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); \ 17078365411SKent Overstreet if (_r == -EINTR) \ 17178365411SKent 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 \ 18078365411SKent 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 20678b77bf8SKent Overstreet 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); 26188b9f8c4SKent Overstreet bset_sector_offset(b, i) < KEY_SIZE(&b->key); 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); 28188b9f8c4SKent Overstreet bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys", 282cafe5635SKent Overstreet err, PTR_BUCKET_NR(b->c, &b->key, 0), 28388b9f8c4SKent Overstreet bset_block_offset(b, i), 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 29378b77bf8SKent Overstreet static 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 48178b77bf8SKent Overstreet /* 48278b77bf8SKent Overstreet * do verify if there was more than one set initially (i.e. we did a 48378b77bf8SKent Overstreet * sort) and we sorted down to a single set: 48478b77bf8SKent Overstreet */ 48578b77bf8SKent Overstreet if (i != b->sets->data && !b->nsets) 48678b77bf8SKent Overstreet bch_btree_verify(b); 48778b77bf8SKent Overstreet 488cafe5635SKent Overstreet if (b->written < btree_blocks(b)) 489cafe5635SKent Overstreet bch_bset_init_next(b); 490cafe5635SKent Overstreet } 491cafe5635SKent Overstreet 492f269af5aSKent Overstreet static void bch_btree_node_write_sync(struct btree *b) 493f269af5aSKent Overstreet { 494f269af5aSKent Overstreet struct closure cl; 495f269af5aSKent Overstreet 496f269af5aSKent Overstreet closure_init_stack(&cl); 497f269af5aSKent Overstreet bch_btree_node_write(b, &cl); 498f269af5aSKent Overstreet closure_sync(&cl); 499f269af5aSKent Overstreet } 500f269af5aSKent Overstreet 50157943511SKent Overstreet static void btree_node_write_work(struct work_struct *w) 502cafe5635SKent Overstreet { 503cafe5635SKent Overstreet struct btree *b = container_of(to_delayed_work(w), struct btree, work); 504cafe5635SKent Overstreet 50557943511SKent Overstreet rw_lock(true, b, b->level); 506cafe5635SKent Overstreet 507cafe5635SKent Overstreet if (btree_node_dirty(b)) 50857943511SKent Overstreet bch_btree_node_write(b, NULL); 50957943511SKent Overstreet rw_unlock(true, b); 510cafe5635SKent Overstreet } 511cafe5635SKent Overstreet 512c18536a7SKent Overstreet static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref) 513cafe5635SKent Overstreet { 514cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 515cafe5635SKent Overstreet struct btree_write *w = btree_current_write(b); 516cafe5635SKent Overstreet 51757943511SKent Overstreet BUG_ON(!b->written); 51857943511SKent Overstreet BUG_ON(!i->keys); 519cafe5635SKent Overstreet 52057943511SKent Overstreet if (!btree_node_dirty(b)) 52157943511SKent Overstreet queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); 52257943511SKent Overstreet 523cafe5635SKent Overstreet set_btree_node_dirty(b); 524cafe5635SKent Overstreet 525c18536a7SKent Overstreet if (journal_ref) { 526cafe5635SKent Overstreet if (w->journal && 527c18536a7SKent Overstreet journal_pin_cmp(b->c, w->journal, journal_ref)) { 528cafe5635SKent Overstreet atomic_dec_bug(w->journal); 529cafe5635SKent Overstreet w->journal = NULL; 530cafe5635SKent Overstreet } 531cafe5635SKent Overstreet 532cafe5635SKent Overstreet if (!w->journal) { 533c18536a7SKent Overstreet w->journal = journal_ref; 534cafe5635SKent Overstreet atomic_inc(w->journal); 535cafe5635SKent Overstreet } 536cafe5635SKent Overstreet } 537cafe5635SKent Overstreet 538cafe5635SKent Overstreet /* Force write if set is too big */ 53957943511SKent Overstreet if (set_bytes(i) > PAGE_SIZE - 48 && 54057943511SKent Overstreet !current->bio_list) 54157943511SKent Overstreet bch_btree_node_write(b, NULL); 542cafe5635SKent Overstreet } 543cafe5635SKent Overstreet 544cafe5635SKent Overstreet /* 545cafe5635SKent Overstreet * Btree in memory cache - allocation/freeing 546cafe5635SKent Overstreet * mca -> memory cache 547cafe5635SKent Overstreet */ 548cafe5635SKent Overstreet 549cafe5635SKent Overstreet static void mca_reinit(struct btree *b) 550cafe5635SKent Overstreet { 551cafe5635SKent Overstreet unsigned i; 552cafe5635SKent Overstreet 553cafe5635SKent Overstreet b->flags = 0; 554cafe5635SKent Overstreet b->written = 0; 555cafe5635SKent Overstreet b->nsets = 0; 556cafe5635SKent Overstreet 557cafe5635SKent Overstreet for (i = 0; i < MAX_BSETS; i++) 558cafe5635SKent Overstreet b->sets[i].size = 0; 559cafe5635SKent Overstreet /* 560cafe5635SKent Overstreet * Second loop starts at 1 because b->sets[0]->data is the memory we 561cafe5635SKent Overstreet * allocated 562cafe5635SKent Overstreet */ 563cafe5635SKent Overstreet for (i = 1; i < MAX_BSETS; i++) 564cafe5635SKent Overstreet b->sets[i].data = NULL; 565cafe5635SKent Overstreet } 566cafe5635SKent Overstreet 567cafe5635SKent Overstreet #define mca_reserve(c) (((c->root && c->root->level) \ 568cafe5635SKent Overstreet ? c->root->level : 1) * 8 + 16) 569cafe5635SKent Overstreet #define mca_can_free(c) \ 570cafe5635SKent Overstreet max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) 571cafe5635SKent Overstreet 572cafe5635SKent Overstreet static void mca_data_free(struct btree *b) 573cafe5635SKent Overstreet { 574cafe5635SKent Overstreet struct bset_tree *t = b->sets; 575cb7a583eSKent Overstreet 576cb7a583eSKent Overstreet BUG_ON(b->io_mutex.count != 1); 577cafe5635SKent Overstreet 578cafe5635SKent Overstreet if (bset_prev_bytes(b) < PAGE_SIZE) 579cafe5635SKent Overstreet kfree(t->prev); 580cafe5635SKent Overstreet else 581cafe5635SKent Overstreet free_pages((unsigned long) t->prev, 582cafe5635SKent Overstreet get_order(bset_prev_bytes(b))); 583cafe5635SKent Overstreet 584cafe5635SKent Overstreet if (bset_tree_bytes(b) < PAGE_SIZE) 585cafe5635SKent Overstreet kfree(t->tree); 586cafe5635SKent Overstreet else 587cafe5635SKent Overstreet free_pages((unsigned long) t->tree, 588cafe5635SKent Overstreet get_order(bset_tree_bytes(b))); 589cafe5635SKent Overstreet 590cafe5635SKent Overstreet free_pages((unsigned long) t->data, b->page_order); 591cafe5635SKent Overstreet 592cafe5635SKent Overstreet t->prev = NULL; 593cafe5635SKent Overstreet t->tree = NULL; 594cafe5635SKent Overstreet t->data = NULL; 595cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache_freed); 596cafe5635SKent Overstreet b->c->bucket_cache_used--; 597cafe5635SKent Overstreet } 598cafe5635SKent Overstreet 599cafe5635SKent Overstreet static void mca_bucket_free(struct btree *b) 600cafe5635SKent Overstreet { 601cafe5635SKent Overstreet BUG_ON(btree_node_dirty(b)); 602cafe5635SKent Overstreet 603cafe5635SKent Overstreet b->key.ptr[0] = 0; 604cafe5635SKent Overstreet hlist_del_init_rcu(&b->hash); 605cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache_freeable); 606cafe5635SKent Overstreet } 607cafe5635SKent Overstreet 608cafe5635SKent Overstreet static unsigned btree_order(struct bkey *k) 609cafe5635SKent Overstreet { 610cafe5635SKent Overstreet return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); 611cafe5635SKent Overstreet } 612cafe5635SKent Overstreet 613cafe5635SKent Overstreet static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) 614cafe5635SKent Overstreet { 615cafe5635SKent Overstreet struct bset_tree *t = b->sets; 616cafe5635SKent Overstreet BUG_ON(t->data); 617cafe5635SKent Overstreet 618cafe5635SKent Overstreet b->page_order = max_t(unsigned, 619cafe5635SKent Overstreet ilog2(b->c->btree_pages), 620cafe5635SKent Overstreet btree_order(k)); 621cafe5635SKent Overstreet 622cafe5635SKent Overstreet t->data = (void *) __get_free_pages(gfp, b->page_order); 623cafe5635SKent Overstreet if (!t->data) 624cafe5635SKent Overstreet goto err; 625cafe5635SKent Overstreet 626cafe5635SKent Overstreet t->tree = bset_tree_bytes(b) < PAGE_SIZE 627cafe5635SKent Overstreet ? kmalloc(bset_tree_bytes(b), gfp) 628cafe5635SKent Overstreet : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); 629cafe5635SKent Overstreet if (!t->tree) 630cafe5635SKent Overstreet goto err; 631cafe5635SKent Overstreet 632cafe5635SKent Overstreet t->prev = bset_prev_bytes(b) < PAGE_SIZE 633cafe5635SKent Overstreet ? kmalloc(bset_prev_bytes(b), gfp) 634cafe5635SKent Overstreet : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); 635cafe5635SKent Overstreet if (!t->prev) 636cafe5635SKent Overstreet goto err; 637cafe5635SKent Overstreet 638cafe5635SKent Overstreet list_move(&b->list, &b->c->btree_cache); 639cafe5635SKent Overstreet b->c->bucket_cache_used++; 640cafe5635SKent Overstreet return; 641cafe5635SKent Overstreet err: 642cafe5635SKent Overstreet mca_data_free(b); 643cafe5635SKent Overstreet } 644cafe5635SKent Overstreet 645cafe5635SKent Overstreet static struct btree *mca_bucket_alloc(struct cache_set *c, 646cafe5635SKent Overstreet struct bkey *k, gfp_t gfp) 647cafe5635SKent Overstreet { 648cafe5635SKent Overstreet struct btree *b = kzalloc(sizeof(struct btree), gfp); 649cafe5635SKent Overstreet if (!b) 650cafe5635SKent Overstreet return NULL; 651cafe5635SKent Overstreet 652cafe5635SKent Overstreet init_rwsem(&b->lock); 653cafe5635SKent Overstreet lockdep_set_novalidate_class(&b->lock); 654cafe5635SKent Overstreet INIT_LIST_HEAD(&b->list); 65557943511SKent Overstreet INIT_DELAYED_WORK(&b->work, btree_node_write_work); 656cafe5635SKent Overstreet b->c = c; 657cb7a583eSKent Overstreet sema_init(&b->io_mutex, 1); 658cafe5635SKent Overstreet 659cafe5635SKent Overstreet mca_data_alloc(b, k, gfp); 660cafe5635SKent Overstreet return b; 661cafe5635SKent Overstreet } 662cafe5635SKent Overstreet 663e8e1d468SKent Overstreet static int mca_reap(struct btree *b, unsigned min_order, bool flush) 664cafe5635SKent Overstreet { 665e8e1d468SKent Overstreet struct closure cl; 666e8e1d468SKent Overstreet 667e8e1d468SKent Overstreet closure_init_stack(&cl); 668cafe5635SKent Overstreet lockdep_assert_held(&b->c->bucket_lock); 669cafe5635SKent Overstreet 670cafe5635SKent Overstreet if (!down_write_trylock(&b->lock)) 671cafe5635SKent Overstreet return -ENOMEM; 672cafe5635SKent Overstreet 673e8e1d468SKent Overstreet BUG_ON(btree_node_dirty(b) && !b->sets[0].data); 674e8e1d468SKent Overstreet 675cb7a583eSKent Overstreet if (b->page_order < min_order) 676cb7a583eSKent Overstreet goto out_unlock; 677cb7a583eSKent Overstreet 678cb7a583eSKent Overstreet if (!flush) { 679cb7a583eSKent Overstreet if (btree_node_dirty(b)) 680cb7a583eSKent Overstreet goto out_unlock; 681cb7a583eSKent Overstreet 682cb7a583eSKent Overstreet if (down_trylock(&b->io_mutex)) 683cb7a583eSKent Overstreet goto out_unlock; 684cb7a583eSKent Overstreet up(&b->io_mutex); 685cafe5635SKent Overstreet } 686cafe5635SKent Overstreet 687f269af5aSKent Overstreet if (btree_node_dirty(b)) 688f269af5aSKent Overstreet bch_btree_node_write_sync(b); 689cafe5635SKent Overstreet 690e8e1d468SKent Overstreet /* wait for any in flight btree write */ 691cb7a583eSKent Overstreet down(&b->io_mutex); 692cb7a583eSKent Overstreet up(&b->io_mutex); 693e8e1d468SKent Overstreet 694cafe5635SKent Overstreet return 0; 695cb7a583eSKent Overstreet out_unlock: 696cb7a583eSKent Overstreet rw_unlock(true, b); 697cb7a583eSKent Overstreet return -ENOMEM; 698cafe5635SKent Overstreet } 699cafe5635SKent Overstreet 7007dc19d5aSDave Chinner static unsigned long bch_mca_scan(struct shrinker *shrink, 7017dc19d5aSDave Chinner struct shrink_control *sc) 702cafe5635SKent Overstreet { 703cafe5635SKent Overstreet struct cache_set *c = container_of(shrink, struct cache_set, shrink); 704cafe5635SKent Overstreet struct btree *b, *t; 705cafe5635SKent Overstreet unsigned long i, nr = sc->nr_to_scan; 7067dc19d5aSDave Chinner unsigned long freed = 0; 707cafe5635SKent Overstreet 708cafe5635SKent Overstreet if (c->shrinker_disabled) 7097dc19d5aSDave Chinner return SHRINK_STOP; 710cafe5635SKent Overstreet 711cafe5635SKent Overstreet if (c->try_harder) 7127dc19d5aSDave Chinner return SHRINK_STOP; 713cafe5635SKent Overstreet 714cafe5635SKent Overstreet /* Return -1 if we can't do anything right now */ 715a698e08cSKent Overstreet if (sc->gfp_mask & __GFP_IO) 716cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 717cafe5635SKent Overstreet else if (!mutex_trylock(&c->bucket_lock)) 718cafe5635SKent Overstreet return -1; 719cafe5635SKent Overstreet 72036c9ea98SKent Overstreet /* 72136c9ea98SKent Overstreet * It's _really_ critical that we don't free too many btree nodes - we 72236c9ea98SKent Overstreet * have to always leave ourselves a reserve. The reserve is how we 72336c9ea98SKent Overstreet * guarantee that allocating memory for a new btree node can always 72436c9ea98SKent Overstreet * succeed, so that inserting keys into the btree can always succeed and 72536c9ea98SKent Overstreet * IO can always make forward progress: 72636c9ea98SKent Overstreet */ 727cafe5635SKent Overstreet nr /= c->btree_pages; 728cafe5635SKent Overstreet nr = min_t(unsigned long, nr, mca_can_free(c)); 729cafe5635SKent Overstreet 730cafe5635SKent Overstreet i = 0; 731cafe5635SKent Overstreet list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { 7327dc19d5aSDave Chinner if (freed >= nr) 733cafe5635SKent Overstreet break; 734cafe5635SKent Overstreet 735cafe5635SKent Overstreet if (++i > 3 && 736e8e1d468SKent Overstreet !mca_reap(b, 0, false)) { 737cafe5635SKent Overstreet mca_data_free(b); 738cafe5635SKent Overstreet rw_unlock(true, b); 7397dc19d5aSDave Chinner