1cafe5635SKent Overstreet /* 2cafe5635SKent Overstreet * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com> 3cafe5635SKent Overstreet * 4cafe5635SKent Overstreet * Uses a block device as cache for other block devices; optimized for SSDs. 5cafe5635SKent Overstreet * All allocation is done in buckets, which should match the erase block size 6cafe5635SKent Overstreet * of the device. 7cafe5635SKent Overstreet * 8cafe5635SKent Overstreet * Buckets containing cached data are kept on a heap sorted by priority; 9cafe5635SKent Overstreet * bucket priority is increased on cache hit, and periodically all the buckets 10cafe5635SKent Overstreet * on the heap have their priority scaled down. This currently is just used as 11cafe5635SKent Overstreet * an LRU but in the future should allow for more intelligent heuristics. 12cafe5635SKent Overstreet * 13cafe5635SKent Overstreet * Buckets have an 8 bit counter; freeing is accomplished by incrementing the 14cafe5635SKent Overstreet * counter. Garbage collection is used to remove stale pointers. 15cafe5635SKent Overstreet * 16cafe5635SKent Overstreet * Indexing is done via a btree; nodes are not necessarily fully sorted, rather 17cafe5635SKent Overstreet * as keys are inserted we only sort the pages that have not yet been written. 18cafe5635SKent Overstreet * When garbage collection is run, we resort the entire node. 19cafe5635SKent Overstreet * 20cafe5635SKent Overstreet * All configuration is done via sysfs; see Documentation/bcache.txt. 21cafe5635SKent Overstreet */ 22cafe5635SKent Overstreet 23cafe5635SKent Overstreet #include "bcache.h" 24cafe5635SKent Overstreet #include "btree.h" 25cafe5635SKent Overstreet #include "debug.h" 26279afbadSKent Overstreet #include "writeback.h" 27cafe5635SKent Overstreet 28cafe5635SKent Overstreet #include <linux/slab.h> 29cafe5635SKent Overstreet #include <linux/bitops.h> 3072a44517SKent Overstreet #include <linux/freezer.h> 31cafe5635SKent Overstreet #include <linux/hash.h> 3272a44517SKent Overstreet #include <linux/kthread.h> 33cd953ed0SGeert Uytterhoeven #include <linux/prefetch.h> 34cafe5635SKent Overstreet #include <linux/random.h> 35cafe5635SKent Overstreet #include <linux/rcupdate.h> 36cafe5635SKent Overstreet #include <trace/events/bcache.h> 37cafe5635SKent Overstreet 38cafe5635SKent Overstreet /* 39cafe5635SKent Overstreet * Todo: 40cafe5635SKent Overstreet * register_bcache: Return errors out to userspace correctly 41cafe5635SKent Overstreet * 42cafe5635SKent Overstreet * Writeback: don't undirty key until after a cache flush 43cafe5635SKent Overstreet * 44cafe5635SKent Overstreet * Create an iterator for key pointers 45cafe5635SKent Overstreet * 46cafe5635SKent Overstreet * On btree write error, mark bucket such that it won't be freed from the cache 47cafe5635SKent Overstreet * 48cafe5635SKent Overstreet * Journalling: 49cafe5635SKent Overstreet * Check for bad keys in replay 50cafe5635SKent Overstreet * Propagate barriers 51cafe5635SKent Overstreet * Refcount journal entries in journal_replay 52cafe5635SKent Overstreet * 53cafe5635SKent Overstreet * Garbage collection: 54cafe5635SKent Overstreet * Finish incremental gc 55cafe5635SKent Overstreet * Gc should free old UUIDs, data for invalid UUIDs 56cafe5635SKent Overstreet * 57cafe5635SKent Overstreet * Provide a way to list backing device UUIDs we have data cached for, and 58cafe5635SKent Overstreet * probably how long it's been since we've seen them, and a way to invalidate 59cafe5635SKent Overstreet * dirty data for devices that will never be attached again 60cafe5635SKent Overstreet * 61cafe5635SKent Overstreet * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so 62cafe5635SKent Overstreet * that based on that and how much dirty data we have we can keep writeback 63cafe5635SKent Overstreet * from being starved 64cafe5635SKent Overstreet * 65cafe5635SKent Overstreet * Add a tracepoint or somesuch to watch for writeback starvation 66cafe5635SKent Overstreet * 67cafe5635SKent Overstreet * When btree depth > 1 and splitting an interior node, we have to make sure 68cafe5635SKent Overstreet * alloc_bucket() cannot fail. This should be true but is not completely 69cafe5635SKent Overstreet * obvious. 70cafe5635SKent Overstreet * 71cafe5635SKent Overstreet * Make sure all allocations get charged to the root cgroup 72cafe5635SKent Overstreet * 73cafe5635SKent Overstreet * Plugging? 74cafe5635SKent Overstreet * 75cafe5635SKent Overstreet * If data write is less than hard sector size of ssd, round up offset in open 76cafe5635SKent Overstreet * bucket to the next whole sector 77cafe5635SKent Overstreet * 78cafe5635SKent Overstreet * Also lookup by cgroup in get_open_bucket() 79cafe5635SKent Overstreet * 80cafe5635SKent Overstreet * Superblock needs to be fleshed out for multiple cache devices 81cafe5635SKent Overstreet * 82cafe5635SKent Overstreet * Add a sysfs tunable for the number of writeback IOs in flight 83cafe5635SKent Overstreet * 84cafe5635SKent Overstreet * Add a sysfs tunable for the number of open data buckets 85cafe5635SKent Overstreet * 86cafe5635SKent Overstreet * IO tracking: Can we track when one process is doing io on behalf of another? 87cafe5635SKent Overstreet * IO tracking: Don't use just an average, weigh more recent stuff higher 88cafe5635SKent Overstreet * 89cafe5635SKent Overstreet * Test module load/unload 90cafe5635SKent Overstreet */ 91cafe5635SKent Overstreet 92df8e8970SKent Overstreet enum { 93df8e8970SKent Overstreet BTREE_INSERT_STATUS_INSERT, 94df8e8970SKent Overstreet BTREE_INSERT_STATUS_BACK_MERGE, 95df8e8970SKent Overstreet BTREE_INSERT_STATUS_OVERWROTE, 96df8e8970SKent Overstreet BTREE_INSERT_STATUS_FRONT_MERGE, 97df8e8970SKent Overstreet }; 98df8e8970SKent Overstreet 99cafe5635SKent Overstreet #define MAX_NEED_GC 64 100cafe5635SKent Overstreet #define MAX_SAVE_PRIO 72 101cafe5635SKent Overstreet 102cafe5635SKent Overstreet #define PTR_DIRTY_BIT (((uint64_t) 1 << 36)) 103cafe5635SKent Overstreet 104cafe5635SKent Overstreet #define PTR_HASH(c, k) \ 105cafe5635SKent Overstreet (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0)) 106cafe5635SKent Overstreet 107cafe5635SKent Overstreet static struct workqueue_struct *btree_io_wq; 108cafe5635SKent Overstreet 109df8e8970SKent Overstreet static inline bool should_split(struct btree *b) 110df8e8970SKent Overstreet { 111df8e8970SKent Overstreet struct bset *i = write_block(b); 112df8e8970SKent Overstreet return b->written >= btree_blocks(b) || 113df8e8970SKent Overstreet (b->written + __set_blocks(i, i->keys + 15, b->c) 114df8e8970SKent Overstreet > btree_blocks(b)); 115df8e8970SKent Overstreet } 116df8e8970SKent Overstreet 117df8e8970SKent Overstreet #define insert_lock(s, b) ((b)->level <= (s)->lock) 118df8e8970SKent Overstreet 119df8e8970SKent Overstreet /* 120df8e8970SKent Overstreet * These macros are for recursing down the btree - they handle the details of 121df8e8970SKent Overstreet * locking and looking up nodes in the cache for you. They're best treated as 122df8e8970SKent Overstreet * mere syntax when reading code that uses them. 123df8e8970SKent Overstreet * 124df8e8970SKent Overstreet * op->lock determines whether we take a read or a write lock at a given depth. 125df8e8970SKent Overstreet * If you've got a read lock and find that you need a write lock (i.e. you're 126df8e8970SKent Overstreet * going to have to split), set op->lock and return -EINTR; btree_root() will 127df8e8970SKent Overstreet * call you again and you'll have the correct lock. 128df8e8970SKent Overstreet */ 129df8e8970SKent Overstreet 130df8e8970SKent Overstreet /** 131df8e8970SKent Overstreet * btree - recurse down the btree on a specified key 132df8e8970SKent Overstreet * @fn: function to call, which will be passed the child node 133df8e8970SKent Overstreet * @key: key to recurse on 134df8e8970SKent Overstreet * @b: parent btree node 135df8e8970SKent Overstreet * @op: pointer to struct btree_op 136df8e8970SKent Overstreet */ 137df8e8970SKent Overstreet #define btree(fn, key, b, op, ...) \ 138df8e8970SKent Overstreet ({ \ 139df8e8970SKent Overstreet int _r, l = (b)->level - 1; \ 140df8e8970SKent Overstreet bool _w = l <= (op)->lock; \ 141df8e8970SKent Overstreet struct btree *_child = bch_btree_node_get((b)->c, key, l, _w); \ 142df8e8970SKent Overstreet if (!IS_ERR(_child)) { \ 143df8e8970SKent Overstreet _child->parent = (b); \ 144df8e8970SKent Overstreet _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \ 145df8e8970SKent Overstreet rw_unlock(_w, _child); \ 146df8e8970SKent Overstreet } else \ 147df8e8970SKent Overstreet _r = PTR_ERR(_child); \ 148df8e8970SKent Overstreet _r; \ 149df8e8970SKent Overstreet }) 150df8e8970SKent Overstreet 151df8e8970SKent Overstreet /** 152df8e8970SKent Overstreet * btree_root - call a function on the root of the btree 153df8e8970SKent Overstreet * @fn: function to call, which will be passed the child node 154df8e8970SKent Overstreet * @c: cache set 155df8e8970SKent Overstreet * @op: pointer to struct btree_op 156df8e8970SKent Overstreet */ 157df8e8970SKent Overstreet #define btree_root(fn, c, op, ...) \ 158df8e8970SKent Overstreet ({ \ 159df8e8970SKent Overstreet int _r = -EINTR; \ 160df8e8970SKent Overstreet do { \ 161df8e8970SKent Overstreet struct btree *_b = (c)->root; \ 162df8e8970SKent Overstreet bool _w = insert_lock(op, _b); \ 163df8e8970SKent Overstreet rw_lock(_w, _b, _b->level); \ 164df8e8970SKent Overstreet if (_b == (c)->root && \ 165df8e8970SKent Overstreet _w == insert_lock(op, _b)) { \ 166df8e8970SKent Overstreet _b->parent = NULL; \ 167df8e8970SKent Overstreet _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \ 168df8e8970SKent Overstreet } \ 169df8e8970SKent Overstreet rw_unlock(_w, _b); \ 170*78365411SKent Overstreet if (_r == -EINTR) \ 171*78365411SKent Overstreet schedule(); \ 172df8e8970SKent Overstreet bch_cannibalize_unlock(c); \ 173df8e8970SKent Overstreet if (_r == -ENOSPC) { \ 174df8e8970SKent Overstreet wait_event((c)->try_wait, \ 175df8e8970SKent Overstreet !(c)->try_harder); \ 176df8e8970SKent Overstreet _r = -EINTR; \ 177df8e8970SKent Overstreet } \ 178df8e8970SKent Overstreet } while (_r == -EINTR); \ 179df8e8970SKent Overstreet \ 180*78365411SKent Overstreet finish_wait(&(c)->bucket_wait, &(op)->wait); \ 181df8e8970SKent Overstreet _r; \ 182df8e8970SKent Overstreet }) 183df8e8970SKent Overstreet 184cafe5635SKent Overstreet /* Btree key manipulation */ 185cafe5635SKent Overstreet 1863a3b6a4eSKent Overstreet void bkey_put(struct cache_set *c, struct bkey *k) 187e7c590ebSKent Overstreet { 188e7c590ebSKent Overstreet unsigned i; 189e7c590ebSKent Overstreet 190e7c590ebSKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) 191e7c590ebSKent Overstreet if (ptr_available(c, k, i)) 192e7c590ebSKent Overstreet atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin); 193e7c590ebSKent Overstreet } 194e7c590ebSKent Overstreet 195cafe5635SKent Overstreet /* Btree IO */ 196cafe5635SKent Overstreet 197cafe5635SKent Overstreet static uint64_t btree_csum_set(struct btree *b, struct bset *i) 198cafe5635SKent Overstreet { 199cafe5635SKent Overstreet uint64_t crc = b->key.ptr[0]; 200cafe5635SKent Overstreet void *data = (void *) i + 8, *end = end(i); 201cafe5635SKent Overstreet 202169ef1cfSKent Overstreet crc = bch_crc64_update(crc, data, end - data); 203c19ed23aSKent Overstreet return crc ^ 0xffffffffffffffffULL; 204cafe5635SKent Overstreet } 205cafe5635SKent Overstreet 206f3059a54SKent Overstreet static void bch_btree_node_read_done(struct btree *b) 207cafe5635SKent Overstreet { 208cafe5635SKent Overstreet const char *err = "bad btree header"; 20957943511SKent Overstreet struct bset *i = b->sets[0].data; 21057943511SKent Overstreet struct btree_iter *iter; 211cafe5635SKent Overstreet 21257943511SKent Overstreet iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT); 21357943511SKent Overstreet iter->size = b->c->sb.bucket_size / b->c->sb.block_size; 214cafe5635SKent Overstreet iter->used = 0; 215cafe5635SKent Overstreet 216280481d0SKent Overstreet #ifdef CONFIG_BCACHE_DEBUG 217280481d0SKent Overstreet iter->b = b; 218280481d0SKent Overstreet #endif 219280481d0SKent Overstreet 22057943511SKent Overstreet if (!i->seq) 221cafe5635SKent Overstreet goto err; 222cafe5635SKent Overstreet 223cafe5635SKent Overstreet for (; 224cafe5635SKent Overstreet b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq; 225cafe5635SKent Overstreet i = write_block(b)) { 226cafe5635SKent Overstreet err = "unsupported bset version"; 227cafe5635SKent Overstreet if (i->version > BCACHE_BSET_VERSION) 228cafe5635SKent Overstreet goto err; 229cafe5635SKent Overstreet 230cafe5635SKent Overstreet err = "bad btree header"; 231cafe5635SKent Overstreet if (b->written + set_blocks(i, b->c) > btree_blocks(b)) 232cafe5635SKent Overstreet goto err; 233cafe5635SKent Overstreet 234cafe5635SKent Overstreet err = "bad magic"; 23581ab4190SKent Overstreet if (i->magic != bset_magic(&b->c->sb)) 236cafe5635SKent Overstreet goto err; 237cafe5635SKent Overstreet 238cafe5635SKent Overstreet err = "bad checksum"; 239cafe5635SKent Overstreet switch (i->version) { 240cafe5635SKent Overstreet case 0: 241cafe5635SKent Overstreet if (i->csum != csum_set(i)) 242cafe5635SKent Overstreet goto err; 243cafe5635SKent Overstreet break; 244cafe5635SKent Overstreet case BCACHE_BSET_VERSION: 245cafe5635SKent Overstreet if (i->csum != btree_csum_set(b, i)) 246cafe5635SKent Overstreet goto err; 247cafe5635SKent Overstreet break; 248cafe5635SKent Overstreet } 249cafe5635SKent Overstreet 250cafe5635SKent Overstreet err = "empty set"; 251cafe5635SKent Overstreet if (i != b->sets[0].data && !i->keys) 252cafe5635SKent Overstreet goto err; 253cafe5635SKent Overstreet 254cafe5635SKent Overstreet bch_btree_iter_push(iter, i->start, end(i)); 255cafe5635SKent Overstreet 256cafe5635SKent Overstreet b->written += set_blocks(i, b->c); 257cafe5635SKent Overstreet } 258cafe5635SKent Overstreet 259cafe5635SKent Overstreet err = "corrupted btree"; 260cafe5635SKent Overstreet for (i = write_block(b); 261cafe5635SKent Overstreet index(i, b) < btree_blocks(b); 262cafe5635SKent Overstreet i = ((void *) i) + block_bytes(b->c)) 263cafe5635SKent Overstreet if (i->seq == b->sets[0].data->seq) 264cafe5635SKent Overstreet goto err; 265cafe5635SKent Overstreet 266cafe5635SKent Overstreet bch_btree_sort_and_fix_extents(b, iter); 267cafe5635SKent Overstreet 268cafe5635SKent Overstreet i = b->sets[0].data; 269cafe5635SKent Overstreet err = "short btree key"; 270cafe5635SKent Overstreet if (b->sets[0].size && 271cafe5635SKent Overstreet bkey_cmp(&b->key, &b->sets[0].end) < 0) 272cafe5635SKent Overstreet goto err; 273cafe5635SKent Overstreet 274cafe5635SKent Overstreet if (b->written < btree_blocks(b)) 275cafe5635SKent Overstreet bch_bset_init_next(b); 276cafe5635SKent Overstreet out: 27757943511SKent Overstreet mempool_free(iter, b->c->fill_iter); 27857943511SKent Overstreet return; 279cafe5635SKent Overstreet err: 280cafe5635SKent Overstreet set_btree_node_io_error(b); 28107e86ccbSKent Overstreet bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys", 282cafe5635SKent Overstreet err, PTR_BUCKET_NR(b->c, &b->key, 0), 283cafe5635SKent Overstreet index(i, b), i->keys); 284cafe5635SKent Overstreet goto out; 285cafe5635SKent Overstreet } 286cafe5635SKent Overstreet 28757943511SKent Overstreet static void btree_node_read_endio(struct bio *bio, int error) 288cafe5635SKent Overstreet { 28957943511SKent Overstreet struct closure *cl = bio->bi_private; 29057943511SKent Overstreet closure_put(cl); 29157943511SKent Overstreet } 292cafe5635SKent Overstreet 29357943511SKent Overstreet void bch_btree_node_read(struct btree *b) 29457943511SKent Overstreet { 29557943511SKent Overstreet uint64_t start_time = local_clock(); 29657943511SKent Overstreet struct closure cl; 29757943511SKent Overstreet struct bio *bio; 298cafe5635SKent Overstreet 299c37511b8SKent Overstreet trace_bcache_btree_read(b); 300c37511b8SKent Overstreet 30157943511SKent Overstreet closure_init_stack(&cl); 302cafe5635SKent Overstreet 30357943511SKent Overstreet bio = bch_bbio_alloc(b->c); 30457943511SKent Overstreet bio->bi_rw = REQ_META|READ_SYNC; 3054f024f37SKent Overstreet bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9; 30657943511SKent Overstreet bio->bi_end_io = btree_node_read_endio; 30757943511SKent Overstreet bio->bi_private = &cl; 30857943511SKent Overstreet 30957943511SKent Overstreet bch_bio_map(bio, b->sets[0].data); 31057943511SKent Overstreet 31157943511SKent Overstreet bch_submit_bbio(bio, b->c, &b->key, 0); 31257943511SKent Overstreet closure_sync(&cl); 31357943511SKent Overstreet 31457943511SKent Overstreet if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) 31557943511SKent Overstreet set_btree_node_io_error(b); 31657943511SKent Overstreet 31757943511SKent Overstreet bch_bbio_free(bio, b->c); 31857943511SKent Overstreet 31957943511SKent Overstreet if (btree_node_io_error(b)) 32057943511SKent Overstreet goto err; 32157943511SKent Overstreet 32257943511SKent Overstreet bch_btree_node_read_done(b); 32357943511SKent Overstreet bch_time_stats_update(&b->c->btree_read_time, start_time); 32457943511SKent Overstreet 32557943511SKent Overstreet return; 32657943511SKent Overstreet err: 32761cbd250SGeert Uytterhoeven bch_cache_set_error(b->c, "io error reading bucket %zu", 32857943511SKent Overstreet PTR_BUCKET_NR(b->c, &b->key, 0)); 329cafe5635SKent Overstreet } 330cafe5635SKent Overstreet 331cafe5635SKent Overstreet static void btree_complete_write(struct btree *b, struct btree_write *w) 332cafe5635SKent Overstreet { 333cafe5635SKent Overstreet if (w->prio_blocked && 334cafe5635SKent Overstreet !