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; |
| 806 | register_shrinker(&c->shrink); |
| 807 | |
| 808 | return 0; |
| 809 | } |
| 810 | |
| 811 | /* Btree in memory cache - hash table */ |
| 812 | |
| 813 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) |
| 814 | { |
| 815 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; |
| 816 | } |
| 817 | |
| 818 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) |
| 819 | { |
| 820 | struct btree *b; |
| 821 | |
| 822 | rcu_read_lock(); |
| 823 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) |
| 824 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) |
| 825 | goto out; |
| 826 | b = NULL; |
| 827 | out: |
| 828 | rcu_read_unlock(); |
| 829 | return b; |
| 830 | } |
| 831 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 832 | static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op) |
| 833 | { |
| 834 | struct task_struct *old; |
| 835 | |
| 836 | old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current); |
| 837 | if (old && old != current) { |
| 838 | if (op) |
| 839 | prepare_to_wait(&c->btree_cache_wait, &op->wait, |
| 840 | TASK_UNINTERRUPTIBLE); |
| 841 | return -EINTR; |
| 842 | } |
| 843 | |
| 844 | return 0; |
| 845 | } |
| 846 | |
| 847 | static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op, |
| 848 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 849 | { |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 850 | struct btree *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 851 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 852 | trace_bcache_btree_cache_cannibalize(c); |
| 853 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 854 | if (mca_cannibalize_lock(c, op)) |
| 855 | return ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 856 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 857 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
| 858 | if (!mca_reap(b, btree_order(k), false)) |
| 859 | return b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 860 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 861 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
| 862 | if (!mca_reap(b, btree_order(k), true)) |
| 863 | return b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 864 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 865 | WARN(1, "btree cache cannibalize failed\n"); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 866 | return ERR_PTR(-ENOMEM); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 867 | } |
| 868 | |
| 869 | /* |
| 870 | * We can only have one thread cannibalizing other cached btree nodes at a time, |
| 871 | * or we'll deadlock. We use an open coded mutex to ensure that, which a |
| 872 | * cannibalize_bucket() will take. This means every time we unlock the root of |
| 873 | * the btree, we need to release this lock if we have it held. |
| 874 | */ |
Kent Overstreet | df8e897 | 2013-07-24 17:37:59 -0700 | [diff] [blame] | 875 | static void bch_cannibalize_unlock(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 876 | { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 877 | if (c->btree_cache_alloc_lock == current) { |
| 878 | c->btree_cache_alloc_lock = NULL; |
| 879 | wake_up(&c->btree_cache_wait); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 880 | } |
| 881 | } |
| 882 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 883 | static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op, |
| 884 | struct bkey *k, int level) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 885 | { |
| 886 | struct btree *b; |
| 887 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 888 | BUG_ON(current->bio_list); |
| 889 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 890 | lockdep_assert_held(&c->bucket_lock); |
| 891 | |
| 892 | if (mca_find(c, k)) |
| 893 | return NULL; |
| 894 | |
| 895 | /* btree_free() doesn't free memory; it sticks the node on the end of |
| 896 | * the list. Check if there's any freed nodes there: |
| 897 | */ |
| 898 | list_for_each_entry(b, &c->btree_cache_freeable, list) |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 899 | if (!mca_reap(b, btree_order(k), false)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 900 | goto out; |
| 901 | |
| 902 | /* We never free struct btree itself, just the memory that holds the on |
| 903 | * disk node. Check the freed list before allocating a new one: |
| 904 | */ |
| 905 | list_for_each_entry(b, &c->btree_cache_freed, list) |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 906 | if (!mca_reap(b, 0, false)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 907 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 908 | if (!b->keys.set[0].data) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 909 | goto err; |
| 910 | else |
| 911 | goto out; |
| 912 | } |
| 913 | |
| 914 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); |
| 915 | if (!b) |
| 916 | goto err; |
| 917 | |
| 918 | BUG_ON(!down_write_trylock(&b->lock)); |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 919 | if (!b->keys.set->data) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 920 | goto err; |
| 921 | out: |
Kent Overstreet | cb7a583 | 2013-12-16 15:27:25 -0800 | [diff] [blame] | 922 | BUG_ON(b->io_mutex.count != 1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 923 | |
| 924 | bkey_copy(&b->key, k); |
| 925 | list_move(&b->list, &c->btree_cache); |
| 926 | hlist_del_init_rcu(&b->hash); |
| 927 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); |
| 928 | |
| 929 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 930 | b->parent = (void *) ~0UL; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 931 | b->flags = 0; |
| 932 | b->written = 0; |
| 933 | b->level = level; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 934 | |
Kent Overstreet | 65d4523 | 2013-12-20 17:22:05 -0800 | [diff] [blame] | 935 | if (!b->level) |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 936 | bch_btree_keys_init(&b->keys, &bch_extent_keys_ops, |
| 937 | &b->c->expensive_debug_checks); |
Kent Overstreet | 65d4523 | 2013-12-20 17:22:05 -0800 | [diff] [blame] | 938 | else |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 939 | bch_btree_keys_init(&b->keys, &bch_btree_keys_ops, |
| 940 | &b->c->expensive_debug_checks); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 941 | |
| 942 | return b; |
| 943 | err: |
| 944 | if (b) |
| 945 | rw_unlock(true, b); |
| 946 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 947 | b = mca_cannibalize(c, op, k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 948 | if (!IS_ERR(b)) |
| 949 | goto out; |
| 950 | |
| 951 | return b; |
| 952 | } |
| 953 | |
| 954 | /** |
| 955 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it |
| 956 | * in from disk if necessary. |
| 957 | * |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 958 | * If IO is necessary and running under generic_make_request, returns -EAGAIN. |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 959 | * |
| 960 | * The btree node will have either a read or a write lock held, depending on |
| 961 | * level and op->lock. |
| 962 | */ |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 963 | 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] | 964 | struct bkey *k, int level, bool write, |
| 965 | struct btree *parent) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 966 | { |
| 967 | int i = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 968 | struct btree *b; |
| 969 | |
| 970 | BUG_ON(level < 0); |
| 971 | retry: |
| 972 | b = mca_find(c, k); |
| 973 | |
| 974 | if (!b) { |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 975 | if (current->bio_list) |
| 976 | return ERR_PTR(-EAGAIN); |
| 977 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 978 | mutex_lock(&c->bucket_lock); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 979 | b = mca_alloc(c, op, k, level); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 980 | mutex_unlock(&c->bucket_lock); |
| 981 | |
| 982 | if (!b) |
| 983 | goto retry; |
| 984 | if (IS_ERR(b)) |
| 985 | return b; |
| 986 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 987 | bch_btree_node_read(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 988 | |
| 989 | if (!write) |
| 990 | downgrade_write(&b->lock); |
| 991 | } else { |
| 992 | rw_lock(write, b, level); |
| 993 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { |
| 994 | rw_unlock(write, b); |
| 995 | goto retry; |
| 996 | } |
| 997 | BUG_ON(b->level != level); |
| 998 | } |
| 999 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1000 | b->parent = parent; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1001 | b->accessed = 1; |
| 1002 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1003 | for (; i <= b->keys.nsets && b->keys.set[i].size; i++) { |
| 1004 | prefetch(b->keys.set[i].tree); |
| 1005 | prefetch(b->keys.set[i].data); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1006 | } |
| 1007 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1008 | for (; i <= b->keys.nsets; i++) |
| 1009 | prefetch(b->keys.set[i].data); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1010 | |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1011 | if (btree_node_io_error(b)) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1012 | rw_unlock(write, b); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1013 | return ERR_PTR(-EIO); |
| 1014 | } |
| 1015 | |
| 1016 | BUG_ON(!b->written); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1017 | |
| 1018 | return b; |
| 1019 | } |
| 1020 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1021 | static void btree_node_prefetch(struct btree *parent, struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1022 | { |
| 1023 | struct btree *b; |
| 1024 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1025 | mutex_lock(&parent->c->bucket_lock); |
| 1026 | b = mca_alloc(parent->c, NULL, k, parent->level - 1); |
| 1027 | mutex_unlock(&parent->c->bucket_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1028 | |