freed++; 740cafe5635SKent Overstreet } 741cafe5635SKent Overstreet } 742cafe5635SKent Overstreet 743b0f32a56SKent Overstreet for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { 744cafe5635SKent Overstreet if (list_empty(&c->btree_cache)) 745cafe5635SKent Overstreet goto out; 746cafe5635SKent Overstreet 747cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache, struct btree, list); 748cafe5635SKent Overstreet list_rotate_left(&c->btree_cache); 749cafe5635SKent Overstreet 750cafe5635SKent Overstreet if (!b->accessed && 751e8e1d468SKent Overstreet !mca_reap(b, 0, false)) { 752cafe5635SKent Overstreet mca_bucket_free(b); 753cafe5635SKent Overstreet mca_data_free(b); 754cafe5635SKent Overstreet rw_unlock(true, b); 7557dc19d5aSDave Chinner freed++; 756cafe5635SKent Overstreet } else 757cafe5635SKent Overstreet b->accessed = 0; 758cafe5635SKent Overstreet } 759cafe5635SKent Overstreet out: 760cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 7617dc19d5aSDave Chinner return freed; 7627dc19d5aSDave Chinner } 7637dc19d5aSDave Chinner 7647dc19d5aSDave Chinner static unsigned long bch_mca_count(struct shrinker *shrink, 7657dc19d5aSDave Chinner struct shrink_control *sc) 7667dc19d5aSDave Chinner { 7677dc19d5aSDave Chinner struct cache_set *c = container_of(shrink, struct cache_set, shrink); 7687dc19d5aSDave Chinner 7697dc19d5aSDave Chinner if (c->shrinker_disabled) 7707dc19d5aSDave Chinner return 0; 7717dc19d5aSDave Chinner 7727dc19d5aSDave Chinner if (c->try_harder) 7737dc19d5aSDave Chinner return 0; 7747dc19d5aSDave Chinner 7757dc19d5aSDave Chinner return mca_can_free(c) * c->btree_pages; 776cafe5635SKent Overstreet } 777cafe5635SKent Overstreet 778cafe5635SKent Overstreet void bch_btree_cache_free(struct cache_set *c) 779cafe5635SKent Overstreet { 780cafe5635SKent Overstreet struct btree *b; 781cafe5635SKent Overstreet struct closure cl; 782cafe5635SKent Overstreet closure_init_stack(&cl); 783cafe5635SKent Overstreet 784cafe5635SKent Overstreet if (c->shrink.list.next) 785cafe5635SKent Overstreet unregister_shrinker(&c->shrink); 786cafe5635SKent Overstreet 787cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 788cafe5635SKent Overstreet 789cafe5635SKent Overstreet #ifdef CONFIG_BCACHE_DEBUG 790cafe5635SKent Overstreet if (c->verify_data) 791cafe5635SKent Overstreet list_move(&c->verify_data->list, &c->btree_cache); 79278b77bf8SKent Overstreet 79378b77bf8SKent Overstreet free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c))); 794cafe5635SKent Overstreet #endif 795cafe5635SKent Overstreet 796cafe5635SKent Overstreet list_splice(&c->btree_cache_freeable, 797cafe5635SKent Overstreet &c->btree_cache); 798cafe5635SKent Overstreet 799cafe5635SKent Overstreet while (!list_empty(&c->btree_cache)) { 800cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache, struct btree, list); 801cafe5635SKent Overstreet 802cafe5635SKent Overstreet if (btree_node_dirty(b)) 803cafe5635SKent Overstreet btree_complete_write(b, btree_current_write(b)); 804cafe5635SKent Overstreet clear_bit(BTREE_NODE_dirty, &b->flags); 805cafe5635SKent Overstreet 806cafe5635SKent Overstreet mca_data_free(b); 807cafe5635SKent Overstreet } 808cafe5635SKent Overstreet 809cafe5635SKent Overstreet while (!list_empty(&c->btree_cache_freed)) { 810cafe5635SKent Overstreet b = list_first_entry(&c->btree_cache_freed, 811cafe5635SKent Overstreet struct btree, list); 812cafe5635SKent Overstreet list_del(&b->list); 813cafe5635SKent Overstreet cancel_delayed_work_sync(&b->work); 814cafe5635SKent Overstreet kfree(b); 815cafe5635SKent Overstreet } 816cafe5635SKent Overstreet 817cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 818cafe5635SKent Overstreet } 819cafe5635SKent Overstreet 820cafe5635SKent Overstreet int bch_btree_cache_alloc(struct cache_set *c) 821cafe5635SKent Overstreet { 822cafe5635SKent Overstreet unsigned i; 823cafe5635SKent Overstreet 824cafe5635SKent Overstreet for (i = 0; i < mca_reserve(c); i++) 82572a44517SKent Overstreet if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL)) 82672a44517SKent Overstreet return -ENOMEM; 827cafe5635SKent Overstreet 828cafe5635SKent Overstreet list_splice_init(&c->btree_cache, 829cafe5635SKent Overstreet &c->btree_cache_freeable); 830cafe5635SKent Overstreet 831cafe5635SKent Overstreet #ifdef CONFIG_BCACHE_DEBUG 832cafe5635SKent Overstreet mutex_init(&c->verify_lock); 833cafe5635SKent Overstreet 83478b77bf8SKent Overstreet c->verify_ondisk = (void *) 83578b77bf8SKent Overstreet __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c))); 83678b77bf8SKent Overstreet 837cafe5635SKent Overstreet c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); 838cafe5635SKent Overstreet 839cafe5635SKent Overstreet if (c->verify_data && 840cafe5635SKent Overstreet c->verify_data->sets[0].data) 841cafe5635SKent Overstreet list_del_init(&c->verify_data->list); 842cafe5635SKent Overstreet else 843cafe5635SKent Overstreet c->verify_data = NULL; 844cafe5635SKent Overstreet #endif 845cafe5635SKent Overstreet 8467dc19d5aSDave Chinner c->shrink.count_objects = bch_mca_count; 8477dc19d5aSDave Chinner c->shrink.scan_objects = bch_mca_scan; 848cafe5635SKent Overstreet c->shrink.seeks = 4; 849cafe5635SKent Overstreet c->shrink.batch = c->btree_pages * 2; 850cafe5635SKent Overstreet register_shrinker(&c->shrink); 851cafe5635SKent Overstreet 852cafe5635SKent Overstreet return 0; 853cafe5635SKent Overstreet } 854cafe5635SKent Overstreet 855cafe5635SKent Overstreet /* Btree in memory cache - hash table */ 856cafe5635SKent Overstreet 857cafe5635SKent Overstreet static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) 858cafe5635SKent Overstreet { 859cafe5635SKent Overstreet return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; 860cafe5635SKent Overstreet } 861cafe5635SKent Overstreet 862cafe5635SKent Overstreet static struct btree *mca_find(struct cache_set *c, struct bkey *k) 863cafe5635SKent Overstreet { 864cafe5635SKent Overstreet struct btree *b; 865cafe5635SKent Overstreet 866cafe5635SKent Overstreet rcu_read_lock(); 867cafe5635SKent Overstreet hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) 868cafe5635SKent Overstreet if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) 869cafe5635SKent Overstreet goto out; 870cafe5635SKent Overstreet b = NULL; 871cafe5635SKent Overstreet out: 872cafe5635SKent Overstreet rcu_read_unlock(); 873cafe5635SKent Overstreet return b; 874cafe5635SKent Overstreet } 875cafe5635SKent Overstreet 876e8e1d468SKent Overstreet static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k) 877cafe5635SKent Overstreet { 878e8e1d468SKent Overstreet struct btree *b; 879cafe5635SKent Overstreet 880c37511b8SKent Overstreet trace_bcache_btree_cache_cannibalize(c); 881c37511b8SKent Overstreet 882e8e1d468SKent Overstreet if (!c->try_harder) { 883e8e1d468SKent Overstreet c->try_harder = current; 884cafe5635SKent Overstreet c->try_harder_start = local_clock(); 885e8e1d468SKent Overstreet } else if (c->try_harder != current) 886e8e1d468SKent Overstreet return ERR_PTR(-ENOSPC); 887cafe5635SKent Overstreet 888e8e1d468SKent Overstreet list_for_each_entry_reverse(b, &c->btree_cache, list) 889e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), false)) 890e8e1d468SKent Overstreet return b; 891cafe5635SKent Overstreet 892e8e1d468SKent Overstreet list_for_each_entry_reverse(b, &c->btree_cache, list) 893e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), true)) 894e8e1d468SKent Overstreet return b; 895e8e1d468SKent Overstreet 896e8e1d468SKent Overstreet return ERR_PTR(-ENOMEM); 897cafe5635SKent Overstreet } 898cafe5635SKent Overstreet 899cafe5635SKent Overstreet /* 900cafe5635SKent Overstreet * We can only have one thread cannibalizing other cached btree nodes at a time, 901cafe5635SKent Overstreet * or we'll deadlock. We use an open coded mutex to ensure that, which a 902cafe5635SKent Overstreet * cannibalize_bucket() will take. This means every time we unlock the root of 903cafe5635SKent Overstreet * the btree, we need to release this lock if we have it held. 904cafe5635SKent Overstreet */ 905df8e8970SKent Overstreet static void bch_cannibalize_unlock(struct cache_set *c) 906cafe5635SKent Overstreet { 907e8e1d468SKent Overstreet if (c->try_harder == current) { 908169ef1cfSKent Overstreet bch_time_stats_update(&c->try_harder_time, c->try_harder_start); 909cafe5635SKent Overstreet c->try_harder = NULL; 910e8e1d468SKent Overstreet wake_up(&c->try_wait); 911cafe5635SKent Overstreet } 912cafe5635SKent Overstreet } 913cafe5635SKent Overstreet 914e8e1d468SKent Overstreet static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level) 915cafe5635SKent Overstreet { 916cafe5635SKent Overstreet struct btree *b; 917cafe5635SKent Overstreet 918e8e1d468SKent Overstreet BUG_ON(current->bio_list); 919e8e1d468SKent Overstreet 920cafe5635SKent Overstreet lockdep_assert_held(&c->bucket_lock); 921cafe5635SKent Overstreet 922cafe5635SKent Overstreet if (mca_find(c, k)) 923cafe5635SKent Overstreet return NULL; 924cafe5635SKent Overstreet 925cafe5635SKent Overstreet /* btree_free() doesn't free memory; it sticks the node on the end of 926cafe5635SKent Overstreet * the list. Check if there's any freed nodes there: 927cafe5635SKent Overstreet */ 928cafe5635SKent Overstreet list_for_each_entry(b, &c->btree_cache_freeable, list) 929e8e1d468SKent Overstreet if (!mca_reap(b, btree_order(k), false)) 930cafe5635SKent Overstreet goto out; 931cafe5635SKent Overstreet 932cafe5635SKent Overstreet /* We never free struct btree itself, just the memory that holds the on 933cafe5635SKent Overstreet * disk node. Check the freed list before allocating a new one: 934cafe5635SKent Overstreet */ 935cafe5635SKent Overstreet list_for_each_entry(b, &c->btree_cache_freed, list) 936e8e1d468SKent Overstreet if (!mca_reap(b, 0, false)) { 937cafe5635SKent Overstreet mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); 938cafe5635SKent Overstreet if (!b->sets[0].data) 939cafe5635SKent Overstreet goto err; 940cafe5635SKent Overstreet else 941cafe5635SKent Overstreet goto out; 942cafe5635SKent Overstreet } 943cafe5635SKent Overstreet 944cafe5635SKent Overstreet b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); 945cafe5635SKent Overstreet if (!b) 946cafe5635SKent Overstreet goto err; 947cafe5635SKent Overstreet 948cafe5635SKent Overstreet BUG_ON(!down_write_trylock(&b->lock)); 949cafe5635SKent Overstreet if (!b->sets->data) 950cafe5635SKent Overstreet goto err; 951cafe5635SKent Overstreet out: 952cb7a583eSKent Overstreet BUG_ON(b->io_mutex.count != 1); 953cafe5635SKent Overstreet 954cafe5635SKent Overstreet bkey_copy(&b->key, k); 955cafe5635SKent Overstreet list_move(&b->list, &c->btree_cache); 956cafe5635SKent Overstreet hlist_del_init_rcu(&b->hash); 957cafe5635SKent Overstreet hlist_add_head_rcu(&b->hash, mca_hash(c, k)); 958cafe5635SKent Overstreet 959cafe5635SKent Overstreet lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); 960cafe5635SKent Overstreet b->level = level; 961d6fd3b11SKent Overstreet b->parent = (void *) ~0UL; 962cafe5635SKent Overstreet 963cafe5635SKent Overstreet mca_reinit(b); 964cafe5635SKent Overstreet 965cafe5635SKent Overstreet return b; 966cafe5635SKent Overstreet err: 967cafe5635SKent Overstreet if (b) 968cafe5635SKent Overstreet rw_unlock(true, b); 969cafe5635SKent Overstreet 970e8e1d468SKent Overstreet b = mca_cannibalize(c, k); 971cafe5635SKent Overstreet if (!IS_ERR(b)) 972cafe5635SKent Overstreet goto out; 973cafe5635SKent Overstreet 974cafe5635SKent Overstreet return b; 975cafe5635SKent Overstreet } 976cafe5635SKent Overstreet 977cafe5635SKent Overstreet /** 978cafe5635SKent Overstreet * bch_btree_node_get - find a btree node in the cache and lock it, reading it 979cafe5635SKent Overstreet * in from disk if necessary. 980cafe5635SKent Overstreet * 981b54d6934SKent Overstreet * If IO is necessary and running under generic_make_request, returns -EAGAIN. 