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) 335119ba0f8SKent Overstreet wake_up_allocators(b->c); 336cafe5635SKent Overstreet 337cafe5635SKent Overstreet if (w->journal) { 338cafe5635SKent Overstreet atomic_dec_bug(w->journal); 339cafe5635SKent Overstreet __closure_wake_up(&b->c->journal.wait); 340cafe5635SKent Overstreet } 341cafe5635SKent Overstreet 342cafe5635SKent Overstreet w->prio_blocked = 0; 343cafe5635SKent Overstreet w->journal = NULL; 344cafe5635SKent Overstreet } 345cafe5635SKent Overstreet 346cb7a583eSKent Overstreet static void btree_node_write_unlock(struct closure *cl) 347cb7a583eSKent Overstreet { 348cb7a583eSKent Overstreet struct btree *b = container_of(cl, struct btree, io); 349cb7a583eSKent Overstreet 350cb7a583eSKent Overstreet up(&b->io_mutex); 351cb7a583eSKent Overstreet } 352cb7a583eSKent Overstreet 35357943511SKent Overstreet static void __btree_node_write_done(struct closure *cl) 354cafe5635SKent Overstreet { 355cb7a583eSKent Overstreet struct btree *b = container_of(cl, struct btree, io); 356cafe5635SKent Overstreet struct btree_write *w = btree_prev_write(b); 357cafe5635SKent Overstreet 358cafe5635SKent Overstreet bch_bbio_free(b->bio, b->c); 359cafe5635SKent Overstreet b->bio = NULL; 360cafe5635SKent Overstreet btree_complete_write(b, w); 361cafe5635SKent Overstreet 362cafe5635SKent Overstreet if (btree_node_dirty(b)) 363cafe5635SKent Overstreet queue_delayed_work(btree_io_wq, &b->work, 364cafe5635SKent Overstreet msecs_to_jiffies(30000)); 365cafe5635SKent Overstreet 366cb7a583eSKent Overstreet closure_return_with_destructor(cl, btree_node_write_unlock); 367cafe5635SKent Overstreet } 368cafe5635SKent Overstreet 36957943511SKent Overstreet static void btree_node_write_done(struct closure *cl) 370cafe5635SKent Overstreet { 371cb7a583eSKent Overstreet struct btree *b = container_of(cl, struct btree, io); 372cafe5635SKent Overstreet struct bio_vec *bv; 373cafe5635SKent Overstreet int n; 374cafe5635SKent Overstreet 3757988613bSKent Overstreet bio_for_each_segment_all(bv, b->bio, n) 376cafe5635SKent Overstreet __free_page(bv->bv_page); 377cafe5635SKent Overstreet 37857943511SKent Overstreet __btree_node_write_done(cl); 379cafe5635SKent Overstreet } 380cafe5635SKent Overstreet 38157943511SKent Overstreet static void btree_node_write_endio(struct bio *bio, int error) 38257943511SKent Overstreet { 38357943511SKent Overstreet struct closure *cl = bio->bi_private; 384cb7a583eSKent Overstreet struct btree *b = container_of(cl, struct btree, io); 38557943511SKent Overstreet 38657943511SKent Overstreet if (error) 38757943511SKent Overstreet set_btree_node_io_error(b); 38857943511SKent Overstreet 38957943511SKent Overstreet bch_bbio_count_io_errors(b->c, bio, error, "writing btree"); 39057943511SKent Overstreet closure_put(cl); 39157943511SKent Overstreet } 39257943511SKent Overstreet 39357943511SKent Overstreet static void do_btree_node_write(struct btree *b) 394cafe5635SKent Overstreet { 395cb7a583eSKent Overstreet struct closure *cl = &b->io; 396cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 397cafe5635SKent Overstreet BKEY_PADDED(key) k; 398cafe5635SKent Overstreet 399cafe5635SKent Overstreet i->version = BCACHE_BSET_VERSION; 400cafe5635SKent Overstreet i->csum = btree_csum_set(b, i); 401cafe5635SKent Overstreet 40257943511SKent Overstreet BUG_ON(b->bio); 40357943511SKent Overstreet b->bio = bch_bbio_alloc(b->c); 40457943511SKent Overstreet 40557943511SKent Overstreet b->bio->bi_end_io = btree_node_write_endio; 406faadf0c9SKent Overstreet b->bio->bi_private = cl; 407e49c7c37SKent Overstreet b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA; 4084f024f37SKent Overstreet b->bio->bi_iter.bi_size = set_blocks(i, b->c) * block_bytes(b->c); 409169ef1cfSKent Overstreet bch_bio_map(b->bio, i); 410cafe5635SKent Overstreet 411e49c7c37SKent Overstreet /* 412e49c7c37SKent Overstreet * If we're appending to a leaf node, we don't technically need FUA - 413e49c7c37SKent Overstreet * this write just needs to be persisted before the next journal write, 414e49c7c37SKent Overstreet * which will be marked FLUSH|FUA. 415e49c7c37SKent Overstreet * 416e49c7c37SKent Overstreet * Similarly if we're writing a new btree root - the pointer is going to 417e49c7c37SKent Overstreet * be in the next journal entry. 418e49c7c37SKent Overstreet * 419e49c7c37SKent Overstreet * But if we're writing a new btree node (that isn't a root) or 420e49c7c37SKent Overstreet * appending to a non leaf btree node, we need either FUA or a flush 421e49c7c37SKent Overstreet * when we write the parent with the new pointer. FUA is cheaper than a 422e49c7c37SKent Overstreet * flush, and writes appending to leaf nodes aren't blocking anything so 423e49c7c37SKent Overstreet * just make all btree node writes FUA to keep things sane. 424e49c7c37SKent Overstreet */ 425e49c7c37SKent Overstreet 426cafe5635SKent Overstreet bkey_copy(&k.key, &b->key); 427cafe5635SKent Overstreet SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); 428cafe5635SKent Overstreet 4298e51e414SKent Overstreet if (!bio_alloc_pages(b->bio, GFP_NOIO)) { 430cafe5635SKent Overstreet int j; 431cafe5635SKent Overstreet struct bio_vec *bv; 432cafe5635SKent Overstreet void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); 433cafe5635SKent Overstreet 4347988613bSKent Overstreet bio_for_each_segment_all(bv, b->bio, j) 435cafe5635SKent Overstreet memcpy(page_address(bv->bv_page), 436cafe5635SKent Overstreet base + j * PAGE_SIZE, PAGE_SIZE); 437cafe5635SKent Overstreet 438cafe5635SKent Overstreet bch_submit_bbio(b->bio, b->c, &k.key, 0); 439cafe5635SKent Overstreet 44057943511SKent Overstreet continue_at(cl, btree_node_write_done, NULL); 441cafe5635SKent Overstreet } else { 442cafe5635SKent Overstreet b->bio->bi_vcnt = 0; 443169ef1cfSKent Overstreet bch_bio_map(b->bio, i); 444cafe5635SKent Overstreet 445cafe5635SKent Overstreet bch_submit_bbio(b->bio, b->c, &k.key, 0); 446cafe5635SKent Overstreet 447cafe5635SKent Overstreet closure_sync(cl); 448cb7a583eSKent Overstreet continue_at_nobarrier(cl, __btree_node_write_done, NULL); 449cafe5635SKent Overstreet } 450cafe5635SKent Overstreet } 451cafe5635SKent Overstreet 45257943511SKent Overstreet void bch_btree_node_write(struct btree *b, struct closure *parent) 453cafe5635SKent Overstreet { 454cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 455cafe5635SKent Overstreet 456c37511b8SKent Overstreet trace_bcache_btree_write(b); 457c37511b8SKent Overstreet 458cafe5635SKent Overstreet BUG_ON(current->bio_list); 45957943511SKent Overstreet BUG_ON(b->written >= btree_blocks(b)); 46057943511SKent Overstreet BUG_ON(b->written && !i->keys); 46157943511SKent Overstreet BUG_ON(b->sets->data->seq != i->seq); 462280481d0SKent Overstreet bch_check_keys(b, "writing"); 463cafe5635SKent Overstreet 464cafe5635SKent Overstreet cancel_delayed_work(&b->work); 465cafe5635SKent Overstreet 46657943511SKent Overstreet /* If caller isn't waiting for write, parent refcount is cache set */ 467cb7a583eSKent Overstreet down(&b->io_mutex); 468cb7a583eSKent Overstreet closure_init(&b->io, parent ?: &b->c->cl); 46957943511SKent Overstreet 470cafe5635SKent Overstreet clear_bit(BTREE_NODE_dirty, &b->flags); 471cafe5635SKent Overstreet change_bit(BTREE_NODE_write_idx, &b->flags); 472cafe5635SKent Overstreet 47357943511SKent Overstreet do_btree_node_write(b); 474cafe5635SKent Overstreet 475cafe5635SKent Overstreet b->written += set_blocks(i, b->c); 476cafe5635SKent Overstreet atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, 477cafe5635SKent Overstreet &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); 478cafe5635SKent Overstreet 479cafe5635SKent Overstreet bch_btree_sort_lazy(b); 480cafe5635SKent Overstreet 481cafe5635SKent Overstreet if (b->written < btree_blocks(b)) 482cafe5635SKent Overstreet bch_bset_init_next(b); 483cafe5635SKent Overstreet } 484cafe5635SKent Overstreet 485f269af5aSKent Overstreet static void bch_btree_node_write_sync(struct btree *b) 486f269af5aSKent Overstreet { 487f269af5aSKent Overstreet struct closure cl; 488f269af5aSKent Overstreet 489f269af5aSKent Overstreet closure_init_stack(&cl); 490f269af5aSKent Overstreet bch_btree_node_write(b, &cl); 491f269af5aSKent Overstreet closure_sync(&cl); 492f269af5aSKent Overstreet } 493f269af5aSKent Overstreet 49457943511SKent Overstreet static void btree_node_write_work(struct work_struct *w) 495cafe5635SKent Overstreet { 496cafe5635SKent Overstreet struct btree *b = container_of(to_delayed_work(w), struct btree, work); 497cafe5635SKent Overstreet 49857943511SKent Overstreet rw_lock(true, b, b->level); 499cafe5635SKent Overstreet 500cafe5635SKent Overstreet if (btree_node_dirty(b)) 50157943511SKent Overstreet bch_btree_node_write(b, NULL); 50257943511SKent Overstreet rw_unlock(true, b); 503cafe5635SKent Overstreet } 504cafe5635SKent Overstreet 505c18536a7SKent Overstreet static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref) 506cafe5635SKent Overstreet { 507cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 508cafe5635SKent Overstreet struct btree_write *w = btree_current_write(b); 509cafe5635SKent Overstreet 51057943511SKent Overstreet BUG_ON(!b->written); 51157943511SKent Overstreet BUG_ON(!i->keys); 512cafe5635SKent Overstreet 51357943511SKent Overstreet if (!btree_node_dirty(b)) 51457943511SKent Overstreet queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); 51557943511SKent Overstreet 516cafe5635SKent Overstreet set_btree_node_dirty(b); 517cafe5635SKent Overstreet 518c18536a7SKent Overstreet if (journal_ref) { 519cafe5635SKent Overstreet if (w->journal && 520c18536a7SKent Overstreet journal_pin_cmp(b->c, w->journal, journal_ref)) { 521cafe5635SKent Overstreet atomic_dec_bug(w->journal); 522cafe5635SKent Overstreet w->journal = NULL; 523cafe5635SKent Overstreet } 524cafe5635SKent Overstreet 525cafe5635SKent Overstreet if (!w->journal) { 526c18536a7SKent Overstreet w->journal = journal_ref; 527cafe5635SKent Overstreet atomic_inc(w->journal); 528cafe5635SKent Overstreet } 529cafe5635SKent Overstreet } 530cafe5635SKent Overstreet 531cafe5635SKent Overstreet /* Force write if set is too big */ 53257943511SKent Overstreet if (set_bytes(i) > PAGE_SIZE - 48 && 53357943511SKent Overstreet !current->bio_list) 53457943511SKent Overstreet bch_btree_node_write(b, NULL); 535cafe5635SKent Overstreet } 536cafe5635SKent Overstreet 537cafe5635SKent Overstreet /* 538cafe5635SKent Overstreet * Btree in memory cache - allocation/freeing 539cafe5635SKent Overstreet * mca -> memory cache 540cafe5635SKent Overstreet */ 541cafe5635SKent Overstreet 542cafe5635SKent Overstreet static void mca_reinit(struct btree *b) 543cafe5635SKent Overstreet { 544cafe5635SKent Overstreet unsigned i; 545cafe5635SKent Overstreet 546cafe5635SKent Overstreet b->flags = 0; 547cafe5635SKent Overstreet b->written = 0; 548cafe5635SKent Overstreet b->nsets = 0; 549cafe5635SKent Overstreet 550cafe5635SKent Overstreet for (i = 0; i < MAX_BSETS; i++) 551cafe5635SKent Overstreet b->sets[i].size = 0; 552cafe5635SKent Overstreet /* 553cafe5635SKent Overstreet * Second loop starts at 1 because b->sets[0]->data is the memory we 554cafe5635SKent Overstreet * allocated 555cafe5635SKent Overstreet */ 556cafe5635SKent Overstreet for (i = 1; i < MAX_BSETS; i++) 557cafe5635SKent Overstreet b->sets[i].data = NULL; 558cafe5635SKent Overstreet } 559cafe5635SKent Overstreet 560cafe5635SKent Overstreet #define mca_reserve(c) (((c->root && c->root->level) \ 561cafe5635SKent Overstreet ? c->root->level : 1) * 8 + 16) 562cafe5635SKent Overstreet #define mca_can_free(c) \ 563cafe5635SKent Overstreet max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) 564cafe5635SKent Overstreet 565cafe5635SKent Overstreet static void mca_data_free(struct btree *b) 566cafe5635SKent Overstreet { 567cafe5635SKent Overstreet struct bset_tree *t = b->sets; 568cb7a583eSKent Overstreet 569cb7a583eSKent Overstreet BUG_ON(b->io_mutex.count != 1); 570cafe5635SKent Overstreet 571cafe5635SKent Overstreet if (bset_prev_bytes(b) < PAGE_SIZE) 572cafe5635SKent Overstreet kfree(t->prev); 573cafe5635SKent Overstreet else 574cafe5635SKent Overstreet free_pages((unsigned long) t->prev, 575cafe5635SKent Overstreet get_order(bset_prev_bytes(b))); 576cafe5635SKent Overstreet 577cafe5635SKent Overstreet if (bset_tree_bytes(b) < PAGE_SIZE) 578cafe5635SKent Overstreet kfree(t->tree); 579cafe5635SKent Overstreet else 580cafe5635SKent Overstreet free_pages((unsigned long) t->tree, 581cafe5635SKent Overstreet get_order(bset_tree_bytes(b))); 582cafe5635SKent Overstreet 583cafe5635SKent Overstreet free_pages((unsigned long) t->data, b->page_order); 584cafe5635SKent Overstreet 585cafe5635SKent Overstreet t->prev = NULL; 586cafe5635SKent Overstreet t->tree = NULL; 587cafe5635SKent Overstreet t->data = NULL; 588cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache_freed); 589cafe5635SKent Overstreet b->c->bucket_cache_used--; 590cafe5635SKent Overstreet } 591cafe5635SKent Overstreet 592cafe5635SKent Overstreet static void mca_bucket_free(struct btree *b) 593cafe5635SKent Overstreet { 594cafe5635SKent Overstreet BUG_ON(btree_node_dirty(b)); 595cafe5635SKent Overstreet 596cafe5635SKent Overstreet b->key.ptr[0] = 0; 597cafe5635SKent Overstreet hlist_del_init_rcu(&b->hash); 598cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache_freeable); 599cafe5635SKent Overstreet } 600cafe5635SKent Overstreet 601cafe5635SKent Overstreet static unsigned btree_order(struct bkey *k) 602cafe5635SKent Overstreet { 603cafe5635SKent Overstreet return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); 604cafe5635SKent Overstreet } 605cafe5635SKent Overstreet 606cafe5635SKent Overstreet static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) 607cafe5635SKent Overstreet { 608cafe5635SKent Overstreet struct bset_tree *t = b->sets; 609cafe5635SKent Overstreet BUG_ON(t->data); 610cafe5635SKent Overstreet 611cafe5635SKent Overstreet b->page_order = max_t(unsigned, 612cafe5635SKent Overstreet ilog2(b->c->btree_pages), 613cafe5635SKent Overstreet btree_order(k)); 614cafe5635SKent Overstreet 615cafe5635SKent Overstreet t->data = (void *) __get_free_pages(gfp, b->page_order); 616cafe5635SKent Overstreet if (!t->data) 617cafe5635SKent Overstreet goto err; 618cafe5635SKent Overstreet 619cafe5635SKent Overstreet t->tree = bset_tree_bytes(b) < PAGE_SIZE 620cafe5635SKent Overstreet ? kmalloc(bset_tree_bytes(b), gfp) 621cafe5635SKent Overstreet : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); 622cafe5635SKent Overstreet if (!t->tree) 623cafe5635SKent Overstreet goto err; 624cafe5635SKent Overstreet 625cafe5635SKent Overstreet t->prev = bset_prev_bytes(b) < PAGE_SIZE 626cafe5635SKent Overstreet ? kmalloc(bset_prev_bytes(b), gfp) 627cafe5635SKent Overstreet : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); 628cafe5635SKent Overstreet if (!t->prev) 629cafe5635SKent Overstreet goto err; 630cafe5635SKent Overstreet 631cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache); 632cafe5635SKent