11802d0beSThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 2f2c870e3STao Ma /* -*- mode: c; c-basic-offset: 8; -*- 3f2c870e3STao Ma * vim: noexpandtab sw=8 ts=8 sts=0: 4f2c870e3STao Ma * 5f2c870e3STao Ma * refcounttree.c 6f2c870e3STao Ma * 7f2c870e3STao Ma * Copyright (C) 2009 Oracle. All rights reserved. 8f2c870e3STao Ma */ 9f2c870e3STao Ma 10e73a819dSTao Ma #include <linux/sort.h> 11f2c870e3STao Ma #include <cluster/masklog.h> 12f2c870e3STao Ma #include "ocfs2.h" 13f2c870e3STao Ma #include "inode.h" 14f2c870e3STao Ma #include "alloc.h" 15f2c870e3STao Ma #include "suballoc.h" 16f2c870e3STao Ma #include "journal.h" 17f2c870e3STao Ma #include "uptodate.h" 18f2c870e3STao Ma #include "super.h" 19f2c870e3STao Ma #include "buffer_head_io.h" 20f2c870e3STao Ma #include "blockcheck.h" 21c732eb16STao Ma #include "refcounttree.h" 228bf396deSTao Ma #include "sysfile.h" 23374a263eSTao Ma #include "dlmglue.h" 24e73a819dSTao Ma #include "extent_map.h" 256f70fa51STao Ma #include "aops.h" 26492a8a33STao Ma #include "xattr.h" 2709bf27a0STao Ma #include "namei.h" 28198aac28STao Ma #include "ocfs2_trace.h" 2929ac8e85SDarrick J. Wong #include "file.h" 306f70fa51STao Ma 316f70fa51STao Ma #include <linux/bio.h> 326f70fa51STao Ma #include <linux/blkdev.h> 336f70fa51STao Ma #include <linux/slab.h> 346f70fa51STao Ma #include <linux/writeback.h> 356f70fa51STao Ma #include <linux/pagevec.h> 366f70fa51STao Ma #include <linux/swap.h> 37bd50873dSTao Ma #include <linux/security.h> 38bd50873dSTao Ma #include <linux/fsnotify.h> 39bd50873dSTao Ma #include <linux/quotaops.h> 40bd50873dSTao Ma #include <linux/namei.h> 41bd50873dSTao Ma #include <linux/mount.h> 42702e5bc6SChristoph Hellwig #include <linux/posix_acl.h> 436f70fa51STao Ma 446f70fa51STao Ma struct ocfs2_cow_context { 456f70fa51STao Ma struct inode *inode; 466f70fa51STao Ma u32 cow_start; 476f70fa51STao Ma u32 cow_len; 48913580b4STao Ma struct ocfs2_extent_tree data_et; 49913580b4STao Ma struct ocfs2_refcount_tree *ref_tree; 506f70fa51STao Ma struct buffer_head *ref_root_bh; 516f70fa51STao Ma struct ocfs2_alloc_context *meta_ac; 526f70fa51STao Ma struct ocfs2_alloc_context *data_ac; 536f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 54492a8a33STao Ma void *cow_object; 55492a8a33STao Ma struct ocfs2_post_refcount *post_refcount; 56492a8a33STao Ma int extra_credits; 57913580b4STao Ma int (*get_clusters)(struct ocfs2_cow_context *context, 58913580b4STao Ma u32 v_cluster, u32 *p_cluster, 59913580b4STao Ma u32 *num_clusters, 60913580b4STao Ma unsigned int *extent_flags); 61913580b4STao Ma int (*cow_duplicate_clusters)(handle_t *handle, 62c7dd3392STiger Yang struct inode *inode, 63913580b4STao Ma u32 cpos, u32 old_cluster, 64913580b4STao Ma u32 new_cluster, u32 new_len); 656f70fa51STao Ma }; 66c732eb16STao Ma 67c732eb16STao Ma static inline struct ocfs2_refcount_tree * 68c732eb16STao Ma cache_info_to_refcount(struct ocfs2_caching_info *ci) 69c732eb16STao Ma { 70c732eb16STao Ma return container_of(ci, struct ocfs2_refcount_tree, rf_ci); 71c732eb16STao Ma } 72f2c870e3STao Ma 73f2c870e3STao Ma static int ocfs2_validate_refcount_block(struct super_block *sb, 74f2c870e3STao Ma struct buffer_head *bh) 75f2c870e3STao Ma { 76f2c870e3STao Ma int rc; 77f2c870e3STao Ma struct ocfs2_refcount_block *rb = 78f2c870e3STao Ma (struct ocfs2_refcount_block *)bh->b_data; 79f2c870e3STao Ma 80198aac28STao Ma trace_ocfs2_validate_refcount_block((unsigned long long)bh->b_blocknr); 81f2c870e3STao Ma 82f2c870e3STao Ma BUG_ON(!buffer_uptodate(bh)); 83f2c870e3STao Ma 84f2c870e3STao Ma /* 85f2c870e3STao Ma * If the ecc fails, we return the error but otherwise 86f2c870e3STao Ma * leave the filesystem running. We know any error is 87f2c870e3STao Ma * local to this block. 88f2c870e3STao Ma */ 89f2c870e3STao Ma rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check); 90f2c870e3STao Ma if (rc) { 91f2c870e3STao Ma mlog(ML_ERROR, "Checksum failed for refcount block %llu\n", 92f2c870e3STao Ma (unsigned long long)bh->b_blocknr); 93f2c870e3STao Ma return rc; 94f2c870e3STao Ma } 95f2c870e3STao Ma 96f2c870e3STao Ma 97f2c870e3STao Ma if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) { 9817a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb, 997ecef14aSJoe Perches "Refcount block #%llu has bad signature %.*s\n", 100f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 7, 101f2c870e3STao Ma rb->rf_signature); 10217a5b9abSGoldwyn Rodrigues goto out; 103f2c870e3STao Ma } 104f2c870e3STao Ma 105f2c870e3STao Ma if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) { 10617a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb, 1077ecef14aSJoe Perches "Refcount block #%llu has an invalid rf_blkno of %llu\n", 108f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 109f2c870e3STao Ma (unsigned long long)le64_to_cpu(rb->rf_blkno)); 11017a5b9abSGoldwyn Rodrigues goto out; 111f2c870e3STao Ma } 112f2c870e3STao Ma 113f2c870e3STao Ma if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) { 11417a5b9abSGoldwyn Rodrigues rc = ocfs2_error(sb, 1157ecef14aSJoe Perches "Refcount block #%llu has an invalid rf_fs_generation of #%u\n", 116f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 117f2c870e3STao Ma le32_to_cpu(rb->rf_fs_generation)); 11817a5b9abSGoldwyn Rodrigues goto out; 119f2c870e3STao Ma } 12017a5b9abSGoldwyn Rodrigues out: 12117a5b9abSGoldwyn Rodrigues return rc; 122f2c870e3STao Ma } 123f2c870e3STao Ma 124f2c870e3STao Ma static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci, 125f2c870e3STao Ma u64 rb_blkno, 126f2c870e3STao Ma struct buffer_head **bh) 127f2c870e3STao Ma { 128f2c870e3STao Ma int rc; 129f2c870e3STao Ma struct buffer_head *tmp = *bh; 130f2c870e3STao Ma 131f2c870e3STao Ma rc = ocfs2_read_block(ci, rb_blkno, &tmp, 132f2c870e3STao Ma ocfs2_validate_refcount_block); 133f2c870e3STao Ma 134f2c870e3STao Ma /* If ocfs2_read_block() got us a new bh, pass it up. */ 135f2c870e3STao Ma if (!rc && !*bh) 136f2c870e3STao Ma *bh = tmp; 137f2c870e3STao Ma 138f2c870e3STao Ma return rc; 139f2c870e3STao Ma } 140c732eb16STao Ma 141c732eb16STao Ma static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci) 142c732eb16STao Ma { 143c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 144c732eb16STao Ma 145c732eb16STao Ma return rf->rf_blkno; 146c732eb16STao Ma } 147c732eb16STao Ma 148c732eb16STao Ma static struct super_block * 149c732eb16STao Ma ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci) 150c732eb16STao Ma { 151c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 152c732eb16STao Ma 153c732eb16STao Ma return rf->rf_sb; 154c732eb16STao Ma } 155c732eb16STao Ma 156c732eb16STao Ma static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci) 157185a7321SJules Irenge __acquires(&rf->rf_lock) 158c732eb16STao Ma { 159c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 160c732eb16STao Ma 161c732eb16STao Ma spin_lock(&rf->rf_lock); 162c732eb16STao Ma } 163c732eb16STao Ma 164c732eb16STao Ma static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci) 165185a7321SJules Irenge __releases(&rf->rf_lock) 166c732eb16STao Ma { 167c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 168c732eb16STao Ma 169c732eb16STao Ma spin_unlock(&rf->rf_lock); 170c732eb16STao Ma } 171c732eb16STao Ma 172c732eb16STao Ma static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci) 173c732eb16STao Ma { 174c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 175c732eb16STao Ma 176c732eb16STao Ma mutex_lock(&rf->rf_io_mutex); 177c732eb16STao Ma } 178c732eb16STao Ma 179c732eb16STao Ma static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci) 180c732eb16STao Ma { 181c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 182c732eb16STao Ma 183c732eb16STao Ma mutex_unlock(&rf->rf_io_mutex); 184c732eb16STao Ma } 185c732eb16STao Ma 186c732eb16STao Ma static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = { 187c732eb16STao Ma .co_owner = ocfs2_refcount_cache_owner, 188c732eb16STao Ma .co_get_super = ocfs2_refcount_cache_get_super, 189c732eb16STao Ma .co_cache_lock = ocfs2_refcount_cache_lock, 190c732eb16STao Ma .co_cache_unlock = ocfs2_refcount_cache_unlock, 191c732eb16STao Ma .co_io_lock = ocfs2_refcount_cache_io_lock, 192c732eb16STao Ma .co_io_unlock = ocfs2_refcount_cache_io_unlock, 193c732eb16STao Ma }; 194374a263eSTao Ma 195374a263eSTao Ma static struct ocfs2_refcount_tree * 196374a263eSTao Ma ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno) 197374a263eSTao Ma { 198374a263eSTao Ma struct rb_node *n = osb->osb_rf_lock_tree.rb_node; 199374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL; 200374a263eSTao Ma 201374a263eSTao Ma while (n) { 202374a263eSTao Ma tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node); 203374a263eSTao Ma 204374a263eSTao Ma if (blkno < tree->rf_blkno) 205374a263eSTao Ma n = n->rb_left; 206374a263eSTao Ma else if (blkno > tree->rf_blkno) 207374a263eSTao Ma n = n->rb_right; 208374a263eSTao Ma else 209374a263eSTao Ma return tree; 210374a263eSTao Ma } 211374a263eSTao Ma 212374a263eSTao Ma return NULL; 213374a263eSTao Ma } 214374a263eSTao Ma 215374a263eSTao Ma /* osb_lock is already locked. */ 216374a263eSTao Ma static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb, 217374a263eSTao Ma struct ocfs2_refcount_tree *new) 218374a263eSTao Ma { 219374a263eSTao Ma u64 rf_blkno = new->rf_blkno; 220374a263eSTao Ma struct rb_node *parent = NULL; 221374a263eSTao Ma struct rb_node **p = &osb->osb_rf_lock_tree.rb_node; 222374a263eSTao Ma struct ocfs2_refcount_tree *tmp; 223374a263eSTao Ma 224374a263eSTao Ma while (*p) { 225374a263eSTao Ma parent = *p; 226374a263eSTao Ma 227374a263eSTao Ma tmp = rb_entry(parent, struct ocfs2_refcount_tree, 228374a263eSTao Ma rf_node); 229374a263eSTao Ma 230374a263eSTao Ma if (rf_blkno < tmp->rf_blkno) 231374a263eSTao Ma p = &(*p)->rb_left; 232374a263eSTao Ma else if (rf_blkno > tmp->rf_blkno) 233374a263eSTao Ma p = &(*p)->rb_right; 234374a263eSTao Ma else { 235374a263eSTao Ma /* This should never happen! */ 236374a263eSTao Ma mlog(ML_ERROR, "Duplicate refcount block %llu found!