982cafe5635SKent Overstreet * 983cafe5635SKent Overstreet * The btree node will have either a read or a write lock held, depending on 984cafe5635SKent Overstreet * level and op->lock. 985cafe5635SKent Overstreet */ 986cafe5635SKent Overstreet struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, 987e8e1d468SKent Overstreet int level, bool write) 988cafe5635SKent Overstreet { 989cafe5635SKent Overstreet int i = 0; 990cafe5635SKent Overstreet struct btree *b; 991cafe5635SKent Overstreet 992cafe5635SKent Overstreet BUG_ON(level < 0); 993cafe5635SKent Overstreet retry: 994cafe5635SKent Overstreet b = mca_find(c, k); 995cafe5635SKent Overstreet 996cafe5635SKent Overstreet if (!b) { 99757943511SKent Overstreet if (current->bio_list) 99857943511SKent Overstreet return ERR_PTR(-EAGAIN); 99957943511SKent Overstreet 1000cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1001e8e1d468SKent Overstreet b = mca_alloc(c, k, level); 1002cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1003cafe5635SKent Overstreet 1004cafe5635SKent Overstreet if (!b) 1005cafe5635SKent Overstreet goto retry; 1006cafe5635SKent Overstreet if (IS_ERR(b)) 1007cafe5635SKent Overstreet return b; 1008cafe5635SKent Overstreet 100957943511SKent Overstreet bch_btree_node_read(b); 1010cafe5635SKent Overstreet 1011cafe5635SKent Overstreet if (!write) 1012cafe5635SKent Overstreet downgrade_write(&b->lock); 1013cafe5635SKent Overstreet } else { 1014cafe5635SKent Overstreet rw_lock(write, b, level); 1015cafe5635SKent Overstreet if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { 1016cafe5635SKent Overstreet rw_unlock(write, b); 1017cafe5635SKent Overstreet goto retry; 1018cafe5635SKent Overstreet } 1019cafe5635SKent Overstreet BUG_ON(b->level != level); 1020cafe5635SKent Overstreet } 1021cafe5635SKent Overstreet 1022cafe5635SKent Overstreet b->accessed = 1; 1023cafe5635SKent Overstreet 1024cafe5635SKent Overstreet for (; i <= b->nsets && b->sets[i].size; i++) { 1025cafe5635SKent Overstreet prefetch(b->sets[i].tree); 1026cafe5635SKent Overstreet prefetch(b->sets[i].data); 1027cafe5635SKent Overstreet } 1028cafe5635SKent Overstreet 1029cafe5635SKent Overstreet for (; i <= b->nsets; i++) 1030cafe5635SKent Overstreet prefetch(b->sets[i].data); 1031cafe5635SKent Overstreet 103257943511SKent Overstreet if (btree_node_io_error(b)) { 1033cafe5635SKent Overstreet rw_unlock(write, b); 103457943511SKent Overstreet return ERR_PTR(-EIO); 103557943511SKent Overstreet } 103657943511SKent Overstreet 1037cafe5635SKent Overstreet BUG_ON(!b->written); 1038cafe5635SKent Overstreet 1039cafe5635SKent Overstreet return b; 1040cafe5635SKent Overstreet } 1041cafe5635SKent Overstreet 1042cafe5635SKent Overstreet static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) 1043cafe5635SKent Overstreet { 1044cafe5635SKent Overstreet struct btree *b; 1045cafe5635SKent Overstreet 1046cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1047e8e1d468SKent Overstreet b = mca_alloc(c, k, level); 1048cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1049cafe5635SKent Overstreet 1050cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(b)) { 105157943511SKent Overstreet bch_btree_node_read(b); 1052cafe5635SKent Overstreet rw_unlock(true, b); 1053cafe5635SKent Overstreet } 1054cafe5635SKent Overstreet } 1055cafe5635SKent Overstreet 1056cafe5635SKent Overstreet /* Btree alloc */ 1057cafe5635SKent Overstreet 1058e8e1d468SKent Overstreet static void btree_node_free(struct btree *b) 1059cafe5635SKent Overstreet { 1060cafe5635SKent Overstreet unsigned i; 1061cafe5635SKent Overstreet 1062c37511b8SKent Overstreet trace_bcache_btree_node_free(b); 1063c37511b8SKent Overstreet 1064cafe5635SKent Overstreet BUG_ON(b == b->c->root); 1065cafe5635SKent Overstreet 1066cafe5635SKent Overstreet if (btree_node_dirty(b)) 1067cafe5635SKent Overstreet btree_complete_write(b, btree_current_write(b)); 1068cafe5635SKent Overstreet clear_bit(BTREE_NODE_dirty, &b->flags); 1069cafe5635SKent Overstreet 1070cafe5635SKent Overstreet cancel_delayed_work(&b->work); 1071cafe5635SKent Overstreet 1072cafe5635SKent Overstreet mutex_lock(&b->c->bucket_lock); 1073cafe5635SKent Overstreet 1074cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&b->key); i++) { 1075cafe5635SKent Overstreet BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); 1076cafe5635SKent Overstreet 1077cafe5635SKent Overstreet bch_inc_gen(PTR_CACHE(b->c, &b->key, i), 1078cafe5635SKent Overstreet PTR_BUCKET(b->c, &b->key, i)); 1079cafe5635SKent Overstreet } 1080cafe5635SKent Overstreet 1081cafe5635SKent Overstreet bch_bucket_free(b->c, &b->key); 1082cafe5635SKent Overstreet mca_bucket_free(b); 1083cafe5635SKent Overstreet mutex_unlock(&b->c->bucket_lock); 1084cafe5635SKent Overstreet } 1085cafe5635SKent Overstreet 1086bc9389eeSKent Overstreet struct btree *bch_btree_node_alloc(struct cache_set *c, int level, bool wait) 1087cafe5635SKent Overstreet { 1088cafe5635SKent Overstreet BKEY_PADDED(key) k; 1089cafe5635SKent Overstreet struct btree *b = ERR_PTR(-EAGAIN); 1090cafe5635SKent Overstreet 1091cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1092cafe5635SKent Overstreet retry: 109378365411SKent Overstreet if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait)) 1094cafe5635SKent Overstreet goto err; 1095cafe5635SKent Overstreet 10963a3b6a4eSKent Overstreet bkey_put(c, &k.key); 1097cafe5635SKent Overstreet SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); 1098cafe5635SKent Overstreet 1099e8e1d468SKent Overstreet b = mca_alloc(c, &k.key, level); 1100cafe5635SKent Overstreet if (IS_ERR(b)) 1101cafe5635SKent Overstreet goto err_free; 1102cafe5635SKent Overstreet 1103cafe5635SKent Overstreet if (!b) { 1104b1a67b0fSKent Overstreet cache_bug(c, 1105b1a67b0fSKent Overstreet "Tried to allocate bucket that was in btree cache"); 1106cafe5635SKent Overstreet goto retry; 1107cafe5635SKent Overstreet } 1108cafe5635SKent Overstreet 1109cafe5635SKent Overstreet b->accessed = 1; 1110cafe5635SKent Overstreet bch_bset_init_next(b); 1111cafe5635SKent Overstreet 1112cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1113c37511b8SKent Overstreet 1114c37511b8SKent Overstreet trace_bcache_btree_node_alloc(b); 1115cafe5635SKent Overstreet return b; 1116cafe5635SKent Overstreet err_free: 1117cafe5635SKent Overstreet bch_bucket_free(c, &k.key); 1118cafe5635SKent Overstreet err: 1119cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1120c37511b8SKent Overstreet 1121c37511b8SKent Overstreet trace_bcache_btree_node_alloc_fail(b); 1122cafe5635SKent Overstreet return b; 1123cafe5635SKent Overstreet } 1124cafe5635SKent Overstreet 1125bc9389eeSKent Overstreet static struct btree *btree_node_alloc_replacement(struct btree *b, bool wait) 1126cafe5635SKent Overstreet { 1127bc9389eeSKent Overstreet struct btree *n = bch_btree_node_alloc(b->c, b->level, wait); 1128cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(n)) 1129cafe5635SKent Overstreet bch_btree_sort_into(b, n); 1130cafe5635SKent Overstreet 1131cafe5635SKent Overstreet return n; 1132cafe5635SKent Overstreet } 1133cafe5635SKent Overstreet 11348835c123SKent Overstreet static void make_btree_freeing_key(struct btree *b, struct bkey *k) 11358835c123SKent Overstreet { 11368835c123SKent Overstreet unsigned i; 11378835c123SKent Overstreet 11388835c123SKent Overstreet bkey_copy(k, &b->key); 11398835c123SKent Overstreet bkey_copy_key(k, &ZERO_KEY); 11408835c123SKent Overstreet 11418835c123SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 11428835c123SKent Overstreet uint8_t g = PTR_BUCKET(b->c, k, i)->gen + 1; 11438835c123SKent Overstreet 11448835c123SKent Overstreet SET_PTR_GEN(k, i, g); 11458835c123SKent Overstreet } 11468835c123SKent Overstreet 11478835c123SKent Overstreet atomic_inc(&b->c->prio_blocked); 11488835c123SKent Overstreet } 11498835c123SKent Overstreet 115078365411SKent Overstreet static int btree_check_reserve(struct btree *b, struct btree_op *op) 115178365411SKent Overstreet { 115278365411SKent Overstreet struct cache_set *c = b->c; 115378365411SKent Overstreet struct cache *ca; 115478365411SKent Overstreet unsigned i, reserve = c->root->level * 2 + 1; 115578365411SKent Overstreet int ret = 0; 115678365411SKent Overstreet 115778365411SKent Overstreet mutex_lock(&c->bucket_lock); 115878365411SKent Overstreet 115978365411SKent Overstreet for_each_cache(ca, c, i) 116078365411SKent Overstreet if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) { 116178365411SKent Overstreet if (op) 116278365411SKent Overstreet prepare_to_wait(&c->bucket_wait, &op->wait, 116378365411SKent Overstreet TASK_UNINTERRUPTIBLE); 116478365411SKent Overstreet ret = -EINTR; 116578365411SKent Overstreet break; 116678365411SKent Overstreet } 116778365411SKent Overstreet 116878365411SKent Overstreet mutex_unlock(&c->bucket_lock); 116978365411SKent Overstreet return ret; 117078365411SKent Overstreet } 117178365411SKent Overstreet 1172cafe5635SKent Overstreet /* Garbage collection */ 1173cafe5635SKent Overstreet 1174cafe5635SKent Overstreet uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) 1175cafe5635SKent Overstreet { 1176cafe5635SKent Overstreet uint8_t stale = 0; 1177cafe5635SKent Overstreet unsigned i; 1178cafe5635SKent Overstreet struct bucket *g; 1179cafe5635SKent Overstreet 1180cafe5635SKent Overstreet /* 1181cafe5635SKent Overstreet * ptr_invalid() can't return true for the keys that mark btree nodes as 1182cafe5635SKent Overstreet * freed, but since ptr_bad() returns true we'll never actually use them 1183cafe5635SKent Overstreet * for anything and thus we don't want mark their pointers here 1184cafe5635SKent Overstreet */ 1185cafe5635SKent Overstreet if (!bkey_cmp(k, &ZERO_KEY)) 1186cafe5635SKent Overstreet return stale; 1187cafe5635SKent Overstreet 1188cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 1189cafe5635SKent Overstreet if (!ptr_available(c, k, i)) 1190cafe5635SKent Overstreet continue; 1191cafe5635SKent Overstreet 1192cafe5635SKent Overstreet g = PTR_BUCKET(c, k, i); 1193cafe5635SKent Overstreet 1194cafe5635SKent Overstreet if (gen_after(g->gc_gen, PTR_GEN(k, i))) 1195cafe5635SKent Overstreet g->gc_gen = PTR_GEN(k, i); 1196cafe5635SKent Overstreet 1197cafe5635SKent Overstreet if (ptr_stale(c, k, i)) { 1198cafe5635SKent Overstreet stale = max(stale, ptr_stale(c, k, i)); 1199cafe5635SKent Overstreet continue; 1200cafe5635SKent Overstreet } 1201cafe5635SKent Overstreet 1202cafe5635SKent Overstreet cache_bug_on(GC_MARK(g) && 1203cafe5635SKent Overstreet (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), 1204cafe5635SKent Overstreet c, "inconsistent ptrs: mark = %llu, level = %i", 1205cafe5635SKent Overstreet GC_MARK(g), level); 1206cafe5635SKent Overstreet 1207cafe5635SKent Overstreet if (level) 1208cafe5635SKent Overstreet SET_GC_MARK(g, GC_MARK_METADATA); 1209cafe5635SKent Overstreet else if (KEY_DIRTY(k)) 1210cafe5635SKent Overstreet SET_GC_MARK(g, GC_MARK_DIRTY); 1211cafe5635SKent Overstreet 1212cafe5635SKent Overstreet /* guard against overflow */ 1213cafe5635SKent Overstreet SET_GC_SECTORS_USED(g, min_t(unsigned, 1214cafe5635SKent Overstreet GC_SECTORS_USED(g) + KEY_SIZE(k), 1215cafe5635SKent Overstreet (1 << 14) - 1)); 1216cafe5635SKent Overstreet 1217cafe5635SKent Overstreet BUG_ON(!GC_SECTORS_USED(g)); 1218cafe5635SKent Overstreet } 1219cafe5635SKent Overstreet 1220cafe5635SKent Overstreet return stale; 1221cafe5635SKent Overstreet } 1222cafe5635SKent Overstreet 1223cafe5635SKent Overstreet #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) 1224cafe5635SKent Overstreet 1225a1f0358bSKent Overstreet static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc) 1226cafe5635SKent Overstreet { 1227cafe5635SKent Overstreet uint8_t stale = 0; 1228a1f0358bSKent Overstreet unsigned keys = 0, good_keys = 0; 1229cafe5635SKent Overstreet struct bkey *k; 1230cafe5635SKent Overstreet struct btree_iter iter; 1231cafe5635SKent Overstreet struct bset_tree *t; 1232cafe5635SKent Overstreet 1233cafe5635SKent Overstreet gc->nodes++; 1234cafe5635SKent Overstreet 1235cafe5635SKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1236cafe5635SKent Overstreet stale = max(stale, btree_mark_key(b, k)); 1237a1f0358bSKent Overstreet keys++; 1238cafe5635SKent Overstreet 1239cafe5635SKent Overstreet if (bch_ptr_bad(b, k)) 1240cafe5635SKent Overstreet continue; 1241cafe5635SKent Overstreet 1242cafe5635SKent Overstreet gc->key_bytes += bkey_u64s(k); 1243cafe5635SKent Overstreet gc->nkeys++; 1244a1f0358bSKent Overstreet good_keys++; 1245cafe5635SKent Overstreet 1246cafe5635SKent Overstreet gc->data += KEY_SIZE(k); 1247cafe5635SKent Overstreet } 1248cafe5635SKent Overstreet 1249cafe5635SKent Overstreet for (t = b->sets; t <= &b->sets[b->nsets]; t++) 1250cafe5635SKent Overstreet btree_bug_on(t->size && 1251cafe5635SKent Overstreet bset_written(b, t) && 1252cafe5635SKent Overstreet bkey_cmp(&b->key, &t->end) < 0, 1253cafe5635SKent Overstreet b, "found short btree key in gc"); 1254cafe5635SKent Overstreet 1255a1f0358bSKent Overstreet if (b->c->gc_always_rewrite) 1256a1f0358bSKent Overstreet return true; 1257a1f0358bSKent Overstreet 1258a1f0358bSKent Overstreet if (stale > 10) 1259a1f0358bSKent Overstreet return true; 1260a1f0358bSKent Overstreet 1261a1f0358bSKent Overstreet if ((keys - good_keys) * 2 > keys) 1262a1f0358bSKent Overstreet return true; 1263a1f0358bSKent Overstreet 1264a1f0358bSKent Overstreet return false; 1265cafe5635SKent Overstreet } 1266cafe5635SKent Overstreet 1267a1f0358bSKent Overstreet #define GC_MERGE_NODES 4U 1268cafe5635SKent Overstreet 1269cafe5635SKent Overstreet struct gc_merge_info { 1270cafe5635SKent Overstreet struct btree *b; 1271cafe5635SKent Overstreet unsigned keys; 1272cafe5635SKent Overstreet }; 1273cafe5635SKent Overstreet 1274a1f0358bSKent Overstreet static int bch_btree_insert_node(struct btree *, struct btree_op *, 1275a1f0358bSKent Overstreet struct keylist *, atomic_t *, struct bkey *); 1276a1f0358bSKent Overstreet 1277a1f0358bSKent Overstreet static int btree_gc_coalesce(struct btree *b, struct btree_op *op, 1278a1f0358bSKent Overstreet struct keylist *keylist, struct gc_stat *gc, 1279e8e1d468SKent Overstreet struct gc_merge_info *r) 1280cafe5635SKent Overstreet { 1281a1f0358bSKent Overstreet unsigned i, nodes = 0, keys = 0, blocks; 1282a1f0358bSKent Overstreet struct btree *new_nodes[GC_MERGE_NODES]; 1283b54d6934SKent Overstreet struct closure cl; 1284a1f0358bSKent Overstreet struct bkey *k; 1285b54d6934SKent Overstreet 1286a1f0358bSKent Overstreet memset(new_nodes, 0, sizeof(new_nodes)); 1287b54d6934SKent Overstreet closure_init_stack(&cl); 1288cafe5635SKent Overstreet 1289a1f0358bSKent Overstreet while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b)) 1290cafe5635SKent Overstreet keys += r[nodes++].keys; 1291cafe5635SKent Overstreet 1292cafe5635SKent Overstreet blocks = btree_default_blocks(b->c) * 