Overstreet b->c->bucket_cache_used++; 633cafe5635SKent Overstreet return; 634cafe5635SKent Overstreet err: 635cafe5635SKent Overstreet mca_data_free(b); 636cafe5635SKent Overstreet } 637cafe5635SKent Overstreet 638cafe5635SKent Overstreet static struct btree *mca_bucket_alloc(struct cache_set *c, 639cafe5635SKent Overstreet struct bkey *k, gfp_t gfp) 640cafe5635SKent Overstreet { 641cafe5635SKent Overstreet struct btree *b = kzalloc(sizeof(struct btree), gfp); 642cafe5635SKent Overstreet if (!b) 643cafe5635SKent Overstreet return NULL; 644cafe5635SKent Overstreet 645cafe5635SKent Overstreet init_rwsem(&b->lock); 646cafe5635SKent Overstreet lockdep_set_novalidate_class(&b->lock); 647cafe5635SKent Overstreet INIT_LIST_HEAD(&b->list); 64857943511SKent Overstreet INIT_DELAYED_WORK(&b->work, btree_node_write_work); 649cafe5635SKent Overstreet b->c = c; 650cb7a583eSKent Overstreet sema_init(&b->io_mutex, 1); 651cafe5635SKent Overstreet 652cafe5635SKent Overstreet mca_data_alloc(b, k, gfp); 653cafe5635SKent Overstreet return b; 654cafe5635SKent Overstreet } 655cafe5635SKent Overstreet 656e8e1d468SKent Overstreet static int mca_reap(struct btree *b, unsigned min_order, bool flush) 657cafe5635SKent Overstreet { 658e8e1d468SKent Overstreet struct closure cl; 659e8e1d468SKent Overstreet 660e8e1d468SKent Overstreet closure_init_stack(&cl); 661cafe5635SKent Overstreet lockdep_assert_held(&b->c->bucket_lock); 662cafe5635SKent Overstreet 663cafe5635SKent Overstreet if (!down_write_trylock(&b->lock)) 664cafe5635SKent Overstreet return -ENOMEM; 665cafe5635SKent Overstreet 666e8e1d468SKent Overstreet BUG_ON(btree_node_dirty(b) && !b->sets[0].data); 667e8e1d468SKent Overstreet 668cb7a583eSKent Overstreet if (b->page_order < min_order) 669cb7a583eSKent Overstreet goto out_unlock; 670cb7a583eSKent Overstreet 671cb7a583eSKent Overstreet if (!flush) { 672cb7a583eSKent Overstreet if (btree_node_dirty(b)) 673cb7a583eSKent Overstreet goto out_unlock; 674cb7a583eSKent Overstreet 675cb7a583eSKent Overstreet if (down_trylock(&b->io_mutex)) 676cb7a583eSKent Overstreet goto out_unlock; 677cb7a583eSKent Overstreet up(&b->io_mutex); 678cafe5635SKent Overstreet } 679cafe5635SKent Overstreet 680f269af5aSKent Overstreet if (btree_node_dirty(b)) 681f269af5aSKent Overstreet bch_btree_node_write_sync(b); 682cafe5635SKent Overstreet 683e8e1d468SKent Overstreet /* wait for any in flight btree write */ 684cb7a583eSKent Overstreet down(&b->io_mutex); 685cb7a583eSKent Overstreet up(&b->io_mutex); 686e8e1d468SKent Overstreet 687cafe5635SKent Overstreet return 0; 688cb7a583eSKent Overstreet out_unlock: 689cb7a583eSKent Overstreet rw_unlock(true, b); 690cb7a583eSKent Overstreet return -ENOMEM; 691cafe5635SKent Overstreet } 692cafe5635SKent Overstreet 6937dc19d5aSDave Chinner static unsigned long bch_mca_scan(struct shrinker *shrink, 6947dc19d5aSDave Chinner struct shrink_control *sc) 695cafe5635SKent Overstreet { 696cafe5635SKent Overstreet struct cache_set *c = container_of(shrink, struct cache_set, shrink); 697cafe5635SKent Overstreet struct btree *b, *t; 698cafe5635SKent Overstreet unsigned long i, nr = sc->nr_to_scan; 6997dc19d5aSDave Chinner unsigned long freed = 0; 700cafe5635SKent Overstreet 701cafe5635SKent Overstreet if (c->shrinker_disabled) 7027dc19d5aSDave Chinner return SHRINK_STOP; 703cafe5635SKent Overstreet 704cafe5635SKent Overstreet if (c->try_harder) 7057dc19d5aSDave Chinner return SHRINK_STOP; 706cafe5635SKent Overstreet 707cafe5635SKent Overstreet /* Return -1 if we can't do anything right now */ 708a698e08cSKent Overstreet if (sc->gfp_mask & __GFP_IO) 709cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 710cafe5635SKent Overstreet else if (!mutex_trylock(&c->bucket_lock)) 711cafe5635SKent Overstreet return -1; 712cafe5635SKent Overstreet 71336c9ea98SKent Overstreet /* 71436c9ea98SKent Overstreet * It's _really_ critical that we don't free too many btree nodes - we 71536c9ea98SKent Overstreet * have to always leave ourselves a reserve. The reserve is how we 71636c9ea98SKent Overstreet * guarantee that allocating memory for a new btree node can always 71736c9ea98SKent Overstreet * succeed, so that inserting keys into the btree can always succeed and 71836c9ea98SKent Overstreet * IO can always make forward progress: 71936c9ea98SKent Overstreet */ 720cafe5635SKent Overstreet nr /= c->btree_pages; 721cafe5635SKent Overstreet nr = min_t(unsigned long, nr, mca_can_free(c)); 722cafe5635SKent Overstreet 723cafe5635SKent Overstreet i = 0; 724cafe5635SKent Overstreet list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { 7257dc19d5aSDave Chinner if (freed >= nr) 726cafe5635SKent Overstreet break; 727cafe5635SKent Overstreet 728cafe5635SKent Overstreet if (++i > 3 && 729e8e1d468SKent Overstreet !mca_reap(b, 0, false)) { 730cafe5635SKent Overstreet mca_data_free(b); 731cafe5635SKent Overstreet rw_unlock(true, b); 7327dc19d5aSDave Chinner freed++; 733cafe5635SKent Overstreet } 734cafe5635SKent Overstreet } 735cafe5635SKent Overstreet 736b0f32a56SKent Overstreet for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { 737cafe5635SKent Overstreet if (list_empty(&c->btree_cache)) 738cafe5635SKent Overstreet goto out; 739cafe5635SKent Overstreet 740cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache, struct btree, list); 741cafe5635SKent Overstreet list_rotate_left(&c->btree_cache); 742cafe5635SKent Overstreet 743cafe5635SKent Overstreet if (!b->accessed && 744e8e1d468SKent Overstreet !mca_reap(b, 0, false)) { 745cafe5635SKent Overstreet mca_bucket_free(b); 746cafe5635SKent Overstreet mca_data_free(b); 747cafe5635SKent Overstreet rw_unlock(true, b); 7487dc19d5aSDave Chinner freed++; 749cafe5635SKent Overstreet } else 750cafe5635SKent Overstreet b->accessed = 0; 751cafe5635SKent Overstreet } 752cafe5635SKent Overstreet out: 753cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 7547dc19d5aSDave Chinner return freed; 7557dc19d5aSDave Chinner } 7567dc19d5aSDave Chinner 7577dc19d5aSDave Chinner static unsigned long bch_mca_count(struct shrinker *shrink, 7587dc19d5aSDave Chinner struct shrink_control *sc) 7597dc19d5aSDave Chinner { 7607dc19d5aSDave Chinner struct cache_set *c = container_of(shrink, struct cache_set, shrink); 7617dc19d5aSDave Chinner 7627dc19d5aSDave Chinner if (c->shrinker_disabled) 7637dc19d5aSDave Chinner return 0; 7647dc19d5aSDave Chinner 7657dc19d5aSDave Chinner if (c->try_harder) 7667dc19d5aSDave Chinner return 0; 7677dc19d5aSDave Chinner 7687dc19d5aSDave Chinner return mca_can_free(c) * c->btree_pages; 769cafe5635SKent Overstreet } 770cafe5635SKent Overstreet 771cafe5635SKent Overstreet void bch_btree_cache_free(struct cache_set *c) 772cafe5635SKent Overstreet { 773cafe5635SKent Overstreet struct btree *b; 774cafe5635SKent Overstreet struct closure cl; 775cafe5635SKent Overstreet closure_init_stack(&cl); 776cafe5635SKent Overstreet 777cafe5635SKent Overstreet if (c->shrink.list.next) 778cafe5635SKent Overstreet unregister_shrinker(&c->shrink); 779cafe5635SKent Overstreet 780cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 781cafe5635SKent Overstreet 782cafe5635SKent Overstreet #ifdef CONFIG_BCACHE_DEBUG 783cafe5635SKent Overstreet if (c->verify_data) 784cafe5635SKent Overstreet list_move(&c->verify_data->list, &c->btree_cache); 785cafe5635SKent Overstreet #endif 786cafe5635SKent Overstreet 787cafe5635SKent Overstreet list_splice(&c->btree_cache_freeable, 788cafe5635SKent Overstreet &c->btree_cache); 789cafe5635SKent Overstreet 790cafe5635SKent Overstreet while (!list_empty(&c->btree_cache)) { 791cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache, struct btree, list); 792cafe5635SKent Overstreet 793cafe5635SKent Overstreet if (btree_node_dirty(b)) 794cafe5635SKent Overstreet btree_complete_write(b, btree_current_write(b)); 795cafe5635SKent Overstreet clear_bit(BTREE_NODE_dirty, &b->flags); 796cafe5635SKent Overstreet 797cafe5635SKent Overstreet mca_data_free(b); 798cafe5635SKent Overstreet } 799cafe5635SKent Overstreet 800cafe5635SKent Overstreet while (!list_empty(&c->btree_cache_freed)) { 801cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache_freed, 802cafe5635SKent Overstreet struct btree, list); 803cafe5635SKent Overstreet list_del(&b->list); 804cafe5635SKent Overstreet cancel_delayed_work_sync(&b->work); 805cafe5635SKent Overstreet kfree(b); 806cafe5635SKent Overstreet } 807cafe5635SKent Overstreet 808cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 809cafe5635SKent Overstreet } 810cafe5635SKent Overstreet 811cafe5635SKent Overstreet int bch_btree_cache_alloc(struct cache_set *c) 812cafe5635SKent Overstreet { 813cafe5635SKent Overstreet unsigned i; 814cafe5635SKent Overstreet 815cafe5635SKent Overstreet for (i = 0; i < mca_reserve(c); i++) 81672a44517SKent Overstreet if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL)) 81772a44517SKent Overstreet return -ENOMEM; 818cafe5635SKent Overstreet 819cafe5635SKent Overstreet list_splice_init(&c->btree_cache, 820cafe5635SKent Overstreet &c->btree_cache_freeable); 821cafe5635SKent Overstreet 822cafe5635SKent Overstreet #ifdef CONFIG_BCACHE_DEBUG 823cafe5635SKent Overstreet mutex_init(&c->verify_lock); 824cafe5635SKent Overstreet 825cafe5635SKent Overstreet c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); 826cafe5635SKent Overstreet 827cafe5635SKent Overstreet if (c->verify_data && 828cafe5635SKent Overstreet c->verify_data->sets[0].data) 829cafe5635SKent Overstreet list_del_init(&c->verify_data->list); 830cafe5635SKent Overstreet else 831cafe5635SKent Overstreet c->verify_data = NULL; 832cafe5635SKent Overstreet #endif 833cafe5635SKent Overstreet 8347dc19d5aSDave Chinner c->shrink.count_objects = bch_mca_count; 8357dc19d5aSDave Chinner c->shrink.scan_objects = bch_mca_scan; 836cafe5635SKent Overstreet c->shrink.seeks = 4; 837cafe5635SKent Overstreet c->shrink.batch = c->btree_pages * 2; 838cafe5635SKent Overstreet register_shrinker(&c->shrink); 839cafe5635SKent Overstreet 840cafe5635SKent Overstreet return 0; 841cafe5635SKent Overstreet } 842cafe5635SKent Overstreet 843cafe5635SKent Overstreet /* Btree in memory cache - hash table */ 844cafe5635SKent Overstreet 845cafe5635SKent Overstreet static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) 846cafe5635SKent Overstreet { 847cafe5635SKent Overstreet return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; 848cafe5635SKent Overstreet } 849cafe5635SKent Overstreet 850cafe5635SKent Overstreet static struct btree *mca_find(struct cache_set *c, struct bkey *k) 851cafe5635SKent Overstreet { 852cafe5635SKent Overstreet struct btree *b; 853cafe5635SKent Overstreet 854cafe5635SKent Overstreet rcu_read_lock(); 855cafe5635SKent Overstreet hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) 856cafe5635SKent Overstreet if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) 857cafe5635SKent Overstreet goto out; 858cafe5635SKent Overstreet b = NULL; 859cafe5635SKent Overstreet out: 860cafe5635SKent Overstreet rcu_read_unlock(); 861cafe5635SKent Overstreet return b; 862cafe5635SKent Overstreet } 863cafe5635SKent Overstreet 864e8e1d468SKent Overstreet static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k) 865cafe5635SKent Overstreet { 866e8e1d468SKent Overstreet struct btree *b; 867cafe5635SKent Overstreet 868c37511b8SKent Overstreet trace_bcache_btree_cache_cannibalize(c); 869c37511b8SKent Overstreet 870e8e1d468SKent Overstreet if (!c->try_harder) { 871e8e1d468SKent Overstreet c->try_harder = current; 872cafe5635SKent Overstreet c->try_harder_start = local_clock(); 873e8e1d468SKent Overstreet } else if (c->try_harder != current) 874e8e1d468SKent Overstreet return ERR_PTR(-ENOSPC); 875cafe5635SKent Overstreet 876e8e1d468SKent Overstreet list_for_each_entry_reverse(b, &c->btree_cache, list) 877e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), false)) 878e8e1d468SKent Overstreet return b; 879cafe5635SKent Overstreet 880e8e1d468SKent Overstreet list_for_each_entry_reverse(b, &c->btree_cache, list) 881e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), true)) 882e8e1d468SKent Overstreet return b; 883e8e1d468SKent Overstreet 884e8e1d468SKent Overstreet return ERR_PTR(-ENOMEM); 885cafe5635SKent Overstreet } 886cafe5635SKent Overstreet 887cafe5635SKent Overstreet /* 888cafe5635SKent Overstreet * We can only have one thread cannibalizing other cached btree nodes at a time, 889cafe5635SKent Overstreet * or we'll deadlock. We use an open coded mutex to ensure that, which a 890cafe5635SKent Overstreet * cannibalize_bucket() will take. This means every time we unlock the root of 891cafe5635SKent Overstreet * the btree, we need to release this lock if we have it held. 892cafe5635SKent Overstreet */ 893df8e8970SKent Overstreet static void bch_cannibalize_unlock(struct cache_set *c) 894cafe5635SKent Overstreet { 895e8e1d468SKent Overstreet if (c->try_harder == current) { 896169ef1cfSKent Overstreet bch_time_stats_update(&c->try_harder_time, c->try_harder_start); 897cafe5635SKent Overstreet c->try_harder = NULL; 898e8e1d468SKent Overstreet wake_up(&c->try_wait); 899cafe5635SKent Overstreet } 900cafe5635SKent Overstreet } 901cafe5635SKent Overstreet 902e8e1d468SKent Overstreet static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level) 903cafe5635SKent Overstreet { 904cafe5635SKent Overstreet struct btree *b; 905cafe5635SKent Overstreet 906e8e1d468SKent Overstreet BUG_ON(current->bio_list); 907e8e1d468SKent Overstreet 908cafe5635SKent Overstreet lockdep_assert_held(&c->bucket_lock); 909cafe5635SKent Overstreet 910cafe5635SKent Overstreet if (mca_find(c, k)) 911cafe5635SKent Overstreet return NULL; 912cafe5635SKent Overstreet 913cafe5635SKent Overstreet /* btree_free() doesn't free memory; it sticks the node on the end of 914cafe5635SKent Overstreet * the list. Check if there's any freed nodes there: 915cafe5635SKent Overstreet */ 916cafe5635SKent Overstreet list_for_each_entry(b, &c->btree_cache_freeable, list) 917e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), false)) 918cafe5635SKent Overstreet goto out; 919cafe5635SKent Overstreet 920cafe5635SKent Overstreet /* We never free struct btree itself, just the memory that holds the on 921cafe5635SKent Overstreet * disk node. Check the freed list before allocating a new one: 922cafe5635SKent Overstreet */ 923cafe5635SKent Overstreet list_for_each_entry(b, &c->btree_cache_freed, list) 924e8e1d468SKent Overstreet if (!mca_reap(b, 0, false)) { 925cafe5635SKent Overstreet mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); 926cafe5635SKent Overstreet if (!b->sets[0].data) 927cafe5635SKent Overstreet goto err; 928cafe5635SKent Overstreet else 929cafe5635SKent Overstreet goto out; 930cafe5635SKent Overstreet } 931cafe5635SKent Overstreet 932cafe5635SKent Overstreet b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); 933cafe5635SKent Overstreet if (!b) 934cafe5635SKent Overstreet goto err; 935cafe5635SKent Overstreet 936cafe5635SKent Overstreet BUG_ON(!down_write_trylock(&b->lock)); 937cafe5635SKent Overstreet if (!b->sets->data) 938cafe5635SKent Overstreet goto err; 939cafe5635SKent Overstreet out: 940cb7a583eSKent Overstreet BUG_ON(b->io_mutex.count != 1); 941cafe5635SKent Overstreet 942cafe5635SKent Overstreet bkey_copy(&b->key, k); 943cafe5635SKent Overstreet list_move(&b->list, &c->btree_cache); 944cafe5635SKent Overstreet hlist_del_init_rcu(&b->hash); 945cafe5635SKent Overstreet hlist_add_head_rcu(&b->hash, mca_hash(c, k)); 946cafe5635SKent Overstreet 947cafe5635SKent Overstreet lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); 948cafe5635SKent Overstreet b->level = level; 949d6fd3b11SKent Overstreet b->parent = (void *) ~0UL; 950cafe5635SKent Overstreet 951cafe5635SKent Overstreet mca_reinit(b); 952cafe5635SKent Overstreet 953cafe5635SKent Overstreet return b; 954cafe5635SKent Overstreet err: 955cafe5635SKent Overstreet if (b) 956cafe5635SKent Overstreet rw_unlock(true, b); 957cafe5635SKent Overstreet 958e8e1d468SKent Overstreet b = mca_cannibalize(c, k); 959cafe5635SKent Overstreet if (!IS_ERR(b)) 960cafe5635SKent Overstreet goto out; 961cafe5635SKent Overstreet 962cafe5635SKent Overstreet return b; 963cafe5635SKent Overstreet } 964cafe5635SKent Overstreet 965cafe5635SKent Overstreet /** 966cafe5635SKent Overstreet * bch_btree_node_get - find a btree node in the cache and lock it, reading it 967cafe5635SKent Overstreet * in from disk if necessary. 