\n", 237374a263eSTao Ma (unsigned long long)rf_blkno); 238374a263eSTao Ma BUG(); 239374a263eSTao Ma } 240374a263eSTao Ma } 241374a263eSTao Ma 242374a263eSTao Ma rb_link_node(&new->rf_node, parent, p); 243374a263eSTao Ma rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree); 244374a263eSTao Ma } 245374a263eSTao Ma 246374a263eSTao Ma static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree) 247374a263eSTao Ma { 248374a263eSTao Ma ocfs2_metadata_cache_exit(&tree->rf_ci); 249374a263eSTao Ma ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres); 250374a263eSTao Ma ocfs2_lock_res_free(&tree->rf_lockres); 251374a263eSTao Ma kfree(tree); 252374a263eSTao Ma } 253374a263eSTao Ma 254374a263eSTao Ma static inline void 255374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb, 256374a263eSTao Ma struct ocfs2_refcount_tree *tree) 257374a263eSTao Ma { 258374a263eSTao Ma rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree); 259374a263eSTao Ma if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree) 260374a263eSTao Ma osb->osb_ref_tree_lru = NULL; 261374a263eSTao Ma } 262374a263eSTao Ma 263374a263eSTao Ma static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb, 264374a263eSTao Ma struct ocfs2_refcount_tree *tree) 265374a263eSTao Ma { 266374a263eSTao Ma spin_lock(&osb->osb_lock); 267374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree); 268374a263eSTao Ma spin_unlock(&osb->osb_lock); 269374a263eSTao Ma } 270374a263eSTao Ma 27112d4cec9STao Ma static void ocfs2_kref_remove_refcount_tree(struct kref *kref) 272374a263eSTao Ma { 273374a263eSTao Ma struct ocfs2_refcount_tree *tree = 274374a263eSTao Ma container_of(kref, struct ocfs2_refcount_tree, rf_getcnt); 275374a263eSTao Ma 276374a263eSTao Ma ocfs2_free_refcount_tree(tree); 277374a263eSTao Ma } 278374a263eSTao Ma 279374a263eSTao Ma static inline void 280374a263eSTao Ma ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree) 281374a263eSTao Ma { 282374a263eSTao Ma kref_get(&tree->rf_getcnt); 283374a263eSTao Ma } 284374a263eSTao Ma 285374a263eSTao Ma static inline void 286374a263eSTao Ma ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree) 287374a263eSTao Ma { 288374a263eSTao Ma kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree); 289374a263eSTao Ma } 290374a263eSTao Ma 291374a263eSTao Ma static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new, 292374a263eSTao Ma struct super_block *sb) 293374a263eSTao Ma { 294374a263eSTao Ma ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops); 295374a263eSTao Ma mutex_init(&new->rf_io_mutex); 296374a263eSTao Ma new->rf_sb = sb; 297374a263eSTao Ma spin_lock_init(&new->rf_lock); 298374a263eSTao Ma } 299374a263eSTao Ma 300374a263eSTao Ma static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb, 301374a263eSTao Ma struct ocfs2_refcount_tree *new, 302374a263eSTao Ma u64 rf_blkno, u32 generation) 303374a263eSTao Ma { 304374a263eSTao Ma init_rwsem(&new->rf_sem); 305374a263eSTao Ma ocfs2_refcount_lock_res_init(&new->rf_lockres, osb, 306374a263eSTao Ma rf_blkno, generation); 307374a263eSTao Ma } 308374a263eSTao Ma 3098bf396deSTao Ma static struct ocfs2_refcount_tree* 3108bf396deSTao Ma ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno) 3118bf396deSTao Ma { 3128bf396deSTao Ma struct ocfs2_refcount_tree *new; 3138bf396deSTao Ma 3148bf396deSTao Ma new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS); 3158bf396deSTao Ma if (!new) 3168bf396deSTao Ma return NULL; 3178bf396deSTao Ma 3188bf396deSTao Ma new->rf_blkno = rf_blkno; 3198bf396deSTao Ma kref_init(&new->rf_getcnt); 3208bf396deSTao Ma ocfs2_init_refcount_tree_ci(new, osb->sb); 3218bf396deSTao Ma 3228bf396deSTao Ma return new; 3238bf396deSTao Ma } 3248bf396deSTao Ma 325374a263eSTao Ma static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno, 326374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree) 327374a263eSTao Ma { 328374a263eSTao Ma int ret = 0; 329374a263eSTao Ma struct ocfs2_refcount_tree *tree, *new = NULL; 330374a263eSTao Ma struct buffer_head *ref_root_bh = NULL; 331374a263eSTao Ma struct ocfs2_refcount_block *ref_rb; 332374a263eSTao Ma 333374a263eSTao Ma spin_lock(&osb->osb_lock); 334374a263eSTao Ma if (osb->osb_ref_tree_lru && 335374a263eSTao Ma osb->osb_ref_tree_lru->rf_blkno == rf_blkno) 336374a263eSTao Ma tree = osb->osb_ref_tree_lru; 337374a263eSTao Ma else 338374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno); 339374a263eSTao Ma if (tree) 340374a263eSTao Ma goto out; 341374a263eSTao Ma 342374a263eSTao Ma spin_unlock(&osb->osb_lock); 343374a263eSTao Ma 3448bf396deSTao Ma new = ocfs2_allocate_refcount_tree(osb, rf_blkno); 345374a263eSTao Ma if (!new) { 346374a263eSTao Ma ret = -ENOMEM; 3478bf396deSTao Ma mlog_errno(ret); 348374a263eSTao Ma return ret; 349374a263eSTao Ma } 350374a263eSTao Ma /* 351374a263eSTao Ma * We need the generation to create the refcount tree lock and since 352374a263eSTao Ma * it isn't changed during the tree modification, we are safe here to 353374a263eSTao Ma * read without protection. 354374a263eSTao Ma * We also have to purge the cache after we create the lock since the 355374a263eSTao Ma * refcount block may have the stale data. It can only be trusted when 356374a263eSTao Ma * we hold the refcount lock. 357374a263eSTao Ma */ 358374a263eSTao Ma ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh); 359374a263eSTao Ma if (ret) { 360374a263eSTao Ma mlog_errno(ret); 361374a263eSTao Ma ocfs2_metadata_cache_exit(&new->rf_ci); 362374a263eSTao Ma kfree(new); 363374a263eSTao Ma return ret; 364374a263eSTao Ma } 365374a263eSTao Ma 366374a263eSTao Ma ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 367374a263eSTao Ma new->rf_generation = le32_to_cpu(ref_rb->rf_generation); 368374a263eSTao Ma ocfs2_init_refcount_tree_lock(osb, new, rf_blkno, 369374a263eSTao Ma new->rf_generation); 370374a263eSTao Ma ocfs2_metadata_cache_purge(&new->rf_ci); 371374a263eSTao Ma 372374a263eSTao Ma spin_lock(&osb->osb_lock); 373374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno); 374374a263eSTao Ma if (tree) 375374a263eSTao Ma goto out; 376374a263eSTao Ma 377374a263eSTao Ma ocfs2_insert_refcount_tree(osb, new); 378374a263eSTao Ma 379374a263eSTao Ma tree = new; 380374a263eSTao Ma new = NULL; 381374a263eSTao Ma 382374a263eSTao Ma out: 383374a263eSTao Ma *ret_tree = tree; 384374a263eSTao Ma 385374a263eSTao Ma osb->osb_ref_tree_lru = tree; 386374a263eSTao Ma 387374a263eSTao Ma spin_unlock(&osb->osb_lock); 388374a263eSTao Ma 389374a263eSTao Ma if (new) 390374a263eSTao Ma ocfs2_free_refcount_tree(new); 391374a263eSTao Ma 392374a263eSTao Ma brelse(ref_root_bh); 393374a263eSTao Ma return ret; 394374a263eSTao Ma } 395374a263eSTao Ma 396374a263eSTao Ma static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno) 397374a263eSTao Ma { 398374a263eSTao Ma int ret; 399374a263eSTao Ma struct buffer_head *di_bh = NULL; 400374a263eSTao Ma struct ocfs2_dinode *di; 401374a263eSTao Ma 402374a263eSTao Ma ret = ocfs2_read_inode_block(inode, &di_bh); 403374a263eSTao Ma if (ret) { 404374a263eSTao Ma mlog_errno(ret); 405374a263eSTao Ma goto out; 406374a263eSTao Ma } 407374a263eSTao Ma 40884e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode)); 409374a263eSTao Ma 410374a263eSTao Ma di = (struct ocfs2_dinode *)di_bh->b_data; 411374a263eSTao Ma *ref_blkno = le64_to_cpu(di->i_refcount_loc); 412374a263eSTao Ma brelse(di_bh); 413374a263eSTao Ma out: 414374a263eSTao Ma return ret; 415374a263eSTao Ma } 416374a263eSTao Ma 417374a263eSTao Ma static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb, 418374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw) 419374a263eSTao Ma { 420374a263eSTao Ma int ret; 421374a263eSTao Ma 422374a263eSTao Ma ret = ocfs2_refcount_lock(tree, rw); 423374a263eSTao Ma if (ret) { 424374a263eSTao Ma mlog_errno(ret); 425374a263eSTao Ma goto out; 426374a263eSTao Ma } 427374a263eSTao Ma 428374a263eSTao Ma if (rw) 429374a263eSTao Ma down_write(&tree->rf_sem); 430374a263eSTao Ma else 431374a263eSTao Ma down_read(&tree->rf_sem); 432374a263eSTao Ma 433374a263eSTao Ma out: 434374a263eSTao Ma return ret; 435374a263eSTao Ma } 436374a263eSTao Ma 437374a263eSTao Ma /* 438374a263eSTao Ma * Lock the refcount tree pointed by ref_blkno and return the tree. 439374a263eSTao Ma * In most case, we lock the tree and read the refcount block. 