2 / 3; 1293cafe5635SKent Overstreet 1294cafe5635SKent Overstreet if (nodes < 2 || 1295cafe5635SKent Overstreet __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) 1296a1f0358bSKent Overstreet return 0; 1297cafe5635SKent Overstreet 1298a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1299bc9389eeSKent Overstreet new_nodes[i] = btree_node_alloc_replacement(r[i].b, false); 1300a1f0358bSKent Overstreet if (IS_ERR_OR_NULL(new_nodes[i])) 1301a1f0358bSKent Overstreet goto out_nocoalesce; 1302cafe5635SKent Overstreet } 1303cafe5635SKent Overstreet 1304cafe5635SKent Overstreet for (i = nodes - 1; i > 0; --i) { 1305a1f0358bSKent Overstreet struct bset *n1 = new_nodes[i]->sets->data; 1306a1f0358bSKent Overstreet struct bset *n2 = new_nodes[i - 1]->sets->data; 1307cafe5635SKent Overstreet struct bkey *k, *last = NULL; 1308cafe5635SKent Overstreet 1309cafe5635SKent Overstreet keys = 0; 1310cafe5635SKent Overstreet 1311a1f0358bSKent Overstreet if (i > 1) { 1312cafe5635SKent Overstreet for (k = n2->start; 1313cafe5635SKent Overstreet k < end(n2); 1314cafe5635SKent Overstreet k = bkey_next(k)) { 1315cafe5635SKent Overstreet if (__set_blocks(n1, n1->keys + keys + 1316cafe5635SKent Overstreet bkey_u64s(k), b->c) > blocks) 1317cafe5635SKent Overstreet break; 1318cafe5635SKent Overstreet 1319cafe5635SKent Overstreet last = k; 1320cafe5635SKent Overstreet keys += bkey_u64s(k); 1321cafe5635SKent Overstreet } 1322a1f0358bSKent Overstreet } else { 1323a1f0358bSKent Overstreet /* 1324a1f0358bSKent Overstreet * Last node we're not getting rid of - we're getting 1325a1f0358bSKent Overstreet * rid of the node at r[0]. Have to try and fit all of 1326a1f0358bSKent Overstreet * the remaining keys into this node; we can't ensure 1327a1f0358bSKent Overstreet * they will always fit due to rounding and variable 1328a1f0358bSKent Overstreet * length keys (shouldn't be possible in practice, 1329a1f0358bSKent Overstreet * though) 1330a1f0358bSKent Overstreet */ 1331a1f0358bSKent Overstreet if (__set_blocks(n1, n1->keys + n2->keys, 1332a1f0358bSKent Overstreet b->c) > btree_blocks(new_nodes[i])) 1333a1f0358bSKent Overstreet goto out_nocoalesce; 1334a1f0358bSKent Overstreet 1335a1f0358bSKent Overstreet keys = n2->keys; 1336a1f0358bSKent Overstreet /* Take the key of the node we're getting rid of */ 1337a1f0358bSKent Overstreet last = &r->b->key; 1338a1f0358bSKent Overstreet } 1339cafe5635SKent Overstreet 1340cafe5635SKent Overstreet BUG_ON(__set_blocks(n1, n1->keys + keys, 1341a1f0358bSKent Overstreet b->c) > btree_blocks(new_nodes[i])); 1342cafe5635SKent Overstreet 1343a1f0358bSKent Overstreet if (last) 1344a1f0358bSKent Overstreet bkey_copy_key(&new_nodes[i]->key, last); 1345cafe5635SKent Overstreet 1346cafe5635SKent Overstreet memcpy(end(n1), 1347cafe5635SKent Overstreet n2->start, 1348cafe5635SKent Overstreet (void *) node(n2, keys) - (void *) n2->start); 1349cafe5635SKent Overstreet 1350cafe5635SKent Overstreet n1->keys += keys; 1351a1f0358bSKent Overstreet r[i].keys = n1->keys; 1352cafe5635SKent Overstreet 1353cafe5635SKent Overstreet memmove(n2->start, 1354cafe5635SKent Overstreet node(n2, keys), 1355cafe5635SKent Overstreet (void *) end(n2) - (void *) node(n2, keys)); 1356cafe5635SKent Overstreet 1357cafe5635SKent Overstreet n2->keys -= keys; 1358cafe5635SKent Overstreet 1359*085d2a3dSKent Overstreet if (__bch_keylist_realloc(keylist, 1360*085d2a3dSKent Overstreet bkey_u64s(&new_nodes[i]->key))) 1361a1f0358bSKent Overstreet goto out_nocoalesce; 1362a1f0358bSKent Overstreet 1363a1f0358bSKent Overstreet bch_btree_node_write(new_nodes[i], &cl); 1364a1f0358bSKent Overstreet bch_keylist_add(keylist, &new_nodes[i]->key); 1365cafe5635SKent Overstreet } 1366cafe5635SKent Overstreet 1367a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1368*085d2a3dSKent Overstreet if (__bch_keylist_realloc(keylist, bkey_u64s(&r[i].b->key))) 1369a1f0358bSKent Overstreet goto out_nocoalesce; 1370a1f0358bSKent Overstreet 1371a1f0358bSKent Overstreet make_btree_freeing_key(r[i].b, keylist->top); 1372a1f0358bSKent Overstreet bch_keylist_push(keylist); 1373a1f0358bSKent Overstreet } 1374a1f0358bSKent Overstreet 1375a1f0358bSKent Overstreet /* We emptied out this node */ 1376a1f0358bSKent Overstreet BUG_ON(new_nodes[0]->sets->data->keys); 1377a1f0358bSKent Overstreet btree_node_free(new_nodes[0]); 1378a1f0358bSKent Overstreet rw_unlock(true, new_nodes[0]); 1379a1f0358bSKent Overstreet 1380a1f0358bSKent Overstreet closure_sync(&cl); 1381a1f0358bSKent Overstreet 1382a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) { 1383a1f0358bSKent Overstreet btree_node_free(r[i].b); 1384a1f0358bSKent Overstreet rw_unlock(true, r[i].b); 1385a1f0358bSKent Overstreet 1386a1f0358bSKent Overstreet r[i].b = new_nodes[i]; 1387a1f0358bSKent Overstreet } 1388a1f0358bSKent Overstreet 1389a1f0358bSKent Overstreet bch_btree_insert_node(b, op, keylist, NULL, NULL); 1390a1f0358bSKent Overstreet BUG_ON(!bch_keylist_empty(keylist)); 1391a1f0358bSKent Overstreet 1392a1f0358bSKent Overstreet memmove(r, r + 1, sizeof(r[0]) * (nodes - 1)); 1393a1f0358bSKent Overstreet r[nodes - 1].b = ERR_PTR(-EINTR); 1394cafe5635SKent Overstreet 1395c37511b8SKent Overstreet trace_bcache_btree_gc_coalesce(nodes); 1396cafe5635SKent Overstreet gc->nodes--; 1397cafe5635SKent Overstreet 1398a1f0358bSKent Overstreet /* Invalidated our iterator */ 1399a1f0358bSKent Overstreet return -EINTR; 1400a1f0358bSKent Overstreet 1401a1f0358bSKent Overstreet out_nocoalesce: 1402a1f0358bSKent Overstreet closure_sync(&cl); 1403a1f0358bSKent Overstreet 1404a1f0358bSKent Overstreet while ((k = bch_keylist_pop(keylist))) 1405a1f0358bSKent Overstreet if (!bkey_cmp(k, &ZERO_KEY)) 1406a1f0358bSKent Overstreet atomic_dec(&b->c->prio_blocked); 1407a1f0358bSKent Overstreet 1408a1f0358bSKent Overstreet for (i = 0; i < nodes; i++) 1409a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(new_nodes[i])) { 1410a1f0358bSKent Overstreet btree_node_free(new_nodes[i]); 1411a1f0358bSKent Overstreet rw_unlock(true, new_nodes[i]); 1412a1f0358bSKent Overstreet } 1413a1f0358bSKent Overstreet return 0; 1414a1f0358bSKent Overstreet } 1415a1f0358bSKent Overstreet 1416a1f0358bSKent Overstreet static unsigned btree_gc_count_keys(struct btree *b) 1417a1f0358bSKent Overstreet { 1418a1f0358bSKent Overstreet struct bkey *k; 1419a1f0358bSKent Overstreet struct btree_iter iter; 1420a1f0358bSKent Overstreet unsigned ret = 0; 1421a1f0358bSKent Overstreet 1422a1f0358bSKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_bad) 1423a1f0358bSKent Overstreet ret += bkey_u64s(k); 1424a1f0358bSKent Overstreet 1425a1f0358bSKent Overstreet return ret; 1426cafe5635SKent Overstreet } 1427cafe5635SKent Overstreet 1428cafe5635SKent Overstreet static int btree_gc_recurse(struct btree *b, struct btree_op *op, 1429cafe5635SKent Overstreet struct closure *writes, struct gc_stat *gc) 1430cafe5635SKent Overstreet { 1431cafe5635SKent Overstreet unsigned i; 1432a1f0358bSKent Overstreet int ret = 0; 1433a1f0358bSKent Overstreet bool should_rewrite; 1434a1f0358bSKent Overstreet struct btree *n; 1435a1f0358bSKent Overstreet struct bkey *k; 1436a1f0358bSKent Overstreet struct keylist keys; 1437a1f0358bSKent Overstreet struct btree_iter iter; 1438cafe5635SKent Overstreet struct gc_merge_info r[GC_MERGE_NODES]; 1439a1f0358bSKent Overstreet struct gc_merge_info *last = r + GC_MERGE_NODES - 1; 1440cafe5635SKent Overstreet 1441a1f0358bSKent Overstreet bch_keylist_init(&keys); 1442a1f0358bSKent Overstreet bch_btree_iter_init(b, &iter, &b->c->gc_done); 1443cafe5635SKent Overstreet 1444a1f0358bSKent Overstreet for (i = 0; i < GC_MERGE_NODES; i++) 1445a1f0358bSKent Overstreet r[i].b = ERR_PTR(-EINTR); 1446cafe5635SKent Overstreet 1447a1f0358bSKent Overstreet while (1) { 1448a1f0358bSKent Overstreet k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); 1449a1f0358bSKent Overstreet if (k) { 1450a1f0358bSKent Overstreet r->b = bch_btree_node_get(b->c, k, b->level - 1, true); 1451cafe5635SKent Overstreet if (IS_ERR(r->b)) { 1452cafe5635SKent Overstreet ret = PTR_ERR(r->b); 1453cafe5635SKent Overstreet break; 1454cafe5635SKent Overstreet } 1455cafe5635SKent Overstreet 1456a1f0358bSKent Overstreet r->keys = btree_gc_count_keys(r->b); 1457cafe5635SKent Overstreet 1458a1f0358bSKent Overstreet ret = btree_gc_coalesce(b, op, &keys, gc, r); 1459a1f0358bSKent Overstreet if (ret) 1460cafe5635SKent Overstreet break; 1461cafe5635SKent Overstreet } 1462cafe5635SKent Overstreet 1463a1f0358bSKent Overstreet if (!last->b) 1464a1f0358bSKent Overstreet break; 1465cafe5635SKent Overstreet 1466a1f0358bSKent Overstreet if (!IS_ERR(last->b)) { 1467a1f0358bSKent Overstreet should_rewrite = btree_gc_mark_node(last->b, gc); 146878365411SKent Overstreet if (should_rewrite && 146978365411SKent Overstreet !btree_check_reserve(b, NULL)) { 1470bc9389eeSKent Overstreet n = btree_node_alloc_replacement(last->b, 1471bc9389eeSKent Overstreet false); 1472cafe5635SKent Overstreet 1473a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(n)) { 1474a1f0358bSKent Overstreet bch_btree_node_write_sync(n); 1475a1f0358bSKent Overstreet bch_keylist_add(&keys, &n->key); 1476cafe5635SKent Overstreet 1477a1f0358bSKent Overstreet make_btree_freeing_key(last->b, 1478a1f0358bSKent Overstreet keys.top); 1479a1f0358bSKent Overstreet bch_keylist_push(&keys); 1480cafe5635SKent Overstreet 1481a1f0358bSKent Overstreet btree_node_free(last->b); 1482a1f0358bSKent Overstreet 1483a1f0358bSKent Overstreet bch_btree_insert_node(b, op, &keys, 1484a1f0358bSKent Overstreet NULL, NULL); 1485a1f0358bSKent Overstreet BUG_ON(!bch_keylist_empty(&keys)); 1486a1f0358bSKent Overstreet 1487a1f0358bSKent Overstreet rw_unlock(true, last->b); 1488a1f0358bSKent Overstreet last->b = n; 1489a1f0358bSKent Overstreet 1490a1f0358bSKent Overstreet /* Invalidated our iterator */ 1491a1f0358bSKent Overstreet ret = -EINTR; 1492a1f0358bSKent Overstreet break; 1493a1f0358bSKent Overstreet } 1494a1f0358bSKent Overstreet } 1495a1f0358bSKent Overstreet 1496a1f0358bSKent Overstreet if (last->b->level) { 1497a1f0358bSKent Overstreet ret = btree_gc_recurse(last->b, op, writes, gc); 1498a1f0358bSKent Overstreet if (ret) 1499a1f0358bSKent Overstreet break; 1500a1f0358bSKent Overstreet } 1501a1f0358bSKent Overstreet 1502a1f0358bSKent Overstreet bkey_copy_key(&b->c->gc_done, &last->b->key); 1503a1f0358bSKent Overstreet 1504a1f0358bSKent Overstreet /* 1505a1f0358bSKent Overstreet * Must flush leaf nodes before gc ends, since replace 1506a1f0358bSKent Overstreet * operations aren't journalled 1507cafe5635SKent Overstreet */ 1508a1f0358bSKent Overstreet if (btree_node_dirty(last->b)) 1509a1f0358bSKent Overstreet bch_btree_node_write(last->b, writes); 1510a1f0358bSKent Overstreet rw_unlock(true, last->b); 1511a1f0358bSKent Overstreet } 1512a1f0358bSKent Overstreet 1513a1f0358bSKent Overstreet memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1)); 1514a1f0358bSKent Overstreet r->b = NULL; 1515a1f0358bSKent Overstreet 1516cafe5635SKent Overstreet if (need_resched()) { 1517cafe5635SKent Overstreet ret = -EAGAIN; 1518cafe5635SKent Overstreet break; 1519cafe5635SKent Overstreet } 1520cafe5635SKent Overstreet } 1521cafe5635SKent Overstreet 1522a1f0358bSKent Overstreet for (i = 0; i < GC_MERGE_NODES; i++) 1523a1f0358bSKent Overstreet if (!IS_ERR_OR_NULL(r[i].b)) { 1524a1f0358bSKent Overstreet if (btree_node_dirty(r[i].b)) 1525a1f0358bSKent Overstreet bch_btree_node_write(r[i].b, writes); 1526a1f0358bSKent Overstreet rw_unlock(true, r[i].b); 1527a1f0358bSKent Overstreet } 1528cafe5635SKent Overstreet 1529a1f0358bSKent Overstreet bch_keylist_free(&keys); 1530cafe5635SKent Overstreet 1531cafe5635SKent Overstreet return ret; 1532cafe5635SKent Overstreet } 1533cafe5635SKent Overstreet 1534cafe5635SKent Overstreet static int bch_btree_gc_root(struct btree *b, struct btree_op *op, 1535cafe5635SKent Overstreet struct closure *writes, struct gc_stat *gc) 1536cafe5635SKent Overstreet { 1537cafe5635SKent Overstreet struct btree *n = NULL; 1538a1f0358bSKent Overstreet int ret = 0; 1539a1f0358bSKent Overstreet bool should_rewrite; 1540cafe5635SKent Overstreet 1541a1f0358bSKent Overstreet should_rewrite = btree_gc_mark_node(b, gc); 1542a1f0358bSKent Overstreet if (should_rewrite) { 1543bc9389eeSKent Overstreet n = btree_node_alloc_replacement(b, false); 1544cafe5635SKent Overstreet 1545cafe5635SKent Overstreet if (!IS_ERR_OR_NULL(n)) { 1546a1f0358bSKent Overstreet bch_btree_node_write_sync(n); 1547a1f0358bSKent Overstreet bch_btree_set_root(n); 1548a1f0358bSKent Overstreet btree_node_free(b); 1549a1f0358bSKent Overstreet rw_unlock(true, n); 1550a1f0358bSKent Overstreet 1551a1f0358bSKent Overstreet return -EINTR; 1552cafe5635SKent Overstreet } 1553a1f0358bSKent Overstreet } 1554a1f0358bSKent Overstreet 1555a1f0358bSKent Overstreet if (b->level) { 1556a1f0358bSKent Overstreet ret = btree_gc_recurse(b, op, writes, gc); 1557a1f0358bSKent Overstreet if (ret) 1558a1f0358bSKent Overstreet return ret; 1559a1f0358bSKent Overstreet } 1560a1f0358bSKent Overstreet 1561a1f0358bSKent Overstreet bkey_copy_key(&b->c->gc_done, &b->key); 1562cafe5635SKent Overstreet 1563cafe5635SKent Overstreet return ret; 1564cafe5635SKent Overstreet } 1565cafe5635SKent Overstreet 1566cafe5635SKent Overstreet static void btree_gc_start(struct cache_set *c) 1567cafe5635SKent Overstreet { 1568cafe5635SKent Overstreet struct cache *ca; 1569cafe5635SKent