968cafe5635SKent Overstreet * 969b54d6934SKent Overstreet * If IO is necessary and running under generic_make_request, returns -EAGAIN. 970cafe5635SKent Overstreet * 971cafe5635SKent Overstreet * The btree node will have either a read or a write lock held, depending on 972cafe5635SKent Overstreet * level and op->lock. 973cafe5635SKent Overstreet */ 974cafe5635SKent Overstreet struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, 975e8e1d468SKent Overstreet int level, bool write) 976cafe5635SKent Overstreet { 977cafe5635SKent Overstreet int i = 0; 978cafe5635SKent Overstreet struct btree *b; 979cafe5635SKent Overstreet 980cafe5635SKent Overstreet BUG_ON(level < 0); 981cafe5635SKent Overstreet retry: 982cafe5635SKent Overstreet b = mca_find(c, k); 983cafe5635SKent Overstreet 984cafe5635SKent Overstreet if (!b) { 98557943511SKent Overstreet if (current->bio_list) 98657943511SKent Overstreet return ERR_PTR(-EAGAIN); 98757943511SKent Overstreet 988cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 989e8e1d468SKent Overstreet b = mca_alloc(c, k, level); 990cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 991cafe5635SKent Overstreet 992cafe5635SKent Overstreet if (!b) 993cafe5635SKent Overstreet goto retry; 994cafe5635SKent Overstreet if (IS_ERR(b)) 995cafe5635SKent Overstreet return b; 996cafe5635SKent Overstreet 99757943511SKent Overstreet bch_btree_node_read(b); 998cafe5635SKent Overstreet 999cafe5635SKent Overstreet if (!write) 1000cafe5635SKent Overstreet downgrade_write(&b->lock); 1001cafe5635SKent Overstreet } else { 1002cafe5635SKent Overstreet rw_lock(write, b, level); 1003cafe5635SKent Overstreet if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { 1004cafe5635SKent Overstreet rw_unlock(write, b); 1005cafe5635SKent Overstreet goto retry; 1006cafe5635SKent Overstreet } 1007cafe5635SKent Overstreet BUG_ON(b->level != level); 1008cafe5635SKent Overstreet } 1009cafe5635SKent Overstreet 1010cafe5635SKent Overstreet b->accessed = 1; 1011cafe5635SKent Overstreet 1012cafe5635SKent Overstreet for (; i <= b->nsets && b->sets[i].size; i++) { 1013cafe5635SKent Overstreet prefetch(b->sets[i].tree); 1014cafe5635SKent Overstreet prefetch(b->sets[i].data); 1015cafe5635SKent Overstreet } 1016cafe5635SKent Overstreet 1017cafe5635SKent Overstreet for (; i <= b->nsets; i++) 1018cafe5635SKent Overstreet prefetch(b->sets[i].data); 1019cafe5635SKent Overstreet 102057943511SKent Overstreet if (btree_node_io_error(b)) { 1021cafe5635SKent Overstreet rw_unlock(write, b); 102257943511SKent Overstreet return ERR_PTR(-EIO); 102357943511SKent Overstreet } 102457943511SKent Overstreet 1025cafe5635SKent Overstreet BUG_ON(!b->written); 1026cafe5635SKent Overstreet 1027cafe5635SKent Overstreet return b; 1028cafe5635SKent Overstreet } 1029cafe5635SKent Overstreet 1030cafe5635SKent Overstreet static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) 1031cafe5635SKent Overstreet { 1032cafe5635SKent Overstreet struct btree *b; 1033cafe5635SKent Overstreet 1034cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1035e8e1d468SKent Overstreet b = mca_alloc(c, k, level); 1036cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1037cafe5635SKent Overstreet 1038cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(b)) { 103957943511SKent Overstreet bch_btree_node_read(b); 1040cafe5635SKent Overstreet rw_unlock(true, b); 1041cafe5635SKent Overstreet } 1042cafe5635SKent Overstreet } 1043cafe5635SKent Overstreet 1044cafe5635SKent Overstreet /* Btree alloc */ 1045cafe5635SKent Overstreet 1046e8e1d468SKent Overstreet static void btree_node_free(struct btree *b) 1047cafe5635SKent Overstreet { 1048cafe5635SKent Overstreet unsigned i; 1049cafe5635SKent Overstreet 1050c37511b8SKent Overstreet trace_bcache_btree_node_free(b); 1051c37511b8SKent Overstreet 1052cafe5635SKent Overstreet BUG_ON(b == b->c->root); 1053cafe5635SKent Overstreet 1054cafe5635SKent Overstreet if (btree_node_dirty(b)) 1055cafe5635SKent Overstreet btree_complete_write(b, btree_current_write(b)); 1056cafe5635SKent Overstreet clear_bit(BTREE_NODE_dirty, &b->flags); 1057cafe5635SKent Overstreet 1058cafe5635SKent Overstreet cancel_delayed_work(&b->work); 1059cafe5635SKent Overstreet 1060cafe5635SKent Overstreet mutex_lock(&b->c->bucket_lock); 1061cafe5635SKent Overstreet 1062cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&b->key); i++) { 1063cafe5635SKent Overstreet BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); 1064cafe5635SKent Overstreet 1065cafe5635SKent Overstreet bch_inc_gen(PTR_CACHE(b->c, &b->key, i), 1066cafe5635SKent Overstreet PTR_BUCKET(b->c, &b->key, i)); 1067cafe5635SKent Overstreet } 1068cafe5635SKent Overstreet 1069cafe5635SKent Overstreet bch_bucket_free(b->c, &b->key); 1070cafe5635SKent Overstreet mca_bucket_free(b); 1071cafe5635SKent Overstreet mutex_unlock(&b->c->bucket_lock); 1072cafe5635SKent Overstreet } 1073cafe5635SKent Overstreet 1074bc9389eeSKent Overstreet struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait) 1075cafe5635SKent Overstreet { 1076cafe5635SKent Overstreet BKEY_PADDED(key) k; 1077cafe5635SKent Overstreet struct btree *b = ERR_PTR(-EAGAIN); 1078cafe5635SKent Overstreet 1079cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1080cafe5635SKent Overstreet retry: 1081*78365411SKent Overstreet if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait)) 1082cafe5635SKent Overstreet goto err; 1083cafe5635SKent Overstreet 10843a3b6a4eSKent Overstreet bkey_put(c, &k.key); 1085cafe5635SKent Overstreet SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); 1086cafe5635SKent Overstreet 1087e8e1d468SKent Overstreet b = mca_alloc(c, &k.key, level); 1088cafe5635SKent Overstreet if (IS_ERR(b)) 1089cafe5635SKent Overstreet goto err_free; 1090cafe5635SKent Overstreet 1091cafe5635SKent Overstreet if (!b) { 1092b1a67b0fSKent Overstreet cache_bug(c, 1093b1a67b0fSKent Overstreet "Tried to allocate bucket that was in btree cache"); 1094cafe5635SKent Overstreet goto retry; 1095cafe5635SKent Overstreet } 1096cafe5635SKent Overstreet 1097cafe5635SKent Overstreet b->accessed = 1; 1098cafe5635SKent Overstreet bch_bset_init_next(b); 1099cafe5635SKent Overstreet 1100cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1101c37511b8SKent Overstreet 1102c37511b8SKent Overstreet trace_bcache_btree_node_alloc(b); 1103cafe5635SKent Overstreet return b; 1104cafe5635SKent Overstreet err_free: 1105cafe5635SKent Overstreet bch_bucket_free(c, &k.key); 1106cafe5635SKent Overstreet err: 1107cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1108c37511b8SKent Overstreet 1109c37511b8SKent Overstreet trace_bcache_btree_node_alloc_fail(b); 1110cafe5635SKent Overstreet return b; 1111cafe5635SKent Overstreet } 1112cafe5635SKent Overstreet 1113bc9389eeSKent Overstreet static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait) 1114cafe5635SKent Overstreet { 1115bc9389eeSKent Overstreet struct btree *n = bch_btree_node_alloc(b->c, b->level, wait); 1116cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(n)) 1117cafe5635SKent Overstreet bch_btree_sort_into(b, n); 1118cafe5635SKent Overstreet 1119cafe5635SKent Overstreet return n; 1120cafe5635SKent Overstreet } 1121cafe5635SKent Overstreet 11228835c123SKent Overstreet static void make_btree_freeing_key(struct btree *b, struct bkey *k) 11238835c123SKent Overstreet { 11248835c123SKent Overstreet unsigned i; 11258835c123SKent Overstreet 11268835c123SKent Overstreet bkey_copy(k, &b->key); 11278835c123SKent Overstreet bkey_copy_key(k, &ZERO_KEY); 11288835c123SKent Overstreet 11298835c123SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 11308835c123SKent Overstreet uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1; 11318835c123SKent Overstreet 11328835c123SKent Overstreet SET_PTR_GEN(k, i, g); 11338835c123SKent Overstreet } 11348835c123SKent Overstreet 11358835c123SKent Overstreet atomic_inc(&b->c->prio_blocked); 11368835c123SKent Overstreet } 11378835c123SKent Overstreet 1138*78365411SKent Overstreet static int btree_check_reserve(struct btree *b, struct btree_op *op) 1139*78365411SKent Overstreet { 1140*78365411SKent Overstreet struct cache_set *c = b->c; 1141*78365411SKent Overstreet struct cache *ca; 1142*78365411SKent Overstreet unsigned i, reserve = c->root->level * 2 + 1; 1143*78365411SKent Overstreet int ret = 0; 1144*78365411SKent Overstreet 1145*78365411SKent Overstreet mutex_lock(&c->bucket_lock); 1146*78365411SKent Overstreet 1147*78365411SKent Overstreet for_each_cache(ca, c, i) 1148*78365411SKent Overstreet if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) { 1149*78365411SKent Overstreet if (op) 1150*78365411SKent Overstreet prepare_to_wait(&c->bucket_wait, &op->wait, 1151*78365411SKent Overstreet TASK_UNINTERRUPTIBLE); 1152*78365411SKent Overstreet ret = -EINTR; 1153*78365411SKent Overstreet break; 1154*78365411SKent Overstreet } 1155*78365411SKent Overstreet 1156*78365411SKent Overstreet mutex_unlock(&c->bucket_lock); 1157*78365411SKent Overstreet return ret; 1158*78365411SKent Overstreet } 1159*78365411SKent Overstreet 1160cafe5635SKent Overstreet /* Garbage collection */ 1161cafe5635SKent Overstreet 1162cafe5635SKent Overstreet uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) 1163cafe5635SKent Overstreet { 1164cafe5635SKent Overstreet uint8_t stale = 0; 1165cafe5635SKent Overstreet unsigned i; 1166cafe5635SKent Overstreet struct bucket *g; 1167cafe5635SKent Overstreet 1168cafe5635SKent Overstreet /* 1169cafe5635SKent Overstreet * ptr_invalid() can't return true for the keys that mark btree nodes as 1170cafe5635SKent Overstreet * freed, but since ptr_bad() returns true we'll never actually use them 1171cafe5635SKent Overstreet * for anything and thus we don't want mark their pointers here 1172cafe5635SKent Overstreet */ 1173cafe5635SKent Overstreet if (!bkey_cmp(k, &ZERO_KEY)) 1174cafe5635SKent Overstreet return stale; 1175cafe5635SKent Overstreet 1176cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 1177cafe5635SKent Overstreet if (!ptr_available(c, k, i)) 1178cafe5635SKent Overstreet continue; 1179cafe5635SKent Overstreet 1180cafe5635SKent Overstreet g = PTR_BUCKET(c, k, i); 1181cafe5635SKent Overstreet 1182cafe5635SKent Overstreet if (gen_after(g->gc_gen, PTR_GEN(k, i))) 1183cafe5635SKent Overstreet g->gc_gen = PTR_GEN(k, i); 1184cafe5635SKent Overstreet 1185cafe5635SKent Overstreet if (ptr_stale(c, k, i)) { 1186cafe5635SKent Overstreet stale = max(stale, ptr_stale(c, k, i)); 1187cafe5635SKent Overstreet continue; 1188cafe5635SKent Overstreet } 1189cafe5635SKent Overstreet 1190cafe5635SKent Overstreet cache_bug_on(GC_MARK(g) && 1191cafe5635SKent Overstreet (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), 1192cafe5635SKent Overstreet c, "inconsistent ptrs: mark = %llu, level = %i", 1193cafe5635SKent Overstreet GC_MARK(g), level); 1194cafe5635SKent Overstreet 1195cafe5635SKent Overstreet if (level) 1196cafe5635SKent Overstreet SET_GC_MARK(g, GC_MARK_METADATA); 1197cafe5635SKent Overstreet else if (KEY_DIRTY(k)) 1198cafe5635SKent Overstreet SET_GC_MARK(g, GC_MARK_DIRTY); 1199cafe5635SKent Overstreet 1200cafe5635SKent Overstreet /* guard against overflow */ 1201cafe5635SKent Overstreet SET_GC_SECTORS_USED(g, min_t(unsigned, 1202cafe5635SKent Overstreet GC_SECTORS_USED(g) + KEY_SIZE(k), 1203cafe5635SKent Overstreet (1 << 14) - 1)); 1204cafe5635SKent Overstreet 1205cafe5635SKent Overstreet BUG_ON(!GC_SECTORS_USED(g)); 1206cafe5635SKent Overstreet } 1207cafe5635SKent Overstreet 1208cafe5635SKent Overstreet return stale; 1209cafe5635SKent Overstreet } 1210cafe5635SKent Overstreet 1211cafe5635SKent Overstreet #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) 1212cafe5635SKent Overstreet 1213a1f0358bSKent Overstreet static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc) 1214cafe5635SKent Overstreet { 1215cafe5635SKent Overstreet uint8_t stale = 0; 1216a1f0358bSKent Overstreet unsigned keys = 0, good_keys = 0; 1217cafe5635SKent Overstreet struct bkey *k; 1218cafe5635SKent Overstreet struct btree_iter iter; 1219cafe5635SKent Overstreet struct bset_tree *t; 1220cafe5635SKent Overstreet 1221cafe5635SKent Overstreet gc->nodes++; 1222cafe5635SKent Overstreet 1223cafe5635SKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1224cafe5635SKent Overstreet stale = max(stale, btree_mark_key(b, k)); 1225a1f0358bSKent Overstreet keys++; 1226cafe5635SKent Overstreet 1227cafe5635SKent Overstreet if (bch_ptr_bad(b, k)) 1228cafe5635SKent Overstreet continue; 1229cafe5635SKent Overstreet 1230cafe5635SKent Overstreet gc->key_bytes += bkey_u64s(k); 1231cafe5635SKent Overstreet gc->nkeys++; 1232a1f0358bSKent Overstreet good_keys++; 1233cafe5635SKent Overstreet 1234cafe5635SKent Overstreet gc->data += KEY_SIZE(k); 1235cafe5635SKent Overstreet } 1236cafe5635SKent Overstreet 1237cafe5635SKent Overstreet for (t = b->sets; t <= &b->sets[b->nsets]; t++) 1238cafe5635SKent Overstreet btree_bug_on(t->size && 1239cafe5635SKent Overstreet bset_written(b, t) && 1240cafe5635SKent Overstreet bkey_cmp(&b->key, &t->end) < 0, 1241cafe5635SKent Overstreet b, "found short btree key in gc"); 1242cafe5635SKent Overstreet 1243a1f0358bSKent Overstreet if (b->c->gc_always_rewrite) 1244a1f0358bSKent Overstreet return true; 1245a1f0358bSKent Overstreet 1246a1f0358bSKent Overstreet if (stale > 10) 1247a1f0358bSKent Overstreet return true; 1248a1f0358bSKent Overstreet 1249a1f0358bSKent Overstreet if ((keys - good_keys) * 2 > keys) 1250a1f0358bSKent Overstreet return true; 1251a1f0358bSKent Overstreet 1252a1f0358bSKent Overstreet return false; 1253cafe5635SKent Overstreet } 1254cafe5635SKent Overstreet 1255a1f0358bSKent Overstreet #define GC_MERGE_NODES 4U 1256cafe5635SKent Overstreet 1257cafe5635SKent Overstreet struct gc_merge_info { 1258cafe5635SKent Overstreet struct btree *b; 1259cafe5635SKent Overstreet unsigned keys; 1260cafe5635SKent Overstreet }; 1261cafe5635SKent Overstreet 1262a1f0358bSKent Overstreet static int bch_btree_insert_node(struct btree *, struct btree_op *, 1263a1f0358bSKent Overstreet struct keylist *, atomic_t *, struct bkey *); 1264a1f0358bSKent Overstreet 1265a1f0358bSKent Overstreet static int btree_gc_coalesce(struct btree *b, struct btree_op *op, 1266a1f0358bSKent Overstreet struct keylist *keylist, struct gc_stat *gc, 1267e8e1d468SKent Overstreet struct gc_merge_info *r) 1268cafe5635SKent Overstreet { 1269a1f0358bSKent Overstreet unsigned i, nodes = 0, keys = 0, blocks; 1270a1f0358bSKent Overstreet struct btree *new_nodes[GC_MERGE_NODES]; 1271b54d6934SKent Overstreet struct closure cl; 1272a1f0358bSKent Overstreet struct bkey *k; 1273b54d6934SKent Overstreet 1274a1f0358bSKent Overstreet memset(new_nodes, 0, sizeof(new_nodes)); 1275b54d6934SKent Overstreet closure_init_stack(&cl); 1276cafe5635SKent Overstreet 1277a1f0358bSKent Overstreet while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b)) 1278cafe5635SKent Overstreet keys += r[nodes++].keys; 1279cafe5635SKent Overstreet 1280cafe5635SKent Overstreet blocks = btree_default_blocks(b->c) * 2 / 3; 1281cafe5635SKent Overstreet 1282cafe5635SKent Overstreet if (nodes < 2 || 1283cafe5635SKent Overstreet __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) 1284a1f0358bSKent Overstreet return 0; 1285cafe5635SKent Overstreet 1286a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1287bc9389eeSKent Overstreet new_nodes[i] = btree_node_alloc_replacement(r[i].b, false); 1288a1f0358bSKent Overstreet if (IS_ERR_OR_NULL(new_nodes[i])) 1289a1f0358bSKent Overstreet goto out_nocoalesce; 1290cafe5635SKent Overstreet } 1291cafe5635SKent Overstreet 1292cafe5635SKent Overstreet for (i = nodes - 1; i > 0; --i) { 1293a1f0358bSKent Overstreet struct bset *n1 = new_nodes[i]->sets->data; 1294a1f0358bSKent Overstreet struct bset *n2 = new_nodes[i - 1]->sets->data; 1295cafe5635SKent Overstreet struct bkey *k, *last = NULL; 1296cafe5635SKent Overstreet 1297cafe5635SKent Overstreet keys = 0; 1298cafe5635SKent Overstreet 1299a1f0358bSKent Overstreet if (i > 1) { 1300cafe5635SKent Overstreet for (k = n2->start; 1301cafe5635SKent Overstreet k < end(n2); 1302cafe5635SKent Overstreet k = bkey_next(k)) { 1303cafe5635SKent Overstreet if (__set_blocks(n1, n1->keys + keys + 1304cafe5635SKent Overstreet bkey_u64s(k), b->c) > blocks) 1305cafe5635SKent Overstreet break; 1306cafe5635SKent Overstreet 1307cafe5635SKent Overstreet last = k; 