440374a263eSTao Ma * So read it here if the caller really needs it. 441374a263eSTao Ma * 442374a263eSTao Ma * If the tree has been re-created by other node, it will free the 443374a263eSTao Ma * old one and re-create it. 444374a263eSTao Ma */ 445374a263eSTao Ma int ocfs2_lock_refcount_tree(struct ocfs2_super *osb, 446374a263eSTao Ma u64 ref_blkno, int rw, 447374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree, 448374a263eSTao Ma struct buffer_head **ref_bh) 449374a263eSTao Ma { 450374a263eSTao Ma int ret, delete_tree = 0; 451374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL; 452374a263eSTao Ma struct buffer_head *ref_root_bh = NULL; 453374a263eSTao Ma struct ocfs2_refcount_block *rb; 454374a263eSTao Ma 455374a263eSTao Ma again: 456374a263eSTao Ma ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree); 457374a263eSTao Ma if (ret) { 458374a263eSTao Ma mlog_errno(ret); 459374a263eSTao Ma return ret; 460374a263eSTao Ma } 461374a263eSTao Ma 462374a263eSTao Ma ocfs2_refcount_tree_get(tree); 463374a263eSTao Ma 464374a263eSTao Ma ret = __ocfs2_lock_refcount_tree(osb, tree, rw); 465374a263eSTao Ma if (ret) { 466374a263eSTao Ma mlog_errno(ret); 467374a263eSTao Ma ocfs2_refcount_tree_put(tree); 468374a263eSTao Ma goto out; 469374a263eSTao Ma } 470374a263eSTao Ma 471374a263eSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno, 472374a263eSTao Ma &ref_root_bh); 473374a263eSTao Ma if (ret) { 474374a263eSTao Ma mlog_errno(ret); 475374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw); 476374a263eSTao Ma goto out; 477374a263eSTao Ma } 478374a263eSTao Ma 479374a263eSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 480374a263eSTao Ma /* 481374a263eSTao Ma * If the refcount block has been freed and re-created, we may need 482374a263eSTao Ma * to recreate the refcount tree also. 483374a263eSTao Ma * 484374a263eSTao Ma * Here we just remove the tree from the rb-tree, and the last 485374a263eSTao Ma * kref holder will unlock and delete this refcount_tree. 486374a263eSTao Ma * Then we goto "again" and ocfs2_get_refcount_tree will create 487374a263eSTao Ma * the new refcount tree for us. 488374a263eSTao Ma */ 489374a263eSTao Ma if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) { 490374a263eSTao Ma if (!tree->rf_removed) { 491374a263eSTao Ma ocfs2_erase_refcount_tree_from_list(osb, tree); 492374a263eSTao Ma tree->rf_removed = 1; 493374a263eSTao Ma delete_tree = 1; 494374a263eSTao Ma } 495374a263eSTao Ma 496374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw); 497374a263eSTao Ma /* 498374a263eSTao Ma * We get an extra reference when we create the refcount 499374a263eSTao Ma * tree, so another put will destroy it. 500374a263eSTao Ma */ 501374a263eSTao Ma if (delete_tree) 502374a263eSTao Ma ocfs2_refcount_tree_put(tree); 503374a263eSTao Ma brelse(ref_root_bh); 504374a263eSTao Ma ref_root_bh = NULL; 505374a263eSTao Ma goto again; 506374a263eSTao Ma } 507374a263eSTao Ma 508374a263eSTao Ma *ret_tree = tree; 509374a263eSTao Ma if (ref_bh) { 510374a263eSTao Ma *ref_bh = ref_root_bh; 511374a263eSTao Ma ref_root_bh = NULL; 512374a263eSTao Ma } 513374a263eSTao Ma out: 514374a263eSTao Ma brelse(ref_root_bh); 515374a263eSTao Ma return ret; 516374a263eSTao Ma } 517374a263eSTao Ma 518374a263eSTao Ma void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb, 519374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw) 520374a263eSTao Ma { 521374a263eSTao Ma if (rw) 522374a263eSTao Ma up_write(&tree->rf_sem); 523374a263eSTao Ma else 524374a263eSTao Ma up_read(&tree->rf_sem); 525374a263eSTao Ma 526374a263eSTao Ma ocfs2_refcount_unlock(tree, rw); 527374a263eSTao Ma ocfs2_refcount_tree_put(tree); 528374a263eSTao Ma } 529374a263eSTao Ma 530374a263eSTao Ma void ocfs2_purge_refcount_trees(struct ocfs2_super *osb) 531374a263eSTao Ma { 532374a263eSTao Ma struct rb_node *node; 533374a263eSTao Ma struct ocfs2_refcount_tree *tree; 534374a263eSTao Ma struct rb_root *root = &osb->osb_rf_lock_tree; 535374a263eSTao Ma 536374a263eSTao Ma while ((node = rb_last(root)) != NULL) { 537374a263eSTao Ma tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node); 538374a263eSTao Ma 539198aac28STao Ma trace_ocfs2_purge_refcount_trees( 540374a263eSTao Ma (unsigned long long) tree->rf_blkno); 541374a263eSTao Ma 542374a263eSTao Ma rb_erase(&tree->rf_node, root); 543374a263eSTao Ma ocfs2_free_refcount_tree(tree); 544374a263eSTao Ma } 545374a263eSTao Ma } 5468bf396deSTao Ma 5478bf396deSTao Ma /* 5488bf396deSTao Ma * Create a refcount tree for an inode. 5498bf396deSTao Ma * We take for granted that the inode is already locked. 5508bf396deSTao Ma */ 5518bf396deSTao Ma static int ocfs2_create_refcount_tree(struct inode *inode, 5528bf396deSTao Ma struct buffer_head *di_bh) 5538bf396deSTao Ma { 5548bf396deSTao Ma int ret; 5558bf396deSTao Ma handle_t *handle = NULL; 5568bf396deSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 5578bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 5588bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 5598bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 5608bf396deSTao Ma struct buffer_head *new_bh = NULL; 5618bf396deSTao Ma struct ocfs2_refcount_block *rb; 5628bf396deSTao Ma struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL; 5638bf396deSTao Ma u16 suballoc_bit_start; 5648bf396deSTao Ma u32 num_got; 5652b6cb576SJoel Becker u64 suballoc_loc, first_blkno; 5668bf396deSTao Ma 56784e40080SDarrick J. Wong BUG_ON(ocfs2_is_refcount_inode(inode)); 5688bf396deSTao Ma 569198aac28STao Ma trace_ocfs2_create_refcount_tree( 570d324cd4cSpiaojun (unsigned long long)oi->ip_blkno); 5718bf396deSTao Ma 5728bf396deSTao Ma ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac); 5738bf396deSTao Ma if (ret) { 5748bf396deSTao Ma mlog_errno(ret); 5758bf396deSTao Ma goto out; 5768bf396deSTao Ma } 5778bf396deSTao Ma 5788bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS); 5798bf396deSTao Ma if (IS_ERR(handle)) { 5808bf396deSTao Ma ret = PTR_ERR(handle); 5818bf396deSTao Ma mlog_errno(ret); 5828bf396deSTao Ma goto out; 5838bf396deSTao Ma } 5848bf396deSTao Ma 5858bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 5868bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 5878bf396deSTao Ma if (ret) { 5888bf396deSTao Ma mlog_errno(ret); 5898bf396deSTao Ma goto out_commit; 5908bf396deSTao Ma } 5918bf396deSTao Ma 5922b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 5938bf396deSTao Ma &suballoc_bit_start, &num_got, 5948bf396deSTao Ma &first_blkno); 5958bf396deSTao Ma if (ret) { 5968bf396deSTao Ma mlog_errno(ret); 5978bf396deSTao Ma goto out_commit; 5988bf396deSTao Ma } 5998bf396deSTao Ma 6008bf396deSTao Ma new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno); 6018bf396deSTao Ma if (!new_tree) { 6028bf396deSTao Ma ret = -ENOMEM; 6038bf396deSTao Ma mlog_errno(ret); 6048bf396deSTao Ma goto out_commit; 6058bf396deSTao Ma } 6068bf396deSTao Ma 6078bf396deSTao Ma new_bh = sb_getblk(inode->i_sb, first_blkno); 60858796207SRui Xiang if (!new_bh) { 60958796207SRui Xiang ret = -ENOMEM; 61058796207SRui Xiang mlog_errno(ret); 61158796207SRui Xiang goto out_commit; 61258796207SRui Xiang } 6138bf396deSTao Ma ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh); 6148bf396deSTao Ma 6158bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh, 6168bf396deSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 6178bf396deSTao Ma if (ret) { 6188bf396deSTao Ma mlog_errno(ret); 6198bf396deSTao Ma goto out_commit; 6208bf396deSTao Ma } 6218bf396deSTao Ma 6228bf396deSTao Ma /* Initialize ocfs2_refcount_block. */ 6238bf396deSTao Ma rb = (struct ocfs2_refcount_block *)new_bh->b_data; 6248bf396deSTao Ma memset(rb, 0, inode->i_sb->s_blocksize); 6258bf396deSTao Ma strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 626b89c5428STiger Yang rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 6272b6cb576SJoel Becker rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 6288bf396deSTao Ma rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 6298bf396deSTao Ma rb->rf_fs_generation = cpu_to_le32(osb->fs_generation); 6308bf396deSTao Ma rb->rf_blkno = cpu_to_le64(first_blkno); 6318bf396deSTao Ma rb->rf_count = cpu_to_le32(1); 