| 1029 | if (!IS_ERR_OR_NULL(b)) { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1030 | b->parent = parent; |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1031 | bch_btree_node_read(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1032 | rw_unlock(true, b); |
| 1033 | } |
| 1034 | } |
| 1035 | |
| 1036 | /* Btree alloc */ |
| 1037 | |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 1038 | static void btree_node_free(struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1039 | { |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1040 | trace_bcache_btree_node_free(b); |
| 1041 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1042 | BUG_ON(b == b->c->root); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1043 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1044 | mutex_lock(&b->write_lock); |
| 1045 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1046 | if (btree_node_dirty(b)) |
| 1047 | btree_complete_write(b, btree_current_write(b)); |
| 1048 | clear_bit(BTREE_NODE_dirty, &b->flags); |
| 1049 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1050 | mutex_unlock(&b->write_lock); |
| 1051 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1052 | cancel_delayed_work(&b->work); |
| 1053 | |
| 1054 | mutex_lock(&b->c->bucket_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1055 | bch_bucket_free(b->c, &b->key); |
| 1056 | mca_bucket_free(b); |
| 1057 | mutex_unlock(&b->c->bucket_lock); |
| 1058 | } |
| 1059 | |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1060 | 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] | 1061 | int level, bool wait, |
| 1062 | struct btree *parent) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1063 | { |
| 1064 | BKEY_PADDED(key) k; |
| 1065 | struct btree *b = ERR_PTR(-EAGAIN); |
| 1066 | |
| 1067 | mutex_lock(&c->bucket_lock); |
| 1068 | retry: |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1069 | if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1070 | goto err; |
| 1071 | |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 1072 | bkey_put(c, &k.key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1073 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); |
| 1074 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1075 | b = mca_alloc(c, op, &k.key, level); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1076 | if (IS_ERR(b)) |
| 1077 | goto err_free; |
| 1078 | |
| 1079 | if (!b) { |
Kent Overstreet | b1a67b0 | 2013-03-25 11:46:44 -0700 | [diff] [blame] | 1080 | cache_bug(c, |
| 1081 | "Tried to allocate bucket that was in btree cache"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1082 | goto retry; |
| 1083 | } |
| 1084 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1085 | b->accessed = 1; |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1086 | b->parent = parent; |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1087 | 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] | 1088 | |
| 1089 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1090 | |
| 1091 | trace_bcache_btree_node_alloc(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1092 | return b; |
| 1093 | err_free: |
| 1094 | bch_bucket_free(c, &k.key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1095 | err: |
| 1096 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1097 | |
Slava Pestov | 913dc33 | 2014-05-23 11:18:35 -0700 | [diff] [blame] | 1098 | trace_bcache_btree_node_alloc_fail(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1099 | return b; |
| 1100 | } |
| 1101 | |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1102 | static struct btree *bch_btree_node_alloc(struct cache_set *c, |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1103 | struct btree_op *op, int level, |
| 1104 | struct btree *parent) |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1105 | { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1106 | return __bch_btree_node_alloc(c, op, level, op != NULL, parent); |
Slava Pestov | c5aa4a3 | 2014-04-21 18:23:12 -0700 | [diff] [blame] | 1107 | } |
| 1108 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1109 | static struct btree *btree_node_alloc_replacement(struct btree *b, |
| 1110 | struct btree_op *op) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1111 | { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1112 | 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] | 1113 | if (!IS_ERR_OR_NULL(n)) { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1114 | mutex_lock(&n->write_lock); |
Kent Overstreet | 89ebb4a | 2013-11-11 18:38:51 -0800 | [diff] [blame] | 1115 | bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort); |
Kent Overstreet | 67539e8 | 2013-09-10 22:53:34 -0700 | [diff] [blame] | 1116 | bkey_copy_key(&n->key, &b->key); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1117 | mutex_unlock(&n->write_lock); |
Kent Overstreet | 67539e8 | 2013-09-10 22:53:34 -0700 | [diff] [blame] | 1118 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1119 | |
| 1120 | return n; |
| 1121 | } |
| 1122 | |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1123 | static void make_btree_freeing_key(struct btree *b, struct bkey *k) |
| 1124 | { |
| 1125 | unsigned i; |
| 1126 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1127 | mutex_lock(&b->c->bucket_lock); |
| 1128 | |
| 1129 | atomic_inc(&b->c->prio_blocked); |
| 1130 | |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1131 | bkey_copy(k, &b->key); |
| 1132 | bkey_copy_key(k, &ZERO_KEY); |
| 1133 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1134 | for (i = 0; i < KEY_PTRS(k); i++) |
| 1135 | SET_PTR_GEN(k, i, |
| 1136 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), |
| 1137 | PTR_BUCKET(b->c, &b->key, i))); |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1138 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1139 | mutex_unlock(&b->c->bucket_lock); |
Kent Overstreet | 8835c12 | 2013-07-24 23:18:05 -0700 | [diff] [blame] | 1140 | } |
| 1141 | |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1142 | static int btree_check_reserve(struct btree *b, struct btree_op *op) |
| 1143 | { |
| 1144 | struct cache_set *c = b->c; |
| 1145 | struct cache *ca; |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1146 | unsigned i, reserve = (c->root->level - b->level) * 2 + 1; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1147 | |
| 1148 | mutex_lock(&c->bucket_lock); |
| 1149 | |
| 1150 | for_each_cache(ca, c, i) |
| 1151 | if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) { |
| 1152 | if (op) |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1153 | prepare_to_wait(&c->btree_cache_wait, &op->wait, |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1154 | TASK_UNINTERRUPTIBLE); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1155 | mutex_unlock(&c->bucket_lock); |
| 1156 | return -EINTR; |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1157 | } |
| 1158 | |
| 1159 | mutex_unlock(&c->bucket_lock); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1160 | |
| 1161 | return mca_cannibalize_lock(b->c, op); |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1162 | } |
| 1163 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1164 | /* Garbage collection */ |
| 1165 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1166 | static uint8_t __bch_btree_mark_key(struct cache_set *c, int level, |
| 1167 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1168 | { |
| 1169 | uint8_t stale = 0; |
| 1170 | unsigned i; |
| 1171 | struct bucket *g; |
| 1172 | |
| 1173 | /* |
| 1174 | * ptr_invalid() can't return true for the keys that mark btree nodes as |
| 1175 | * freed, but since ptr_bad() returns true we'll never actually use them |
| 1176 | * for anything and thus we don't want mark their pointers here |
| 1177 | */ |
| 1178 | if (!bkey_cmp(k, &ZERO_KEY)) |
| 1179 | return stale; |
| 1180 | |
| 1181 | for (i = 0; i < KEY_PTRS(k); i++) { |
| 1182 | if (!ptr_available(c, k, i)) |
| 1183 | continue; |
| 1184 | |
| 1185 | g = PTR_BUCKET(c, k, i); |
| 1186 | |
Kent Overstreet | 3a2fd9d | 2014-02-27 17:51:12 -0800 | [diff] [blame] | 1187 | if (gen_after(g->last_gc, PTR_GEN(k, i))) |
| 1188 | g->last_gc = PTR_GEN(k, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1189 | |
| 1190 | if (ptr_stale(c, k, i)) { |
| 1191 | stale = max(stale, ptr_stale(c, k, i)); |
| 1192 | continue; |
| 1193 | } |
| 1194 | |
| 1195 | cache_bug_on(GC_MARK(g) && |
| 1196 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), |
| 1197 | c, "inconsistent ptrs: mark = %llu, level = %i", |
| 1198 | GC_MARK(g), level); |
| 1199 | |
| 1200 | if (level) |
| 1201 | SET_GC_MARK(g, GC_MARK_METADATA); |
| 1202 | else if (KEY_DIRTY(k)) |
| 1203 | SET_GC_MARK(g, GC_MARK_DIRTY); |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1204 | else if (!GC_MARK(g)) |
| 1205 | SET_GC_MARK(g, GC_MARK_RECLAIMABLE); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1206 | |
| 1207 | /* guard against overflow */ |
| 1208 | SET_GC_SECTORS_USED(g, min_t(unsigned, |
| 1209 | GC_SECTORS_USED(g) + KEY_SIZE(k), |
Darrick J. Wong | 9471744 | 2014-01-28 16:57:39 -0800 | [diff] [blame] | 1210 | MAX_GC_SECTORS_USED)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1211 | |
| 1212 | BUG_ON(!GC_SECTORS_USED(g)); |
| 1213 | } |
| 1214 | |
| 1215 | return stale; |
| 1216 | } |
| 1217 | |
| 1218 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) |
| 1219 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1220 | void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k) |
| 1221 | { |
| 1222 | unsigned i; |
| 1223 | |
| 1224 | for (i = 0; i < KEY_PTRS(k); i++) |
| 1225 | if (ptr_available(c, k, i) && |
| 1226 | !ptr_stale(c, k, i)) { |
| 1227 | struct bucket *b = PTR_BUCKET(c, k, i); |
| 1228 | |
| 1229 | b->gen = PTR_GEN(k, i); |
| 1230 | |
| 1231 | if (level && bkey_cmp(k, &ZERO_KEY)) |
| 1232 | b->prio = BTREE_PRIO; |
| 1233 | else if (!level && b->prio == BTREE_PRIO) |
| 1234 | b->prio = INITIAL_PRIO; |
| 1235 | } |
| 1236 | |
| 1237 | __bch_btree_mark_key(c, level, k); |
| 1238 | } |
| 1239 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1240 | 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] | 1241 | { |