1308cafe5635SKent Overstreet keys += bkey_u64s(k); 1309cafe5635SKent Overstreet } 1310a1f0358bSKent Overstreet } else { 1311a1f0358bSKent Overstreet /* 1312a1f0358bSKent Overstreet * Last node we're not getting rid of - we're getting 1313a1f0358bSKent Overstreet * rid of the node at r[0]. Have to try and fit all of 1314a1f0358bSKent Overstreet * the remaining keys into this node; we can't ensure 1315a1f0358bSKent Overstreet * they will always fit due to rounding and variable 1316a1f0358bSKent Overstreet * length keys (shouldn't be possible in practice, 1317a1f0358bSKent Overstreet * though) 1318a1f0358bSKent Overstreet */ 1319a1f0358bSKent Overstreet if (__set_blocks(n1, n1->keys + n2->keys, 1320a1f0358bSKent Overstreet b->c) > btree_blocks(new_nodes[i])) 1321a1f0358bSKent Overstreet goto out_nocoalesce; 1322a1f0358bSKent Overstreet 1323a1f0358bSKent Overstreet keys = n2->keys; 1324a1f0358bSKent Overstreet /* Take the key of the node we're getting rid of */ 1325a1f0358bSKent Overstreet last = &r->b->key; 1326a1f0358bSKent Overstreet } 1327cafe5635SKent Overstreet 1328cafe5635SKent Overstreet BUG_ON(__set_blocks(n1, n1->keys + keys, 1329a1f0358bSKent Overstreet b->c) > btree_blocks(new_nodes[i])); 1330cafe5635SKent Overstreet 1331a1f0358bSKent Overstreet if (last) 1332a1f0358bSKent Overstreet bkey_copy_key(&new_nodes[i]->key, last); 1333cafe5635SKent Overstreet 1334cafe5635SKent Overstreet memcpy(end(n1), 1335cafe5635SKent Overstreet n2->start, 1336cafe5635SKent Overstreet (void *) node(n2, keys) - (void *) n2->start); 1337cafe5635SKent Overstreet 1338cafe5635SKent Overstreet n1->keys += keys; 1339a1f0358bSKent Overstreet r[i].keys = n1->keys; 1340cafe5635SKent Overstreet 1341cafe5635SKent Overstreet memmove(n2->start, 1342cafe5635SKent Overstreet node(n2, keys), 1343cafe5635SKent Overstreet (void *) end(n2) - (void *) node(n2, keys)); 1344cafe5635SKent Overstreet 1345cafe5635SKent Overstreet n2->keys -= keys; 1346cafe5635SKent Overstreet 1347a1f0358bSKent Overstreet if (bch_keylist_realloc(keylist, 1348a1f0358bSKent Overstreet KEY_PTRS(&new_nodes[i]->key), b->c)) 1349a1f0358bSKent Overstreet goto out_nocoalesce; 1350a1f0358bSKent Overstreet 1351a1f0358bSKent Overstreet bch_btree_node_write(new_nodes[i], &cl); 1352a1f0358bSKent Overstreet bch_keylist_add(keylist, &new_nodes[i]->key); 1353cafe5635SKent Overstreet } 1354cafe5635SKent Overstreet 1355a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1356a1f0358bSKent Overstreet if (bch_keylist_realloc(keylist, KEY_PTRS(&r[i].b->key), b->c)) 1357a1f0358bSKent Overstreet goto out_nocoalesce; 1358a1f0358bSKent Overstreet 1359a1f0358bSKent Overstreet make_btree_freeing_key(r[i].b, keylist->top); 1360a1f0358bSKent Overstreet bch_keylist_push(keylist); 1361a1f0358bSKent Overstreet } 1362a1f0358bSKent Overstreet 1363a1f0358bSKent Overstreet /* We emptied out this node */ 1364a1f0358bSKent Overstreet BUG_ON(new_nodes[0]->sets->data->keys); 1365a1f0358bSKent Overstreet btree_node_free(new_nodes[0]); 1366a1f0358bSKent Overstreet rw_unlock(true, new_nodes[0]); 1367a1f0358bSKent Overstreet 1368a1f0358bSKent Overstreet closure_sync(&cl); 1369a1f0358bSKent Overstreet 1370a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1371a1f0358bSKent Overstreet btree_node_free(r[i].b); 1372a1f0358bSKent Overstreet rw_unlock(true, r[i].b); 1373a1f0358bSKent Overstreet 1374a1f0358bSKent Overstreet r[i].b = new_nodes[i]; 1375a1f0358bSKent Overstreet } 1376a1f0358bSKent Overstreet 1377a1f0358bSKent Overstreet bch_btree_insert_node(b, op, keylist, NULL, NULL); 1378a1f0358bSKent Overstreet BUG_ON(!bch_keylist_empty(keylist)); 1379a1f0358bSKent Overstreet 1380a1f0358bSKent Overstreet memmove(r, r + 1, sizeof(r[0]) * (nodes - 1)); 1381a1f0358bSKent Overstreet r[nodes - 1].b = ERR_PTR(-EINTR); 1382cafe5635SKent Overstreet 1383c37511b8SKent Overstreet trace_bcache_btree_gc_coalesce(nodes); 1384cafe5635SKent Overstreet gc->nodes--; 1385cafe5635SKent Overstreet 1386a1f0358bSKent Overstreet /* Invalidated our iterator */ 1387a1f0358bSKent Overstreet return -EINTR; 1388a1f0358bSKent Overstreet 1389a1f0358bSKent Overstreet out_nocoalesce: 1390a1f0358bSKent Overstreet closure_sync(&cl); 1391a1f0358bSKent Overstreet 1392a1f0358bSKent Overstreet while ((k = bch_keylist_pop(keylist))) 1393a1f0358bSKent Overstreet if (!bkey_cmp(k, &ZERO_KEY)) 1394a1f0358bSKent Overstreet atomic_dec(&b->c->prio_blocked); 1395a1f0358bSKent Overstreet 1396a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) 1397a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(new_nodes[i])) { 1398a1f0358bSKent Overstreet btree_node_free(new_nodes[i]); 1399a1f0358bSKent Overstreet rw_unlock(true, new_nodes[i]); 1400a1f0358bSKent Overstreet } 1401a1f0358bSKent Overstreet return 0; 1402a1f0358bSKent Overstreet } 1403a1f0358bSKent Overstreet 1404a1f0358bSKent Overstreet static unsigned btree_gc_count_keys(struct btree *b) 1405a1f0358bSKent Overstreet { 1406a1f0358bSKent Overstreet struct bkey *k; 1407a1f0358bSKent Overstreet struct btree_iter iter; 1408a1f0358bSKent Overstreet unsigned ret = 0; 1409a1f0358bSKent Overstreet 1410a1f0358bSKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_bad) 1411a1f0358bSKent Overstreet ret += bkey_u64s(k); 1412a1f0358bSKent Overstreet 1413a1f0358bSKent Overstreet return ret; 1414cafe5635SKent Overstreet } 1415cafe5635SKent Overstreet 1416cafe5635SKent Overstreet static int btree_gc_recurse(struct btree *b, struct btree_op *op, 1417cafe5635SKent Overstreet struct closure *writes, struct gc_stat *gc) 1418cafe5635SKent Overstreet { 1419cafe5635SKent Overstreet unsigned i; 1420a1f0358bSKent Overstreet int ret = 0; 1421a1f0358bSKent Overstreet bool should_rewrite; 1422a1f0358bSKent Overstreet struct btree *n; 1423a1f0358bSKent Overstreet struct bkey *k; 1424a1f0358bSKent Overstreet struct keylist keys; 1425a1f0358bSKent Overstreet struct btree_iter iter; 1426cafe5635SKent Overstreet struct gc_merge_info r[GC_MERGE_NODES]; 1427a1f0358bSKent Overstreet struct gc_merge_info *last = r + GC_MERGE_NODES - 1; 1428cafe5635SKent Overstreet 1429a1f0358bSKent Overstreet bch_keylist_init(&keys); 1430a1f0358bSKent Overstreet bch_btree_iter_init(b, &iter, &b->c->gc_done); 1431cafe5635SKent Overstreet 1432a1f0358bSKent Overstreet for (i = 0; i < GC_MERGE_NODES; i++) 1433a1f0358bSKent Overstreet r[i].b = ERR_PTR(-EINTR); 1434cafe5635SKent Overstreet 1435a1f0358bSKent Overstreet while (1) { 1436a1f0358bSKent Overstreet k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); 1437a1f0358bSKent Overstreet if (k) { 1438a1f0358bSKent Overstreet r->b = bch_btree_node_get(b->c, k, b->level - 1, true); 1439cafe5635SKent Overstreet if (IS_ERR(r->b)) { 1440cafe5635SKent Overstreet ret = PTR_ERR(r->b); 1441cafe5635SKent Overstreet break; 1442cafe5635SKent Overstreet } 1443cafe5635SKent Overstreet 1444a1f0358bSKent Overstreet r->keys = btree_gc_count_keys(r->b); 1445cafe5635SKent Overstreet 1446a1f0358bSKent Overstreet ret = btree_gc_coalesce(b, op, &keys, gc, r); 1447a1f0358bSKent Overstreet if (ret) 1448cafe5635SKent Overstreet break; 1449cafe5635SKent Overstreet } 1450cafe5635SKent Overstreet 1451a1f0358bSKent Overstreet if (!last->b) 1452a1f0358bSKent Overstreet break; 1453cafe5635SKent Overstreet 1454a1f0358bSKent Overstreet if (!IS_ERR(last->b)) { 1455a1f0358bSKent Overstreet should_rewrite = btree_gc_mark_node(last->b, gc); 1456*78365411SKent Overstreet if (should_rewrite && 1457*78365411SKent Overstreet !btree_check_reserve(b, NULL)) { 1458bc9389eeSKent Overstreet n = btree_node_alloc_replacement(last->b, 1459bc9389eeSKent Overstreet false); 1460cafe5635SKent Overstreet 1461a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(n)) { 1462a1f0358bSKent Overstreet bch_btree_node_write_sync(n); 1463a1f0358bSKent Overstreet bch_keylist_add(&keys, &n->key); 1464cafe5635SKent Overstreet 1465a1f0358bSKent Overstreet make_btree_freeing_key(last->b, 1466a1f0358bSKent Overstreet keys.top); 1467a1f0358bSKent Overstreet bch_keylist_push(&keys); 1468cafe5635SKent Overstreet 1469a1f0358bSKent Overstreet btree_node_free(last->b); 1470a1f0358bSKent Overstreet 1471a1f0358bSKent Overstreet bch_btree_insert_node(b, op, &keys, 1472a1f0358bSKent Overstreet NULL, NULL); 1473a1f0358bSKent Overstreet BUG_ON(!bch_keylist_empty(&keys)); 1474a1f0358bSKent Overstreet 1475a1f0358bSKent Overstreet rw_unlock(true, last->b); 1476a1f0358bSKent Overstreet last->b = n; 1477a1f0358bSKent Overstreet 1478a1f0358bSKent Overstreet /* Invalidated our iterator */ 1479a1f0358bSKent Overstreet ret = -EINTR; 1480a1f0358bSKent Overstreet break; 1481a1f0358bSKent Overstreet } 1482a1f0358bSKent Overstreet } 1483a1f0358bSKent Overstreet 1484a1f0358bSKent Overstreet if (last->b->level) { 1485a1f0358bSKent Overstreet ret = btree_gc_recurse(last->b, op, writes, gc); 1486a1f0358bSKent Overstreet if (ret) 1487a1f0358bSKent Overstreet break; 1488a1f0358bSKent Overstreet } 1489a1f0358bSKent Overstreet 1490a1f0358bSKent Overstreet bkey_copy_key(&b->c->gc_done, &last->b->key); 1491a1f0358bSKent Overstreet 1492a1f0358bSKent Overstreet /* 1493a1f0358bSKent Overstreet * Must flush leaf nodes before gc ends, since replace 1494a1f0358bSKent Overstreet * operations aren't journalled 1495cafe5635SKent Overstreet */ 1496a1f0358bSKent Overstreet if (btree_node_dirty(last->b)) 1497a1f0358bSKent Overstreet bch_btree_node_write(last->b, writes); 1498a1f0358bSKent Overstreet rw_unlock(true, last->b); 1499a1f0358bSKent Overstreet } 1500a1f0358bSKent Overstreet 1501a1f0358bSKent Overstreet memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1)); 1502a1f0358bSKent Overstreet r->b = NULL; 1503a1f0358bSKent Overstreet 1504cafe5635SKent Overstreet if (need_resched()) { 1505cafe5635SKent Overstreet ret = -EAGAIN; 1506cafe5635SKent Overstreet break; 1507cafe5635SKent Overstreet } 1508cafe5635SKent Overstreet } 1509cafe5635SKent Overstreet 1510a1f0358bSKent Overstreet for (i = 0; i < GC_MERGE_NODES; i++) 1511a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(r[i].b)) { 1512a1f0358bSKent Overstreet if (btree_node_dirty(r[i].b)) 1513a1f0358bSKent Overstreet bch_btree_node_write(r[i].b, writes); 1514a1f0358bSKent Overstreet rw_unlock(true, r[i].b); 1515a1f0358bSKent Overstreet } 1516cafe5635SKent Overstreet 1517a1f0358bSKent Overstreet bch_keylist_free(&keys); 1518cafe5635SKent Overstreet 1519cafe5635SKent Overstreet return ret; 1520cafe5635SKent Overstreet } 1521cafe5635SKent Overstreet 1522cafe5635SKent Overstreet static int bch_btree_gc_root(struct btree *b, struct btree_op *op, 1523cafe5635SKent Overstreet struct closure *writes, struct gc_stat *gc) 1524cafe5635SKent Overstreet { 1525cafe5635SKent Overstreet struct btree *n = NULL; 1526a1f0358bSKent Overstreet int ret = 0; 1527a1f0358bSKent Overstreet bool should_rewrite; 1528cafe5635SKent Overstreet 1529a1f0358bSKent Overstreet should_rewrite = btree_gc_mark_node(b, gc); 1530a1f0358bSKent Overstreet if (should_rewrite) { 1531bc9389eeSKent Overstreet n = btree_node_alloc_replacement(b, false); 1532cafe5635SKent Overstreet 1533cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(n)) { 1534a1f0358bSKent Overstreet bch_btree_node_write_sync(n); 1535a1f0358bSKent Overstreet bch_btree_set_root(n); 1536a1f0358bSKent Overstreet btree_node_free(b); 1537a1f0358bSKent Overstreet rw_unlock(true, n); 1538a1f0358bSKent Overstreet 1539a1f0358bSKent Overstreet return -EINTR; 1540cafe5635SKent Overstreet } 1541a1f0358bSKent Overstreet } 1542a1f0358bSKent Overstreet 1543a1f0358bSKent Overstreet if (b->level) { 1544a1f0358bSKent Overstreet ret = btree_gc_recurse(b, op, writes, gc); 1545a1f0358bSKent Overstreet if (ret) 1546a1f0358bSKent Overstreet return ret; 1547a1f0358bSKent Overstreet } 1548a1f0358bSKent Overstreet 1549a1f0358bSKent Overstreet bkey_copy_key(&b->c->gc_done, &b->key); 1550cafe5635SKent Overstreet 1551cafe5635SKent Overstreet return ret; 1552cafe5635SKent Overstreet } 1553cafe5635SKent Overstreet 1554cafe5635SKent Overstreet static void btree_gc_start(struct cache_set *c) 1555cafe5635SKent Overstreet { 1556cafe5635SKent Overstreet struct cache *ca; 1557cafe5635SKent Overstreet struct bucket *b; 1558cafe5635SKent Overstreet unsigned i; 1559cafe5635SKent Overstreet 1560cafe5635SKent Overstreet if (!c->gc_mark_valid) 1561cafe5635SKent Overstreet return; 1562cafe5635SKent Overstreet 1563cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1564cafe5635SKent Overstreet 1565cafe5635SKent Overstreet c->gc_mark_valid = 0; 1566cafe5635SKent Overstreet c->gc_done = ZERO_KEY; 1567cafe5635SKent Overstreet 1568cafe5635SKent Overstreet for_each_cache(ca, c, i) 1569cafe5635SKent Overstreet for_each_bucket(b, ca) { 1570cafe5635SKent Overstreet b->gc_gen = b->gen; 157129ebf465SKent Overstreet if (!atomic_read(&b->pin)) { 1572cafe5635SKent Overstreet SET_GC_MARK(b, GC_MARK_RECLAIMABLE); 157329ebf465SKent Overstreet SET_GC_SECTORS_USED(b, 0); 157429ebf465SKent Overstreet } 1575cafe5635SKent Overstreet } 1576cafe5635SKent Overstreet 1577cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1578cafe5635SKent Overstreet } 1579cafe5635SKent Overstreet 1580cafe5635SKent Overstreet size_t bch_btree_gc_finish(struct cache_set *c) 1581cafe5635SKent Overstreet { 1582cafe5635SKent Overstreet size_t available = 0; 1583cafe5635SKent Overstreet struct bucket *b; 1584cafe5635SKent Overstreet struct cache *ca; 1585cafe5635SKent Overstreet unsigned i; 1586cafe5635SKent Overstreet 1587cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1588cafe5635SKent Overstreet 1589cafe5635SKent Overstreet set_gc_sectors(c); 1590cafe5635SKent Overstreet c->gc_mark_valid = 1; 1591cafe5635SKent Overstreet c->need_gc = 0; 1592cafe5635SKent Overstreet 1593cafe5635SKent Overstreet if (c->root) 1594cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&c->root->key); i++) 1595cafe5635SKent Overstreet SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), 1596cafe5635SKent Overstreet GC_MARK_METADATA); 1597cafe5635SKent Overstreet 1598cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) 1599cafe5635SKent Overstreet SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), 1600cafe5635SKent Overstreet GC_MARK_METADATA); 1601cafe5635SKent Overstreet 1602bf0a628aSNicholas Swenson /* don't reclaim buckets to which writeback keys point */ 1603bf0a628aSNicholas Swenson rcu_read_lock(); 1604bf0a628aSNicholas Swenson for (i = 0; i < c->nr_uuids; i++) { 1605bf0a628aSNicholas Swenson struct bcache_device *d = c->devices[i]; 1606bf0a628aSNicholas Swenson struct cached_dev *dc; 1607bf0a628aSNicholas Swenson struct keybuf_key *w, *n; 1608bf0a628aSNicholas Swenson unsigned j; 1609bf0a628aSNicholas Swenson 1610bf0a628aSNicholas Swenson if (!d || UUID_FLASH_ONLY(&c->uuids[i])) 1611bf0a628aSNicholas Swenson continue; 1612bf0a628aSNicholas Swenson dc = container_of(d, struct cached_dev, disk); 1613bf0a628aSNicholas Swenson 1614bf0a628aSNicholas Swenson spin_lock(&dc->writeback_keys.lock); 1615bf0a628aSNicholas Swenson rbtree_postorder_for_each_entry_safe(w, n, 1616bf0a628aSNicholas Swenson &dc->writeback_keys.keys, node) 1617bf0a628aSNicholas Swenson for (j = 0; j < KEY_PTRS(&w->key); j++) 1618bf0a628aSNicholas Swenson SET_GC_MARK(PTR_BUCKET(c, &w->key, j), 1619bf0a628aSNicholas Swenson GC_MARK_DIRTY); 1620bf0a628aSNicholas Swenson spin_unlock(&dc->writeback_keys.lock); 1621bf0a628aSNicholas Swenson } 1622bf0a628aSNicholas Swenson rcu_read_unlock(); 1623bf0a628aSNicholas Swenson 1624cafe5635SKent Overstreet for_each_cache(ca, c, i) { 1625cafe5635SKent Overstreet uint64_t *i; 1626cafe5635SKent Overstreet 1627cafe5635SKent Overstreet ca->invalidate_needs_gc = 0; 1628cafe5635SKent Overstreet 1629cafe5635SKent Overstreet for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) 1630cafe5635SKent Overstreet SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1631cafe5635SKent Overstreet 1632cafe5635SKent Overstreet for (i = ca->prio_buckets; 1633cafe5635SKent Overstreet i < ca->prio_buckets + prio_buckets(ca) * 2; i++) 1634cafe5635SKent Overstreet SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1635cafe5635SKent Overstreet 1636cafe5635SKent Overstreet for_each_bucket(b, ca) { 1637cafe5635SKent Overstreet b->last_gc = b->gc_gen; 1638cafe5635SKent Overstreet c->need_gc = max(c->need_gc, bucket_gc_gen(b)); 1639cafe5635SKent Overstreet 1640cafe5635SKent Overstreet if (!atomic_read(&b->pin) && 1641cafe5635SKent Overstreet