6328bf396deSTao Ma rb->rf_records.rl_count = 6338bf396deSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb)); 6348bf396deSTao Ma spin_lock(&osb->osb_lock); 6358bf396deSTao Ma rb->rf_generation = osb->s_next_generation++; 6368bf396deSTao Ma spin_unlock(&osb->osb_lock); 6378bf396deSTao Ma 6388bf396deSTao Ma ocfs2_journal_dirty(handle, new_bh); 6398bf396deSTao Ma 6408bf396deSTao Ma spin_lock(&oi->ip_lock); 6418bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 6428bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 6438bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(first_blkno); 6448bf396deSTao Ma spin_unlock(&oi->ip_lock); 6458bf396deSTao Ma 646198aac28STao Ma trace_ocfs2_create_refcount_tree_blkno((unsigned long long)first_blkno); 6478bf396deSTao Ma 6488bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 6498bf396deSTao Ma 6508bf396deSTao Ma /* 6518bf396deSTao Ma * We have to init the tree lock here since it will use 6528bf396deSTao Ma * the generation number to create it. 6538bf396deSTao Ma */ 6548bf396deSTao Ma new_tree->rf_generation = le32_to_cpu(rb->rf_generation); 6558bf396deSTao Ma ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno, 6568bf396deSTao Ma new_tree->rf_generation); 6578bf396deSTao Ma 6588bf396deSTao Ma spin_lock(&osb->osb_lock); 6598bf396deSTao Ma tree = ocfs2_find_refcount_tree(osb, first_blkno); 6608bf396deSTao Ma 6618bf396deSTao Ma /* 6628bf396deSTao Ma * We've just created a new refcount tree in this block. If 6638bf396deSTao Ma * we found a refcount tree on the ocfs2_super, it must be 6648bf396deSTao Ma * one we just deleted. We free the old tree before 6658bf396deSTao Ma * inserting the new tree. 6668bf396deSTao Ma */ 6678bf396deSTao Ma BUG_ON(tree && tree->rf_generation == new_tree->rf_generation); 6688bf396deSTao Ma if (tree) 6698bf396deSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree); 6708bf396deSTao Ma ocfs2_insert_refcount_tree(osb, new_tree); 6718bf396deSTao Ma spin_unlock(&osb->osb_lock); 6728bf396deSTao Ma new_tree = NULL; 6738bf396deSTao Ma if (tree) 6748bf396deSTao Ma ocfs2_refcount_tree_put(tree); 6758bf396deSTao Ma 6768bf396deSTao Ma out_commit: 6778bf396deSTao Ma ocfs2_commit_trans(osb, handle); 6788bf396deSTao Ma 6798bf396deSTao Ma out: 6808bf396deSTao Ma if (new_tree) { 6818bf396deSTao Ma ocfs2_metadata_cache_exit(&new_tree->rf_ci); 6828bf396deSTao Ma kfree(new_tree); 6838bf396deSTao Ma } 6848bf396deSTao Ma 6858bf396deSTao Ma brelse(new_bh); 6868bf396deSTao Ma if (meta_ac) 6878bf396deSTao Ma ocfs2_free_alloc_context(meta_ac); 6888bf396deSTao Ma 6898bf396deSTao Ma return ret; 6908bf396deSTao Ma } 6918bf396deSTao Ma 6928bf396deSTao Ma static int ocfs2_set_refcount_tree(struct inode *inode, 6938bf396deSTao Ma struct buffer_head *di_bh, 6948bf396deSTao Ma u64 refcount_loc) 6958bf396deSTao Ma { 6968bf396deSTao Ma int ret; 6978bf396deSTao Ma handle_t *handle = NULL; 6988bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 6998bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7008bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7018bf396deSTao Ma struct buffer_head *ref_root_bh = NULL; 7028bf396deSTao Ma struct ocfs2_refcount_block *rb; 7038bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7048bf396deSTao Ma 70584e40080SDarrick J. Wong BUG_ON(ocfs2_is_refcount_inode(inode)); 7068bf396deSTao Ma 7078bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1, 7088bf396deSTao Ma &ref_tree, &ref_root_bh); 7098bf396deSTao Ma if (ret) { 7108bf396deSTao Ma mlog_errno(ret); 7118bf396deSTao Ma return ret; 7128bf396deSTao Ma } 7138bf396deSTao Ma 7148bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS); 7158bf396deSTao Ma if (IS_ERR(handle)) { 7168bf396deSTao Ma ret = PTR_ERR(handle); 7178bf396deSTao Ma mlog_errno(ret); 7188bf396deSTao Ma goto out; 7198bf396deSTao Ma } 7208bf396deSTao Ma 7218bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 7228bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7238bf396deSTao Ma if (ret) { 7248bf396deSTao Ma mlog_errno(ret); 7258bf396deSTao Ma goto out_commit; 7268bf396deSTao Ma } 7278bf396deSTao Ma 7288bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh, 7298bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7308bf396deSTao Ma if (ret) { 7318bf396deSTao Ma mlog_errno(ret); 7328bf396deSTao Ma goto out_commit; 7338bf396deSTao Ma } 7348bf396deSTao Ma 7358bf396deSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 7368bf396deSTao Ma le32_add_cpu(&rb->rf_count, 1); 7378bf396deSTao Ma 7388bf396deSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 7398bf396deSTao Ma 7408bf396deSTao Ma spin_lock(&oi->ip_lock); 7418bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 7428bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 7438bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(refcount_loc); 7448bf396deSTao Ma spin_unlock(&oi->ip_lock); 7458bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 7468bf396deSTao Ma 7478bf396deSTao Ma out_commit: 7488bf396deSTao Ma ocfs2_commit_trans(osb, handle); 7498bf396deSTao Ma out: 7508bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 7518bf396deSTao Ma brelse(ref_root_bh); 7528bf396deSTao Ma 7538bf396deSTao Ma return ret; 7548bf396deSTao Ma } 7558bf396deSTao Ma 7568bf396deSTao Ma int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh) 7578bf396deSTao Ma { 7588bf396deSTao Ma int ret, delete_tree = 0; 7598bf396deSTao Ma handle_t *handle = NULL; 7608bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7618bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7628bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7638bf396deSTao Ma struct ocfs2_refcount_block *rb; 7648bf396deSTao Ma struct inode *alloc_inode = NULL; 7658bf396deSTao Ma struct buffer_head *alloc_bh = NULL; 7668bf396deSTao Ma struct buffer_head *blk_bh = NULL; 7678bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7688bf396deSTao Ma int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS; 7698bf396deSTao Ma u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc); 7708bf396deSTao Ma u16 bit = 0; 7718bf396deSTao Ma 77284e40080SDarrick J. Wong if (!ocfs2_is_refcount_inode(inode)) 7738bf396deSTao Ma return 0; 7748bf396deSTao Ma 7758bf396deSTao Ma BUG_ON(!ref_blkno); 7768bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh); 7778bf396deSTao Ma if (ret) { 7788bf396deSTao Ma mlog_errno(ret); 7798bf396deSTao Ma return ret; 7808bf396deSTao Ma } 7818bf396deSTao Ma 7828bf396deSTao Ma rb = (struct ocfs2_refcount_block *)blk_bh->b_data; 7838bf396deSTao Ma 7848bf396deSTao Ma /* 7858bf396deSTao Ma * If we are the last user, we need to free the block. 7868bf396deSTao Ma * So lock the allocator ahead. 7878bf396deSTao Ma */ 7888bf396deSTao Ma if (le32_to_cpu(rb->rf_count) == 1) { 7898bf396deSTao Ma blk = le64_to_cpu(rb->rf_blkno); 7908bf396deSTao Ma bit = le16_to_cpu(rb->rf_suballoc_bit); 79174380c47STao Ma if (rb->rf_suballoc_loc) 79274380c47STao Ma bg_blkno = le64_to_cpu(rb->rf_suballoc_loc); 79374380c47STao Ma else 7948bf396deSTao Ma bg_blkno = ocfs2_which_suballoc_group(blk, bit); 7958bf396deSTao Ma 7968bf396deSTao Ma alloc_inode = ocfs2_get_system_file_inode(osb, 7978bf396deSTao Ma EXTENT_ALLOC_SYSTEM_INODE, 7988bf396deSTao Ma le16_to_cpu(rb->rf_suballoc_slot)); 7998bf396deSTao Ma if (!alloc_inode) { 8008bf396deSTao Ma ret = -ENOMEM; 8018bf396deSTao Ma mlog_errno(ret); 8028bf396deSTao Ma goto out; 8038bf396deSTao Ma } 8045955102cSAl Viro inode_lock(alloc_inode); 8058bf396deSTao Ma 8068bf396deSTao Ma ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1); 8078bf396deSTao Ma if (ret) { 8088bf396deSTao Ma mlog_errno(ret); 8098bf396deSTao Ma goto out_mutex; 8108bf396deSTao Ma } 8118bf396deSTao Ma 8128bf396deSTao Ma credits += OCFS2_SUBALLOC_FREE; 8138bf396deSTao Ma } 8148bf396deSTao Ma 8158bf396deSTao Ma handle = ocfs2_start_trans(osb, credits); 8168bf396deSTao Ma if (IS_ERR(handle)) { 8178bf396deSTao Ma ret = PTR_ERR(handle); 8188bf396deSTao Ma mlog_errno(ret); 8198bf396deSTao Ma goto out_unlock; 8208bf396deSTao Ma } 8218bf396deSTao Ma 8228bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 8238bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8248bf396deSTao Ma if (ret) { 8258bf396deSTao Ma mlog_errno(ret); 8268bf396deSTao Ma goto out_commit; 8278bf396deSTao Ma } 8288bf396deSTao Ma 8298bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh, 8308bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8318bf396deSTao Ma if (ret) { 8328bf396deSTao Ma mlog_errno(ret); 8338bf396deSTao Ma goto out_commit; 8348bf396deSTao Ma } 8358bf396deSTao Ma 8368bf396deSTao Ma spin_lock(&oi->ip_lock); 8378bf396deSTao Ma oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL; 8388bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 8398bf396deSTao Ma di->i_refcount_loc = 0; 8408bf396deSTao Ma spin_unlock(&oi->ip_lock); 8418bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 