Overstreet struct bucket *b; 1570cafe5635SKent Overstreet unsigned i; 1571cafe5635SKent Overstreet 1572cafe5635SKent Overstreet if (!c->gc_mark_valid) 1573cafe5635SKent Overstreet return; 1574cafe5635SKent Overstreet 1575cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1576cafe5635SKent Overstreet 1577cafe5635SKent Overstreet c->gc_mark_valid = 0; 1578cafe5635SKent Overstreet c->gc_done = ZERO_KEY; 1579cafe5635SKent Overstreet 1580cafe5635SKent Overstreet for_each_cache(ca, c, i) 1581cafe5635SKent Overstreet for_each_bucket(b, ca) { 1582cafe5635SKent Overstreet b->gc_gen = b->gen; 158329ebf465SKent Overstreet if (!atomic_read(&b->pin)) { 1584cafe5635SKent Overstreet SET_GC_MARK(b, GC_MARK_RECLAIMABLE); 158529ebf465SKent Overstreet SET_GC_SECTORS_USED(b, 0); 158629ebf465SKent Overstreet } 1587cafe5635SKent Overstreet } 1588cafe5635SKent Overstreet 1589cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1590cafe5635SKent Overstreet } 1591cafe5635SKent Overstreet 1592cafe5635SKent Overstreet size_t bch_btree_gc_finish(struct cache_set *c) 1593cafe5635SKent Overstreet { 1594cafe5635SKent Overstreet size_t available = 0; 1595cafe5635SKent Overstreet struct bucket *b; 1596cafe5635SKent Overstreet struct cache *ca; 1597cafe5635SKent Overstreet unsigned i; 1598cafe5635SKent Overstreet 1599cafe5635SKent Overstreet mutex_lock(&c->bucket_lock); 1600cafe5635SKent Overstreet 1601cafe5635SKent Overstreet set_gc_sectors(c); 1602cafe5635SKent Overstreet c->gc_mark_valid = 1; 1603cafe5635SKent Overstreet c->need_gc = 0; 1604cafe5635SKent Overstreet 1605cafe5635SKent Overstreet if (c->root) 1606cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&c->root->key); i++) 1607cafe5635SKent Overstreet SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), 1608cafe5635SKent Overstreet GC_MARK_METADATA); 1609cafe5635SKent Overstreet 1610cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) 1611cafe5635SKent Overstreet SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), 1612cafe5635SKent Overstreet GC_MARK_METADATA); 1613cafe5635SKent Overstreet 1614bf0a628aSNicholas Swenson /* don't reclaim buckets to which writeback keys point */ 1615bf0a628aSNicholas Swenson rcu_read_lock(); 1616bf0a628aSNicholas Swenson for (i = 0; i < c->nr_uuids; i++) { 1617bf0a628aSNicholas Swenson struct bcache_device *d = c->devices[i]; 1618bf0a628aSNicholas Swenson struct cached_dev *dc; 1619bf0a628aSNicholas Swenson struct keybuf_key *w, *n; 1620bf0a628aSNicholas Swenson unsigned j; 1621bf0a628aSNicholas Swenson 1622bf0a628aSNicholas Swenson if (!d || UUID_FLASH_ONLY(&c->uuids[i])) 1623bf0a628aSNicholas Swenson continue; 1624bf0a628aSNicholas Swenson dc = container_of(d, struct cached_dev, disk); 1625bf0a628aSNicholas Swenson 1626bf0a628aSNicholas Swenson spin_lock(&dc->writeback_keys.lock); 1627bf0a628aSNicholas Swenson rbtree_postorder_for_each_entry_safe(w, n, 1628bf0a628aSNicholas Swenson &dc->writeback_keys.keys, node) 1629bf0a628aSNicholas Swenson for (j = 0; j < KEY_PTRS(&w->key); j++) 1630bf0a628aSNicholas Swenson SET_GC_MARK(PTR_BUCKET(c, &w->key, j), 1631bf0a628aSNicholas Swenson GC_MARK_DIRTY); 1632bf0a628aSNicholas Swenson spin_unlock(&dc->writeback_keys.lock); 1633bf0a628aSNicholas Swenson } 1634bf0a628aSNicholas Swenson rcu_read_unlock(); 1635bf0a628aSNicholas Swenson 1636cafe5635SKent Overstreet for_each_cache(ca, c, i) { 1637cafe5635SKent Overstreet uint64_t *i; 1638cafe5635SKent Overstreet 1639cafe5635SKent Overstreet ca->invalidate_needs_gc = 0; 1640cafe5635SKent Overstreet 1641cafe5635SKent Overstreet for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) 1642cafe5635SKent Overstreet SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1643cafe5635SKent Overstreet 1644cafe5635SKent Overstreet for (i = ca->prio_buckets; 1645cafe5635SKent Overstreet i < ca->prio_buckets + prio_buckets(ca) * 2; i++) 1646cafe5635SKent Overstreet SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); 1647cafe5635SKent Overstreet 1648cafe5635SKent Overstreet for_each_bucket(b, ca) { 1649cafe5635SKent Overstreet b->last_gc = b->gc_gen; 1650cafe5635SKent Overstreet c->need_gc = max(c->need_gc, bucket_gc_gen(b)); 1651cafe5635SKent Overstreet 1652cafe5635SKent Overstreet if (!atomic_read(&b->pin) && 1653cafe5635SKent Overstreet GC_MARK(b) == GC_MARK_RECLAIMABLE) { 1654cafe5635SKent Overstreet available++; 1655cafe5635SKent Overstreet if (!GC_SECTORS_USED(b)) 1656cafe5635SKent Overstreet bch_bucket_add_unused(ca, b); 1657cafe5635SKent Overstreet } 1658cafe5635SKent Overstreet } 1659cafe5635SKent Overstreet } 1660cafe5635SKent Overstreet 1661cafe5635SKent Overstreet mutex_unlock(&c->bucket_lock); 1662cafe5635SKent Overstreet return available; 1663cafe5635SKent Overstreet } 1664cafe5635SKent Overstreet 166572a44517SKent Overstreet static void bch_btree_gc(struct cache_set *c) 1666cafe5635SKent Overstreet { 1667cafe5635SKent Overstreet int ret; 1668cafe5635SKent Overstreet unsigned long available; 1669cafe5635SKent Overstreet struct gc_stat stats; 1670cafe5635SKent Overstreet struct closure writes; 1671cafe5635SKent Overstreet struct btree_op op; 1672cafe5635SKent Overstreet uint64_t start_time = local_clock(); 167357943511SKent Overstreet 1674c37511b8SKent Overstreet trace_bcache_gc_start(c); 1675cafe5635SKent Overstreet 1676cafe5635SKent Overstreet memset(&stats, 0, sizeof(struct gc_stat)); 1677cafe5635SKent Overstreet closure_init_stack(&writes); 1678b54d6934SKent Overstreet bch_btree_op_init(&op, SHRT_MAX); 1679cafe5635SKent Overstreet 1680cafe5635SKent Overstreet btree_gc_start(c); 1681cafe5635SKent Overstreet 1682a1f0358bSKent Overstreet do { 1683cafe5635SKent Overstreet ret = btree_root(gc_root, c, &op, &writes, &stats); 1684cafe5635SKent Overstreet closure_sync(&writes); 1685cafe5635SKent Overstreet 1686a1f0358bSKent Overstreet if (ret && ret != -EAGAIN) 1687cafe5635SKent Overstreet pr_warn("gc failed!"); 1688a1f0358bSKent Overstreet } while (ret); 1689cafe5635SKent Overstreet 1690cafe5635SKent Overstreet available = bch_btree_gc_finish(c); 169157943511SKent Overstreet wake_up_allocators(c); 169257943511SKent Overstreet 1693169ef1cfSKent Overstreet bch_time_stats_update(&c->btree_gc_time, start_time); 1694cafe5635SKent Overstreet 1695cafe5635SKent Overstreet stats.key_bytes *= sizeof(uint64_t); 1696cafe5635SKent Overstreet stats.data <<= 9; 1697cafe5635SKent Overstreet stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; 1698cafe5635SKent Overstreet memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); 1699cafe5635SKent Overstreet 1700c37511b8SKent Overstreet trace_bcache_gc_end(c); 1701cafe5635SKent Overstreet 170272a44517SKent Overstreet bch_moving_gc(c); 1703cafe5635SKent Overstreet } 1704cafe5635SKent Overstreet 170572a44517SKent Overstreet static int bch_gc_thread(void *arg) 1706cafe5635SKent Overstreet { 170772a44517SKent Overstreet struct cache_set *c = arg; 1708a1f0358bSKent Overstreet struct cache *ca; 1709a1f0358bSKent Overstreet unsigned i; 171072a44517SKent Overstreet 171172a44517SKent Overstreet while (1) { 1712a1f0358bSKent Overstreet again: 171372a44517SKent Overstreet bch_btree_gc(c); 171472a44517SKent Overstreet 171572a44517SKent Overstreet set_current_state(TASK_INTERRUPTIBLE); 171672a44517SKent Overstreet if (kthread_should_stop()) 171772a44517SKent Overstreet break; 171872a44517SKent Overstreet 1719a1f0358bSKent Overstreet mutex_lock(&c->bucket_lock); 1720a1f0358bSKent Overstreet 1721a1f0358bSKent Overstreet for_each_cache(ca, c, i) 1722a1f0358bSKent Overstreet if (ca->invalidate_needs_gc) { 1723a1f0358bSKent Overstreet mutex_unlock(&c->bucket_lock); 1724a1f0358bSKent Overstreet set_current_state(TASK_RUNNING); 1725a1f0358bSKent Overstreet goto again; 1726a1f0358bSKent Overstreet } 1727a1f0358bSKent Overstreet 1728a1f0358bSKent Overstreet mutex_unlock(&c->bucket_lock); 1729a1f0358bSKent Overstreet 173072a44517SKent Overstreet try_to_freeze(); 173172a44517SKent Overstreet schedule(); 173272a44517SKent Overstreet } 173372a44517SKent Overstreet 173472a44517SKent Overstreet return 0; 173572a44517SKent Overstreet } 173672a44517SKent Overstreet 173772a44517SKent Overstreet int bch_gc_thread_start(struct cache_set *c) 173872a44517SKent Overstreet { 173972a44517SKent Overstreet c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc"); 174072a44517SKent Overstreet if (IS_ERR(c->gc_thread)) 174172a44517SKent Overstreet return PTR_ERR(c->gc_thread); 174272a44517SKent Overstreet 174372a44517SKent Overstreet set_task_state(c->gc_thread, TASK_INTERRUPTIBLE); 174472a44517SKent Overstreet return 0; 1745cafe5635SKent Overstreet } 1746cafe5635SKent Overstreet 1747cafe5635SKent Overstreet /* Initial partial gc */ 1748cafe5635SKent Overstreet 1749cafe5635SKent Overstreet static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, 1750cafe5635SKent Overstreet unsigned long **seen) 1751cafe5635SKent Overstreet { 175250310164SKent Overstreet int ret = 0; 1753cafe5635SKent Overstreet unsigned i; 175450310164SKent Overstreet struct bkey *k, *p = NULL; 1755cafe5635SKent Overstreet struct bucket *g; 1756cafe5635SKent Overstreet struct btree_iter iter; 1757cafe5635SKent Overstreet 1758cafe5635SKent Overstreet for_each_key_filter(b, k, &iter, bch_ptr_invalid) { 1759cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(k); i++) { 1760cafe5635SKent Overstreet if (!ptr_available(b->c, k, i)) 1761cafe5635SKent Overstreet continue; 1762cafe5635SKent Overstreet 1763cafe5635SKent Overstreet g = PTR_BUCKET(b->c, k, i); 1764cafe5635SKent Overstreet 1765cafe5635SKent Overstreet if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), 1766cafe5635SKent Overstreet seen[PTR_DEV(k, i)]) || 1767cafe5635SKent Overstreet !ptr_stale(b->c, k, i)) { 1768cafe5635SKent Overstreet g->gen = PTR_GEN(k, i); 1769cafe5635SKent Overstreet 1770cafe5635SKent Overstreet if (b->level) 1771cafe5635SKent Overstreet g->prio = BTREE_PRIO; 1772cafe5635SKent Overstreet else if (g->prio == BTREE_PRIO) 1773cafe5635SKent Overstreet g->prio = INITIAL_PRIO; 1774cafe5635SKent Overstreet } 1775cafe5635SKent Overstreet } 1776cafe5635SKent Overstreet 1777cafe5635SKent Overstreet btree_mark_key(b, k); 1778cafe5635SKent Overstreet } 1779cafe5635SKent Overstreet 1780cafe5635SKent Overstreet if (b->level) { 178150310164SKent Overstreet bch_btree_iter_init(b, &iter, NULL); 1782cafe5635SKent Overstreet 178350310164SKent Overstreet do { 178450310164SKent Overstreet k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); 178550310164SKent Overstreet if (k) 178650310164SKent Overstreet btree_node_prefetch(b->c, k, b->level - 1); 178750310164SKent Overstreet 1788cafe5635SKent Overstreet if (p) 178950310164SKent Overstreet ret = btree(check_recurse, p, b, op, seen); 1790cafe5635SKent Overstreet 179150310164SKent Overstreet p = k; 179250310164SKent Overstreet } while (p && !ret); 1793cafe5635SKent Overstreet } 1794cafe5635SKent Overstreet 1795cafe5635SKent Overstreet return 0; 1796cafe5635SKent Overstreet } 1797cafe5635SKent Overstreet 1798c18536a7SKent Overstreet int bch_btree_check(struct cache_set *c) 1799cafe5635SKent Overstreet { 1800cafe5635SKent Overstreet int ret = -ENOMEM; 1801cafe5635SKent Overstreet unsigned i; 1802cafe5635SKent Overstreet unsigned long *seen[MAX_CACHES_PER_SET]; 1803c18536a7SKent Overstreet struct btree_op op; 1804cafe5635SKent Overstreet 1805cafe5635SKent Overstreet memset(seen, 0, sizeof(seen)); 1806b54d6934SKent Overstreet bch_btree_op_init(&op, SHRT_MAX); 1807cafe5635SKent Overstreet 1808cafe5635SKent Overstreet for (i = 0; c->cache[i]; i++) { 1809cafe5635SKent Overstreet size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); 1810cafe5635SKent Overstreet seen[i] = kmalloc(n, GFP_KERNEL); 1811cafe5635SKent Overstreet if (!seen[i]) 1812cafe5635SKent Overstreet goto err; 1813cafe5635SKent Overstreet 1814cafe5635SKent Overstreet /* Disables the seen array until prio_read() uses it too */ 1815cafe5635SKent Overstreet memset(seen[i], 0xFF, n); 1816cafe5635SKent Overstreet } 1817cafe5635SKent Overstreet 1818c18536a7SKent Overstreet ret = btree_root(check_recurse, c, &op, seen); 1819cafe5635SKent Overstreet err: 1820cafe5635SKent Overstreet for (i = 0; i < MAX_CACHES_PER_SET; i++) 1821cafe5635SKent Overstreet kfree(seen[i]); 1822cafe5635SKent Overstreet return ret; 1823cafe5635SKent Overstreet } 1824cafe5635SKent Overstreet 1825cafe5635SKent Overstreet /* Btree insertion */ 1826cafe5635SKent Overstreet 1827cafe5635SKent Overstreet static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) 1828cafe5635SKent Overstreet { 1829cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 1830cafe5635SKent Overstreet 1831cafe5635SKent Overstreet memmove((uint64_t *) where + bkey_u64s(insert), 1832cafe5635SKent Overstreet where, 1833cafe5635SKent Overstreet (void *) end(i) - (void *) where); 1834cafe5635SKent Overstreet 1835cafe5635SKent Overstreet i->keys += bkey_u64s(insert); 1836cafe5635SKent Overstreet bkey_copy(where, insert); 1837cafe5635SKent Overstreet bch_bset_fix_lookup_table(b, where); 1838cafe5635SKent Overstreet } 1839cafe5635SKent Overstreet 18401b207d80SKent Overstreet static bool fix_overlapping_extents(struct btree *b, struct bkey *insert, 1841cafe5635SKent Overstreet struct btree_iter *iter, 18421b207d80SKent Overstreet struct bkey *replace_key) 1843cafe5635SKent Overstreet { 1844279afbadSKent Overstreet void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) 1845cafe5635SKent Overstreet { 1846279afbadSKent Overstreet if (KEY_DIRTY(k)) 1847279afbadSKent Overstreet bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 1848279afbadSKent Overstreet offset, -sectors); 1849cafe5635SKent Overstreet } 1850cafe5635SKent Overstreet 1851279afbadSKent Overstreet uint64_t old_offset; 1852cafe5635SKent Overstreet unsigned old_size, sectors_found = 0; 1853cafe5635SKent Overstreet 1854cafe5635SKent Overstreet while (1) { 1855cafe5635SKent Overstreet struct bkey *k = bch_btree_iter_next(iter); 1856cafe5635SKent Overstreet if (!k || 1857cafe5635SKent Overstreet bkey_cmp(&START_KEY(k), insert) >= 0) 1858cafe5635SKent Overstreet break; 1859cafe5635SKent Overstreet 1860cafe5635SKent Overstreet if (bkey_cmp(k, &START_KEY(insert)) <= 0) 1861cafe5635SKent Overstreet continue; 1862cafe5635SKent Overstreet 1863279afbadSKent Overstreet old_offset = KEY_START(k); 1864cafe5635SKent Overstreet old_size = KEY_SIZE(k); 1865cafe5635SKent Overstreet 1866cafe5635SKent Overstreet /* 1867cafe5635SKent Overstreet * We might overlap with 0 size extents; we can't skip these 1868cafe5635SKent Overstreet * because if they're in the set we're inserting to we have to 1869cafe5635SKent Overstreet * adjust them so they don't overlap with the key we're 18701b207d80SKent Overstreet * inserting. But we don't want to check them for replace 1871cafe5635SKent Overstreet * operations. 