GC_MARK(b) == GC_MARK_RECLAIMABLE) { 1642cafe5635SKent Overstreet available++; 1643cafe5635SKent Overstreet if (!GC_SECTORS_USED(b)) 1644cafe5635SKent Overstreet bch_bucket_add_unused(ca, b); 1645cafe5635SKent Overstreet } 1646cafe5635SKent Overstreet } 1647cafe5635SKent Overstreet } 1648cafe5635SKent Overstreet 1649cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1650cafe5635SKent Overstreet return available; 1651cafe5635SKent Overstreet } 1652cafe5635SKent Overstreet 165372a44517SKent Overstreet static void bch_btree_gc(struct cache_set *c) 1654cafe5635SKent Overstreet { 1655cafe5635SKent Overstreet int ret; 1656cafe5635SKent Overstreet unsigned long available; 1657cafe5635SKent Overstreet struct gc_stat stats; 1658cafe5635SKent Overstreet struct closure writes; 1659cafe5635SKent Overstreet struct btree_op op; 1660cafe5635SKent Overstreet uint64_t start_time = local_clock(); 166157943511SKent Overstreet 1662c37511b8SKent Overstreet trace_bcache_gc_start(c); 1663cafe5635SKent Overstreet 1664cafe5635SKent Overstreet memset(&stats, 0, sizeof(struct gc_stat)); 1665cafe5635SKent Overstreet closure_init_stack(&writes); 1666b54d6934SKent Overstreet bch_btree_op_init(&op, SHRT_MAX); 1667cafe5635SKent Overstreet 1668cafe5635SKent Overstreet btree_gc_start(c); 1669cafe5635SKent Overstreet 1670a1f0358bSKent Overstreet do { 1671cafe5635SKent Overstreet ret = btree_root(gc_root, c, &op, &writes, &stats); 1672cafe5635SKent Overstreet closure_sync(&writes); 1673cafe5635SKent Overstreet 1674a1f0358bSKent Overstreet if (ret && ret != -EAGAIN) 1675cafe5635SKent Overstreet pr_warn("gc failed!"); 1676a1f0358bSKent Overstreet } while (ret); 1677cafe5635SKent Overstreet 1678cafe5635SKent Overstreet available = bch_btree_gc_finish(c); 167957943511SKent Overstreet wake_up_allocators(c); 168057943511SKent Overstreet 1681169ef1cfSKent Overstreet bch_time_stats_update(&c->btree_gc_time, start_time); 1682cafe5635SKent Overstreet 1683cafe5635SKent Overstreet stats.key_bytes *= sizeof(uint64_t); 1684cafe5635SKent Overstreet stats.data <<= 9; 1685cafe5635SKent Overstreet stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; 1686cafe5635SKent Overstreet memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); 1687cafe5635SKent Overstreet 1688c37511b8SKent Overstreet trace_bcache_gc_end(c); 1689cafe5635SKent Overstreet 169072a44517SKent Overstreet bch_moving_gc(c); 1691cafe5635SKent Overstreet } 1692cafe5635SKent Overstreet 169372a44517SKent Overstreet static int bch_gc_thread(void *arg) 1694cafe5635SKent Overstreet { 169572a44517SKent Overstreet struct cache_set *c = arg; 1696a1f0358bSKent Overstreet struct cache *ca; 1697a1f0358bSKent Overstreet unsigned i; 169872a44517SKent Overstreet 169972a44517SKent Overstreet while (1) { 1700a1f0358bSKent Overstreet again: 170172a44517SKent Overstreet bch_btree_gc(c); 170272a44517SKent Overstreet 170372a44517SKent Overstreet set_current_state(TASK_INTERRUPTIBLE); 170472a44517SKent Overstreet if (kthread_should_stop()) 170572a44517SKent Overstreet break; 170672a44517SKent Overstreet 1707a1f0358bSKent Overstreet mutex_lock(&c->bucket_lock); 1708a1f0358bSKent Overstreet 1709a1f0358bSKent Overstreet for_each_cache(ca, c, i) 1710a1f0358bSKent Overstreet if (ca->invalidate_needs_gc) { 1711a1f0358bSKent Overstreet mutex_unlock(&c->bucket_lock); 1712a1f0358bSKent Overstreet set_current_state(TASK_RUNNING); 1713a1f0358bSKent Overstreet goto again; 1714a1f0358bSKent Overstreet } 1715a1f0358bSKent Overstreet 1716a1f0358bSKent Overstreet mutex_unlock(&c->bucket_lock); 1717a1f0358bSKent Overstreet 171872a44517SKent Overstreet try_to_freeze(); 171972a44517SKent Overstreet schedule(); 172072a44517SKent Overstreet } 172172a44517SKent Overstreet 172272a44517SKent Overstreet return 0; 172372a44517SKent Overstreet } 172472a44517SKent Overstreet 172572a44517SKent Overstreet int bch_gc_thread_start(struct cache_set *c) 172672a44517SKent Overstreet { 172772a44517SKent Overstreet c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc"); 172872a44517SKent Overstreet if (IS_ERR(c->gc_thread)) 172972a44517SKent Overstreet return PTR_ERR(c->gc_thread); 173072a44517SKent Overstreet 173172a44517SKent Overstreet set_task_state(c->gc_thread, TASK_INTERRUPTIBLE); 173272a44517SKent Overstreet return 0; 1733cafe5635SKent Overstreet } 1734cafe5635SKent Overstreet 1735cafe5635SKent Overstreet /* Initial partial gc */ 1736cafe5635SKent Overstreet 1737cafe5635SKent Overstreet static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, 1738cafe5635SKent Overstreet unsigned long **seen) 1739cafe5635SKent Overstreet { 174050310164SKent Overstreet int ret = 0; 1741cafe5635SKent Overstreet unsigned i; 174250310164SKent Overstreet struct bkey *k, *p = NULL; 1743cafe5635SKent Overstreet struct bucket *g; 1744cafe5635SKent Overstreet struct btree_iter iter; 1745cafe5635SKent Overstreet 1746cafe5635SKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1747cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 1748cafe5635SKent Overstreet if (!ptr_available(b->c, k, i)) 1749cafe5635SKent Overstreet continue; 1750cafe5635SKent Overstreet 1751cafe5635SKent Overstreet g = PTR_BUCKET(b->c, k, i); 1752cafe5635SKent Overstreet 1753cafe5635SKent Overstreet if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), 1754cafe5635SKent Overstreet seen[PTR_DEV(k, i)]) || 1755cafe5635SKent Overstreet !ptr_stale(b->c, k, i)) { 1756cafe5635SKent Overstreet g->gen = PTR_GEN(k, i); 1757cafe5635SKent Overstreet 1758cafe5635SKent Overstreet if (b->level) 1759cafe5635SKent Overstreet g->prio = BTREE_PRIO; 1760cafe5635SKent Overstreet else if (g->prio == BTREE_PRIO) 1761cafe5635SKent Overstreet g->prio = INITIAL_PRIO; 1762cafe5635SKent Overstreet } 1763cafe5635SKent Overstreet } 1764cafe5635SKent Overstreet 1765cafe5635SKent Overstreet btree_mark_key(b, k); 1766cafe5635SKent Overstreet } 1767cafe5635SKent Overstreet 1768cafe5635SKent Overstreet if (b->level) { 176950310164SKent Overstreet bch_btree_iter_init(b, &iter, NULL); 1770cafe5635SKent Overstreet 177150310164SKent Overstreet do { 177250310164SKent Overstreet k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); 177350310164SKent Overstreet if (k) 177450310164SKent Overstreet btree_node_prefetch(b->c, k, b->level - 1); 177550310164SKent Overstreet 1776cafe5635SKent Overstreet if (p) 177750310164SKent Overstreet ret = btree(check_recurse, p, b, op, seen); 1778cafe5635SKent Overstreet 177950310164SKent Overstreet p = k; 178050310164SKent Overstreet } while (p && !ret); 1781cafe5635SKent Overstreet } 1782cafe5635SKent Overstreet 1783cafe5635SKent Overstreet return 0; 1784cafe5635SKent Overstreet } 1785cafe5635SKent Overstreet 1786c18536a7SKent Overstreet int bch_btree_check(struct cache_set *c) 1787cafe5635SKent Overstreet { 1788cafe5635SKent Overstreet int ret = -ENOMEM; 1789cafe5635SKent Overstreet unsigned i; 1790cafe5635SKent Overstreet unsigned long *seen[MAX_CACHES_PER_SET]; 1791c18536a7SKent Overstreet struct btree_op op; 1792cafe5635SKent Overstreet 1793cafe5635SKent Overstreet memset(seen, 0, sizeof(seen)); 1794b54d6934SKent Overstreet bch_btree_op_init(&op, SHRT_MAX); 1795cafe5635SKent Overstreet 1796cafe5635SKent Overstreet for (i = 0; c->cache[i]; i++) { 1797cafe5635SKent Overstreet size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); 1798cafe5635SKent Overstreet seen[i] = kmalloc(n, GFP_KERNEL); 1799cafe5635SKent Overstreet if (!seen[i]) 1800cafe5635SKent Overstreet goto err; 1801cafe5635SKent Overstreet 1802cafe5635SKent Overstreet /* Disables the seen array until prio_read() uses it too */ 1803cafe5635SKent Overstreet memset(seen[i], 0xFF, n); 1804cafe5635SKent Overstreet } 1805cafe5635SKent Overstreet 1806c18536a7SKent Overstreet ret = btree_root(check_recurse, c, &op, seen); 1807cafe5635SKent Overstreet err: 1808cafe5635SKent Overstreet for (i = 0; i < MAX_CACHES_PER_SET; i++) 1809cafe5635SKent Overstreet kfree(seen[i]); 1810cafe5635SKent Overstreet return ret; 1811cafe5635SKent Overstreet } 1812cafe5635SKent Overstreet 1813cafe5635SKent Overstreet /* Btree insertion */ 1814cafe5635SKent Overstreet 1815cafe5635SKent Overstreet static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) 1816cafe5635SKent Overstreet { 1817cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 1818cafe5635SKent Overstreet 1819cafe5635SKent Overstreet memmove((uint64_t *) where + bkey_u64s(insert), 1820cafe5635SKent Overstreet where, 1821cafe5635SKent Overstreet (void *) end(i) - (void *) where); 1822cafe5635SKent Overstreet 1823cafe5635SKent Overstreet i->keys += bkey_u64s(insert); 1824cafe5635SKent Overstreet bkey_copy(where, insert); 1825cafe5635SKent Overstreet bch_bset_fix_lookup_table(b, where); 1826cafe5635SKent Overstreet } 1827cafe5635SKent Overstreet 18281b207d80SKent Overstreet static bool fix_overlapping_extents(struct btree *b, struct bkey *insert, 1829cafe5635SKent Overstreet struct btree_iter *iter, 18301b207d80SKent Overstreet struct bkey *replace_key) 1831cafe5635SKent Overstreet { 1832279afbadSKent Overstreet void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) 1833cafe5635SKent Overstreet { 1834279afbadSKent Overstreet if (KEY_DIRTY(k)) 1835279afbadSKent Overstreet bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 1836279afbadSKent Overstreet offset, -sectors); 1837cafe5635SKent Overstreet } 1838cafe5635SKent Overstreet 1839279afbadSKent Overstreet uint64_t old_offset; 1840cafe5635SKent Overstreet unsigned old_size, sectors_found = 0; 1841cafe5635SKent Overstreet 1842cafe5635SKent Overstreet while (1) { 1843cafe5635SKent Overstreet struct bkey *k = bch_btree_iter_next(iter); 1844cafe5635SKent Overstreet if (!k || 1845cafe5635SKent Overstreet bkey_cmp(&START_KEY(k), insert) >= 0) 1846cafe5635SKent Overstreet break; 1847cafe5635SKent Overstreet 1848cafe5635SKent Overstreet if (bkey_cmp(k, &START_KEY(insert)) <= 0) 1849cafe5635SKent Overstreet continue; 1850cafe5635SKent Overstreet 1851279afbadSKent Overstreet old_offset = KEY_START(k); 1852cafe5635SKent Overstreet old_size = KEY_SIZE(k); 1853cafe5635SKent Overstreet 1854cafe5635SKent Overstreet /* 1855cafe5635SKent Overstreet * We might overlap with 0 size extents; we can't skip these 1856cafe5635SKent Overstreet * because if they're in the set we're inserting to we have to 1857cafe5635SKent Overstreet * adjust them so they don't overlap with the key we're 18581b207d80SKent Overstreet * inserting. But we don't want to check them for replace 1859cafe5635SKent Overstreet * operations. 1860cafe5635SKent Overstreet */ 1861cafe5635SKent Overstreet 18621b207d80SKent Overstreet if (replace_key && KEY_SIZE(k)) { 1863cafe5635SKent Overstreet /* 1864cafe5635SKent Overstreet * k might have been split since we inserted/found the 1865cafe5635SKent Overstreet * key we're replacing 1866cafe5635SKent Overstreet */ 1867cafe5635SKent Overstreet unsigned i; 1868cafe5635SKent Overstreet uint64_t offset = KEY_START(k) - 18691b207d80SKent Overstreet KEY_START(replace_key); 1870cafe5635SKent Overstreet 1871cafe5635SKent Overstreet /* But it must be a subset of the replace key */ 18721b207d80SKent Overstreet if (KEY_START(k) < KEY_START(replace_key) || 18731b207d80SKent Overstreet KEY_OFFSET(k) > KEY_OFFSET(replace_key)) 1874cafe5635SKent Overstreet goto check_failed; 1875cafe5635SKent Overstreet 1876cafe5635SKent Overstreet /* We didn't find a key that we were supposed to */ 1877cafe5635SKent Overstreet if (KEY_START(k) > KEY_START(insert) + sectors_found) 1878cafe5635SKent Overstreet goto check_failed; 1879cafe5635SKent Overstreet 1880d24a6e10SKent Overstreet if (KEY_PTRS(k) != KEY_PTRS(replace_key) || 1881d24a6e10SKent Overstreet KEY_DIRTY(k) != KEY_DIRTY(replace_key)) 1882cafe5635SKent Overstreet goto check_failed; 1883cafe5635SKent Overstreet 1884cafe5635SKent Overstreet /* skip past gen */ 1885cafe5635SKent Overstreet offset <<= 8; 1886cafe5635SKent Overstreet 18871b207d80SKent Overstreet BUG_ON(!KEY_PTRS(replace_key)); 1888cafe5635SKent Overstreet 18891b207d80SKent Overstreet for (i = 0; i < KEY_PTRS(replace_key); i++) 18901b207d80SKent Overstreet if (k->ptr[i] != replace_key->ptr[i] + offset) 1891cafe5635SKent Overstreet goto check_failed; 1892cafe5635SKent Overstreet 1893cafe5635SKent Overstreet sectors_found = KEY_OFFSET(k) - KEY_START(insert); 1894cafe5635SKent Overstreet } 1895cafe5635SKent Overstreet 1896cafe5635SKent Overstreet if (bkey_cmp(insert, k) < 0 && 1897cafe5635SKent Overstreet bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { 1898cafe5635SKent Overstreet /* 1899cafe5635SKent Overstreet * We overlapped in the middle of an existing key: that 1900cafe5635SKent Overstreet * means we have to split the old key. But we have to do 1901cafe5635SKent Overstreet * slightly different things depending on whether the 1902cafe5635SKent Overstreet * old key has been written out yet. 1903cafe5635SKent Overstreet */ 1904cafe5635SKent Overstreet 1905cafe5635SKent Overstreet struct bkey *top; 1906cafe5635SKent Overstreet 1907279afbadSKent Overstreet subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); 1908cafe5635SKent Overstreet 1909cafe5635SKent Overstreet if (bkey_written(b, k)) { 1910cafe5635SKent Overstreet /* 1911cafe5635SKent Overstreet * We insert a new key to cover the top of the 1912cafe5635SKent Overstreet * old key, and the old key is modified in place 1913cafe5635SKent Overstreet * to represent the bottom split. 1914cafe5635SKent Overstreet * 1915cafe5635SKent Overstreet * It's completely arbitrary whether the new key 1916cafe5635SKent Overstreet * is the top or the bottom, but it has to match 1917cafe5635SKent Overstreet * up with what btree_sort_fixup() does - it 1918cafe5635SKent Overstreet * doesn't check for this kind of overlap, it 1919cafe5635SKent Overstreet * depends on us inserting a new key for the top 1920cafe5635SKent Overstreet * here. 1921cafe5635SKent Overstreet */ 1922cafe5635SKent Overstreet top = bch_bset_search(b, &b->sets[b->nsets], 1923cafe5635SKent Overstreet insert); 1924cafe5635SKent Overstreet shift_keys(b, top, k); 1925cafe5635SKent Overstreet } else { 1926cafe5635SKent Overstreet BKEY_PADDED(key) temp; 1927cafe5635SKent Overstreet bkey_copy(&temp.key, k); 1928cafe5635SKent Overstreet shift_keys(b, k, &temp.key); 1929cafe5635SKent Overstreet top = bkey_next(k); 1930cafe5635SKent Overstreet } 1931cafe5635SKent Overstreet 1932cafe5635SKent Overstreet bch_cut_front(insert, top); 1933cafe5635SKent Overstreet bch_cut_back(&START_KEY(insert), k); 1934cafe5635SKent Overstreet bch_bset_fix_invalidated_key(b, k); 1935cafe5635SKent Overstreet return false; 1936cafe5635SKent Overstreet } 1937cafe5635SKent Overstreet 1938cafe5635SKent Overstreet if (bkey_cmp(insert, k) < 0) { 1939cafe5635SKent Overstreet bch_cut_front(insert, k); 1940cafe5635SKent Overstreet } else { 19411fa8455dSKent Overstreet if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) 19421fa8455dSKent Overstreet old_offset = KEY_START(insert); 19431fa8455dSKent Overstreet 1944cafe5635SKent Overstreet if (bkey_written(b, k) && 1945cafe5635SKent Overstreet bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { 1946cafe5635SKent Overstreet /* 1947cafe5635SKent Overstreet * Completely overwrote, so we don't have to 1948cafe5635SKent Overstreet * invalidate the binary search tree 1949cafe5635SKent Overstreet */ 1950cafe5635SKent Overstreet bch_cut_front(k, k); 1951cafe5635SKent Overstreet } else { 1952cafe5635SKent Overstreet __bch_cut_back(&START_KEY(insert), k); 1953cafe5635SKent Overstreet bch_bset_fix_invalidated_key(b, k); 1954cafe5635SKent Overstreet } 1955cafe5635SKent Overstreet } 1956cafe5635SKent Overstreet 1957279afbadSKent Overstreet subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); 1958cafe5635SKent Overstreet } 1959cafe5635SKent Overstreet 1960cafe5635SKent Overstreet check_failed: 19611b207d80SKent Overstreet if (replace_key) { 1962cafe5635SKent Overstreet if (!sectors_found) { 1963cafe5635SKent Overstreet return true; 1964cafe5635SKent Overstreet } else if (sectors_found < KEY_SIZE(insert)) { 1965cafe5635SKent Overstreet SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - 1966cafe5635SKent Overstreet (KEY_SIZE(insert) - sectors_found)); 1967cafe5635SKent Overstreet SET_KEY_SIZE(insert, sectors_found); 1968cafe5635SKent Overstreet } 1969cafe5635SKent Overstreet } 1970cafe5635SKent Overstreet 1971cafe5635SKent Overstreet return false; 