8428bf396deSTao Ma 8438bf396deSTao Ma le32_add_cpu(&rb->rf_count , -1); 8448bf396deSTao Ma ocfs2_journal_dirty(handle, blk_bh); 8458bf396deSTao Ma 8468bf396deSTao Ma if (!rb->rf_count) { 8478bf396deSTao Ma delete_tree = 1; 8488bf396deSTao Ma ocfs2_erase_refcount_tree_from_list(osb, ref_tree); 8498bf396deSTao Ma ret = ocfs2_free_suballoc_bits(handle, alloc_inode, 8508bf396deSTao Ma alloc_bh, bit, bg_blkno, 1); 8518bf396deSTao Ma if (ret) 8528bf396deSTao Ma mlog_errno(ret); 8538bf396deSTao Ma } 8548bf396deSTao Ma 8558bf396deSTao Ma out_commit: 8568bf396deSTao Ma ocfs2_commit_trans(osb, handle); 8578bf396deSTao Ma out_unlock: 8588bf396deSTao Ma if (alloc_inode) { 8598bf396deSTao Ma ocfs2_inode_unlock(alloc_inode, 1); 8608bf396deSTao Ma brelse(alloc_bh); 8618bf396deSTao Ma } 8628bf396deSTao Ma out_mutex: 8638bf396deSTao Ma if (alloc_inode) { 8645955102cSAl Viro inode_unlock(alloc_inode); 8658bf396deSTao Ma iput(alloc_inode); 8668bf396deSTao Ma } 8678bf396deSTao Ma out: 8688bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 8698bf396deSTao Ma if (delete_tree) 8708bf396deSTao Ma ocfs2_refcount_tree_put(ref_tree); 8718bf396deSTao Ma brelse(blk_bh); 8728bf396deSTao Ma 8738bf396deSTao Ma return ret; 8748bf396deSTao Ma } 875e73a819dSTao Ma 876e73a819dSTao Ma static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci, 877e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 878e73a819dSTao Ma u64 cpos, unsigned int len, 879e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 880e73a819dSTao Ma int *index) 881e73a819dSTao Ma { 882e73a819dSTao Ma int i = 0; 883e73a819dSTao Ma struct ocfs2_refcount_block *rb = 884e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 885e73a819dSTao Ma struct ocfs2_refcount_rec *rec = NULL; 886e73a819dSTao Ma 887e73a819dSTao Ma for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) { 888e73a819dSTao Ma rec = &rb->rf_records.rl_recs[i]; 889e73a819dSTao Ma 890e73a819dSTao Ma if (le64_to_cpu(rec->r_cpos) + 891e73a819dSTao Ma le32_to_cpu(rec->r_clusters) <= cpos) 892e73a819dSTao Ma continue; 893e73a819dSTao Ma else if (le64_to_cpu(rec->r_cpos) > cpos) 894e73a819dSTao Ma break; 895e73a819dSTao Ma 896e73a819dSTao Ma /* ok, cpos fail in this rec. Just return. */ 897e73a819dSTao Ma if (ret_rec) 898e73a819dSTao Ma *ret_rec = *rec; 899e73a819dSTao Ma goto out; 900e73a819dSTao Ma } 901e73a819dSTao Ma 902e73a819dSTao Ma if (ret_rec) { 903e73a819dSTao Ma /* We meet with a hole here, so fake the rec. */ 904e73a819dSTao Ma ret_rec->r_cpos = cpu_to_le64(cpos); 905e73a819dSTao Ma ret_rec->r_refcount = 0; 906e73a819dSTao Ma if (i < le16_to_cpu(rb->rf_records.rl_used) && 907e73a819dSTao Ma le64_to_cpu(rec->r_cpos) < cpos + len) 908e73a819dSTao Ma ret_rec->r_clusters = 909e73a819dSTao Ma cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos); 910e73a819dSTao Ma else 911e73a819dSTao Ma ret_rec->r_clusters = cpu_to_le32(len); 912e73a819dSTao Ma } 913e73a819dSTao Ma 914e73a819dSTao Ma out: 915e73a819dSTao Ma *index = i; 916e73a819dSTao Ma } 917e73a819dSTao Ma 918e73a819dSTao Ma /* 9198b2c0dbaSTao Ma * Try to remove refcount tree. The mechanism is: 9208b2c0dbaSTao Ma * 1) Check whether i_clusters == 0, if no, exit. 9218b2c0dbaSTao Ma * 2) check whether we have i_xattr_loc in dinode. if yes, exit. 9228b2c0dbaSTao Ma * 3) Check whether we have inline xattr stored outside, if yes, exit. 9238b2c0dbaSTao Ma * 4) Remove the tree. 9248b2c0dbaSTao Ma */ 9258b2c0dbaSTao Ma int ocfs2_try_remove_refcount_tree(struct inode *inode, 9268b2c0dbaSTao Ma struct buffer_head *di_bh) 9278b2c0dbaSTao Ma { 9288b2c0dbaSTao Ma int ret; 9298b2c0dbaSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 9308b2c0dbaSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 9318b2c0dbaSTao Ma 9328b2c0dbaSTao Ma down_write(&oi->ip_xattr_sem); 9338b2c0dbaSTao Ma down_write(&oi->ip_alloc_sem); 9348b2c0dbaSTao Ma 9358b2c0dbaSTao Ma if (oi->ip_clusters) 9368b2c0dbaSTao Ma goto out; 9378b2c0dbaSTao Ma 9388b2c0dbaSTao Ma if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc) 9398b2c0dbaSTao Ma goto out; 9408b2c0dbaSTao Ma 9418b2c0dbaSTao Ma if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL && 9428b2c0dbaSTao Ma ocfs2_has_inline_xattr_value_outside(inode, di)) 9438b2c0dbaSTao Ma goto out; 9448b2c0dbaSTao Ma 9458b2c0dbaSTao Ma ret = ocfs2_remove_refcount_tree(inode, di_bh); 9468b2c0dbaSTao Ma if (ret) 9478b2c0dbaSTao Ma mlog_errno(ret); 9488b2c0dbaSTao Ma out: 9498b2c0dbaSTao Ma up_write(&oi->ip_alloc_sem); 9508b2c0dbaSTao Ma up_write(&oi->ip_xattr_sem); 9518b2c0dbaSTao Ma return 0; 9528b2c0dbaSTao Ma } 9538b2c0dbaSTao Ma 9548b2c0dbaSTao Ma /* 95538a04e43STao Ma * Find the end range for a leaf refcount block indicated by 95638a04e43STao Ma * el->l_recs[index].e_blkno. 95738a04e43STao Ma */ 95838a04e43STao Ma static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci, 95938a04e43STao Ma struct buffer_head *ref_root_bh, 96038a04e43STao Ma struct ocfs2_extent_block *eb, 96138a04e43STao Ma struct ocfs2_extent_list *el, 96238a04e43STao Ma int index, u32 *cpos_end) 96338a04e43STao Ma { 96438a04e43STao Ma int ret, i, subtree_root; 96538a04e43STao Ma u32 cpos; 96638a04e43STao Ma u64 blkno; 96738a04e43STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 96838a04e43STao Ma struct ocfs2_path *left_path = NULL, *right_path = NULL; 96938a04e43STao Ma struct ocfs2_extent_tree et; 97038a04e43STao Ma struct ocfs2_extent_list *tmp_el; 97138a04e43STao Ma 97238a04e43STao Ma if (index < le16_to_cpu(el->l_next_free_rec) - 1) { 97338a04e43STao Ma /* 97438a04e43STao Ma * We have a extent rec after index, so just use the e_cpos 97538a04e43STao Ma * of the next extent rec. 97638a04e43STao Ma */ 97738a04e43STao Ma *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos); 97838a04e43STao Ma return 0; 97938a04e43STao Ma } 98038a04e43STao Ma 98138a04e43STao Ma if (!eb || (eb && !eb->h_next_leaf_blk)) { 98238a04e43STao Ma /* 98338a04e43STao Ma * We are the last extent rec, so any high cpos should 98438a04e43STao Ma * be stored in this leaf refcount block. 98538a04e43STao Ma */ 98638a04e43STao Ma *cpos_end = UINT_MAX; 98738a04e43STao Ma return 0; 98838a04e43STao Ma } 98938a04e43STao Ma 99038a04e43STao Ma /* 99138a04e43STao Ma * If the extent block isn't the last one, we have to find 99238a04e43STao Ma * the subtree root between this extent block and the next 99338a04e43STao Ma * leaf extent block and get the corresponding e_cpos from 99438a04e43STao Ma * the subroot. Otherwise we may corrupt the b-tree. 99538a04e43STao Ma */ 99638a04e43STao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 99738a04e43STao Ma 99838a04e43STao Ma left_path = ocfs2_new_path_from_et(&et); 99938a04e43STao Ma if (!left_path) { 100038a04e43STao Ma ret = -ENOMEM; 100138a04e43STao Ma mlog_errno(ret); 100238a04e43STao Ma goto out; 100338a04e43STao Ma } 100438a04e43STao Ma 100538a04e43STao Ma cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos); 100638a04e43STao Ma ret = ocfs2_find_path(ci, left_path, cpos); 100738a04e43STao Ma if (ret) { 100838a04e43STao Ma mlog_errno(ret); 100938a04e43STao Ma goto out; 101038a04e43STao Ma } 101138a04e43STao Ma 101238a04e43STao Ma right_path = ocfs2_new_path_from_path(left_path); 101338a04e43STao Ma if (!right_path) { 101438a04e43STao Ma ret = -ENOMEM; 101538a04e43STao Ma mlog_errno(ret); 101638a04e43STao Ma goto out; 101738a04e43STao Ma } 101838a04e43STao Ma 101938a04e43STao Ma ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos); 102038a04e43STao Ma if (ret) { 102138a04e43STao Ma mlog_errno(ret); 102238a04e43STao Ma goto out; 102338a04e43STao Ma } 102438a04e43STao Ma 102538a04e43STao Ma ret = ocfs2_find_path(ci, right_path, cpos); 102638a04e43STao Ma if (ret) { 102738a04e43STao Ma mlog_errno(ret); 102838a04e43STao Ma goto out; 102938a04e43STao Ma } 103038a04e43STao Ma 103138a04e43STao Ma subtree_root = ocfs2_find_subtree_root(&et, left_path, 103238a04e43STao Ma right_path); 103338a04e43STao Ma 103438a04e43STao Ma tmp_el = left_path->p_node[subtree_root].el; 103538a04e43STao Ma blkno = left_path->p_node[subtree_root+1].bh->b_blocknr; 10363a251f04SAl Viro for (i = 0; i < le16_to_cpu(tmp_el->l_next_free_rec); i++) { 103738a04e43STao Ma if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) { 103838a04e43STao Ma *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos); 103938a04e43STao Ma break; 104038a04e43STao Ma } 104138a04e43STao Ma } 104238a04e43STao Ma 10433a251f04SAl Viro BUG_ON(i == le16_to_cpu(tmp_el->l_next_free_rec)); 104438a04e43STao Ma 104538a04e43STao Ma out: 104638a04e43STao Ma ocfs2_free_path(left_path); 104738a04e43STao Ma ocfs2_free_path(right_path); 104838a04e43STao Ma return ret; 104938a04e43STao Ma } 105038a04e43STao Ma 105138a04e43STao Ma /* 1052e73a819dSTao Ma * Given a cpos and len, try to find the refcount record which contains cpos. 1053e73a819dSTao Ma * 1. If cpos can be found in one refcount record, return the record. 