2362 | { |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2363 | struct refill *refill = container_of(op, struct refill, op); |
| 2364 | struct keybuf *buf = refill->buf; |
| 2365 | int ret = MAP_CONTINUE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2366 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2367 | if (bkey_cmp(k, refill->end) >= 0) { |
| 2368 | ret = MAP_DONE; |
| 2369 | goto out; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2370 | } |
| 2371 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2372 | if (!KEY_SIZE(k)) /* end key */ |
| 2373 | goto out; |
| 2374 | |
| 2375 | if (refill->pred(buf, k)) { |
| 2376 | struct keybuf_key *w; |
| 2377 | |
| 2378 | spin_lock(&buf->lock); |
| 2379 | |
| 2380 | w = array_alloc(&buf->freelist); |
| 2381 | if (!w) { |
| 2382 | spin_unlock(&buf->lock); |
| 2383 | return MAP_DONE; |
| 2384 | } |
| 2385 | |
| 2386 | w->private = NULL; |
| 2387 | bkey_copy(&w->key, k); |
| 2388 | |
| 2389 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) |
| 2390 | array_free(&buf->freelist, w); |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2391 | else |
| 2392 | refill->nr_found++; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2393 | |
| 2394 | if (array_freelist_empty(&buf->freelist)) |
| 2395 | ret = MAP_DONE; |
| 2396 | |
| 2397 | spin_unlock(&buf->lock); |
| 2398 | } |
| 2399 | out: |
| 2400 | buf->last_scanned = *k; |
| 2401 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2402 | } |
| 2403 | |
| 2404 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2405 | struct bkey *end, keybuf_pred_fn *pred) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2406 | { |
| 2407 | struct bkey start = buf->last_scanned; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2408 | struct refill refill; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2409 | |
| 2410 | cond_resched(); |
| 2411 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2412 | bch_btree_op_init(&refill.op, -1); |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2413 | refill.nr_found = 0; |
| 2414 | refill.buf = buf; |
| 2415 | refill.end = end; |
| 2416 | refill.pred = pred; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2417 | |
| 2418 | bch_btree_map_keys(&refill.op, c, &buf->last_scanned, |
| 2419 | refill_keybuf_fn, MAP_END_KEY); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2420 | |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2421 | trace_bcache_keyscan(refill.nr_found, |
| 2422 | KEY_INODE(&start), KEY_OFFSET(&start), |
| 2423 | KEY_INODE(&buf->last_scanned), |
| 2424 | KEY_OFFSET(&buf->last_scanned)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2425 | |
| 2426 | spin_lock(&buf->lock); |
| 2427 | |
| 2428 | if (!RB_EMPTY_ROOT(&buf->keys)) { |
| 2429 | struct keybuf_key *w; |
| 2430 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); |
| 2431 | buf->start = START_KEY(&w->key); |
| 2432 | |
| 2433 | w = RB_LAST(&buf->keys, struct keybuf_key, node); |
| 2434 | buf->end = w->key; |
| 2435 | } else { |
| 2436 | buf->start = MAX_KEY; |
| 2437 | buf->end = MAX_KEY; |
| 2438 | } |
| 2439 | |
| 2440 | spin_unlock(&buf->lock); |
| 2441 | } |
| 2442 | |
| 2443 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) |
| 2444 | { |
| 2445 | rb_erase(&w->node, &buf->keys); |
| 2446 | array_free(&buf->freelist, w); |
| 2447 | } |
| 2448 | |
| 2449 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) |
| 2450 | { |
| 2451 | spin_lock(&buf->lock); |
| 2452 | __bch_keybuf_del(buf, w); |
| 2453 | spin_unlock(&buf->lock); |
| 2454 | } |
| 2455 | |
| 2456 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, |
| 2457 | struct bkey *end) |
| 2458 | { |
| 2459 | bool ret = false; |
| 2460 | struct keybuf_key *p, *w, s; |
| 2461 | s.key = *start; |
| 2462 | |
| 2463 | if (bkey_cmp(end, &buf->start) <= 0 || |
| 2464 | bkey_cmp(start, &buf->end) >= 0) |
| 2465 | return false; |
| 2466 | |
| 2467 | spin_lock(&buf->lock); |
| 2468 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); |
| 2469 | |
| 2470 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { |
| 2471 | p = w; |
| 2472 | w = RB_NEXT(w, node); |
| 2473 | |
| 2474 | if (p->private) |
| 2475 | ret = true; |
| 2476 | else |
| 2477 | __bch_keybuf_del(buf, p); |
| 2478 | } |
| 2479 | |
| 2480 | spin_unlock(&buf->lock); |
| 2481 | return ret; |
| 2482 | } |
| 2483 | |
| 2484 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) |
| 2485 | { |
| 2486 | struct keybuf_key *w; |
| 2487 | spin_lock(&buf->lock); |
| 2488 | |
| 2489 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); |
| 2490 | |
| 2491 | while (w && w->private) |
| 2492 | w = RB_NEXT(w, node); |
| 2493 | |
| 2494 | if (w) |
| 2495 | w->private = ERR_PTR(-EINTR); |
| 2496 | |
| 2497 | spin_unlock(&buf->lock); |
| 2498 | return w; |
| 2499 | } |
| 2500 | |
| 2501 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2502 | struct keybuf *buf, |
| 2503 | struct bkey *end, |
| 2504 | keybuf_pred_fn *pred) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2505 | { |
| 2506 | struct keybuf_key *ret; |
| 2507 | |
| 2508 | while (1) { |
| 2509 | ret = bch_keybuf_next(buf); |
| 2510 | if (ret) |
| 2511 | break; |
| 2512 | |
| 2513 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { |
| 2514 | pr_debug("scan finished"); |
| 2515 | break; |
| 2516 | } |
| 2517 | |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2518 | bch_refill_keybuf(c, buf, end, pred); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2519 | } |
| 2520 | |
| 2521 | return ret; |
| 2522 | } |
| 2523 | |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2524 | void bch_keybuf_init(struct keybuf *buf) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2525 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2526 | buf->last_scanned = MAX_KEY; |
| 2527 | buf->keys = RB_ROOT; |
| 2528 | |
| 2529 | spin_lock_init(&buf->lock); |
| 2530 | array_allocator_init(&buf->freelist); |
| 2531 | } |