| 1242 | uint8_t stale = 0; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1243 | unsigned keys = 0, good_keys = 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1244 | struct bkey *k; |
| 1245 | struct btree_iter iter; |
| 1246 | struct bset_tree *t; |
| 1247 | |
| 1248 | gc->nodes++; |
| 1249 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1250 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1251 | stale = max(stale, btree_mark_key(b, k)); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1252 | keys++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1253 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1254 | if (bch_ptr_bad(&b->keys, k)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1255 | continue; |
| 1256 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1257 | gc->key_bytes += bkey_u64s(k); |
| 1258 | gc->nkeys++; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1259 | good_keys++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1260 | |
| 1261 | gc->data += KEY_SIZE(k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1262 | } |
| 1263 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1264 | 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] | 1265 | btree_bug_on(t->size && |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1266 | bset_written(&b->keys, t) && |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1267 | bkey_cmp(&b->key, &t->end) < 0, |
| 1268 | b, "found short btree key in gc"); |
| 1269 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1270 | if (b->c->gc_always_rewrite) |
| 1271 | return true; |
| 1272 | |
| 1273 | if (stale > 10) |
| 1274 | return true; |
| 1275 | |
| 1276 | if ((keys - good_keys) * 2 > keys) |
| 1277 | return true; |
| 1278 | |
| 1279 | return false; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1280 | } |
| 1281 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1282 | #define GC_MERGE_NODES 4U |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1283 | |
| 1284 | struct gc_merge_info { |
| 1285 | struct btree *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1286 | unsigned keys; |
| 1287 | }; |
| 1288 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1289 | static int bch_btree_insert_node(struct btree *, struct btree_op *, |
| 1290 | struct keylist *, atomic_t *, struct bkey *); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1291 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1292 | static int btree_gc_coalesce(struct btree *b, struct btree_op *op, |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1293 | struct gc_stat *gc, struct gc_merge_info *r) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1294 | { |
| 1295 | unsigned i, nodes = 0, keys = 0, blocks; |
| 1296 | struct btree *new_nodes[GC_MERGE_NODES]; |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1297 | struct keylist keylist; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1298 | struct closure cl; |
| 1299 | struct bkey *k; |
| 1300 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1301 | bch_keylist_init(&keylist); |
| 1302 | |
| 1303 | if (btree_check_reserve(b, NULL)) |
| 1304 | return 0; |
| 1305 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1306 | memset(new_nodes, 0, sizeof(new_nodes)); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1307 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1308 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1309 | while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b)) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1310 | keys += r[nodes++].keys; |
| 1311 | |
| 1312 | blocks = btree_default_blocks(b->c) * 2 / 3; |
| 1313 | |
| 1314 | if (nodes < 2 || |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1315 | __set_blocks(b->keys.set[0].data, keys, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1316 | block_bytes(b->c)) > blocks * (nodes - 1)) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1317 | return 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1318 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1319 | for (i = 0; i < nodes; i++) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1320 | new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1321 | if (IS_ERR_OR_NULL(new_nodes[i])) |
| 1322 | goto out_nocoalesce; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1323 | } |
| 1324 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1325 | /* |
| 1326 | * We have to check the reserve here, after we've allocated our new |
| 1327 | * nodes, to make sure the insert below will succeed - we also check |
| 1328 | * before as an optimization to potentially avoid a bunch of expensive |
| 1329 | * allocs/sorts |
| 1330 | */ |
| 1331 | if (btree_check_reserve(b, NULL)) |
| 1332 | goto out_nocoalesce; |
| 1333 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1334 | for (i = 0; i < nodes; i++) |
| 1335 | mutex_lock(&new_nodes[i]->write_lock); |
| 1336 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1337 | for (i = nodes - 1; i > 0; --i) { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1338 | struct bset *n1 = btree_bset_first(new_nodes[i]); |
| 1339 | struct bset *n2 = btree_bset_first(new_nodes[i - 1]); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1340 | struct bkey *k, *last = NULL; |
| 1341 | |
| 1342 | keys = 0; |
| 1343 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1344 | if (i > 1) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1345 | for (k = n2->start; |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1346 | k < bset_bkey_last(n2); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1347 | k = bkey_next(k)) { |
| 1348 | if (__set_blocks(n1, n1->keys + keys + |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1349 | bkey_u64s(k), |
| 1350 | block_bytes(b->c)) > blocks) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1351 | break; |
| 1352 | |
| 1353 | last = k; |
| 1354 | keys += bkey_u64s(k); |
| 1355 | } |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1356 | } else { |
| 1357 | /* |
| 1358 | * Last node we're not getting rid of - we're getting |
| 1359 | * rid of the node at r[0]. Have to try and fit all of |
| 1360 | * the remaining keys into this node; we can't ensure |
| 1361 | * they will always fit due to rounding and variable |
| 1362 | * length keys (shouldn't be possible in practice, |
| 1363 | * though) |
| 1364 | */ |
| 1365 | if (__set_blocks(n1, n1->keys + n2->keys, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1366 | block_bytes(b->c)) > |
| 1367 | btree_blocks(new_nodes[i])) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1368 | goto out_nocoalesce; |
| 1369 | |
| 1370 | keys = n2->keys; |
| 1371 | /* Take the key of the node we're getting rid of */ |
| 1372 | last = &r->b->key; |
| 1373 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1374 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1375 | BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) > |
| 1376 | btree_blocks(new_nodes[i])); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1377 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1378 | if (last) |
| 1379 | bkey_copy_key(&new_nodes[i]->key, last); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1380 | |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1381 | memcpy(bset_bkey_last(n1), |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1382 | n2->start, |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1383 | (void *) bset_bkey_idx(n2, keys) - (void *) n2->start); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1384 | |
| 1385 | n1->keys += keys; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1386 | r[i].keys = n1->keys; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1387 | |
| 1388 | memmove(n2->start, |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 1389 | bset_bkey_idx(n2, keys), |
| 1390 | (void *) bset_bkey_last(n2) - |
| 1391 | (void *) bset_bkey_idx(n2, keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1392 | |
| 1393 | n2->keys -= keys; |
| 1394 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1395 | if (__bch_keylist_realloc(&keylist, |
Kent Overstreet | 085d2a3 | 2013-11-11 18:20:51 -0800 | [diff] [blame] | 1396 | bkey_u64s(&new_nodes[i]->key))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1397 | goto out_nocoalesce; |
| 1398 | |
| 1399 | bch_btree_node_write(new_nodes[i], &cl); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1400 | bch_keylist_add(&keylist, &new_nodes[i]->key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1401 | } |
| 1402 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1403 | for (i = 0; i < nodes; i++) |
| 1404 | mutex_unlock(&new_nodes[i]->write_lock); |
| 1405 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1406 | closure_sync(&cl); |
| 1407 | |
| 1408 | /* We emptied out this node */ |
| 1409 | BUG_ON(btree_bset_first(new_nodes[0])->keys); |
| 1410 | btree_node_free(new_nodes[0]); |
| 1411 | rw_unlock(true, new_nodes[0]); |
Slava Pestov | 400ffaa | 2014-07-12 21:53:11 -0700 | [diff] [blame] | 1412 | new_nodes[0] = NULL; |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 1413 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1414 | for (i = 0; i < nodes; i++) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1415 | if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1416 | goto out_nocoalesce; |
| 1417 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1418 | make_btree_freeing_key(r[i].b, keylist.top); |
| 1419 | bch_keylist_push(&keylist); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1420 | } |
| 1421 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1422 | bch_btree_insert_node(b, op, &keylist, NULL, NULL); |
| 1423 | BUG_ON(!bch_keylist_empty(&keylist)); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1424 | |
| 1425 | for (i = 0; i < nodes; i++) { |
| 1426 | btree_node_free(r[i].b); |
| 1427 | rw_unlock(true, r[i].b); |
| 1428 | |
| 1429 | r[i].b = new_nodes[i]; |
| 1430 | } |
| 1431 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1432 | memmove(r, r + 1, sizeof(r[0]) * (nodes - 1)); |