1054e73a819dSTao Ma * 2. If cpos can't be found, return a fake record which start from cpos 1055e73a819dSTao Ma * and end at a small value between cpos+len and start of the next record. 1056e73a819dSTao Ma * This fake record has r_refcount = 0. 1057e73a819dSTao Ma */ 1058e73a819dSTao Ma static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci, 1059e73a819dSTao Ma struct buffer_head *ref_root_bh, 1060e73a819dSTao Ma u64 cpos, unsigned int len, 1061e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 1062e73a819dSTao Ma int *index, 1063e73a819dSTao Ma struct buffer_head **ret_bh) 1064e73a819dSTao Ma { 1065e73a819dSTao Ma int ret = 0, i, found; 106638a04e43STao Ma u32 low_cpos, uninitialized_var(cpos_end); 1067e73a819dSTao Ma struct ocfs2_extent_list *el; 106838a04e43STao Ma struct ocfs2_extent_rec *rec = NULL; 106938a04e43STao Ma struct ocfs2_extent_block *eb = NULL; 1070e73a819dSTao Ma struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL; 1071e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1072e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1073e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1074e73a819dSTao Ma 1075e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) { 1076e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len, 1077e73a819dSTao Ma ret_rec, index); 1078e73a819dSTao Ma *ret_bh = ref_root_bh; 1079e73a819dSTao Ma get_bh(ref_root_bh); 1080e73a819dSTao Ma return 0; 1081e73a819dSTao Ma } 1082e73a819dSTao Ma 1083e73a819dSTao Ma el = &rb->rf_list; 1084e73a819dSTao Ma low_cpos = cpos & OCFS2_32BIT_POS_MASK; 1085e73a819dSTao Ma 1086e73a819dSTao Ma if (el->l_tree_depth) { 1087e73a819dSTao Ma ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh); 1088e73a819dSTao Ma if (ret) { 1089e73a819dSTao Ma mlog_errno(ret); 1090e73a819dSTao Ma goto out; 1091e73a819dSTao Ma } 1092e73a819dSTao Ma 1093e73a819dSTao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 1094e73a819dSTao Ma el = &eb->h_list; 1095e73a819dSTao Ma 1096e73a819dSTao Ma if (el->l_tree_depth) { 109717a5b9abSGoldwyn Rodrigues ret = ocfs2_error(sb, 10987ecef14aSJoe Perches "refcount tree %llu has non zero tree depth in leaf btree tree block %llu\n", 1099e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1100e73a819dSTao Ma (unsigned long long)eb_bh->b_blocknr); 1101e73a819dSTao Ma goto out; 1102e73a819dSTao Ma } 1103e73a819dSTao Ma } 1104e73a819dSTao Ma 1105e73a819dSTao Ma found = 0; 1106e73a819dSTao Ma for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) { 1107e73a819dSTao Ma rec = &el->l_recs[i]; 1108e73a819dSTao Ma 1109e73a819dSTao Ma if (le32_to_cpu(rec->e_cpos) <= low_cpos) { 1110e73a819dSTao Ma found = 1; 1111e73a819dSTao Ma break; 1112e73a819dSTao Ma } 1113e73a819dSTao Ma } 1114e73a819dSTao Ma 111538a04e43STao Ma if (found) { 111638a04e43STao Ma ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh, 111738a04e43STao Ma eb, el, i, &cpos_end); 111838a04e43STao Ma if (ret) { 111938a04e43STao Ma mlog_errno(ret); 112038a04e43STao Ma goto out; 112138a04e43STao Ma } 1122e73a819dSTao Ma 112338a04e43STao Ma if (cpos_end < low_cpos + len) 112438a04e43STao Ma len = cpos_end - low_cpos; 1125e73a819dSTao Ma } 1126e73a819dSTao Ma 1127e73a819dSTao Ma ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno), 1128e73a819dSTao Ma &ref_leaf_bh); 1129e73a819dSTao Ma if (ret) { 1130e73a819dSTao Ma mlog_errno(ret); 1131e73a819dSTao Ma goto out; 1132e73a819dSTao Ma } 1133e73a819dSTao Ma 1134e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len, 1135e73a819dSTao Ma ret_rec, index); 1136e73a819dSTao Ma *ret_bh = ref_leaf_bh; 1137e73a819dSTao Ma out: 1138e73a819dSTao Ma brelse(eb_bh); 1139e73a819dSTao Ma return ret; 1140e73a819dSTao Ma } 1141e73a819dSTao Ma 1142e73a819dSTao Ma enum ocfs2_ref_rec_contig { 1143e73a819dSTao Ma REF_CONTIG_NONE = 0, 1144e73a819dSTao Ma REF_CONTIG_LEFT, 1145e73a819dSTao Ma REF_CONTIG_RIGHT, 1146e73a819dSTao Ma REF_CONTIG_LEFTRIGHT, 1147e73a819dSTao Ma }; 1148e73a819dSTao Ma 1149e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1150e73a819dSTao Ma ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb, 1151e73a819dSTao Ma int index) 1152e73a819dSTao Ma { 1153e73a819dSTao Ma if ((rb->rf_records.rl_recs[index].r_refcount == 1154e73a819dSTao Ma rb->rf_records.rl_recs[index + 1].r_refcount) && 1155e73a819dSTao Ma (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) + 1156e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) == 1157e73a819dSTao Ma le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos))) 1158e73a819dSTao Ma return REF_CONTIG_RIGHT; 1159e73a819dSTao Ma 1160e73a819dSTao Ma return REF_CONTIG_NONE; 1161e73a819dSTao Ma } 1162e73a819dSTao Ma 1163e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1164e73a819dSTao Ma ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb, 1165e73a819dSTao Ma int index) 1166e73a819dSTao Ma { 1167e73a819dSTao Ma enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE; 1168e73a819dSTao Ma 1169e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 1) 1170e73a819dSTao Ma ret = ocfs2_refcount_rec_adjacent(rb, index); 1171e73a819dSTao Ma 1172e73a819dSTao Ma if (index > 0) { 1173e73a819dSTao Ma enum ocfs2_ref_rec_contig tmp; 1174e73a819dSTao Ma 1175e73a819dSTao Ma tmp = ocfs2_refcount_rec_adjacent(rb, index - 1); 1176e73a819dSTao Ma 1177e73a819dSTao Ma if (tmp == REF_CONTIG_RIGHT) { 1178e73a819dSTao Ma if (ret == REF_CONTIG_RIGHT) 1179e73a819dSTao Ma ret = REF_CONTIG_LEFTRIGHT; 1180e73a819dSTao Ma else 1181e73a819dSTao Ma ret = REF_CONTIG_LEFT; 1182e73a819dSTao Ma } 1183e73a819dSTao Ma } 1184e73a819dSTao Ma 1185e73a819dSTao Ma return ret; 1186e73a819dSTao Ma } 1187e73a819dSTao Ma 1188e73a819dSTao Ma static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb, 1189e73a819dSTao Ma int index) 1190e73a819dSTao Ma { 1191e73a819dSTao Ma BUG_ON(rb->rf_records.rl_recs[index].r_refcount != 1192e73a819dSTao Ma rb->rf_records.rl_recs[index+1].r_refcount); 1193e73a819dSTao Ma 1194e73a819dSTao Ma le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters, 1195e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters)); 1196e73a819dSTao Ma 1197e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 2) 1198e73a819dSTao Ma memmove(&rb->rf_records.rl_recs[index + 1], 1199e73a819dSTao Ma &rb->rf_records.rl_recs[index + 2], 1200e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec) * 1201e73a819dSTao Ma (le16_to_cpu(rb->rf_records.rl_used) - index - 2)); 1202e73a819dSTao Ma 1203e73a819dSTao Ma memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1], 1204e73a819dSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 1205e73a819dSTao Ma le16_add_cpu(&rb->rf_records.rl_used, -1); 1206e73a819dSTao Ma } 1207e73a819dSTao Ma 1208e73a819dSTao Ma /* 1209e73a819dSTao Ma * Merge the refcount rec if we are contiguous with the adjacent recs. 1210e73a819dSTao Ma */ 1211e73a819dSTao Ma static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb, 1212e73a819dSTao Ma int index) 1213e73a819dSTao Ma { 1214e73a819dSTao Ma enum ocfs2_ref_rec_contig contig = 1215e73a819dSTao Ma ocfs2_refcount_rec_contig(rb, index); 1216e73a819dSTao Ma 1217e73a819dSTao Ma if (contig == REF_CONTIG_NONE) 1218e73a819dSTao Ma return; 1219e73a819dSTao Ma 1220e73a819dSTao Ma if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) { 1221e73a819dSTao Ma BUG_ON(index == 0); 1222e73a819dSTao Ma index--; 1223e73a819dSTao Ma } 1224e73a819dSTao Ma 1225e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1226e73a819dSTao Ma 1227e73a819dSTao Ma if (contig == REF_CONTIG_LEFTRIGHT) 1228e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1229e73a819dSTao Ma } 1230e73a819dSTao Ma 12311823cb0bSTao Ma /* 12321823cb0bSTao Ma * Change the refcount indexed by "index" in ref_bh. 12331823cb0bSTao Ma * If refcount reaches 0, remove it. 12341823cb0bSTao Ma */ 1235e73a819dSTao Ma static int ocfs2_change_refcount_rec(handle_t *handle, 1236e73a819dSTao Ma struct ocfs2_caching_info *ci, 1237e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 12387540c1a7STao Ma int index, int merge, int change) 1239e73a819dSTao Ma { 1240e73a819dSTao Ma int ret; 1241e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1242e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 12431823cb0bSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 12441823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rl->rl_recs[index]; 1245e73a819dSTao Ma 1246e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1247e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1248e73a819dSTao Ma if (ret) { 1249e73a819dSTao Ma mlog_errno(ret); 1250e73a819dSTao Ma goto out; 1251e73a819dSTao Ma } 1252e73a819dSTao Ma 1253198aac28STao Ma trace_ocfs2_change_refcount_rec( 1254198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1255198aac28STao Ma