1972cafe5635SKent Overstreet } 1973cafe5635SKent Overstreet 1974cafe5635SKent Overstreet static bool btree_insert_key(struct btree *b, struct btree_op *op, 19751b207d80SKent Overstreet struct bkey *k, struct bkey *replace_key) 1976cafe5635SKent Overstreet { 1977cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 1978cafe5635SKent Overstreet struct bkey *m, *prev; 197985b1492eSKent Overstreet unsigned status = BTREE_INSERT_STATUS_INSERT; 1980cafe5635SKent Overstreet 1981cafe5635SKent Overstreet BUG_ON(bkey_cmp(k, &b->key) > 0); 1982cafe5635SKent Overstreet BUG_ON(b->level && !KEY_PTRS(k)); 1983cafe5635SKent Overstreet BUG_ON(!b->level && !KEY_OFFSET(k)); 1984cafe5635SKent Overstreet 1985cafe5635SKent Overstreet if (!b->level) { 1986cafe5635SKent Overstreet struct btree_iter iter; 1987cafe5635SKent Overstreet 1988cafe5635SKent Overstreet /* 1989cafe5635SKent Overstreet * bset_search() returns the first key that is strictly greater 1990cafe5635SKent Overstreet * than the search key - but for back merging, we want to find 19910eacac22SKent Overstreet * the previous key. 1992cafe5635SKent Overstreet */ 1993cafe5635SKent Overstreet prev = NULL; 19940eacac22SKent Overstreet m = bch_btree_iter_init(b, &iter, PRECEDING_KEY(&START_KEY(k))); 1995cafe5635SKent Overstreet 19961b207d80SKent Overstreet if (fix_overlapping_extents(b, k, &iter, replace_key)) { 19971b207d80SKent Overstreet op->insert_collision = true; 1998cafe5635SKent Overstreet return false; 19991b207d80SKent Overstreet } 2000cafe5635SKent Overstreet 20011fa8455dSKent Overstreet if (KEY_DIRTY(k)) 20021fa8455dSKent Overstreet bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 20031fa8455dSKent Overstreet KEY_START(k), KEY_SIZE(k)); 20041fa8455dSKent Overstreet 2005cafe5635SKent Overstreet while (m != end(i) && 2006cafe5635SKent Overstreet bkey_cmp(k, &START_KEY(m)) > 0) 2007cafe5635SKent Overstreet prev = m, m = bkey_next(m); 2008cafe5635SKent Overstreet 2009cafe5635SKent Overstreet if (key_merging_disabled(b->c)) 2010cafe5635SKent Overstreet goto insert; 2011cafe5635SKent Overstreet 2012cafe5635SKent Overstreet /* prev is in the tree, if we merge we're done */ 201385b1492eSKent Overstreet status = BTREE_INSERT_STATUS_BACK_MERGE; 2014cafe5635SKent Overstreet if (prev && 2015cafe5635SKent Overstreet bch_bkey_try_merge(b, prev, k)) 2016cafe5635SKent Overstreet goto merged; 2017cafe5635SKent Overstreet 201885b1492eSKent Overstreet status = BTREE_INSERT_STATUS_OVERWROTE; 2019cafe5635SKent Overstreet if (m != end(i) && 2020cafe5635SKent Overstreet KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) 2021cafe5635SKent Overstreet goto copy; 2022cafe5635SKent Overstreet 202385b1492eSKent Overstreet status = BTREE_INSERT_STATUS_FRONT_MERGE; 2024cafe5635SKent Overstreet if (m != end(i) && 2025cafe5635SKent Overstreet bch_bkey_try_merge(b, k, m)) 2026cafe5635SKent Overstreet goto copy; 20271b207d80SKent Overstreet } else { 20281b207d80SKent Overstreet BUG_ON(replace_key); 2029cafe5635SKent Overstreet m = bch_bset_search(b, &b->sets[b->nsets], k); 20301b207d80SKent Overstreet } 2031cafe5635SKent Overstreet 2032cafe5635SKent Overstreet insert: shift_keys(b, m, k); 2033cafe5635SKent Overstreet copy: bkey_copy(m, k); 2034cafe5635SKent Overstreet merged: 20351b207d80SKent Overstreet bch_check_keys(b, "%u for %s", status, 20361b207d80SKent Overstreet replace_key ? "replace" : "insert"); 2037cafe5635SKent Overstreet 2038cafe5635SKent Overstreet if (b->level && !KEY_OFFSET(k)) 203957943511SKent Overstreet btree_current_write(b)->prio_blocked++; 2040cafe5635SKent Overstreet 20411b207d80SKent Overstreet trace_bcache_btree_insert_key(b, k, replace_key != NULL, status); 2042cafe5635SKent Overstreet 2043cafe5635SKent Overstreet return true; 2044cafe5635SKent Overstreet } 2045cafe5635SKent Overstreet 204626c949f8SKent Overstreet static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, 20471b207d80SKent Overstreet struct keylist *insert_keys, 20481b207d80SKent Overstreet struct bkey *replace_key) 2049cafe5635SKent Overstreet { 2050cafe5635SKent Overstreet bool ret = false; 2051280481d0SKent Overstreet int oldsize = bch_count_data(b); 2052cafe5635SKent Overstreet 205326c949f8SKent Overstreet while (!bch_keylist_empty(insert_keys)) { 2054403b6cdeSKent Overstreet struct bset *i = write_block(b); 2055c2f95ae2SKent Overstreet struct bkey *k = insert_keys->keys; 205626c949f8SKent Overstreet 2057403b6cdeSKent Overstreet if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c) 2058403b6cdeSKent Overstreet > btree_blocks(b)) 2059403b6cdeSKent Overstreet break; 2060403b6cdeSKent Overstreet 2061403b6cdeSKent Overstreet if (bkey_cmp(k, &b->key) <= 0) { 20623a3b6a4eSKent Overstreet if (!b->level) 20633a3b6a4eSKent Overstreet bkey_put(b->c, k); 206426c949f8SKent Overstreet 20651b207d80SKent Overstreet ret |= btree_insert_key(b, op, k, replace_key); 206626c949f8SKent Overstreet bch_keylist_pop_front(insert_keys); 206726c949f8SKent Overstreet } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { 206826c949f8SKent Overstreet BKEY_PADDED(key) temp; 2069c2f95ae2SKent Overstreet bkey_copy(&temp.key, insert_keys->keys); 207026c949f8SKent Overstreet 207126c949f8SKent Overstreet bch_cut_back(&b->key, &temp.key); 2072c2f95ae2SKent Overstreet bch_cut_front(&b->key, insert_keys->keys); 207326c949f8SKent Overstreet 20741b207d80SKent Overstreet ret |= btree_insert_key(b, op, &temp.key, replace_key); 207526c949f8SKent Overstreet break; 207626c949f8SKent Overstreet } else { 207726c949f8SKent Overstreet break; 207826c949f8SKent Overstreet } 2079cafe5635SKent Overstreet } 2080cafe5635SKent Overstreet 2081403b6cdeSKent Overstreet BUG_ON(!bch_keylist_empty(insert_keys) && b->level); 2082403b6cdeSKent Overstreet 2083cafe5635SKent Overstreet BUG_ON(bch_count_data(b) < oldsize); 2084cafe5635SKent Overstreet return ret; 2085cafe5635SKent Overstreet } 2086cafe5635SKent Overstreet 208726c949f8SKent Overstreet static int btree_split(struct btree *b, struct btree_op *op, 208826c949f8SKent Overstreet struct keylist *insert_keys, 20891b207d80SKent Overstreet struct bkey *replace_key) 2090cafe5635SKent Overstreet { 2091d6fd3b11SKent Overstreet bool split; 2092cafe5635SKent Overstreet struct btree *n1, *n2 = NULL, *n3 = NULL; 2093cafe5635SKent Overstreet uint64_t start_time = local_clock(); 2094b54d6934SKent Overstreet struct closure cl; 209517e21a9fSKent Overstreet struct keylist parent_keys; 2096b54d6934SKent Overstreet 2097b54d6934SKent Overstreet closure_init_stack(&cl); 209817e21a9fSKent Overstreet bch_keylist_init(&parent_keys); 2099cafe5635SKent Overstreet 2100*78365411SKent Overstreet if (!b->level && 2101*78365411SKent Overstreet btree_check_reserve(b, op)) 2102*78365411SKent Overstreet return -EINTR; 2103*78365411SKent Overstreet 2104bc9389eeSKent Overstreet n1 = btree_node_alloc_replacement(b, true); 2105cafe5635SKent Overstreet if (IS_ERR(n1)) 2106cafe5635SKent Overstreet goto err; 2107cafe5635SKent Overstreet 2108cafe5635SKent Overstreet split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; 2109cafe5635SKent Overstreet 2110cafe5635SKent Overstreet if (split) { 2111cafe5635SKent Overstreet unsigned keys = 0; 2112cafe5635SKent Overstreet 2113c37511b8SKent Overstreet trace_bcache_btree_node_split(b, n1->sets[0].data->keys); 2114c37511b8SKent Overstreet 2115bc9389eeSKent Overstreet n2 = bch_btree_node_alloc(b->c, b->level, true); 2116cafe5635SKent Overstreet if (IS_ERR(n2)) 2117cafe5635SKent Overstreet goto err_free1; 2118cafe5635SKent Overstreet 2119d6fd3b11SKent Overstreet if (!b->parent) { 2120bc9389eeSKent Overstreet n3 = bch_btree_node_alloc(b->c, b->level + 1, true); 2121cafe5635SKent Overstreet if (IS_ERR(n3)) 2122cafe5635SKent Overstreet goto err_free2; 2123cafe5635SKent Overstreet } 2124cafe5635SKent Overstreet 21251b207d80SKent Overstreet bch_btree_insert_keys(n1, op, insert_keys, replace_key); 2126cafe5635SKent Overstreet 2127d6fd3b11SKent Overstreet /* 2128d6fd3b11SKent Overstreet * Has to be a linear search because we don't have an auxiliary 2129cafe5635SKent Overstreet * search tree yet 2130cafe5635SKent Overstreet */ 2131cafe5635SKent Overstreet 2132cafe5635SKent Overstreet while (keys < (n1->sets[0].data->keys * 3) / 5) 2133cafe5635SKent Overstreet keys += bkey_u64s(node(n1->sets[0].data, keys)); 2134cafe5635SKent Overstreet 2135cafe5635SKent Overstreet bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); 2136cafe5635SKent Overstreet keys += bkey_u64s(node(n1->sets[0].data, keys)); 2137cafe5635SKent Overstreet 2138cafe5635SKent Overstreet n2->sets[0].data->keys = n1->sets[0].data->keys - keys; 2139cafe5635SKent Overstreet n1->sets[0].data->keys = keys; 2140cafe5635SKent Overstreet 2141cafe5635SKent Overstreet memcpy(n2->sets[0].data->start, 2142cafe5635SKent Overstreet end(n1->sets[0].data), 2143cafe5635SKent Overstreet n2->sets[0].data->keys * sizeof(uint64_t)); 2144cafe5635SKent Overstreet 2145cafe5635SKent Overstreet bkey_copy_key(&n2->key, &b->key); 2146cafe5635SKent Overstreet 214717e21a9fSKent Overstreet bch_keylist_add(&parent_keys, &n2->key); 2148b54d6934SKent Overstreet bch_btree_node_write(n2, &cl); 2149cafe5635SKent Overstreet rw_unlock(true, n2); 2150c37511b8SKent Overstreet } else { 2151c37511b8SKent Overstreet trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); 2152c37511b8SKent Overstreet 21531b207d80SKent Overstreet bch_btree_insert_keys(n1, op, insert_keys, replace_key); 2154c37511b8SKent Overstreet } 2155cafe5635SKent Overstreet 215617e21a9fSKent Overstreet bch_keylist_add(&parent_keys, &n1->key); 2157b54d6934SKent Overstreet bch_btree_node_write(n1, &cl); 2158cafe5635SKent Overstreet 2159cafe5635SKent Overstreet if (n3) { 2160d6fd3b11SKent Overstreet /* Depth increases, make a new root */ 2161cafe5635SKent Overstreet bkey_copy_key(&n3->key, &MAX_KEY); 216217e21a9fSKent Overstreet bch_btree_insert_keys(n3, op, &parent_keys, NULL); 2163b54d6934SKent Overstreet bch_btree_node_write(n3, &cl); 2164cafe5635SKent Overstreet 2165b54d6934SKent Overstreet closure_sync(&cl); 2166cafe5635SKent Overstreet bch_btree_set_root(n3); 2167cafe5635SKent Overstreet rw_unlock(true, n3); 216817e21a9fSKent Overstreet 216917e21a9fSKent Overstreet btree_node_free(b); 2170d6fd3b11SKent Overstreet } else if (!b->parent) { 2171d6fd3b11SKent Overstreet /* Root filled up but didn't need to be split */ 2172b54d6934SKent Overstreet closure_sync(&cl); 2173cafe5635SKent Overstreet bch_btree_set_root(n1); 217417e21a9fSKent Overstreet 217517e21a9fSKent Overstreet btree_node_free(b); 2176cafe5635SKent Overstreet } else { 217717e21a9fSKent Overstreet /* Split a non root node */ 2178b54d6934SKent Overstreet closure_sync(&cl); 217917e21a9fSKent Overstreet make_btree_freeing_key(b, parent_keys.top); 218017e21a9fSKent Overstreet bch_keylist_push(&parent_keys); 218117e21a9fSKent Overstreet 218217e21a9fSKent Overstreet btree_node_free(b); 218317e21a9fSKent Overstreet 218417e21a9fSKent Overstreet bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL); 218517e21a9fSKent Overstreet BUG_ON(!bch_keylist_empty(&parent_keys)); 2186cafe5635SKent Overstreet } 2187cafe5635SKent Overstreet 2188cafe5635SKent Overstreet rw_unlock(true, n1); 2189cafe5635SKent Overstreet 2190169ef1cfSKent Overstreet bch_time_stats_update(&b->c->btree_split_time, start_time); 2191cafe5635SKent Overstreet 2192cafe5635SKent Overstreet return 0; 2193cafe5635SKent Overstreet err_free2: 2194e8e1d468SKent Overstreet btree_node_free(n2); 2195cafe5635SKent Overstreet rw_unlock(true, n2); 2196cafe5635SKent Overstreet err_free1: 2197e8e1d468SKent Overstreet btree_node_free(n1); 2198cafe5635SKent Overstreet rw_unlock(true, n1); 2199cafe5635SKent Overstreet err: 2200cafe5635SKent Overstreet if (n3 == ERR_PTR(-EAGAIN) || 2201cafe5635SKent Overstreet n2 == ERR_PTR(-EAGAIN) || 2202cafe5635SKent Overstreet n1 == ERR_PTR(-EAGAIN)) 2203cafe5635SKent Overstreet return -EAGAIN; 2204cafe5635SKent Overstreet 2205cafe5635SKent Overstreet pr_warn("couldn't split"); 2206cafe5635SKent Overstreet return -ENOMEM; 2207cafe5635SKent Overstreet } 2208cafe5635SKent Overstreet 220926c949f8SKent Overstreet static int bch_btree_insert_node(struct btree *b, struct btree_op *op, 2210c18536a7SKent Overstreet struct keylist *insert_keys, 22111b207d80SKent Overstreet atomic_t *journal_ref, 22121b207d80SKent Overstreet struct bkey *replace_key) 221326c949f8SKent Overstreet { 22141b207d80SKent Overstreet BUG_ON(b->level && replace_key); 22151b207d80SKent Overstreet 221626c949f8SKent Overstreet if (should_split(b)) { 221726c949f8SKent Overstreet if (current->bio_list) { 221826c949f8SKent Overstreet op->lock = b->c->root->level + 1; 221917e21a9fSKent Overstreet return -EAGAIN; 222026c949f8SKent Overstreet } else if (op->lock <= b->c->root->level) { 222126c949f8SKent Overstreet op->lock = b->c->root->level + 1; 222217e21a9fSKent Overstreet return -EINTR; 222326c949f8SKent Overstreet } else { 222417e21a9fSKent Overstreet /* Invalidated all iterators */ 222517e21a9fSKent Overstreet return btree_split(b, op, insert_keys, replace_key) ?: 222617e21a9fSKent Overstreet -EINTR; 222726c949f8SKent Overstreet } 222826c949f8SKent Overstreet } else { 222926c949f8SKent Overstreet BUG_ON(write_block(b) != b->sets[b->nsets].data); 223026c949f8SKent Overstreet 223117e21a9fSKent Overstreet if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) { 2232f269af5aSKent Overstreet if (!b->level) 2233c18536a7SKent Overstreet bch_btree_leaf_dirty(b, journal_ref); 2234f269af5aSKent Overstreet else 2235f269af5aSKent Overstreet bch_btree_node_write_sync(b); 223626c949f8SKent Overstreet } 223726c949f8SKent Overstreet 223817e21a9fSKent Overstreet return 0; 223917e21a9fSKent Overstreet } 224026c949f8SKent Overstreet } 224126c949f8SKent Overstreet 2242e7c590ebSKent Overstreet int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, 2243e7c590ebSKent Overstreet struct bkey *check_key) 2244e7c590ebSKent Overstreet { 2245e7c590ebSKent Overstreet int ret = -EINTR; 2246e7c590ebSKent Overstreet uint64_t btree_ptr = b->key.ptr[0]; 2247e7c590ebSKent Overstreet unsigned long seq = b->seq; 2248e7c590ebSKent Overstreet struct keylist insert; 2249e7c590ebSKent Overstreet bool upgrade = op->lock == -1; 2250e7c590ebSKent Overstreet 2251e7c590ebSKent Overstreet bch_keylist_init(&insert); 2252e7c590ebSKent Overstreet 2253e7c590ebSKent Overstreet if (upgrade) { 2254e7c590ebSKent Overstreet rw_unlock(false, b); 2255e7c590ebSKent Overstreet rw_lock(true, b, b->level); 2256e7c590ebSKent Overstreet 2257e7c590ebSKent Overstreet if (b->key.ptr[0] != btree_ptr || 2258e7c590ebSKent Overstreet b->seq != seq + 1) 2259e7c590ebSKent Overstreet goto out; 2260e7c590ebSKent Overstreet } 2261e7c590ebSKent Overstreet 2262e7c590ebSKent Overstreet SET_KEY_PTRS(check_key, 1); 2263e7c590ebSKent Overstreet get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); 2264e7c590ebSKent Overstreet 2265e7c590ebSKent Overstreet SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); 2266e7c590ebSKent Overstreet 2267e7c590ebSKent Overstreet bch_keylist_add(&insert, check_key); 2268e7c590ebSKent Overstreet 22691b207d80SKent Overstreet ret = bch_btree_insert_node(b, op, &insert, NULL, NULL); 2270e7c590ebSKent Overstreet 2271e7c590ebSKent Overstreet BUG_ON(!ret && !bch_keylist_empty(&insert)); 2272e7c590ebSKent Overstreet out: 2273e7c590ebSKent Overstreet if (upgrade) 2274e7c590ebSKent Overstreet downgrade_write(&b->lock); 2275e7c590ebSKent Overstreet return ret; 2276e7c590ebSKent Overstreet } 2277e7c590ebSKent Overstreet 2278cc7b8819SKent Overstreet struct btree_insert_op { 2279cc7b8819SKent Overstreet struct btree_op op; 2280cc7b8819SKent Overstreet struct keylist *keys; 2281cc7b8819SKent Overstreet atomic_t *journal_ref; 2282cc7b8819SKent Overstreet struct bkey *replace_key; 2283cc7b8819SKent Overstreet }; 2284cc7b8819SKent Overstreet 228508239ca2SWei Yongjun static int btree_insert_fn(struct btree_op *b_op, struct btree *b) 2286cafe5635SKent Overstreet { 2287cc7b8819SKent Overstreet struct btree_insert_op *op = container_of(b_op, 2288cc7b8819SKent Overstreet struct btree_insert_op, op); 2289403b6cdeSKent Overstreet 2290cc7b8819SKent Overstreet int ret = bch_btree_insert_node(b, &op->op, op->keys, 2291cc7b8819SKent Overstreet op->journal_ref, op->replace_key); 2292cc7b8819SKent Overstreet if (ret && !bch_keylist_empty(op->keys)) 2293cc7b8819SKent Overstreet return ret; 