1872cafe5635SKent Overstreet */ 1873cafe5635SKent Overstreet 18741b207d80SKent Overstreet if (replace_key && KEY_SIZE(k)) { 1875cafe5635SKent Overstreet /* 1876cafe5635SKent Overstreet * k might have been split since we inserted/found the 1877cafe5635SKent Overstreet * key we're replacing 1878cafe5635SKent Overstreet */ 1879cafe5635SKent Overstreet unsigned i; 1880cafe5635SKent Overstreet uint64_t offset = KEY_START(k) - 18811b207d80SKent Overstreet KEY_START(replace_key); 1882cafe5635SKent Overstreet 1883cafe5635SKent Overstreet /* But it must be a subset of the replace key */ 18841b207d80SKent Overstreet if (KEY_START(k) < KEY_START(replace_key) || 18851b207d80SKent Overstreet KEY_OFFSET(k) > KEY_OFFSET(replace_key)) 1886cafe5635SKent Overstreet goto check_failed; 1887cafe5635SKent Overstreet 1888cafe5635SKent Overstreet /* We didn't find a key that we were supposed to */ 1889cafe5635SKent Overstreet if (KEY_START(k) > KEY_START(insert) + sectors_found) 1890cafe5635SKent Overstreet goto check_failed; 1891cafe5635SKent Overstreet 1892d24a6e10SKent Overstreet if (KEY_PTRS(k) != KEY_PTRS(replace_key) || 1893d24a6e10SKent Overstreet KEY_DIRTY(k) != KEY_DIRTY(replace_key)) 1894cafe5635SKent Overstreet goto check_failed; 1895cafe5635SKent Overstreet 1896cafe5635SKent Overstreet /* skip past gen */ 1897cafe5635SKent Overstreet offset <<= 8; 1898cafe5635SKent Overstreet 18991b207d80SKent Overstreet BUG_ON(!KEY_PTRS(replace_key)); 1900cafe5635SKent Overstreet 19011b207d80SKent Overstreet for (i = 0; i < KEY_PTRS(replace_key); i++) 19021b207d80SKent Overstreet if (k->ptr[i] != replace_key->ptr[i] + offset) 1903cafe5635SKent Overstreet goto check_failed; 1904cafe5635SKent Overstreet 1905cafe5635SKent Overstreet sectors_found = KEY_OFFSET(k) - KEY_START(insert); 1906cafe5635SKent Overstreet } 1907cafe5635SKent Overstreet 1908cafe5635SKent Overstreet if (bkey_cmp(insert, k) < 0 && 1909cafe5635SKent Overstreet bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { 1910cafe5635SKent Overstreet /* 1911cafe5635SKent Overstreet * We overlapped in the middle of an existing key: that 1912cafe5635SKent Overstreet * means we have to split the old key. But we have to do 1913cafe5635SKent Overstreet * slightly different things depending on whether the 1914cafe5635SKent Overstreet * old key has been written out yet. 1915cafe5635SKent Overstreet */ 1916cafe5635SKent Overstreet 1917cafe5635SKent Overstreet struct bkey *top; 1918cafe5635SKent Overstreet 1919279afbadSKent Overstreet subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); 1920cafe5635SKent Overstreet 1921cafe5635SKent Overstreet if (bkey_written(b, k)) { 1922cafe5635SKent Overstreet /* 1923cafe5635SKent Overstreet * We insert a new key to cover the top of the 1924cafe5635SKent Overstreet * old key, and the old key is modified in place 1925cafe5635SKent Overstreet * to represent the bottom split. 1926cafe5635SKent Overstreet * 1927cafe5635SKent Overstreet * It's completely arbitrary whether the new key 1928cafe5635SKent Overstreet * is the top or the bottom, but it has to match 1929cafe5635SKent Overstreet * up with what btree_sort_fixup() does - it 1930cafe5635SKent Overstreet * doesn't check for this kind of overlap, it 1931cafe5635SKent Overstreet * depends on us inserting a new key for the top 1932cafe5635SKent Overstreet * here. 1933cafe5635SKent Overstreet */ 1934cafe5635SKent Overstreet top = bch_bset_search(b, &b->sets[b->nsets], 1935cafe5635SKent Overstreet insert); 1936cafe5635SKent Overstreet shift_keys(b, top, k); 1937cafe5635SKent Overstreet } else { 1938cafe5635SKent Overstreet BKEY_PADDED(key) temp; 1939cafe5635SKent Overstreet bkey_copy(&temp.key, k); 1940cafe5635SKent Overstreet shift_keys(b, k, &temp.key); 1941cafe5635SKent Overstreet top = bkey_next(k); 1942cafe5635SKent Overstreet } 1943cafe5635SKent Overstreet 1944cafe5635SKent Overstreet bch_cut_front(insert, top); 1945cafe5635SKent Overstreet bch_cut_back(&START_KEY(insert), k); 1946cafe5635SKent Overstreet bch_bset_fix_invalidated_key(b, k); 1947cafe5635SKent Overstreet return false; 1948cafe5635SKent Overstreet } 1949cafe5635SKent Overstreet 1950cafe5635SKent Overstreet if (bkey_cmp(insert, k) < 0) { 1951cafe5635SKent Overstreet bch_cut_front(insert, k); 1952cafe5635SKent Overstreet } else { 19531fa8455dSKent Overstreet if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) 19541fa8455dSKent Overstreet old_offset = KEY_START(insert); 19551fa8455dSKent Overstreet 1956cafe5635SKent Overstreet if (bkey_written(b, k) && 1957cafe5635SKent Overstreet bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { 1958cafe5635SKent Overstreet /* 1959cafe5635SKent Overstreet * Completely overwrote, so we don't have to 1960cafe5635SKent Overstreet * invalidate the binary search tree 1961cafe5635SKent Overstreet */ 1962cafe5635SKent Overstreet bch_cut_front(k, k); 1963cafe5635SKent Overstreet } else { 1964cafe5635SKent Overstreet __bch_cut_back(&START_KEY(insert), k); 1965cafe5635SKent Overstreet bch_bset_fix_invalidated_key(b, k); 1966cafe5635SKent Overstreet } 1967cafe5635SKent Overstreet } 1968cafe5635SKent Overstreet 1969279afbadSKent Overstreet subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); 1970cafe5635SKent Overstreet } 1971cafe5635SKent Overstreet 1972cafe5635SKent Overstreet check_failed: 19731b207d80SKent Overstreet if (replace_key) { 1974cafe5635SKent Overstreet if (!sectors_found) { 1975cafe5635SKent Overstreet return true; 1976cafe5635SKent Overstreet } else if (sectors_found < KEY_SIZE(insert)) { 1977cafe5635SKent Overstreet SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - 1978cafe5635SKent Overstreet (KEY_SIZE(insert) - sectors_found)); 1979cafe5635SKent Overstreet SET_KEY_SIZE(insert, sectors_found); 1980cafe5635SKent Overstreet } 1981cafe5635SKent Overstreet } 1982cafe5635SKent Overstreet 1983cafe5635SKent Overstreet return false; 1984cafe5635SKent Overstreet } 1985cafe5635SKent Overstreet 1986cafe5635SKent Overstreet static bool btree_insert_key(struct btree *b, struct btree_op *op, 19871b207d80SKent Overstreet struct bkey *k, struct bkey *replace_key) 1988cafe5635SKent Overstreet { 1989cafe5635SKent Overstreet struct bset *i = b->sets[b->nsets].data; 1990cafe5635SKent Overstreet struct bkey *m, *prev; 199185b1492eSKent Overstreet unsigned status = BTREE_INSERT_STATUS_INSERT; 1992cafe5635SKent Overstreet 1993cafe5635SKent Overstreet BUG_ON(bkey_cmp(k, &b->key) > 0); 1994cafe5635SKent Overstreet BUG_ON(b->level && !KEY_PTRS(k)); 1995cafe5635SKent Overstreet BUG_ON(!b->level && !KEY_OFFSET(k)); 1996cafe5635SKent Overstreet 1997cafe5635SKent Overstreet if (!b->level) { 1998cafe5635SKent Overstreet struct btree_iter iter; 1999cafe5635SKent Overstreet 2000cafe5635SKent Overstreet /* 2001cafe5635SKent Overstreet * bset_search() returns the first key that is strictly greater 2002cafe5635SKent Overstreet * than the search key - but for back merging, we want to find 20030eacac22SKent Overstreet * the previous key. 2004cafe5635SKent Overstreet */ 2005cafe5635SKent Overstreet prev = NULL; 20060eacac22SKent Overstreet m = bch_btree_iter_init(b, &iter, PRECEDING_KEY(&START_KEY(k))); 2007cafe5635SKent Overstreet 20081b207d80SKent Overstreet if (fix_overlapping_extents(b, k, &iter, replace_key)) { 20091b207d80SKent Overstreet op->insert_collision = true; 2010cafe5635SKent Overstreet return false; 20111b207d80SKent Overstreet } 2012cafe5635SKent Overstreet 20131fa8455dSKent Overstreet if (KEY_DIRTY(k)) 20141fa8455dSKent Overstreet bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), 20151fa8455dSKent Overstreet KEY_START(k), KEY_SIZE(k)); 20161fa8455dSKent Overstreet 2017cafe5635SKent Overstreet while (m != end(i) && 2018cafe5635SKent Overstreet bkey_cmp(k, &START_KEY(m)) > 0) 2019cafe5635SKent Overstreet prev = m, m = bkey_next(m); 2020cafe5635SKent Overstreet 2021cafe5635SKent Overstreet if (key_merging_disabled(b->c)) 2022cafe5635SKent Overstreet goto insert; 2023cafe5635SKent Overstreet 2024cafe5635SKent Overstreet /* prev is in the tree, if we merge we're done */ 202585b1492eSKent Overstreet status = BTREE_INSERT_STATUS_BACK_MERGE; 2026cafe5635SKent Overstreet if (prev && 2027cafe5635SKent Overstreet bch_bkey_try_merge(b, prev, k)) 2028cafe5635SKent Overstreet goto merged; 2029cafe5635SKent Overstreet 203085b1492eSKent Overstreet status = BTREE_INSERT_STATUS_OVERWROTE; 2031cafe5635SKent Overstreet if (m != end(i) && 2032cafe5635SKent Overstreet KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) 2033cafe5635SKent Overstreet goto copy; 2034cafe5635SKent Overstreet 203585b1492eSKent Overstreet status = BTREE_INSERT_STATUS_FRONT_MERGE; 2036cafe5635SKent Overstreet if (m != end(i) && 2037cafe5635SKent Overstreet bch_bkey_try_merge(b, k, m)) 2038cafe5635SKent Overstreet goto copy; 20391b207d80SKent Overstreet } else { 20401b207d80SKent Overstreet BUG_ON(replace_key); 2041cafe5635SKent Overstreet m = bch_bset_search(b, &b->sets[b->nsets], k); 20421b207d80SKent Overstreet } 2043cafe5635SKent Overstreet 2044cafe5635SKent Overstreet insert: shift_keys(b, m, k); 2045cafe5635SKent Overstreet copy: bkey_copy(m, k); 2046cafe5635SKent Overstreet merged: 20471b207d80SKent Overstreet bch_check_keys(b, "%u for %s", status, 20481b207d80SKent Overstreet replace_key ? "replace" : "insert"); 2049cafe5635SKent Overstreet 2050cafe5635SKent Overstreet if (b->level && !KEY_OFFSET(k)) 205157943511SKent Overstreet btree_current_write(b)->prio_blocked++; 2052cafe5635SKent Overstreet 20531b207d80SKent Overstreet trace_bcache_btree_insert_key(b, k, replace_key != NULL, status); 2054cafe5635SKent Overstreet 2055cafe5635SKent Overstreet return true; 2056cafe5635SKent Overstreet } 2057cafe5635SKent Overstreet 205826c949f8SKent Overstreet static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, 20591b207d80SKent Overstreet struct keylist *insert_keys, 20601b207d80SKent Overstreet struct bkey *replace_key) 2061cafe5635SKent Overstreet { 2062cafe5635SKent Overstreet bool ret = false; 2063280481d0SKent Overstreet int oldsize = bch_count_data(b); 2064cafe5635SKent Overstreet 206526c949f8SKent Overstreet while (!bch_keylist_empty(insert_keys)) { 2066403b6cdeSKent Overstreet struct bset *i = write_block(b); 2067c2f95ae2SKent Overstreet struct bkey *k = insert_keys->keys; 206826c949f8SKent Overstreet 2069403b6cdeSKent Overstreet if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c) 2070403b6cdeSKent Overstreet > btree_blocks(b)) 2071403b6cdeSKent Overstreet break; 2072403b6cdeSKent Overstreet 2073403b6cdeSKent Overstreet if (bkey_cmp(k, &b->key) <= 0) { 20743a3b6a4eSKent Overstreet if (!b->level) 20753a3b6a4eSKent Overstreet bkey_put(b->c, k); 207626c949f8SKent Overstreet 20771b207d80SKent Overstreet ret |= btree_insert_key(b, op, k, replace_key); 207826c949f8SKent Overstreet bch_keylist_pop_front(insert_keys); 207926c949f8SKent Overstreet } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { 208026c949f8SKent Overstreet BKEY_PADDED(key) temp; 2081c2f95ae2SKent Overstreet bkey_copy(&temp.key, insert_keys->keys); 208226c949f8SKent Overstreet 208326c949f8SKent Overstreet bch_cut_back(&b->key, &temp.key); 2084c2f95ae2SKent Overstreet bch_cut_front(&b->key, insert_keys->keys); 208526c949f8SKent Overstreet 20861b207d80SKent Overstreet ret |= btree_insert_key(b, op, &temp.key, replace_key); 208726c949f8SKent Overstreet break; 208826c949f8SKent Overstreet } else { 208926c949f8SKent Overstreet break; 209026c949f8SKent Overstreet } 2091cafe5635SKent Overstreet } 2092cafe5635SKent Overstreet 2093403b6cdeSKent Overstreet BUG_ON(!bch_keylist_empty(insert_keys) && b->level); 2094403b6cdeSKent Overstreet 2095cafe5635SKent Overstreet BUG_ON(bch_count_data(b) < oldsize); 2096cafe5635SKent Overstreet return ret; 2097cafe5635SKent Overstreet } 2098cafe5635SKent Overstreet 209926c949f8SKent Overstreet static int btree_split(struct btree *b, struct btree_op *op, 210026c949f8SKent Overstreet struct keylist *insert_keys, 21011b207d80SKent Overstreet struct bkey *replace_key) 2102cafe5635SKent Overstreet { 2103d6fd3b11SKent Overstreet bool split; 2104cafe5635SKent Overstreet struct btree *n1, *n2 = NULL, *n3 = NULL; 2105cafe5635SKent Overstreet uint64_t start_time = local_clock(); 2106b54d6934SKent Overstreet struct closure cl; 210717e21a9fSKent Overstreet struct keylist parent_keys; 2108b54d6934SKent Overstreet 2109b54d6934SKent Overstreet closure_init_stack(&cl); 211017e21a9fSKent Overstreet bch_keylist_init(&parent_keys); 2111cafe5635SKent Overstreet 211278365411SKent Overstreet if (!b->level && 211378365411SKent Overstreet btree_check_reserve(b, op)) 211478365411SKent Overstreet return -EINTR; 211578365411SKent Overstreet 2116bc9389eeSKent Overstreet n1 = btree_node_alloc_replacement(b, true); 2117cafe5635SKent Overstreet if (IS_ERR(n1)) 2118cafe5635SKent Overstreet goto err; 2119cafe5635SKent Overstreet 2120cafe5635SKent Overstreet split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; 2121cafe5635SKent Overstreet 2122cafe5635SKent Overstreet if (split) { 2123cafe5635SKent Overstreet unsigned keys = 0; 2124cafe5635SKent Overstreet 2125c37511b8SKent Overstreet trace_bcache_btree_node_split(b, n1->sets[0].data->keys); 2126c37511b8SKent Overstreet 2127bc9389eeSKent Overstreet n2 = bch_btree_node_alloc(b->c, b->level, true); 2128cafe5635SKent Overstreet if (IS_ERR(n2)) 2129cafe5635SKent Overstreet goto err_free1; 2130cafe5635SKent Overstreet 2131d6fd3b11SKent Overstreet if (!b->parent) { 2132bc9389eeSKent Overstreet n3 = bch_btree_node_alloc(b->c, b->level + 1, true); 2133cafe5635SKent Overstreet if (IS_ERR(n3)) 2134cafe5635SKent Overstreet goto err_free2; 2135cafe5635SKent Overstreet } 2136cafe5635SKent Overstreet 21371b207d80SKent Overstreet bch_btree_insert_keys(n1, op, insert_keys, replace_key); 2138cafe5635SKent Overstreet 2139d6fd3b11SKent Overstreet /* 2140d6fd3b11SKent Overstreet * Has to be a linear search because we don't have an auxiliary 2141cafe5635SKent Overstreet * search tree yet 2142cafe5635SKent Overstreet */ 2143cafe5635SKent Overstreet 2144cafe5635SKent Overstreet while (keys < (n1->sets[0].data->keys * 3) / 5) 2145cafe5635SKent Overstreet keys += bkey_u64s(node(n1->sets[0].data, keys)); 2146cafe5635SKent Overstreet 2147cafe5635SKent Overstreet bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); 2148cafe5635SKent Overstreet keys += bkey_u64s(node(n1->sets[0].data, keys)); 2149cafe5635SKent Overstreet 2150cafe5635SKent Overstreet n2->sets[0].data->keys = n1->sets[0].data->keys - keys; 2151cafe5635SKent Overstreet n1->sets[0].data->keys = keys; 2152cafe5635SKent Overstreet 2153cafe5635SKent Overstreet memcpy(n2->sets[0].data->start, 2154cafe5635SKent Overstreet end(n1->sets[0].data), 2155cafe5635SKent Overstreet n2->sets[0].data->keys * sizeof(uint64_t)); 2156cafe5635SKent Overstreet 2157cafe5635SKent Overstreet bkey_copy_key(&n2->key, &b->key); 2158cafe5635SKent Overstreet 215917e21a9fSKent Overstreet bch_keylist_add(&parent_keys, &n2->key); 2160b54d6934SKent Overstreet bch_btree_node_write(n2, &cl); 2161cafe5635SKent Overstreet rw_unlock(true, n2); 2162c37511b8SKent Overstreet } else { 2163c37511b8SKent Overstreet trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); 2164c37511b8SKent Overstreet 21651b207d80SKent Overstreet bch_btree_insert_keys(n1, op, insert_keys, replace_key); 2166c37511b8SKent Overstreet } 2167cafe5635SKent Overstreet 216817e21a9fSKent Overstreet bch_keylist_add(&parent_keys, &n1->key); 2169b54d6934SKent Overstreet bch_btree_node_write(n1, &cl); 2170cafe5635SKent Overstreet 2171cafe5635SKent Overstreet if (n3) { 2172d6fd3b11SKent Overstreet /* Depth increases, make a new root */ 2173cafe5635SKent Overstreet bkey_copy_key(&n3->key, &MAX_KEY); 217417e21a9fSKent Overstreet bch_btree_insert_keys(n3, op, &parent_keys, NULL); 2175b54d6934SKent Overstreet bch_btree_node_write(n3, &cl); 2176cafe5635SKent Overstreet 2177b54d6934SKent Overstreet closure_sync(&cl); 2178cafe5635SKent Overstreet bch_btree_set_root(n3); 2179cafe5635SKent Overstreet rw_unlock(true, n3); 218017e21a9fSKent Overstreet 218117e21a9fSKent Overstreet btree_node_free(b); 2182d6fd3b11SKent Overstreet } else if (!b->parent) { 2183d6fd3b11SKent Overstreet /* Root filled up but didn't need to be split */ 2184b54d6934SKent Overstreet closure_sync(&cl); 2185cafe5635SKent Overstreet bch_btree_set_root(n1); 218617e21a9fSKent Overstreet 218717e21a9fSKent Overstreet btree_node_free(b); 2188cafe5635SKent Overstreet } else { 218917e21a9fSKent Overstreet /* Split a non root node */ 2190b54d6934SKent Overstreet closure_sync(&cl); 219117e21a9fSKent Overstreet make_btree_freeing_key(b, parent_keys.top); 219217e21a9fSKent Overstreet bch_keylist_push(&parent_keys); 219317e21a9fSKent Overstreet 219417e21a9fSKent Overstreet btree_node_free(b); 219517e21a9fSKent Overstreet 219617e21a9fSKent Overstreet bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL); 219717e21a9fSKent Overstreet BUG_ON(!bch_keylist_empty(&parent_keys)); 2198cafe5635SKent Overstreet } 2199cafe5635SKent Overstreet 2200cafe5635SKent Overstreet rw_unlock(true, n1); 2201cafe5635SKent Overstreet 2202169ef1cfSKent Overstreet bch_time_stats_update(&b->c->btree_split_time, start_time); 2203cafe5635SKent Overstreet 2204cafe5635SKent Overstreet return 0; 2205cafe5635SKent Overstreet err_free2: 22065f5837d2SKent Overstreet bkey_put(b->c, &n2->key); 2207e8e1d468SKent Overstreet btree_node_free(n2); 2208cafe5635SKent Overstreet rw_unlock(true, n2); 2209cafe5635SKent Overstreet err_free1: 22105f5837d2SKent Overstreet bkey_put(b->c, &n1->key); 2211e8e1d468SKent Overstreet btree_node_free(n1); 2212cafe5635SKent Overstreet rw_unlock(true, n1); 2213cafe5635SKent Overstreet err: 22145f5837d2SKent Overstreet WARN(1, "bcache: btree split failed"); 22155f5837d2SKent Overstreet 2216cafe5635SKent Overstreet if (n3 == ERR_PTR(-EAGAIN) || 2217cafe5635SKent Overstreet n2 == ERR_PTR(-EAGAIN) || 2218cafe5635SKent Overstreet n1 == ERR_PTR(-EAGAIN)) 2219cafe5635SKent Overstreet return -EAGAIN; 2220cafe5635SKent Overstreet 2221cafe5635SKent Overstreet return -ENOMEM; 2222cafe5635SKent Overstreet } 2223cafe5635SKent Overstreet 222426c949f8SKent Overstreet static int bch_btree_insert_node(struct btree *b, struct btree_op *op, 2225c18536a7SKent Overstreet struct keylist *insert_keys, 22261b207d80SKent Overstreet atomic_t *journal_ref, 22271b207d80SKent Overstreet struct bkey *replace_key) 222826c949f8SKent Overstreet { 22291b207d80SKent Overstreet BUG_ON(b->level && replace_key); 22301b207d80SKent Overstreet 223126c949f8SKent Overstreet if (should_split(b)) { 223226c949f8SKent Overstreet if (current->bio_list) { 223326c949f8SKent Overstreet op->lock = b->c->root->level + 1; 223417e21a9fSKent Overstreet return -EAGAIN; 223526c949f8SKent Overstreet } else if (op->lock <= b->c->root->level) { 223626c949f8SKent Overstreet op->lock = b->c->root->level + 1; 223717e21a9fSKent Overstreet return -EINTR; 223826c949f8SKent Overstreet } else { 223917e21a9fSKent Overstreet /* Invalidated all iterators */ 224017e21a9fSKent Overstreet return btree_split(b, op, insert_keys, replace_key) ?: 224117e21a9fSKent Overstreet -EINTR; 224226c949f8SKent Overstreet } 224326c949f8SKent Overstreet } else { 224426c949f8SKent Overstreet BUG_ON(write_block(b) != b->sets[b->nsets].data); 224526c949f8SKent Overstreet 224617e21a9fSKent Overstreet if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) { 2247f269af5aSKent Overstreet if (!b->level) 2248c18536a7SKent Overstreet bch_btree_leaf_dirty(b, journal_ref); 2249f269af5aSKent Overstreet else 2250f269af5aSKent Overstreet bch_btree_node_write_sync(b); 225126c949f8SKent Overstreet } 225226c949f8SKent Overstreet 225317e21a9fSKent Overstreet return 0; 225417e21a9fSKent Overstreet } 225526c949f8SKent Overstreet } 225626c949f8SKent Overstreet 2257e7c590ebSKent Overstreet int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, 2258e7c590ebSKent Overstreet struct bkey *check_key) 2259e7c590ebSKent Overstreet { 2260e7c590ebSKent Overstreet int ret = -EINTR; 2261e7c590ebSKent Overstreet uint64_t btree_ptr = b->key.ptr[0]; 2262e7c590ebSKent Overstreet unsigned long seq = b->seq; 2263e7c590ebSKent Overstreet struct keylist insert; 2264e7c590ebSKent Overstreet bool upgrade = op->lock == -1; 2265e7c590ebSKent Overstreet 2266e7c590ebSKent Overstreet bch_keylist_init(&insert); 2267e7c590ebSKent Overstreet 2268e7c590ebSKent Overstreet if (upgrade) { 2269e7c590ebSKent Overstreet rw_unlock(false, b); 2270e7c590ebSKent Overstreet rw_lock(true, b, b->level); 2271e7c590ebSKent Overstreet 2272e7c590ebSKent Overstreet if (b->key.ptr[0] != btree_ptr || 2273e7c590ebSKent Overstreet b->seq != seq + 1) 2274e7c590ebSKent Overstreet goto out; 2275e7c590ebSKent Overstreet } 2276e7c590ebSKent Overstreet 2277e7c590ebSKent Overstreet SET_KEY_PTRS(check_key, 1); 2278e7c590ebSKent Overstreet get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); 2279e7c590ebSKent Overstreet 2280e7c590ebSKent Overstreet SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); 2281e7c590ebSKent Overstreet 2282e7c590ebSKent Overstreet bch_keylist_add(&insert, check_key); 2283e7c590ebSKent Overstreet 22841b207d80SKent Overstreet ret = bch_btree_insert_node(b, op, &insert, NULL, NULL); 2285e7c590ebSKent Overstreet 2286e7c590ebSKent Overstreet BUG_ON(!ret && !bch_keylist_empty(&insert)); 2287e7c590ebSKent Overstreet out: 2288e7c590ebSKent Overstreet if (upgrade) 2289e7c590ebSKent Overstreet downgrade_write(&b->lock); 2290e7c590ebSKent Overstreet return ret; 2291e7c590ebSKent Overstreet } 2292e7c590ebSKent Overstreet 2293cc7b8819SKent Overstreet struct btree_insert_op { 2294cc7b8819SKent Overstreet struct btree_op op; 2295cc7b8819SKent Overstreet struct keylist *keys; 2296cc7b8819SKent Overstreet atomic_t *journal_ref; 2297cc7b8819SKent Overstreet struct bkey *replace_key; 2298cc7b8819SKent Overstreet }; 2299cc7b8819SKent Overstreet 230008239ca2SWei Yongjun static int btree_insert_fn(struct btree_op *b_op, struct btree *b) 2301cafe5635SKent Overstreet { 2302cc7b8819SKent Overstreet struct btree_insert_op *op = container_of(b_op, 2303cc7b8819SKent Overstreet struct btree_insert_op, op); 2304403b6cdeSKent Overstreet 2305cc7b8819SKent Overstreet int ret = bch_btree_insert_node(b, &op->op, op->keys, 2306cc7b8819SKent Overstreet op->journal_ref, op->replace_key); 2307cc7b8819SKent Overstreet if (ret && !bch_keylist_empty(op->keys)) 2308cc7b8819SKent Overstreet return ret; 2309cc7b8819SKent Overstreet else 2310cc7b8819SKent Overstreet return MAP_DONE; 2311cafe5635SKent Overstreet } 2312cafe5635SKent Overstreet 2313cc7b8819SKent Overstreet int bch_btree_insert(struct cache_set *c, struct keylist *keys, 2314cc7b8819SKent Overstreet atomic_t *journal_ref, struct bkey *replace_key) 2315cafe5635SKent Overstreet { 2316cc7b8819SKent Overstreet struct btree_insert_op op; 2317cafe5635SKent Overstreet int ret = 0; 2318cafe5635SKent Overstreet 2319cc7b8819SKent Overstreet BUG_ON(current->bio_list); 23204f3d4014SKent Overstreet BUG_ON(bch_keylist_empty(keys)); 2321cafe5635SKent Overstreet 2322cc7b8819SKent Overstreet bch_btree_op_init(&op.op, 0); 2323cc7b8819SKent Overstreet op.keys = keys; 2324cc7b8819SKent Overstreet op.journal_ref = journal_ref; 2325cc7b8819SKent Overstreet op.replace_key = replace_key; 2326cafe5635SKent Overstreet 2327cc7b8819SKent Overstreet while (!ret && !bch_keylist_empty(keys)) { 2328cc7b8819SKent Overstreet op.op.lock = 0; 2329cc7b8819SKent Overstreet ret = bch_btree_map_leaf_nodes(&op.op, c, 2330cc7b8819SKent Overstreet &START_KEY(keys->keys), 2331cc7b8819SKent Overstreet btree_insert_fn); 2332cc7b8819SKent Overstreet } 2333cc7b8819SKent Overstreet 2334cc7b8819SKent Overstreet if (ret) { 2335cafe5635SKent Overstreet struct bkey *k; 2336cafe5635SKent Overstreet 23371b207d80SKent Overstreet pr_err("error %i", ret); 2338cafe5635SKent Overstreet 23394f3d4014SKent Overstreet while ((k = bch_keylist_pop(keys))) 23403a3b6a4eSKent Overstreet bkey_put(c, k); 2341cc7b8819SKent Overstreet } else if (op.op.insert_collision) 2342cc7b8819SKent Overstreet ret = -ESRCH; 23436054c6d4SKent Overstreet 2344cafe5635SKent Overstreet return ret; 2345cafe5635SKent Overstreet } 2346cafe5635SKent Overstreet 2347cafe5635SKent Overstreet void bch_btree_set_root(struct btree *b) 2348cafe5635SKent Overstreet { 2349cafe5635SKent Overstreet unsigned i; 2350e49c7c37SKent Overstreet struct closure cl; 2351e49c7c37SKent Overstreet 2352e49c7c37SKent Overstreet closure_init_stack(&cl); 2353cafe5635SKent Overstreet 2354c37511b8SKent Overstreet trace_bcache_btree_set_root(b); 2355c37511b8SKent Overstreet 2356cafe5635SKent Overstreet BUG_ON(!b->written); 2357cafe5635SKent Overstreet 2358cafe5635SKent Overstreet for (i = 0; i < KEY_PTRS(&b->key); i++) 2359cafe5635SKent Overstreet BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); 2360cafe5635SKent Overstreet 2361cafe5635SKent Overstreet mutex_lock(&b->c->bucket_lock); 2362cafe5635SKent Overstreet list_del_init(&b->list); 2363cafe5635SKent Overstreet mutex_unlock(&b->c->bucket_lock); 2364cafe5635SKent Overstreet 2365cafe5635SKent Overstreet b->c->root = b; 2366cafe5635SKent Overstreet 2367e49c7c37SKent Overstreet bch_journal_meta(b->c, &cl); 2368e49c7c37SKent Overstreet closure_sync(&cl); 2369cafe5635SKent Overstreet } 2370cafe5635SKent Overstreet 237148dad8baSKent Overstreet /* Map across nodes or keys */ 237248dad8baSKent Overstreet 237348dad8baSKent