index, le32_to_cpu(rec->r_refcount), change); 1256e73a819dSTao Ma le32_add_cpu(&rec->r_refcount, change); 1257e73a819dSTao Ma 12581823cb0bSTao Ma if (!rec->r_refcount) { 12591823cb0bSTao Ma if (index != le16_to_cpu(rl->rl_used) - 1) { 12601823cb0bSTao Ma memmove(rec, rec + 1, 12611823cb0bSTao Ma (le16_to_cpu(rl->rl_used) - index - 1) * 12621823cb0bSTao Ma sizeof(struct ocfs2_refcount_rec)); 12631823cb0bSTao Ma memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1], 12641823cb0bSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 12651823cb0bSTao Ma } 12661823cb0bSTao Ma 12671823cb0bSTao Ma le16_add_cpu(&rl->rl_used, -1); 12687540c1a7STao Ma } else if (merge) 1269e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1270e73a819dSTao Ma 1271ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1272e73a819dSTao Ma out: 1273e73a819dSTao Ma return ret; 1274e73a819dSTao Ma } 1275e73a819dSTao Ma 1276e73a819dSTao Ma static int ocfs2_expand_inline_ref_root(handle_t *handle, 1277e73a819dSTao Ma struct ocfs2_caching_info *ci, 1278e73a819dSTao Ma struct buffer_head *ref_root_bh, 1279e73a819dSTao Ma struct buffer_head **ref_leaf_bh, 1280e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1281e73a819dSTao Ma { 1282e73a819dSTao Ma int ret; 1283e73a819dSTao Ma u16 suballoc_bit_start; 1284e73a819dSTao Ma u32 num_got; 12852b6cb576SJoel Becker u64 suballoc_loc, blkno; 1286e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1287e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1288e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1289e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1290e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1291e73a819dSTao Ma 1292e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1293e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1294e73a819dSTao Ma if (ret) { 1295e73a819dSTao Ma mlog_errno(ret); 1296e73a819dSTao Ma goto out; 1297e73a819dSTao Ma } 1298e73a819dSTao Ma 12992b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 1300e73a819dSTao Ma &suballoc_bit_start, &num_got, 1301e73a819dSTao Ma &blkno); 1302e73a819dSTao Ma if (ret) { 1303e73a819dSTao Ma mlog_errno(ret); 1304e73a819dSTao Ma goto out; 1305e73a819dSTao Ma } 1306e73a819dSTao Ma 1307e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1308e73a819dSTao Ma if (new_bh == NULL) { 13097391a294SRui Xiang ret = -ENOMEM; 1310e73a819dSTao Ma mlog_errno(ret); 1311e73a819dSTao Ma goto out; 1312e73a819dSTao Ma } 1313e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1314e73a819dSTao Ma 1315e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1316e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1317e73a819dSTao Ma if (ret) { 1318e73a819dSTao Ma mlog_errno(ret); 1319e73a819dSTao Ma goto out; 1320e73a819dSTao Ma } 1321e73a819dSTao Ma 1322e73a819dSTao Ma /* 1323e73a819dSTao Ma * Initialize ocfs2_refcount_block. 1324e73a819dSTao Ma * It should contain the same information as the old root. 1325e73a819dSTao Ma * so just memcpy it and change the corresponding field. 1326e73a819dSTao Ma */ 1327e73a819dSTao Ma memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize); 1328e73a819dSTao Ma 1329e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1330b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 13312b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 1332e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1333e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1334e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(0); 1335e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1336e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1337e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1338e73a819dSTao Ma 1339e73a819dSTao Ma /* Now change the root. */ 1340e73a819dSTao Ma memset(&root_rb->rf_list, 0, sb->s_blocksize - 1341e73a819dSTao Ma offsetof(struct ocfs2_refcount_block, rf_list)); 1342e73a819dSTao Ma root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb)); 1343e73a819dSTao Ma root_rb->rf_clusters = cpu_to_le32(1); 1344e73a819dSTao Ma root_rb->rf_list.l_next_free_rec = cpu_to_le16(1); 1345e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno); 1346e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1); 1347e73a819dSTao Ma root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL); 1348e73a819dSTao Ma 1349e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 1350e73a819dSTao Ma 1351198aac28STao Ma trace_ocfs2_expand_inline_ref_root((unsigned long long)blkno, 1352e73a819dSTao Ma le16_to_cpu(new_rb->rf_records.rl_used)); 1353e73a819dSTao Ma 1354e73a819dSTao Ma *ref_leaf_bh = new_bh; 1355e73a819dSTao Ma new_bh = NULL; 1356e73a819dSTao Ma out: 1357e73a819dSTao Ma brelse(new_bh); 1358e73a819dSTao Ma return ret; 1359e73a819dSTao Ma } 1360e73a819dSTao Ma 1361e73a819dSTao Ma static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev, 1362e73a819dSTao Ma struct ocfs2_refcount_rec *next) 1363e73a819dSTao Ma { 1364e73a819dSTao Ma if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <= 1365e73a819dSTao Ma ocfs2_get_ref_rec_low_cpos(next)) 1366e73a819dSTao Ma return 1; 1367e73a819dSTao Ma 1368e73a819dSTao Ma return 0; 1369e73a819dSTao Ma } 1370e73a819dSTao Ma 1371e73a819dSTao Ma static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b) 1372e73a819dSTao Ma { 1373e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1374e73a819dSTao Ma u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l); 1375e73a819dSTao Ma u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r); 1376e73a819dSTao Ma 1377e73a819dSTao Ma if (l_cpos > r_cpos) 1378e73a819dSTao Ma return 1; 1379e73a819dSTao Ma if (l_cpos < r_cpos) 1380e73a819dSTao Ma return -1; 1381e73a819dSTao Ma return 0; 1382e73a819dSTao Ma } 1383e73a819dSTao Ma 1384e73a819dSTao Ma static int cmp_refcount_rec_by_cpos(const void *a, const void *b) 1385e73a819dSTao Ma { 1386e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1387e73a819dSTao Ma u64 l_cpos = le64_to_cpu(l->r_cpos); 1388e73a819dSTao Ma u64 r_cpos = le64_to_cpu(r->r_cpos); 1389e73a819dSTao Ma 1390e73a819dSTao Ma if (l_cpos > r_cpos) 1391e73a819dSTao Ma return 1; 1392e73a819dSTao Ma if (l_cpos < r_cpos) 1393e73a819dSTao Ma return -1; 1394e73a819dSTao Ma return 0; 1395e73a819dSTao Ma } 1396e73a819dSTao Ma 1397e73a819dSTao Ma static void swap_refcount_rec(void *a, void *b, int size) 1398e73a819dSTao Ma { 1399a612543fSFabian Frederick struct ocfs2_refcount_rec *l = a, *r = b; 1400e73a819dSTao Ma 1401a612543fSFabian Frederick swap(*l, *r); 1402e73a819dSTao Ma } 1403e73a819dSTao Ma 1404e73a819dSTao Ma /* 1405e73a819dSTao Ma * The refcount cpos are ordered by their 64bit cpos, 1406e73a819dSTao Ma * But we will use the low 32 bit to be the e_cpos in the b-tree. 1407e73a819dSTao Ma * So we need to make sure that this pos isn't intersected with others. 1408e73a819dSTao Ma * 1409e73a819dSTao Ma * Note: The refcount block is already sorted by their low 32 bit cpos, 1410e73a819dSTao Ma * So just try the middle pos first, and we will exit when we find 1411e73a819dSTao Ma * the good position. 1412e73a819dSTao Ma */ 1413e73a819dSTao Ma static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl, 1414e73a819dSTao Ma u32 *split_pos, int *split_index) 1415e73a819dSTao Ma { 1416e73a819dSTao Ma int num_used = le16_to_cpu(rl->rl_used); 1417e73a819dSTao Ma int delta, middle = num_used / 2; 1418e73a819dSTao Ma 1419e73a819dSTao Ma for (delta = 0; delta < middle; delta++) { 1420e73a819dSTao Ma /* Let's check delta earlier than middle */ 1421e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1422e73a819dSTao Ma &rl->rl_recs[middle - delta - 1], 1423e73a819dSTao Ma &rl->rl_recs[middle - delta])) { 1424e73a819dSTao Ma *split_index = middle - delta; 1425e73a819dSTao Ma break; 1426e73a819dSTao Ma } 1427e73a819dSTao Ma 1428e73a819dSTao Ma /* For even counts, don't walk off the end */ 1429e73a819dSTao Ma if ((middle + delta + 1) == num_used) 1430e73a819dSTao Ma continue; 1431e73a819dSTao Ma 1432e73a819dSTao Ma /* Now try delta past middle */ 1433e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1434e73a819dSTao Ma &rl->rl_recs[middle + delta], 1435e73a819dSTao Ma &rl->rl_recs[middle + delta + 1])) { 1436e73a819dSTao Ma *split_index = middle + delta + 1; 1437e73a819dSTao Ma break; 1438e73a819dSTao Ma } 1439e73a819dSTao Ma } 1440e73a819dSTao Ma 1441e73a819dSTao Ma if (delta >= middle) 1442e73a819dSTao Ma return -ENOSPC; 1443e73a819dSTao Ma 1444e73a819dSTao Ma *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]); 1445e73a819dSTao Ma return 0; 1446e73a819dSTao Ma } 1447e73a819dSTao Ma 1448e73a819dSTao Ma static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh, 1449e73a819dSTao Ma struct buffer_head *new_bh, 1450e73a819dSTao Ma u32 *split_cpos) 1451e73a819dSTao Ma { 1452e73a819dSTao Ma int split_index = 0, num_moved, ret; 1453e73a819dSTao Ma u32 cpos = 0; 1454e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1455e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1456e73a819dSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 1457e73a819dSTao Ma struct ocfs2_refcount_block *new_rb = 1458e73a819dSTao Ma (struct ocfs2_refcount_block *)new_bh->b_data; 1459e73a819dSTao Ma struct ocfs2_refcount_list *new_rl = &new_rb->rf_records; 1460e73a819dSTao Ma 1461198aac28STao Ma trace_ocfs2_divide_leaf_refcount_block( 1462e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, 146328748b32SAl Viro le16_to_cpu(rl->rl_count), le16_to_cpu(rl->rl_used)); 1464e73a819dSTao Ma 1465e73a819dSTao Ma /* 1466e73a819dSTao Ma * XXX: Improvement later. 