| 1433 | r[nodes - 1].b = ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1434 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1435 | trace_bcache_btree_gc_coalesce(nodes); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1436 | gc->nodes--; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1437 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1438 | bch_keylist_free(&keylist); |
| 1439 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1440 | /* Invalidated our iterator */ |
| 1441 | return -EINTR; |
| 1442 | |
| 1443 | out_nocoalesce: |
| 1444 | closure_sync(&cl); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1445 | bch_keylist_free(&keylist); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1446 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1447 | while ((k = bch_keylist_pop(&keylist))) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1448 | if (!bkey_cmp(k, &ZERO_KEY)) |
| 1449 | atomic_dec(&b->c->prio_blocked); |
| 1450 | |
| 1451 | for (i = 0; i < nodes; i++) |
| 1452 | if (!IS_ERR_OR_NULL(new_nodes[i])) { |
| 1453 | btree_node_free(new_nodes[i]); |
| 1454 | rw_unlock(true, new_nodes[i]); |
| 1455 | } |
| 1456 | return 0; |
| 1457 | } |
| 1458 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1459 | static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op, |
| 1460 | struct btree *replace) |
| 1461 | { |
| 1462 | struct keylist keys; |
| 1463 | struct btree *n; |
| 1464 | |
| 1465 | if (btree_check_reserve(b, NULL)) |
| 1466 | return 0; |
| 1467 | |
| 1468 | n = btree_node_alloc_replacement(replace, NULL); |
| 1469 | |
| 1470 | /* recheck reserve after allocating replacement node */ |
| 1471 | if (btree_check_reserve(b, NULL)) { |
| 1472 | btree_node_free(n); |
| 1473 | rw_unlock(true, n); |
| 1474 | return 0; |
| 1475 | } |
| 1476 | |
| 1477 | bch_btree_node_write_sync(n); |
| 1478 | |
| 1479 | bch_keylist_init(&keys); |
| 1480 | bch_keylist_add(&keys, &n->key); |
| 1481 | |
| 1482 | make_btree_freeing_key(replace, keys.top); |
| 1483 | bch_keylist_push(&keys); |
| 1484 | |
| 1485 | bch_btree_insert_node(b, op, &keys, NULL, NULL); |
| 1486 | BUG_ON(!bch_keylist_empty(&keys)); |
| 1487 | |
| 1488 | btree_node_free(replace); |
| 1489 | rw_unlock(true, n); |
| 1490 | |
| 1491 | /* Invalidated our iterator */ |
| 1492 | return -EINTR; |
| 1493 | } |
| 1494 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1495 | static unsigned btree_gc_count_keys(struct btree *b) |
| 1496 | { |
| 1497 | struct bkey *k; |
| 1498 | struct btree_iter iter; |
| 1499 | unsigned ret = 0; |
| 1500 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1501 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1502 | ret += bkey_u64s(k); |
| 1503 | |
| 1504 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1505 | } |
| 1506 | |
| 1507 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, |
| 1508 | struct closure *writes, struct gc_stat *gc) |
| 1509 | { |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1510 | int ret = 0; |
| 1511 | bool should_rewrite; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1512 | struct bkey *k; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1513 | struct btree_iter iter; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1514 | struct gc_merge_info r[GC_MERGE_NODES]; |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1515 | struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1516 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1517 | bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1518 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1519 | for (i = r; i < r + ARRAY_SIZE(r); i++) |
| 1520 | i->b = ERR_PTR(-EINTR); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1521 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1522 | while (1) { |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1523 | k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1524 | if (k) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1525 | 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] | 1526 | true, b); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1527 | if (IS_ERR(r->b)) { |
| 1528 | ret = PTR_ERR(r->b); |
| 1529 | break; |
| 1530 | } |
| 1531 | |
| 1532 | r->keys = btree_gc_count_keys(r->b); |
| 1533 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1534 | ret = btree_gc_coalesce(b, op, gc, r); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1535 | if (ret) |
| 1536 | break; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1537 | } |
| 1538 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1539 | if (!last->b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1540 | break; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1541 | |
| 1542 | if (!IS_ERR(last->b)) { |
| 1543 | should_rewrite = btree_gc_mark_node(last->b, gc); |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1544 | if (should_rewrite) { |
| 1545 | ret = btree_gc_rewrite_node(b, op, last->b); |
| 1546 | if (ret) |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1547 | break; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1548 | } |
| 1549 | |
| 1550 | if (last->b->level) { |
| 1551 | ret = btree_gc_recurse(last->b, op, writes, gc); |
| 1552 | if (ret) |
| 1553 | break; |
| 1554 | } |
| 1555 | |
| 1556 | bkey_copy_key(&b->c->gc_done, &last->b->key); |
| 1557 | |
| 1558 | /* |
| 1559 | * Must flush leaf nodes before gc ends, since replace |
| 1560 | * operations aren't journalled |
| 1561 | */ |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1562 | mutex_lock(&last->b->write_lock); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1563 | if (btree_node_dirty(last->b)) |
| 1564 | bch_btree_node_write(last->b, writes); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1565 | mutex_unlock(&last->b->write_lock); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1566 | rw_unlock(true, last->b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1567 | } |
| 1568 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1569 | memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1)); |
| 1570 | r->b = NULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1571 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1572 | if (need_resched()) { |
| 1573 | ret = -EAGAIN; |
| 1574 | break; |
| 1575 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1576 | } |
| 1577 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1578 | for (i = r; i < r + ARRAY_SIZE(r); i++) |
| 1579 | if (!IS_ERR_OR_NULL(i->b)) { |
| 1580 | mutex_lock(&i->b->write_lock); |
| 1581 | if (btree_node_dirty(i->b)) |
| 1582 | bch_btree_node_write(i->b, writes); |
| 1583 | mutex_unlock(&i->b->write_lock); |
| 1584 | rw_unlock(true, i->b); |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1585 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1586 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1587 | return ret; |
| 1588 | } |
| 1589 | |
| 1590 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, |
| 1591 | struct closure *writes, struct gc_stat *gc) |
| 1592 | { |
| 1593 | struct btree *n = NULL; |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1594 | int ret = 0; |
| 1595 | bool should_rewrite; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1596 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1597 | should_rewrite = btree_gc_mark_node(b, gc); |
| 1598 | if (should_rewrite) { |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1599 | n = btree_node_alloc_replacement(b, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1600 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1601 | if (!IS_ERR_OR_NULL(n)) { |
| 1602 | bch_btree_node_write_sync(n); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1603 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1604 | bch_btree_set_root(n); |
| 1605 | btree_node_free(b); |
| 1606 | rw_unlock(true, n); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1607 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1608 | return -EINTR; |
| 1609 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1610 | } |
| 1611 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1612 | __bch_btree_mark_key(b->c, b->level + 1, &b->key); |
| 1613 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1614 | if (b->level) { |
| 1615 | ret = btree_gc_recurse(b, op, writes, gc); |
| 1616 | if (ret) |
| 1617 | return ret; |
| 1618 | } |
| 1619 | |
| 1620 | bkey_copy_key(&b->c->gc_done, &b->key); |
| 1621 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1622 | return ret; |
| 1623 | } |
| 1624 | |
| 1625 | static void btree_gc_start(struct cache_set *c) |
| 1626 | { |
| 1627 | struct cache *ca; |
| 1628 | struct bucket *b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1629 | unsigned i; |
| 1630 | |
| 1631 | if (!c->gc_mark_valid) |
| 1632 | return; |
| 1633 | |
| 1634 | mutex_lock(&c->bucket_lock); |
| 1635 | |
| 1636 | c->gc_mark_valid = 0; |
| 1637 | c->gc_done = ZERO_KEY; |
| 1638 | |
| 1639 | for_each_cache(ca, c, i) |
| 1640 | for_each_bucket(b, ca) { |
Kent Overstreet | 3a2fd9d | 2014-02-27 17:51:12 -0800 | [diff] [blame] | 1641 | b->last_gc = b->gen; |
Kent Overstreet | 29ebf46 | 2013-07-11 19:43:21 -0700 | [diff] [blame] | 1642 | if (!atomic_read(&b->pin)) { |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1643 | SET_GC_MARK(b, 0); |
Kent Overstreet | 29ebf46 | 2013-07-11 19:43:21 -0700 | [diff] [blame] | 1644 | SET_GC_SECTORS_USED(b, 0); |
| 1645 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1646 | } |
| 1647 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1648 | mutex_unlock(&c->bucket_lock); |
| 1649 | } |
| 1650 | |
Kent Overstreet | 2531d9ee | 2014-03-17 16:55:55 -0700 | [diff] [blame] | 1651 | static size_t bch_btree_gc_finish(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1652 | { |
| 1653 | size_t available = 0; |
| 1654 | struct bucket *b; |
| 1655 | struct cache *ca; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1656 | unsigned i; |
| 1657 | |
| 1658 | mutex_lock(&c->bucket_lock); |
| 1659 | |
| 1660 | set_gc_sectors(c); |
| 1661 | c->gc_mark_valid = 1; |
| 1662 | c->need_gc = 0; |