2294cc7b8819SKent Overstreet else 2295cc7b8819SKent Overstreet return MAP_DONE; 2296cafe5635SKent Overstreet } 2297cafe5635SKent Overstreet 2298cc7b8819SKent Overstreet int bch_btree_insert(struct cache_set *c, struct keylist *keys, 2299cc7b8819SKent Overstreet atomic_t *journal_ref, struct bkey *replace_key) 2300cafe5635SKent Overstreet { 2301cc7b8819SKent Overstreet struct btree_insert_op op; 2302cafe5635SKent Overstreet int ret = 0; 2303cafe5635SKent Overstreet 2304cc7b8819SKent Overstreet BUG_ON(current->bio_list); 23054f3d4014SKent Overstreet BUG_ON(bch_keylist_empty(keys)); 2306cafe5635SKent Overstreet 2307cc7b8819SKent Overstreet bch_btree_op_init(&op.op, 0); 2308cc7b8819SKent Overstreet op.keys = keys; 2309cc7b8819SKent Overstreet op.journal_ref = journal_ref; 2310cc7b8819SKent Overstreet op.replace_key = replace_key; 2311cafe5635SKent Overstreet 2312cc7b8819SKent Overstreet while (!ret && !bch_keylist_empty(keys)) { 2313cc7b8819SKent Overstreet op.op.lock = 0; 2314cc7b8819SKent Overstreet ret = bch_btree_map_leaf_nodes(&op.op, c, 2315cc7b8819SKent Overstreet &START_KEY(keys->keys), 2316cc7b8819SKent Overstreet btree_insert_fn); 2317cc7b8819SKent Overstreet } 2318cc7b8819SKent Overstreet 2319cc7b8819SKent Overstreet if (ret) { 2320cafe5635SKent Overstreet struct bkey *k; 2321cafe5635SKent Overstreet 23221b207d80SKent Overstreet pr_err("error %i", ret); 2323cafe5635SKent Overstreet 23244f3d4014SKent Overstreet while ((k = bch_keylist_pop(keys))) 23253a3b6a4eSKent Overstreet bkey_put(c, k); 2326cc7b8819SKent Overstreet } else if (op.op.insert_collision) 2327cc7b8819SKent Overstreet ret = -ESRCH; 23286054c6d4SKent Overstreet 2329cafe5635SKent Overstreet return ret; 2330cafe5635SKent Overstreet } 2331cafe5635SKent Overstreet 2332cafe5635SKent Overstreet void bch_btree_set_root(struct btree *b) 2333cafe5635SKent Overstreet { 2334cafe5635SKent Overstreet unsigned i; 2335e49c7c37SKent Overstreet struct closure cl; 2336e49c7c37SKent Overstreet 2337e49c7c37SKent Overstreet closure_init_stack(&cl); 2338cafe5635SKent Overstreet 2339c37511b8SKent Overstreet trace_bcache_btree_set_root(b); 2340c37511b8SKent Overstreet 2341cafe5635SKent Overstreet BUG_ON(!b->written); 2342cafe5635SKent Overstreet 2343cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&b->key); i++) 2344cafe5635SKent Overstreet BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); 2345cafe5635SKent Overstreet 2346cafe5635SKent Overstreet mutex_lock(&b->c->bucket_lock); 2347cafe5635SKent Overstreet list_del_init(&b->list); 2348cafe5635SKent Overstreet mutex_unlock(&b->c->bucket_lock); 2349cafe5635SKent Overstreet 2350cafe5635SKent Overstreet b->c->root = b; 2351cafe5635SKent Overstreet 2352e49c7c37SKent Overstreet bch_journal_meta(b->c, &cl); 2353e49c7c37SKent Overstreet closure_sync(&cl); 2354cafe5635SKent Overstreet } 2355cafe5635SKent Overstreet 235648dad8baSKent Overstreet /* Map across nodes or keys */ 235748dad8baSKent Overstreet 235848dad8baSKent Overstreet static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, 235948dad8baSKent Overstreet struct bkey *from, 236048dad8baSKent Overstreet btree_map_nodes_fn *fn, int flags) 236148dad8baSKent Overstreet { 236248dad8baSKent Overstreet int ret = MAP_CONTINUE; 236348dad8baSKent Overstreet 236448dad8baSKent Overstreet if (b->level) { 236548dad8baSKent Overstreet struct bkey *k; 236648dad8baSKent Overstreet struct btree_iter iter; 236748dad8baSKent Overstreet 236848dad8baSKent Overstreet bch_btree_iter_init(b, &iter, from); 236948dad8baSKent Overstreet 237048dad8baSKent Overstreet while ((k = bch_btree_iter_next_filter(&iter, b, 237148dad8baSKent Overstreet bch_ptr_bad))) { 237248dad8baSKent Overstreet ret = btree(map_nodes_recurse, k, b, 237348dad8baSKent Overstreet op, from, fn, flags); 237448dad8baSKent Overstreet from = NULL; 237548dad8baSKent Overstreet 237648dad8baSKent Overstreet if (ret != MAP_CONTINUE) 237748dad8baSKent Overstreet return ret; 237848dad8baSKent Overstreet } 237948dad8baSKent Overstreet } 238048dad8baSKent Overstreet 238148dad8baSKent Overstreet if (!b->level || flags == MAP_ALL_NODES) 238248dad8baSKent Overstreet ret = fn(op, b); 238348dad8baSKent Overstreet 238448dad8baSKent Overstreet return ret; 238548dad8baSKent Overstreet } 238648dad8baSKent Overstreet 238748dad8baSKent Overstreet int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, 238848dad8baSKent Overstreet struct bkey *from, btree_map_nodes_fn *fn, int flags) 238948dad8baSKent Overstreet { 2390b54d6934SKent Overstreet return btree_root(map_nodes_recurse, c, op, from, fn, flags); 239148dad8baSKent Overstreet } 239248dad8baSKent Overstreet 239348dad8baSKent Overstreet static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, 239448dad8baSKent Overstreet struct bkey *from, btree_map_keys_fn *fn, 239548dad8baSKent Overstreet int flags) 239648dad8baSKent Overstreet { 239748dad8baSKent Overstreet int ret = MAP_CONTINUE; 239848dad8baSKent Overstreet struct bkey *k; 239948dad8baSKent Overstreet struct btree_iter iter; 240048dad8baSKent Overstreet 240148dad8baSKent Overstreet bch_btree_iter_init(b, &iter, from); 240248dad8baSKent Overstreet 240348dad8baSKent Overstreet while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) { 240448dad8baSKent Overstreet ret = !b->level 240548dad8baSKent Overstreet ? fn(op, b, k) 240648dad8baSKent Overstreet : btree(map_keys_recurse, k, b, op, from, fn, flags); 240748dad8baSKent Overstreet from = NULL; 240848dad8baSKent Overstreet 240948dad8baSKent Overstreet if (ret != MAP_CONTINUE) 241048dad8baSKent Overstreet return ret; 241148dad8baSKent Overstreet } 241248dad8baSKent Overstreet 241348dad8baSKent Overstreet if (!b->level && (flags & MAP_END_KEY)) 241448dad8baSKent Overstreet ret = fn(op, b, &KEY(KEY_INODE(&b->key), 241548dad8baSKent Overstreet KEY_OFFSET(&b->key), 0)); 241648dad8baSKent Overstreet 241748dad8baSKent Overstreet return ret; 241848dad8baSKent Overstreet } 241948dad8baSKent Overstreet 242048dad8baSKent Overstreet int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, 242148dad8baSKent Overstreet struct bkey *from, btree_map_keys_fn *fn, int flags) 242248dad8baSKent Overstreet { 2423b54d6934SKent Overstreet return btree_root(map_keys_recurse, c, op, from, fn, flags); 242448dad8baSKent Overstreet } 242548dad8baSKent Overstreet 2426cafe5635SKent Overstreet /* Keybuf code */ 2427cafe5635SKent Overstreet 2428cafe5635SKent Overstreet static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) 2429cafe5635SKent Overstreet { 2430cafe5635SKent Overstreet /* Overlapping keys compare equal */ 2431cafe5635SKent Overstreet if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) 2432cafe5635SKent Overstreet return -1; 2433cafe5635SKent Overstreet if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) 2434cafe5635SKent Overstreet return 1; 2435cafe5635SKent Overstreet return 0; 2436cafe5635SKent Overstreet } 2437cafe5635SKent Overstreet 2438cafe5635SKent Overstreet static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, 2439cafe5635SKent Overstreet struct keybuf_key *r) 2440cafe5635SKent Overstreet { 2441cafe5635SKent Overstreet return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); 2442cafe5635SKent Overstreet } 2443cafe5635SKent Overstreet 244448dad8baSKent Overstreet struct refill { 244548dad8baSKent Overstreet struct btree_op op; 244648a915a8SKent Overstreet unsigned nr_found; 244748dad8baSKent Overstreet struct keybuf *buf; 244848dad8baSKent Overstreet struct bkey *end; 244948dad8baSKent Overstreet keybuf_pred_fn *pred; 245048dad8baSKent Overstreet }; 245148dad8baSKent Overstreet 245248dad8baSKent Overstreet static int refill_keybuf_fn(struct btree_op *op, struct btree *b, 245348dad8baSKent Overstreet struct bkey *k) 2454cafe5635SKent Overstreet { 245548dad8baSKent Overstreet struct refill *refill = container_of(op, struct refill, op); 245648dad8baSKent Overstreet struct keybuf *buf = refill->buf; 245748dad8baSKent Overstreet int ret = MAP_CONTINUE; 2458cafe5635SKent Overstreet 245948dad8baSKent Overstreet if (bkey_cmp(k, refill->end) >= 0) { 246048dad8baSKent Overstreet ret = MAP_DONE; 246148dad8baSKent Overstreet goto out; 2462cafe5635SKent Overstreet } 2463cafe5635SKent Overstreet 246448dad8baSKent Overstreet if (!KEY_SIZE(k)) /* end key */ 246548dad8baSKent Overstreet goto out; 2466cafe5635SKent Overstreet 246748dad8baSKent Overstreet if (refill->pred(buf, k)) { 2468cafe5635SKent Overstreet struct keybuf_key *w; 2469cafe5635SKent Overstreet 2470cafe5635SKent Overstreet spin_lock(&buf->lock); 2471cafe5635SKent Overstreet 2472cafe5635SKent Overstreet w = array_alloc(&buf->freelist); 247348dad8baSKent Overstreet if (!w) { 247448dad8baSKent Overstreet spin_unlock(&buf->lock); 247548dad8baSKent Overstreet return MAP_DONE; 247648dad8baSKent Overstreet } 2477cafe5635SKent Overstreet 2478cafe5635SKent Overstreet w->private = NULL; 2479cafe5635SKent Overstreet bkey_copy(&w->key, k); 2480cafe5635SKent Overstreet 2481cafe5635SKent Overstreet if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) 2482cafe5635SKent Overstreet array_free(&buf->freelist, w); 248348a915a8SKent Overstreet else 248448a915a8SKent Overstreet refill->nr_found++; 2485cafe5635SKent Overstreet 248648dad8baSKent Overstreet if (array_freelist_empty(&buf->freelist)) 248748dad8baSKent Overstreet ret = MAP_DONE; 248848dad8baSKent Overstreet 2489cafe5635SKent Overstreet spin_unlock(&buf->lock); 2490cafe5635SKent Overstreet } 249148dad8baSKent Overstreet out: 249248dad8baSKent Overstreet buf->last_scanned = *k; 249348dad8baSKent Overstreet return ret; 2494cafe5635SKent Overstreet } 2495cafe5635SKent Overstreet 2496cafe5635SKent Overstreet void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, 249772c27061SKent Overstreet struct bkey *end, keybuf_pred_fn *pred) 2498cafe5635SKent Overstreet { 2499cafe5635SKent Overstreet struct bkey start = buf->last_scanned; 250048dad8baSKent Overstreet struct refill refill; 2501cafe5635SKent Overstreet 2502cafe5635SKent Overstreet cond_resched(); 2503cafe5635SKent Overstreet 2504b54d6934SKent Overstreet bch_btree_op_init(&refill.op, -1); 250548a915a8SKent Overstreet refill.nr_found = 0; 250648dad8baSKent Overstreet refill.buf = buf; 250748dad8baSKent Overstreet refill.end = end; 250848dad8baSKent Overstreet refill.pred = pred; 250948dad8baSKent Overstreet 251048dad8baSKent Overstreet bch_btree_map_keys(&refill.op, c, &buf->last_scanned, 251148dad8baSKent Overstreet refill_keybuf_fn, MAP_END_KEY); 2512cafe5635SKent Overstreet 251348a915a8SKent Overstreet trace_bcache_keyscan(refill.nr_found, 2514cafe5635SKent Overstreet KEY_INODE(&start), KEY_OFFSET(&start), 251548a915a8SKent Overstreet KEY_INODE(&buf->last_scanned), 251648a915a8SKent Overstreet KEY_OFFSET(&buf->last_scanned)); 2517cafe5635SKent Overstreet 2518cafe5635SKent Overstreet spin_lock(&buf->lock); 2519cafe5635SKent Overstreet 2520cafe5635SKent Overstreet if (!RB_EMPTY_ROOT(&buf->keys)) { 2521cafe5635SKent Overstreet struct keybuf_key *w; 2522cafe5635SKent Overstreet w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2523cafe5635SKent Overstreet buf->start = START_KEY(&w->key); 2524cafe5635SKent Overstreet 2525cafe5635SKent Overstreet w = RB_LAST(&buf->keys, struct keybuf_key, node); 2526cafe5635SKent Overstreet buf->end = w->key; 2527cafe5635SKent Overstreet } else { 2528cafe5635SKent Overstreet buf->start = MAX_KEY; 2529cafe5635SKent Overstreet buf->end = MAX_KEY; 2530cafe5635SKent Overstreet } 2531cafe5635SKent Overstreet 2532cafe5635SKent Overstreet spin_unlock(&buf->lock); 2533cafe5635SKent Overstreet } 2534cafe5635SKent Overstreet 2535cafe5635SKent Overstreet static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2536cafe5635SKent Overstreet { 2537cafe5635SKent Overstreet rb_erase(&w->node, &buf->keys); 2538cafe5635SKent Overstreet array_free(&buf->freelist, w); 2539cafe5635SKent Overstreet } 2540cafe5635SKent Overstreet 2541cafe5635SKent Overstreet void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2542cafe5635SKent Overstreet { 2543cafe5635SKent Overstreet spin_lock(&buf->lock); 2544cafe5635SKent Overstreet __bch_keybuf_del(buf, w); 2545cafe5635SKent Overstreet spin_unlock(&buf->lock); 2546cafe5635SKent Overstreet } 2547cafe5635SKent Overstreet 2548cafe5635SKent Overstreet bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, 2549cafe5635SKent Overstreet struct bkey *end) 2550cafe5635SKent Overstreet { 2551cafe5635SKent Overstreet bool ret = false; 2552cafe5635SKent Overstreet struct keybuf_key *p, *w, s; 2553cafe5635SKent Overstreet s.key = *start; 2554cafe5635SKent Overstreet 2555cafe5635SKent Overstreet if (bkey_cmp(end, &buf->start) <= 0 || 2556cafe5635SKent Overstreet bkey_cmp(start, &buf->end) >= 0) 2557cafe5635SKent Overstreet return false; 2558cafe5635SKent Overstreet 2559cafe5635SKent Overstreet spin_lock(&buf->lock); 2560cafe5635SKent Overstreet w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); 2561cafe5635SKent Overstreet 2562cafe5635SKent Overstreet while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { 2563cafe5635SKent Overstreet p = w; 2564cafe5635SKent Overstreet w = RB_NEXT(w, node); 2565cafe5635SKent Overstreet 2566cafe5635SKent Overstreet if (p->private) 2567cafe5635SKent Overstreet ret = true; 2568cafe5635SKent Overstreet else 2569cafe5635SKent Overstreet __bch_keybuf_del(buf, p); 2570cafe5635SKent Overstreet } 2571cafe5635SKent Overstreet 2572cafe5635SKent Overstreet spin_unlock(&buf->lock); 2573cafe5635SKent Overstreet return ret; 2574cafe5635SKent Overstreet } 2575cafe5635SKent Overstreet 2576cafe5635SKent Overstreet struct keybuf_key *bch_keybuf_next(struct keybuf *buf) 2577cafe5635SKent Overstreet { 2578cafe5635SKent Overstreet struct keybuf_key *w; 2579cafe5635SKent Overstreet spin_lock(&buf->lock); 2580cafe5635SKent Overstreet 2581cafe5635SKent Overstreet w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2582cafe5635SKent Overstreet 2583cafe5635SKent Overstreet while (w && w->private) 2584cafe5635SKent Overstreet w = RB_NEXT(w, node); 2585cafe5635SKent Overstreet 2586cafe5635SKent Overstreet if (w) 2587cafe5635SKent Overstreet w->private = ERR_PTR(-EINTR); 2588cafe5635SKent Overstreet 2589cafe5635SKent Overstreet spin_unlock(&buf->lock); 2590cafe5635SKent Overstreet return w; 2591cafe5635SKent Overstreet } 2592cafe5635SKent Overstreet 2593cafe5635SKent Overstreet struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, 2594cafe5635SKent Overstreet struct keybuf *buf, 259572c27061SKent Overstreet struct bkey *end, 259672c27061SKent Overstreet keybuf_pred_fn *pred) 2597cafe5635SKent Overstreet { 2598cafe5635SKent Overstreet struct keybuf_key *ret; 2599cafe5635SKent Overstreet 2600cafe5635SKent Overstreet while (1) { 2601cafe5635SKent Overstreet ret = bch_keybuf_next(buf); 2602cafe5635SKent Overstreet if (ret) 2603cafe5635SKent Overstreet break; 2604cafe5635SKent Overstreet 2605cafe5635SKent Overstreet if (bkey_cmp(&buf->last_scanned, end) >= 0) { 2606cafe5635SKent Overstreet pr_debug("scan finished"); 2607cafe5635SKent Overstreet break; 2608cafe5635SKent Overstreet } 2609cafe5635SKent Overstreet 261072c27061SKent Overstreet bch_refill_keybuf(c, buf, end, pred); 2611cafe5635SKent Overstreet } 2612cafe5635SKent Overstreet 2613cafe5635SKent Overstreet return ret; 2614cafe5635SKent Overstreet } 2615cafe5635SKent Overstreet 261672c27061SKent Overstreet void bch_keybuf_init(struct keybuf *buf) 2617cafe5635SKent Overstreet { 2618cafe5635SKent Overstreet buf->last_scanned = MAX_KEY; 2619cafe5635SKent Overstreet buf->keys = RB_ROOT; 2620cafe5635SKent Overstreet 2621cafe5635SKent Overstreet spin_lock_init(&buf->lock); 2622cafe5635SKent Overstreet array_allocator_init(&buf->freelist); 2623cafe5635SKent Overstreet } 2624cafe5635SKent Overstreet 2625cafe5635SKent Overstreet void bch_btree_exit(void) 2626cafe5635SKent Overstreet { 2627cafe5635SKent Overstreet if (btree_io_wq) 2628cafe5635SKent Overstreet destroy_workqueue(btree_io_wq); 2629cafe5635SKent Overstreet } 2630cafe5635SKent Overstreet 2631cafe5635SKent Overstreet int __init bch_btree_init(void) 2632cafe5635SKent Overstreet { 263372a44517SKent Overstreet btree_io_wq = create_singlethread_workqueue("bch_btree_io"); 263472a44517SKent Overstreet if (!btree_io_wq) 2635cafe5635SKent Overstreet return -ENOMEM; 2636cafe5635SKent Overstreet 2637cafe5635SKent Overstreet return 0; 2638cafe5635SKent Overstreet } 2639