Overstreet static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, 237448dad8baSKent Overstreet struct bkey *from, 237548dad8baSKent Overstreet btree_map_nodes_fn *fn, int flags) 237648dad8baSKent Overstreet { 237748dad8baSKent Overstreet int ret = MAP_CONTINUE; 237848dad8baSKent Overstreet 237948dad8baSKent Overstreet if (b->level) { 238048dad8baSKent Overstreet struct bkey *k; 238148dad8baSKent Overstreet struct btree_iter iter; 238248dad8baSKent Overstreet 238348dad8baSKent Overstreet bch_btree_iter_init(b, &iter, from); 238448dad8baSKent Overstreet 238548dad8baSKent Overstreet while ((k = bch_btree_iter_next_filter(&iter, b, 238648dad8baSKent Overstreet bch_ptr_bad))) { 238748dad8baSKent Overstreet ret = btree(map_nodes_recurse, k, b, 238848dad8baSKent Overstreet op, from, fn, flags); 238948dad8baSKent Overstreet from = NULL; 239048dad8baSKent Overstreet 239148dad8baSKent Overstreet if (ret != MAP_CONTINUE) 239248dad8baSKent Overstreet return ret; 239348dad8baSKent Overstreet } 239448dad8baSKent Overstreet } 239548dad8baSKent Overstreet 239648dad8baSKent Overstreet if (!b->level || flags == MAP_ALL_NODES) 239748dad8baSKent Overstreet ret = fn(op, b); 239848dad8baSKent Overstreet 239948dad8baSKent Overstreet return ret; 240048dad8baSKent Overstreet } 240148dad8baSKent Overstreet 240248dad8baSKent Overstreet int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, 240348dad8baSKent Overstreet struct bkey *from, btree_map_nodes_fn *fn, int flags) 240448dad8baSKent Overstreet { 2405b54d6934SKent Overstreet return btree_root(map_nodes_recurse, c, op, from, fn, flags); 240648dad8baSKent Overstreet } 240748dad8baSKent Overstreet 240848dad8baSKent Overstreet static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, 240948dad8baSKent Overstreet struct bkey *from, btree_map_keys_fn *fn, 241048dad8baSKent Overstreet int flags) 241148dad8baSKent Overstreet { 241248dad8baSKent Overstreet int ret = MAP_CONTINUE; 241348dad8baSKent Overstreet struct bkey *k; 241448dad8baSKent Overstreet struct btree_iter iter; 241548dad8baSKent Overstreet 241648dad8baSKent Overstreet bch_btree_iter_init(b, &iter, from); 241748dad8baSKent Overstreet 241848dad8baSKent Overstreet while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) { 241948dad8baSKent Overstreet ret = !b->level 242048dad8baSKent Overstreet ? fn(op, b, k) 242148dad8baSKent Overstreet : btree(map_keys_recurse, k, b, op, from, fn, flags); 242248dad8baSKent Overstreet from = NULL; 242348dad8baSKent Overstreet 242448dad8baSKent Overstreet if (ret != MAP_CONTINUE) 242548dad8baSKent Overstreet return ret; 242648dad8baSKent Overstreet } 242748dad8baSKent Overstreet 242848dad8baSKent Overstreet if (!b->level && (flags & MAP_END_KEY)) 242948dad8baSKent Overstreet ret = fn(op, b, &KEY(KEY_INODE(&b->key), 243048dad8baSKent Overstreet KEY_OFFSET(&b->key), 0)); 243148dad8baSKent Overstreet 243248dad8baSKent Overstreet return ret; 243348dad8baSKent Overstreet } 243448dad8baSKent Overstreet 243548dad8baSKent Overstreet int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, 243648dad8baSKent Overstreet struct bkey *from, btree_map_keys_fn *fn, int flags) 243748dad8baSKent Overstreet { 2438b54d6934SKent Overstreet return btree_root(map_keys_recurse, c, op, from, fn, flags); 243948dad8baSKent Overstreet } 244048dad8baSKent Overstreet 2441cafe5635SKent Overstreet /* Keybuf code */ 2442cafe5635SKent Overstreet 2443cafe5635SKent Overstreet static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) 2444cafe5635SKent Overstreet { 2445cafe5635SKent Overstreet /* Overlapping keys compare equal */ 2446cafe5635SKent Overstreet if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) 2447cafe5635SKent Overstreet return -1; 2448cafe5635SKent Overstreet if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) 2449cafe5635SKent Overstreet return 1; 2450cafe5635SKent Overstreet return 0; 2451cafe5635SKent Overstreet } 2452cafe5635SKent Overstreet 2453cafe5635SKent Overstreet static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, 2454cafe5635SKent Overstreet struct keybuf_key *r) 2455cafe5635SKent Overstreet { 2456cafe5635SKent Overstreet return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); 2457cafe5635SKent Overstreet } 2458cafe5635SKent Overstreet 245948dad8baSKent Overstreet struct refill { 246048dad8baSKent Overstreet struct btree_op op; 246148a915a8SKent Overstreet unsigned nr_found; 246248dad8baSKent Overstreet struct keybuf *buf; 246348dad8baSKent Overstreet struct bkey *end; 246448dad8baSKent Overstreet keybuf_pred_fn *pred; 246548dad8baSKent Overstreet }; 246648dad8baSKent Overstreet 246748dad8baSKent Overstreet static int refill_keybuf_fn(struct btree_op *op, struct btree *b, 246848dad8baSKent Overstreet struct bkey *k) 2469cafe5635SKent Overstreet { 247048dad8baSKent Overstreet struct refill *refill = container_of(op, struct refill, op); 247148dad8baSKent Overstreet struct keybuf *buf = refill->buf; 247248dad8baSKent Overstreet int ret = MAP_CONTINUE; 2473cafe5635SKent Overstreet 247448dad8baSKent Overstreet if (bkey_cmp(k, refill->end) >= 0) { 247548dad8baSKent Overstreet ret = MAP_DONE; 247648dad8baSKent Overstreet goto out; 2477cafe5635SKent Overstreet } 2478cafe5635SKent Overstreet 247948dad8baSKent Overstreet if (!KEY_SIZE(k)) /* end key */ 248048dad8baSKent Overstreet goto out; 2481cafe5635SKent Overstreet 248248dad8baSKent Overstreet if (refill->pred(buf, k)) { 2483cafe5635SKent Overstreet struct keybuf_key *w; 2484cafe5635SKent Overstreet 2485cafe5635SKent Overstreet spin_lock(&buf->lock); 2486cafe5635SKent Overstreet 2487cafe5635SKent Overstreet w = array_alloc(&buf->freelist); 248848dad8baSKent Overstreet if (!w) { 248948dad8baSKent Overstreet spin_unlock(&buf->lock); 249048dad8baSKent Overstreet return MAP_DONE; 249148dad8baSKent Overstreet } 2492cafe5635SKent Overstreet 2493cafe5635SKent Overstreet w->private = NULL; 2494cafe5635SKent Overstreet bkey_copy(&w->key, k); 2495cafe5635SKent Overstreet 2496cafe5635SKent Overstreet if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) 2497cafe5635SKent Overstreet array_free(&buf->freelist, w); 249848a915a8SKent Overstreet else 249948a915a8SKent Overstreet refill->nr_found++; 2500cafe5635SKent Overstreet 250148dad8baSKent Overstreet if (array_freelist_empty(&buf->freelist)) 250248dad8baSKent Overstreet ret = MAP_DONE; 250348dad8baSKent Overstreet 2504cafe5635SKent Overstreet spin_unlock(&buf->lock); 2505cafe5635SKent Overstreet } 250648dad8baSKent Overstreet out: 250748dad8baSKent Overstreet buf->last_scanned = *k; 250848dad8baSKent Overstreet return ret; 2509cafe5635SKent Overstreet } 2510cafe5635SKent Overstreet 2511cafe5635SKent Overstreet void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, 251272c27061SKent Overstreet struct bkey *end, keybuf_pred_fn *pred) 2513cafe5635SKent Overstreet { 2514cafe5635SKent Overstreet struct bkey start = buf->last_scanned; 251548dad8baSKent Overstreet struct refill refill; 2516cafe5635SKent Overstreet 2517cafe5635SKent Overstreet cond_resched(); 2518cafe5635SKent Overstreet 2519b54d6934SKent Overstreet bch_btree_op_init(&refill.op, -1); 252048a915a8SKent Overstreet refill.nr_found = 0; 252148dad8baSKent Overstreet refill.buf = buf; 252248dad8baSKent Overstreet refill.end = end; 252348dad8baSKent Overstreet refill.pred = pred; 252448dad8baSKent Overstreet 252548dad8baSKent Overstreet bch_btree_map_keys(&refill.op, c, &buf->last_scanned, 252648dad8baSKent Overstreet refill_keybuf_fn, MAP_END_KEY); 2527cafe5635SKent Overstreet 252848a915a8SKent Overstreet trace_bcache_keyscan(refill.nr_found, 2529cafe5635SKent Overstreet KEY_INODE(&start), KEY_OFFSET(&start), 253048a915a8SKent Overstreet KEY_INODE(&buf->last_scanned), 253148a915a8SKent Overstreet KEY_OFFSET(&buf->last_scanned)); 2532cafe5635SKent Overstreet 2533cafe5635SKent Overstreet spin_lock(&buf->lock); 2534cafe5635SKent Overstreet 2535cafe5635SKent Overstreet if (!RB_EMPTY_ROOT(&buf->keys)) { 2536cafe5635SKent Overstreet struct keybuf_key *w; 2537cafe5635SKent Overstreet w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2538cafe5635SKent Overstreet buf->start = START_KEY(&w->key); 2539cafe5635SKent Overstreet 2540cafe5635SKent Overstreet w = RB_LAST(&buf->keys, struct keybuf_key, node); 2541cafe5635SKent Overstreet buf->end = w->key; 2542cafe5635SKent Overstreet } else { 2543cafe5635SKent Overstreet buf->start = MAX_KEY; 2544cafe5635SKent Overstreet buf->end = MAX_KEY; 2545cafe5635SKent Overstreet } 2546cafe5635SKent Overstreet 2547cafe5635SKent Overstreet spin_unlock(&buf->lock); 2548cafe5635SKent Overstreet } 2549cafe5635SKent Overstreet 2550cafe5635SKent Overstreet static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2551cafe5635SKent Overstreet { 2552cafe5635SKent Overstreet rb_erase(&w->node, &buf->keys); 2553cafe5635SKent Overstreet array_free(&buf->freelist, w); 2554cafe5635SKent Overstreet } 2555cafe5635SKent Overstreet 2556cafe5635SKent Overstreet void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) 2557cafe5635SKent Overstreet { 2558cafe5635SKent Overstreet spin_lock(&buf->lock); 2559cafe5635SKent Overstreet __bch_keybuf_del(buf, w); 2560cafe5635SKent Overstreet spin_unlock(&buf->lock); 2561cafe5635SKent Overstreet } 2562cafe5635SKent Overstreet 2563cafe5635SKent Overstreet bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, 2564cafe5635SKent Overstreet struct bkey *end) 2565cafe5635SKent Overstreet { 2566cafe5635SKent Overstreet bool ret = false; 2567cafe5635SKent Overstreet struct keybuf_key *p, *w, s; 2568cafe5635SKent Overstreet s.key = *start; 2569cafe5635SKent Overstreet 2570cafe5635SKent Overstreet if (bkey_cmp(end, &buf->start) <= 0 || 2571cafe5635SKent Overstreet bkey_cmp(start, &buf->end) >= 0) 2572cafe5635SKent Overstreet return false; 2573cafe5635SKent Overstreet 2574cafe5635SKent Overstreet spin_lock(&buf->lock); 2575cafe5635SKent Overstreet w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); 2576cafe5635SKent Overstreet 2577cafe5635SKent Overstreet while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { 2578cafe5635SKent Overstreet p = w; 2579cafe5635SKent Overstreet w = RB_NEXT(w, node); 2580cafe5635SKent Overstreet 2581cafe5635SKent Overstreet if (p->private) 2582cafe5635SKent Overstreet ret = true; 2583cafe5635SKent Overstreet else 2584cafe5635SKent Overstreet __bch_keybuf_del(buf, p); 2585cafe5635SKent Overstreet } 2586cafe5635SKent Overstreet 2587cafe5635SKent Overstreet spin_unlock(&buf->lock); 2588cafe5635SKent Overstreet return ret; 2589cafe5635SKent Overstreet } 2590cafe5635SKent Overstreet 2591cafe5635SKent Overstreet struct keybuf_key *bch_keybuf_next(struct keybuf *buf) 2592cafe5635SKent Overstreet { 2593cafe5635SKent Overstreet struct keybuf_key *w; 2594cafe5635SKent Overstreet spin_lock(&buf->lock); 2595cafe5635SKent Overstreet 2596cafe5635SKent Overstreet w = RB_FIRST(&buf->keys, struct keybuf_key, node); 2597cafe5635SKent Overstreet 2598cafe5635SKent Overstreet while (w && w->private) 2599cafe5635SKent Overstreet w = RB_NEXT(w, node); 2600cafe5635SKent Overstreet 2601cafe5635SKent Overstreet if (w) 2602cafe5635SKent Overstreet w->private = ERR_PTR(-EINTR); 2603cafe5635SKent Overstreet 2604cafe5635SKent Overstreet spin_unlock(&buf->lock); 2605cafe5635SKent Overstreet return w; 2606cafe5635SKent Overstreet } 2607cafe5635SKent Overstreet 2608cafe5635SKent Overstreet struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, 2609cafe5635SKent Overstreet struct keybuf *buf, 261072c27061SKent Overstreet struct bkey *end, 261172c27061SKent Overstreet keybuf_pred_fn *pred) 2612cafe5635SKent Overstreet { 2613cafe5635SKent Overstreet struct keybuf_key *ret; 2614cafe5635SKent Overstreet 2615cafe5635SKent Overstreet while (1) { 2616cafe5635SKent Overstreet ret = bch_keybuf_next(buf); 2617cafe5635SKent Overstreet if (ret) 2618cafe5635SKent Overstreet break; 2619cafe5635SKent Overstreet 2620cafe5635SKent Overstreet if (bkey_cmp(&buf->last_scanned, end) >= 0) { 2621cafe5635SKent Overstreet pr_debug("scan finished"); 2622cafe5635SKent Overstreet break; 2623cafe5635SKent Overstreet } 2624cafe5635SKent Overstreet 262572c27061SKent Overstreet bch_refill_keybuf(c, buf, end, pred); 2626cafe5635SKent Overstreet } 2627cafe5635SKent Overstreet 2628cafe5635SKent Overstreet return ret; 2629cafe5635SKent Overstreet } 2630cafe5635SKent Overstreet 263172c27061SKent Overstreet void bch_keybuf_init(struct keybuf *buf) 2632cafe5635SKent Overstreet { 2633cafe5635SKent Overstreet buf->last_scanned = MAX_KEY; 2634cafe5635SKent Overstreet buf->keys = RB_ROOT; 2635cafe5635SKent Overstreet 2636cafe5635SKent Overstreet spin_lock_init(&buf->lock); 2637cafe5635SKent Overstreet array_allocator_init(&buf->freelist); 2638cafe5635SKent Overstreet } 2639cafe5635SKent Overstreet 2640cafe5635SKent Overstreet void bch_btree_exit(void) 2641cafe5635SKent Overstreet { 2642cafe5635SKent Overstreet if (btree_io_wq) 2643cafe5635SKent Overstreet destroy_workqueue(btree_io_wq); 2644cafe5635SKent Overstreet } 2645cafe5635SKent Overstreet 2646cafe5635SKent Overstreet int __init bch_btree_init(void) 2647cafe5635SKent Overstreet { 264872a44517SKent Overstreet btree_io_wq = create_singlethread_workqueue("bch_btree_io"); 264972a44517SKent Overstreet if (!btree_io_wq) 2650cafe5635SKent Overstreet return -ENOMEM; 2651cafe5635SKent Overstreet 2652cafe5635SKent Overstreet return 0; 2653cafe5635SKent Overstreet } 2654