| 1663 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1664 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) |
| 1665 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), |
| 1666 | GC_MARK_METADATA); |
| 1667 | |
Nicholas Swenson | bf0a628 | 2013-11-26 19:14:23 -0800 | [diff] [blame] | 1668 | /* don't reclaim buckets to which writeback keys point */ |
| 1669 | rcu_read_lock(); |
| 1670 | for (i = 0; i < c->nr_uuids; i++) { |
| 1671 | struct bcache_device *d = c->devices[i]; |
| 1672 | struct cached_dev *dc; |
| 1673 | struct keybuf_key *w, *n; |
| 1674 | unsigned j; |
| 1675 | |
| 1676 | if (!d || UUID_FLASH_ONLY(&c->uuids[i])) |
| 1677 | continue; |
| 1678 | dc = container_of(d, struct cached_dev, disk); |
| 1679 | |
| 1680 | spin_lock(&dc->writeback_keys.lock); |
| 1681 | rbtree_postorder_for_each_entry_safe(w, n, |
| 1682 | &dc->writeback_keys.keys, node) |
| 1683 | for (j = 0; j < KEY_PTRS(&w->key); j++) |
| 1684 | SET_GC_MARK(PTR_BUCKET(c, &w->key, j), |
| 1685 | GC_MARK_DIRTY); |
| 1686 | spin_unlock(&dc->writeback_keys.lock); |
| 1687 | } |
| 1688 | rcu_read_unlock(); |
| 1689 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1690 | for_each_cache(ca, c, i) { |
| 1691 | uint64_t *i; |
| 1692 | |
| 1693 | ca->invalidate_needs_gc = 0; |
| 1694 | |
| 1695 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) |
| 1696 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); |
| 1697 | |
| 1698 | for (i = ca->prio_buckets; |
| 1699 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) |
| 1700 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); |
| 1701 | |
| 1702 | for_each_bucket(b, ca) { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1703 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); |
| 1704 | |
Kent Overstreet | 4fe6a81 | 2014-03-13 13:46:29 -0700 | [diff] [blame] | 1705 | if (atomic_read(&b->pin)) |
| 1706 | continue; |
| 1707 | |
| 1708 | BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b)); |
| 1709 | |
| 1710 | if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1711 | available++; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1712 | } |
| 1713 | } |
| 1714 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1715 | mutex_unlock(&c->bucket_lock); |
| 1716 | return available; |
| 1717 | } |
| 1718 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1719 | static void bch_btree_gc(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1720 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1721 | int ret; |
| 1722 | unsigned long available; |
| 1723 | struct gc_stat stats; |
| 1724 | struct closure writes; |
| 1725 | struct btree_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1726 | uint64_t start_time = local_clock(); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1727 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1728 | trace_bcache_gc_start(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1729 | |
| 1730 | memset(&stats, 0, sizeof(struct gc_stat)); |
| 1731 | closure_init_stack(&writes); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1732 | bch_btree_op_init(&op, SHRT_MAX); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1733 | |
| 1734 | btree_gc_start(c); |
| 1735 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1736 | do { |
| 1737 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
| 1738 | closure_sync(&writes); |
Kent Overstreet | c5f1e5a | 2015-11-29 17:18:33 -0800 | [diff] [blame] | 1739 | cond_resched(); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1740 | |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1741 | if (ret && ret != -EAGAIN) |
| 1742 | pr_warn("gc failed!"); |
| 1743 | } while (ret); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1744 | |
| 1745 | available = bch_btree_gc_finish(c); |
Kent Overstreet | 5794351 | 2013-04-25 13:58:35 -0700 | [diff] [blame] | 1746 | wake_up_allocators(c); |
| 1747 | |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 1748 | bch_time_stats_update(&c->btree_gc_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1749 | |
| 1750 | stats.key_bytes *= sizeof(uint64_t); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1751 | stats.data <<= 9; |
| 1752 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; |
| 1753 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1754 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1755 | trace_bcache_gc_end(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1756 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1757 | bch_moving_gc(c); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1758 | } |
| 1759 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1760 | static bool gc_should_run(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1761 | { |
Kent Overstreet | a1f0358 | 2013-09-10 19:07:00 -0700 | [diff] [blame] | 1762 | struct cache *ca; |
| 1763 | unsigned i; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1764 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1765 | for_each_cache(ca, c, i) |
| 1766 | if (ca->invalidate_needs_gc) |
| 1767 | return true; |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1768 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1769 | if (atomic_read(&c->sectors_to_gc) < 0) |
| 1770 | return true; |
| 1771 | |
| 1772 | return false; |
| 1773 | } |
| 1774 | |
| 1775 | static int bch_gc_thread(void *arg) |
| 1776 | { |
| 1777 | struct cache_set *c = arg; |
| 1778 | |
| 1779 | while (1) { |
| 1780 | wait_event_interruptible(c->gc_wait, |
| 1781 | kthread_should_stop() || gc_should_run(c)); |
| 1782 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1783 | if (kthread_should_stop()) |
| 1784 | break; |
| 1785 | |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1786 | set_gc_sectors(c); |
| 1787 | bch_btree_gc(c); |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1788 | } |
| 1789 | |
| 1790 | return 0; |
| 1791 | } |
| 1792 | |
| 1793 | int bch_gc_thread_start(struct cache_set *c) |
| 1794 | { |
Kent Overstreet | 8629aed | 2016-10-26 20:31:17 -0700 | [diff] [blame] | 1795 | c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc"); |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1796 | if (IS_ERR(c->gc_thread)) |
| 1797 | return PTR_ERR(c->gc_thread); |
| 1798 | |
Kent Overstreet | 72a4451 | 2013-10-24 17:19:26 -0700 | [diff] [blame] | 1799 | return 0; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1800 | } |
| 1801 | |
| 1802 | /* Initial partial gc */ |
| 1803 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1804 | 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] | 1805 | { |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1806 | int ret = 0; |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1807 | struct bkey *k, *p = NULL; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1808 | struct btree_iter iter; |
| 1809 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1810 | for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) |
| 1811 | bch_initial_mark_key(b->c, b->level, k); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1812 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1813 | bch_initial_mark_key(b->c, b->level + 1, &b->key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1814 | |
| 1815 | if (b->level) { |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 1816 | bch_btree_iter_init(&b->keys, &iter, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1817 | |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1818 | do { |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 1819 | k = bch_btree_iter_next_filter(&iter, &b->keys, |
| 1820 | bch_ptr_bad); |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1821 | if (k) |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1822 | btree_node_prefetch(b, k); |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1823 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1824 | if (p) |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1825 | ret = btree(check_recurse, p, b, op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1826 | |
Kent Overstreet | 5031016 | 2013-09-10 17:18:59 -0700 | [diff] [blame] | 1827 | p = k; |
| 1828 | } while (p && !ret); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1829 | } |
| 1830 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1831 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1832 | } |
| 1833 | |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 1834 | int bch_btree_check(struct cache_set *c) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1835 | { |
Kent Overstreet | c18536a | 2013-07-24 17:44:17 -0700 | [diff] [blame] | 1836 | struct btree_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1837 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1838 | bch_btree_op_init(&op, SHRT_MAX); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1839 | |
Kent Overstreet | 487dded | 2014-03-17 15:13:26 -0700 | [diff] [blame] | 1840 | return btree_root(check_recurse, c, &op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1841 | } |
| 1842 | |
Kent Overstreet | 2531d9ee | 2014-03-17 16:55:55 -0700 | [diff] [blame] | 1843 | void bch_initial_gc_finish(struct cache_set *c) |
| 1844 | { |
| 1845 | struct cache *ca; |
| 1846 | struct bucket *b; |
| 1847 | unsigned i; |
| 1848 | |
| 1849 | bch_btree_gc_finish(c); |
| 1850 | |
| 1851 | mutex_lock(&c->bucket_lock); |
| 1852 | |
| 1853 | /* |
| 1854 | * We need to put some unused buckets directly on the prio freelist in |
| 1855 | * order to get the allocator thread started - it needs freed buckets in |
| 1856 | * order to rewrite the prios and gens, and it needs to rewrite prios |
| 1857 | * and gens in order to free buckets. |
| 1858 | * |
| 1859 | * This is only safe for buckets that have no live data in them, which |
| 1860 | * there should always be some of. |
| 1861 | */ |
| 1862 | for_each_cache(ca, c, i) { |
| 1863 | for_each_bucket(b, ca) { |
| 1864 | if (fifo_full(&ca->free[RESERVE_PRIO])) |
| 1865 | break; |
| 1866 | |
| 1867 | if (bch_can_invalidate_bucket(ca, b) && |
| 1868 | !GC_MARK(b)) { |
| 1869 | __bch_invalidate_one_bucket(ca, b); |
| 1870 | fifo_push(&ca->free[RESERVE_PRIO], |
| 1871 | b - ca->buckets); |
| 1872 | } |
| 1873 | } |
| 1874 | } |
| 1875 | |
| 1876 | mutex_unlock(&c->bucket_lock); |
| 1877 | } |
| 1878 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1879 | /* Btree insertion */ |
| 1880 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1881 | static bool btree_insert_key(struct btree *b, struct bkey *k, |
| 1882 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1883 | { |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1884 | unsigned status; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1885 | |
| 1886 | BUG_ON(bkey_cmp(k, &b->key) > 0); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1887 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1888 | status = bch_btree_insert_key(&b->keys, k, replace_key); |
| 1889 | if (status != BTREE_INSERT_STATUS_NO_INSERT) { |
| 1890 | bch_check_keys(&b->keys, "%u for %s", status, |
| 1891 | replace_key ? "replace" : "insert"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1892 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1893 | trace_bcache_btree_insert_key(b, k, replace_key != NULL, |
| 1894 | status); |
| 1895 | return true; |
| 1896 | } else |
| 1897 | return false; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1898 | } |
| 1899 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1900 | static size_t insert_u64s_remaining(struct btree *b) |
| 1901 | { |
Kent Overstreet | 3572324 | 2014-01-10 18:53:02 -0800 | [diff] [blame] | 1902 | long ret = bch_btree_keys_u64s_remaining(&b->keys); |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1903 | |
| 1904 | /* |
| 1905 | * Might land in the middle of an existing extent and have to split it |
| 1906 | */ |
| 1907 | if (b->keys.ops->is_extents) |
| 1908 | ret -= KEY_MAX_U64S; |
| 1909 | |
| 1910 | return max(ret, 0L); |
| 1911 | } |
| 1912 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1913 | 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] | 1914 | struct keylist *insert_keys, |
| 1915 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1916 | { |
| 1917 | bool ret = false; |
Kent Overstreet | dc9d98d | 2013-12-17 23:47:33 -0800 | [diff] [blame] | 1918 | int oldsize = bch_count_data(&b->keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1919 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1920 | while (!bch_keylist_empty(insert_keys)) { |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1921 | struct bkey *k = insert_keys->keys; |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1922 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 1923 | if (bkey_u64s(k) > insert_u64s_remaining(b)) |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 1924 | break; |
| 1925 | |
| 1926 | if (bkey_cmp(k, &b->key) <= 0) { |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 1927 | if (!b->level) |
| 1928 | bkey_put(b->c, k); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1929 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1930 | ret |= btree_insert_key(b, k, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1931 | bch_keylist_pop_front(insert_keys); |
| 1932 | } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1933 | BKEY_PADDED(key) temp; |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1934 | bkey_copy(&temp.key, insert_keys->keys); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1935 | |
| 1936 | bch_cut_back(&b->key, &temp.key); |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 1937 | bch_cut_front(&b->key, insert_keys->keys); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1938 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1939 | ret |= btree_insert_key(b, &temp.key, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1940 | break; |
| 1941 | } else { |
| 1942 | break; |
| 1943 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1944 | } |
| 1945 | |
Kent Overstreet | 829a60b | 2013-11-11 17:02:31 -0800 | [diff] [blame] | 1946 | if (!ret) |
| 1947 | op->insert_collision = true; |
| 1948 | |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 1949 | BUG_ON(!bch_keylist_empty(insert_keys) && b->level); |
| 1950 | |
Kent Overstreet | dc9d98d | 2013-12-17 23:47:33 -0800 | [diff] [blame] | 1951 | BUG_ON(bch_count_data(&b->keys) < oldsize); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1952 | return ret; |
| 1953 | } |
| 1954 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 1955 | static int btree_split(struct btree *b, struct btree_op *op, |
| 1956 | struct keylist *insert_keys, |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 1957 | struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1958 | { |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 1959 | bool split; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1960 | struct btree *n1, *n2 = NULL, *n3 = NULL; |
| 1961 | uint64_t start_time = local_clock(); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1962 | struct closure cl; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 1963 | struct keylist parent_keys; |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1964 | |
| 1965 | closure_init_stack(&cl); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 1966 | bch_keylist_init(&parent_keys); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1967 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1968 | if (btree_check_reserve(b, op)) { |
| 1969 | if (!b->level) |
| 1970 | return -EINTR; |
| 1971 | else |
| 1972 | WARN(1, "insufficient reserve for split\n"); |
| 1973 | } |
Kent Overstreet | 7836541 | 2013-12-17 01:29:34 -0800 | [diff] [blame] | 1974 | |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 1975 | n1 = btree_node_alloc_replacement(b, op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1976 | if (IS_ERR(n1)) |
| 1977 | goto err; |
| 1978 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1979 | split = set_blocks(btree_bset_first(n1), |
| 1980 | block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1981 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1982 | if (split) { |
| 1983 | unsigned keys = 0; |
| 1984 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 1985 | trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 1986 | |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1987 | n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1988 | if (IS_ERR(n2)) |
| 1989 | goto err_free1; |
| 1990 | |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 1991 | if (!b->parent) { |
Slava Pestov | 2452cc8 | 2014-07-12 00:22:53 -0700 | [diff] [blame] | 1992 | n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1993 | if (IS_ERR(n3)) |
| 1994 | goto err_free2; |
| 1995 | } |
| 1996 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 1997 | mutex_lock(&n1->write_lock); |
| 1998 | mutex_lock(&n2->write_lock); |
| 1999 | |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2000 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2001 | |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2002 | /* |
| 2003 | * 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] | 2004 | * search tree yet |
| 2005 | */ |
| 2006 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2007 | while (keys < (btree_bset_first(n1)->keys * 3) / 5) |
| 2008 | keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 2009 | keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2010 | |
Kent Overstreet | fafff81 | 2013-12-17 21:56:21 -0800 | [diff] [blame] | 2011 | bkey_copy_key(&n1->key, |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2012 | bset_bkey_idx(btree_bset_first(n1), keys)); |
| 2013 | keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2014 | |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2015 | btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys; |
| 2016 | btree_bset_first(n1)->keys = 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 | memcpy(btree_bset_first(n2)->start, |
| 2019 | bset_bkey_last(btree_bset_first(n1)), |
| 2020 | btree_bset_first(n2)->keys * sizeof(uint64_t)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2021 | |
| 2022 | bkey_copy_key(&n2->key, &b->key); |
| 2023 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2024 | bch_keylist_add(&parent_keys, &n2->key); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2025 | bch_btree_node_write(n2, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2026 | mutex_unlock(&n2->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2027 | rw_unlock(true, n2); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2028 | } else { |
Kent Overstreet | ee81128 | 2013-12-17 23:49:49 -0800 | [diff] [blame] | 2029 | trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2030 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2031 | mutex_lock(&n1->write_lock); |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2032 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2033 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2034 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2035 | bch_keylist_add(&parent_keys, &n1->key); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2036 | bch_btree_node_write(n1, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2037 | mutex_unlock(&n1->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2038 | |
| 2039 | if (n3) { |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2040 | /* Depth increases, make a new root */ |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2041 | mutex_lock(&n3->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2042 | bkey_copy_key(&n3->key, &MAX_KEY); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2043 | bch_btree_insert_keys(n3, op, &parent_keys, NULL); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2044 | bch_btree_node_write(n3, &cl); |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2045 | mutex_unlock(&n3->write_lock); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2046 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2047 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2048 | bch_btree_set_root(n3); |
| 2049 | rw_unlock(true, n3); |
Kent Overstreet | d6fd3b1 | 2013-07-24 17:20:19 -0700 | [diff] [blame] | 2050 | } else if (!b->parent) { |
| 2051 | /* Root filled up but didn't need to be split */ |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2052 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2053 | bch_btree_set_root(n1); |
| 2054 | } else { |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2055 | /* Split a non root node */ |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2056 | closure_sync(&cl); |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2057 | make_btree_freeing_key(b, parent_keys.top); |
| 2058 | bch_keylist_push(&parent_keys); |
| 2059 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2060 | bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL); |
| 2061 | BUG_ON(!bch_keylist_empty(&parent_keys)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2062 | } |
| 2063 | |
Kent Overstreet | 05335cf | 2014-03-17 18:22:34 -0700 | [diff] [blame] | 2064 | btree_node_free(b); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2065 | rw_unlock(true, n1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2066 | |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 2067 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2068 | |
| 2069 | return 0; |
| 2070 | err_free2: |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2071 | bkey_put(b->c, &n2->key); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 2072 | btree_node_free(n2); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2073 | rw_unlock(true, n2); |
| 2074 | err_free1: |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2075 | bkey_put(b->c, &n1->key); |
Kent Overstreet | e8e1d46 | 2013-07-24 17:27:07 -0700 | [diff] [blame] | 2076 | btree_node_free(n1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2077 | rw_unlock(true, n1); |
| 2078 | err: |
Kent Overstreet | 0a63b66 | 2014-03-17 17:15:53 -0700 | [diff] [blame] | 2079 | WARN(1, "bcache: btree split failed (level %u)", b->level); |
Kent Overstreet | 5f5837d | 2013-12-16 16:38:49 -0800 | [diff] [blame] | 2080 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2081 | if (n3 == ERR_PTR(-EAGAIN) || |
| 2082 | n2 == ERR_PTR(-EAGAIN) || |
| 2083 | n1 == ERR_PTR(-EAGAIN)) |
| 2084 | return -EAGAIN; |
| 2085 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2086 | return -ENOMEM; |
| 2087 | } |
| 2088 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2089 | 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] | 2090 | struct keylist *insert_keys, |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2091 | atomic_t *journal_ref, |
| 2092 | struct bkey *replace_key) |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2093 | { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2094 | struct closure cl; |
| 2095 | |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2096 | BUG_ON(b->level && replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2097 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2098 | closure_init_stack(&cl); |
| 2099 | |
| 2100 | mutex_lock(&b->write_lock); |
| 2101 | |
| 2102 | if (write_block(b) != btree_bset_last(b) && |
| 2103 | b->keys.last_set_unwritten) |
| 2104 | bch_btree_init_next(b); /* just wrote a set */ |
| 2105 | |
Kent Overstreet | 59158fd | 2013-11-11 19:03:54 -0800 | [diff] [blame] | 2106 | if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2107 | mutex_unlock(&b->write_lock); |
| 2108 | goto split; |
| 2109 | } |
Kent Overstreet | 3b3e9e5 | 2013-12-07 03:57:58 -0800 | [diff] [blame] | 2110 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2111 | BUG_ON(write_block(b) != btree_bset_last(b)); |
| 2112 | |
| 2113 | if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) { |
| 2114 | if (!b->level) |
| 2115 | bch_btree_leaf_dirty(b, journal_ref); |
| 2116 | else |
| 2117 | bch_btree_node_write(b, &cl); |
| 2118 | } |
| 2119 | |
| 2120 | mutex_unlock(&b->write_lock); |
| 2121 | |
| 2122 | /* wait for btree node write if necessary, after unlock */ |
| 2123 | closure_sync(&cl); |
| 2124 | |
| 2125 | return 0; |
| 2126 | split: |
| 2127 | if (current->bio_list) { |
| 2128 | op->lock = b->c->root->level + 1; |
| 2129 | return -EAGAIN; |
| 2130 | } else if (op->lock <= b->c->root->level) { |
| 2131 | op->lock = b->c->root->level + 1; |
| 2132 | return -EINTR; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2133 | } else { |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2134 | /* Invalidated all iterators */ |
| 2135 | int ret = btree_split(b, op, insert_keys, replace_key); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2136 | |
Kent Overstreet | 2a28568 | 2014-03-04 16:42:42 -0800 | [diff] [blame] | 2137 | if (bch_keylist_empty(insert_keys)) |
| 2138 | return 0; |
| 2139 | else if (!ret) |
| 2140 | return -EINTR; |
| 2141 | return ret; |
Kent Overstreet | 17e21a9 | 2013-07-26 12:32:38 -0700 | [diff] [blame] | 2142 | } |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 2143 | } |
| 2144 | |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2145 | int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, |
| 2146 | struct bkey *check_key) |
| 2147 | { |
| 2148 | int ret = -EINTR; |
| 2149 | uint64_t btree_ptr = b->key.ptr[0]; |
| 2150 | unsigned long seq = b->seq; |
| 2151 | struct keylist insert; |
| 2152 | bool upgrade = op->lock == -1; |
| 2153 | |
| 2154 | bch_keylist_init(&insert); |
| 2155 | |
| 2156 | if (upgrade) { |
| 2157 | rw_unlock(false, b); |
| 2158 | rw_lock(true, b, b->level); |
| 2159 | |
| 2160 | if (b->key.ptr[0] != btree_ptr || |
Zheng Liu | 2ef9ccb | 2015-11-29 17:17:05 -0800 | [diff] [blame] | 2161 | b->seq != seq + 1) { |
| 2162 | op->lock = b->level; |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2163 | goto out; |
Zheng Liu | 2ef9ccb | 2015-11-29 17:17:05 -0800 | [diff] [blame] | 2164 | } |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2165 | } |
| 2166 | |
| 2167 | SET_KEY_PTRS(check_key, 1); |
| 2168 | get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); |
| 2169 | |
| 2170 | SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); |
| 2171 | |
| 2172 | bch_keylist_add(&insert, check_key); |
| 2173 | |
Kent Overstreet | 1b207d8 | 2013-09-10 18:52:54 -0700 | [diff] [blame] | 2174 | ret = bch_btree_insert_node(b, op, &insert, NULL, NULL); |
Kent Overstreet | e7c590e | 2013-09-10 18:39:16 -0700 | [diff] [blame] | 2175 | |
| 2176 | BUG_ON(!ret && !bch_keylist_empty(&insert)); |
| 2177 | out: |
| 2178 | if (upgrade) |
| 2179 | downgrade_write(&b->lock); |
| 2180 | return ret; |
| 2181 | } |
| 2182 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2183 | struct btree_insert_op { |
| 2184 | struct btree_op op; |
| 2185 | struct keylist *keys; |
| 2186 | atomic_t *journal_ref; |
| 2187 | struct bkey *replace_key; |
| 2188 | }; |
| 2189 | |
Wei Yongjun | 08239ca | 2013-11-28 10:31:35 +0800 | [diff] [blame] | 2190 | 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] | 2191 | { |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2192 | struct btree_insert_op *op = container_of(b_op, |
| 2193 | struct btree_insert_op, op); |
Kent Overstreet | 403b6cd | 2013-07-24 17:22:44 -0700 | [diff] [blame] | 2194 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2195 | int ret = bch_btree_insert_node(b, &op->op, op->keys, |
| 2196 | op->journal_ref, op->replace_key); |
| 2197 | if (ret && !bch_keylist_empty(op->keys)) |
| 2198 | return ret; |
| 2199 | else |
| 2200 | return MAP_DONE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2201 | } |
| 2202 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2203 | int bch_btree_insert(struct cache_set *c, struct keylist *keys, |
| 2204 | atomic_t *journal_ref, struct bkey *replace_key) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2205 | { |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2206 | struct btree_insert_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2207 | int ret = 0; |
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 | BUG_ON(current->bio_list); |
Kent Overstreet | 4f3d401 | 2013-09-10 18:46:36 -0700 | [diff] [blame] | 2210 | BUG_ON(bch_keylist_empty(keys)); |
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 | bch_btree_op_init(&op.op, 0); |
| 2213 | op.keys = keys; |
| 2214 | op.journal_ref = journal_ref; |
| 2215 | op.replace_key = replace_key; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2216 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2217 | while (!ret && !bch_keylist_empty(keys)) { |
| 2218 | op.op.lock = 0; |
| 2219 | ret = bch_btree_map_leaf_nodes(&op.op, c, |
| 2220 | &START_KEY(keys->keys), |
| 2221 | btree_insert_fn); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2222 | } |
| 2223 | |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2224 | if (ret) { |
| 2225 | struct bkey *k; |
| 2226 | |
| 2227 | pr_err("error %i", ret); |
| 2228 | |
| 2229 | while ((k = bch_keylist_pop(keys))) |
Kent Overstreet | 3a3b6a4 | 2013-07-24 16:46:42 -0700 | [diff] [blame] | 2230 | bkey_put(c, k); |
Kent Overstreet | cc7b881 | 2013-07-24 18:07:22 -0700 | [diff] [blame] | 2231 | } else if (op.op.insert_collision) |
| 2232 | ret = -ESRCH; |
Kent Overstreet | 6054c6d | 2013-07-24 18:06:22 -0700 | [diff] [blame] | 2233 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2234 | return ret; |
| 2235 | } |
| 2236 | |
| 2237 | void bch_btree_set_root(struct btree *b) |
| 2238 | { |
| 2239 | unsigned i; |
Kent Overstreet | e49c7c3 | 2013-06-26 17:25:38 -0700 | [diff] [blame] | 2240 | struct closure cl; |
| 2241 | |
| 2242 | closure_init_stack(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2243 | |
Kent Overstreet | c37511b | 2013-04-26 15:39:55 -0700 | [diff] [blame] | 2244 | trace_bcache_btree_set_root(b); |
| 2245 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2246 | BUG_ON(!b->written); |
| 2247 | |
| 2248 | for (i = 0; i < KEY_PTRS(&b->key); i++) |
| 2249 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); |
| 2250 | |
| 2251 | mutex_lock(&b->c->bucket_lock); |
| 2252 | list_del_init(&b->list); |
| 2253 | mutex_unlock(&b->c->bucket_lock); |
| 2254 | |
| 2255 | b->c->root = b; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2256 | |
Kent Overstreet | e49c7c3 | 2013-06-26 17:25:38 -0700 | [diff] [blame] | 2257 | bch_journal_meta(b->c, &cl); |
| 2258 | closure_sync(&cl); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2259 | } |
| 2260 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2261 | /* Map across nodes or keys */ |
| 2262 | |
| 2263 | static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, |
| 2264 | struct bkey *from, |
| 2265 | btree_map_nodes_fn *fn, int flags) |
| 2266 | { |
| 2267 | int ret = MAP_CONTINUE; |
| 2268 | |
| 2269 | if (b->level) { |
| 2270 | struct bkey *k; |
| 2271 | struct btree_iter iter; |
| 2272 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 2273 | bch_btree_iter_init(&b->keys, &iter, from); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2274 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 2275 | while ((k = bch_btree_iter_next_filter(&iter, &b->keys, |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2276 | bch_ptr_bad))) { |
| 2277 | ret = btree(map_nodes_recurse, k, b, |
| 2278 | op, from, fn, flags); |
| 2279 | from = NULL; |
| 2280 | |
| 2281 | if (ret != MAP_CONTINUE) |
| 2282 | return ret; |
| 2283 | } |
| 2284 | } |
| 2285 | |
| 2286 | if (!b->level || flags == MAP_ALL_NODES) |
| 2287 | ret = fn(op, b); |
| 2288 | |
| 2289 | return ret; |
| 2290 | } |
| 2291 | |
| 2292 | int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, |
| 2293 | struct bkey *from, btree_map_nodes_fn *fn, int flags) |
| 2294 | { |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2295 | return btree_root(map_nodes_recurse, c, op, from, fn, flags); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2296 | } |
| 2297 | |
| 2298 | static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, |
| 2299 | struct bkey *from, btree_map_keys_fn *fn, |
| 2300 | int flags) |
| 2301 | { |
| 2302 | int ret = MAP_CONTINUE; |
| 2303 | struct bkey *k; |
| 2304 | struct btree_iter iter; |
| 2305 | |
Kent Overstreet | c052dd9 | 2013-11-11 17:35:24 -0800 | [diff] [blame] | 2306 | bch_btree_iter_init(&b->keys, &iter, from); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2307 | |
Kent Overstreet | a85e968 | 2013-12-20 17:28:16 -0800 | [diff] [blame] | 2308 | 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] | 2309 | ret = !b->level |
| 2310 | ? fn(op, b, k) |
| 2311 | : btree(map_keys_recurse, k, b, op, from, fn, flags); |
| 2312 | from = NULL; |
| 2313 | |
| 2314 | if (ret != MAP_CONTINUE) |
| 2315 | return ret; |
| 2316 | } |
| 2317 | |
| 2318 | if (!b->level && (flags & MAP_END_KEY)) |
| 2319 | ret = fn(op, b, &KEY(KEY_INODE(&b->key), |
| 2320 | KEY_OFFSET(&b->key), 0)); |
| 2321 | |
| 2322 | return ret; |
| 2323 | } |
| 2324 | |
| 2325 | int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, |
| 2326 | struct bkey *from, btree_map_keys_fn *fn, int flags) |
| 2327 | { |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2328 | return btree_root(map_keys_recurse, c, op, from, fn, flags); |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2329 | } |
| 2330 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2331 | /* Keybuf code */ |
| 2332 | |
| 2333 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) |
| 2334 | { |
| 2335 | /* Overlapping keys compare equal */ |
| 2336 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) |
| 2337 | return -1; |
| 2338 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) |
| 2339 | return 1; |
| 2340 | return 0; |
| 2341 | } |
| 2342 | |
| 2343 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, |
| 2344 | struct keybuf_key *r) |
| 2345 | { |
| 2346 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); |
| 2347 | } |
| 2348 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2349 | struct refill { |
| 2350 | struct btree_op op; |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2351 | unsigned nr_found; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2352 | struct keybuf *buf; |
| 2353 | struct bkey *end; |
| 2354 | keybuf_pred_fn *pred; |
| 2355 | }; |
| 2356 | |
| 2357 | static int refill_keybuf_fn(struct btree_op *op, struct btree *b, |
| 2358 | struct bkey *k) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2359 | { |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2360 | struct refill *refill = container_of(op, struct refill, op); |
| 2361 | struct keybuf *buf = refill->buf; |
| 2362 | int ret = MAP_CONTINUE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2363 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2364 | if (bkey_cmp(k, refill->end) >= 0) { |
| 2365 | ret = MAP_DONE; |
| 2366 | goto out; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2367 | } |
| 2368 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2369 | if (!KEY_SIZE(k)) /* end key */ |
| 2370 | goto out; |
| 2371 | |
| 2372 | if (refill->pred(buf, k)) { |
| 2373 | struct keybuf_key *w; |
| 2374 | |
| 2375 | spin_lock(&buf->lock); |
| 2376 | |
| 2377 | w = array_alloc(&buf->freelist); |
| 2378 | if (!w) { |
| 2379 | spin_unlock(&buf->lock); |
| 2380 | return MAP_DONE; |
| 2381 | } |
| 2382 | |
| 2383 | w->private = NULL; |
| 2384 | bkey_copy(&w->key, k); |
| 2385 | |
| 2386 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) |
| 2387 | array_free(&buf->freelist, w); |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2388 | else |
| 2389 | refill->nr_found++; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2390 | |
| 2391 | if (array_freelist_empty(&buf->freelist)) |
| 2392 | ret = MAP_DONE; |
| 2393 | |
| 2394 | spin_unlock(&buf->lock); |
| 2395 | } |
| 2396 | out: |
| 2397 | buf->last_scanned = *k; |
| 2398 | return ret; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2399 | } |
| 2400 | |
| 2401 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2402 | struct bkey *end, keybuf_pred_fn *pred) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2403 | { |
| 2404 | struct bkey start = buf->last_scanned; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2405 | struct refill refill; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2406 | |
| 2407 | cond_resched(); |
| 2408 | |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 2409 | bch_btree_op_init(&refill.op, -1); |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2410 | refill.nr_found = 0; |
| 2411 | refill.buf = buf; |
| 2412 | refill.end = end; |
| 2413 | refill.pred = pred; |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2414 | |
| 2415 | bch_btree_map_keys(&refill.op, c, &buf->last_scanned, |
| 2416 | refill_keybuf_fn, MAP_END_KEY); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2417 | |
Kent Overstreet | 48a915a | 2013-10-31 15:43:22 -0700 | [diff] [blame] | 2418 | trace_bcache_keyscan(refill.nr_found, |
| 2419 | KEY_INODE(&start), KEY_OFFSET(&start), |
| 2420 | KEY_INODE(&buf->last_scanned), |
| 2421 | KEY_OFFSET(&buf->last_scanned)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2422 | |
| 2423 | spin_lock(&buf->lock); |
| 2424 | |
| 2425 | if (!RB_EMPTY_ROOT(&buf->keys)) { |
| 2426 | struct keybuf_key *w; |
| 2427 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); |
| 2428 | buf->start = START_KEY(&w->key); |
| 2429 | |
| 2430 | w = RB_LAST(&buf->keys, struct keybuf_key, node); |
| 2431 | buf->end = w->key; |
| 2432 | } else { |
| 2433 | buf->start = MAX_KEY; |
| 2434 | buf->end = MAX_KEY; |
| 2435 | } |
| 2436 | |
| 2437 | spin_unlock(&buf->lock); |
| 2438 | } |
| 2439 | |
| 2440 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) |
| 2441 | { |
| 2442 | rb_erase(&w->node, &buf->keys); |
| 2443 | array_free(&buf->freelist, w); |
| 2444 | } |
| 2445 | |
| 2446 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) |
| 2447 | { |
| 2448 | spin_lock(&buf->lock); |
| 2449 | __bch_keybuf_del(buf, w); |
| 2450 | spin_unlock(&buf->lock); |
| 2451 | } |
| 2452 | |
| 2453 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, |
| 2454 | struct bkey *end) |
| 2455 | { |
| 2456 | bool ret = false; |
| 2457 | struct keybuf_key *p, *w, s; |
| 2458 | s.key = *start; |
| 2459 | |
| 2460 | if (bkey_cmp(end, &buf->start) <= 0 || |
| 2461 | bkey_cmp(start, &buf->end) >= 0) |
| 2462 | return false; |
| 2463 | |
| 2464 | spin_lock(&buf->lock); |
| 2465 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); |
| 2466 | |
| 2467 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { |
| 2468 | p = w; |
| 2469 | w = RB_NEXT(w, node); |
| 2470 | |
| 2471 | if (p->private) |
| 2472 | ret = true; |
| 2473 | else |
| 2474 | __bch_keybuf_del(buf, p); |
| 2475 | } |
| 2476 | |
| 2477 | spin_unlock(&buf->lock); |
| 2478 | return ret; |
| 2479 | } |
| 2480 | |
| 2481 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) |
| 2482 | { |
| 2483 | struct keybuf_key *w; |
| 2484 | spin_lock(&buf->lock); |
| 2485 | |
| 2486 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); |
| 2487 | |
| 2488 | while (w && w->private) |
| 2489 | w = RB_NEXT(w, node); |
| 2490 | |
| 2491 | if (w) |
| 2492 | w->private = ERR_PTR(-EINTR); |
| 2493 | |
| 2494 | spin_unlock(&buf->lock); |
| 2495 | return w; |
| 2496 | } |
| 2497 | |
| 2498 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 2499 | struct keybuf *buf, |
| 2500 | struct bkey *end, |
| 2501 | keybuf_pred_fn *pred) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2502 | { |
| 2503 | struct keybuf_key *ret; |
| 2504 | |
| 2505 | while (1) { |
| 2506 | ret = bch_keybuf_next(buf); |
| 2507 | if (ret) |
| 2508 | break; |
| 2509 | |
| 2510 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { |
| 2511 | pr_debug("scan finished"); |
| 2512 | break; |
| 2513 | } |
| 2514 | |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2515 | bch_refill_keybuf(c, buf, end, pred); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2516 | } |
| 2517 | |
| 2518 | return ret; |
| 2519 | } |
| 2520 | |
Kent Overstreet | 72c2706 | 2013-06-05 06:24:39 -0700 | [diff] [blame] | 2521 | void bch_keybuf_init(struct keybuf *buf) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2522 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 2523 | buf->last_scanned = MAX_KEY; |
| 2524 | buf->keys = RB_ROOT; |
| 2525 | |
| 2526 | spin_lock_init(&buf->lock); |
| 2527 | array_allocator_init(&buf->freelist); |
| 2528 | } |