1467e73a819dSTao Ma * If we know all the high 32 bit cpos is the same, no need to sort. 1468e73a819dSTao Ma * 1469e73a819dSTao Ma * In order to make the whole process safe, we do: 1470e73a819dSTao Ma * 1. sort the entries by their low 32 bit cpos first so that we can 1471e73a819dSTao Ma * find the split cpos easily. 1472e73a819dSTao Ma * 2. call ocfs2_insert_extent to insert the new refcount block. 1473e73a819dSTao Ma * 3. move the refcount rec to the new block. 1474e73a819dSTao Ma * 4. sort the entries by their 64 bit cpos. 1475e73a819dSTao Ma * 5. dirty the new_rb and rb. 1476e73a819dSTao Ma */ 1477e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1478e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1479e73a819dSTao Ma cmp_refcount_rec_by_low_cpos, swap_refcount_rec); 1480e73a819dSTao Ma 1481e73a819dSTao Ma ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index); 1482e73a819dSTao Ma if (ret) { 1483e73a819dSTao Ma mlog_errno(ret); 1484e73a819dSTao Ma return ret; 1485e73a819dSTao Ma } 1486e73a819dSTao Ma 1487e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(cpos); 1488e73a819dSTao Ma 1489e73a819dSTao Ma /* move refcount records starting from split_index to the new block. */ 1490e73a819dSTao Ma num_moved = le16_to_cpu(rl->rl_used) - split_index; 1491e73a819dSTao Ma memcpy(new_rl->rl_recs, &rl->rl_recs[split_index], 1492e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1493e73a819dSTao Ma 1494e73a819dSTao Ma /*ok, remove the entries we just moved over to the other block. */ 1495e73a819dSTao Ma memset(&rl->rl_recs[split_index], 0, 1496e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1497e73a819dSTao Ma 1498e73a819dSTao Ma /* change old and new rl_used accordingly. */ 1499e73a819dSTao Ma le16_add_cpu(&rl->rl_used, -num_moved); 150012d4cec9STao Ma new_rl->rl_used = cpu_to_le16(num_moved); 1501e73a819dSTao Ma 1502e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1503e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1504e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1505e73a819dSTao Ma 1506e73a819dSTao Ma sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used), 1507e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1508e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1509e73a819dSTao Ma 1510e73a819dSTao Ma *split_cpos = cpos; 1511e73a819dSTao Ma return 0; 1512e73a819dSTao Ma } 1513e73a819dSTao Ma 1514e73a819dSTao Ma static int ocfs2_new_leaf_refcount_block(handle_t *handle, 1515e73a819dSTao Ma struct ocfs2_caching_info *ci, 1516e73a819dSTao Ma struct buffer_head *ref_root_bh, 1517e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1518e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1519e73a819dSTao Ma { 1520e73a819dSTao Ma int ret; 1521e73a819dSTao Ma u16 suballoc_bit_start; 1522e73a819dSTao Ma u32 num_got, new_cpos; 15232b6cb576SJoel Becker u64 suballoc_loc, blkno; 1524e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1525e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1526e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1527e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1528e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1529e73a819dSTao Ma struct ocfs2_extent_tree ref_et; 1530e73a819dSTao Ma 1531e73a819dSTao Ma BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)); 1532e73a819dSTao Ma 1533e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1534e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1535e73a819dSTao Ma if (ret) { 1536e73a819dSTao Ma mlog_errno(ret); 1537e73a819dSTao Ma goto out; 1538e73a819dSTao Ma } 1539e73a819dSTao Ma 1540e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1541e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1542e73a819dSTao Ma if (ret) { 1543e73a819dSTao Ma mlog_errno(ret); 1544e73a819dSTao Ma goto out; 1545e73a819dSTao Ma } 1546e73a819dSTao Ma 15472b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 1548e73a819dSTao Ma &suballoc_bit_start, &num_got, 1549e73a819dSTao Ma &blkno); 1550e73a819dSTao Ma if (ret) { 1551e73a819dSTao Ma mlog_errno(ret); 1552e73a819dSTao Ma goto out; 1553e73a819dSTao Ma } 1554e73a819dSTao Ma 1555e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1556e73a819dSTao Ma if (new_bh == NULL) { 15577391a294SRui Xiang ret = -ENOMEM; 1558e73a819dSTao Ma mlog_errno(ret); 1559e73a819dSTao Ma goto out; 1560e73a819dSTao Ma } 1561e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1562e73a819dSTao Ma 1563e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1564e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1565e73a819dSTao Ma if (ret) { 1566e73a819dSTao Ma mlog_errno(ret); 1567e73a819dSTao Ma goto out; 1568e73a819dSTao Ma } 1569e73a819dSTao Ma 1570e73a819dSTao Ma /* Initialize ocfs2_refcount_block. */ 1571e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1572e73a819dSTao Ma memset(new_rb, 0, sb->s_blocksize); 1573e73a819dSTao Ma strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 1574b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 15752b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 1576e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1577e73a819dSTao Ma new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation); 1578e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1579e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1580e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1581e73a819dSTao Ma new_rb->rf_records.rl_count = 1582e73a819dSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 1583e73a819dSTao Ma new_rb->rf_generation = root_rb->rf_generation; 1584e73a819dSTao Ma 1585e73a819dSTao Ma ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos); 1586e73a819dSTao Ma if (ret) { 1587e73a819dSTao Ma mlog_errno(ret); 1588e73a819dSTao Ma goto out; 1589e73a819dSTao Ma } 1590e73a819dSTao Ma 1591e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1592e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1593e73a819dSTao Ma 1594e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh); 1595e73a819dSTao Ma 1596198aac28STao Ma trace_ocfs2_new_leaf_refcount_block( 1597e73a819dSTao Ma (unsigned long long)new_bh->b_blocknr, new_cpos); 1598e73a819dSTao Ma 1599e73a819dSTao Ma /* Insert the new leaf block with the specific offset cpos. */ 1600e73a819dSTao Ma ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr, 1601e73a819dSTao Ma 1, 0, meta_ac); 1602e73a819dSTao Ma if (ret) 1603e73a819dSTao Ma mlog_errno(ret); 1604e73a819dSTao Ma 1605e73a819dSTao Ma out: 1606e73a819dSTao Ma brelse(new_bh); 1607e73a819dSTao Ma return ret; 1608e73a819dSTao Ma } 1609e73a819dSTao Ma 1610e73a819dSTao Ma static int ocfs2_expand_refcount_tree(handle_t *handle, 1611e73a819dSTao Ma struct ocfs2_caching_info *ci, 1612e73a819dSTao Ma struct buffer_head *ref_root_bh, 1613e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1614e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1615e73a819dSTao Ma { 1616e73a819dSTao Ma int ret; 1617e73a819dSTao Ma struct buffer_head *expand_bh = NULL; 1618e73a819dSTao Ma 1619e73a819dSTao Ma if (ref_root_bh == ref_leaf_bh) { 1620e73a819dSTao Ma /* 1621e73a819dSTao Ma * the old root bh hasn't been expanded to a b-tree, 1622e73a819dSTao Ma * so expand it first. 1623e73a819dSTao Ma */ 1624e73a819dSTao Ma ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh, 1625e73a819dSTao Ma &expand_bh, meta_ac); 1626e73a819dSTao Ma if (ret) { 1627e73a819dSTao Ma mlog_errno(ret); 1628e73a819dSTao Ma goto out; 1629e73a819dSTao Ma } 1630e73a819dSTao Ma } else { 1631e73a819dSTao Ma expand_bh = ref_leaf_bh; 1632e73a819dSTao Ma get_bh(expand_bh); 1633e73a819dSTao Ma } 1634e73a819dSTao Ma 1635e73a819dSTao Ma 1636e73a819dSTao Ma /* Now add a new refcount block into the tree.*/ 1637e73a819dSTao Ma ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh, 1638e73a819dSTao Ma expand_bh, meta_ac); 1639e73a819dSTao Ma if (ret) 1640e73a819dSTao Ma mlog_errno(ret); 1641e73a819dSTao Ma out: 1642e73a819dSTao Ma brelse(expand_bh); 1643e73a819dSTao Ma return ret; 1644e73a819dSTao Ma } 1645e73a819dSTao Ma 1646e73a819dSTao Ma /* 1647e73a819dSTao Ma * Adjust the extent rec in b-tree representing ref_leaf_bh. 1648e73a819dSTao Ma * 1649e73a819dSTao Ma * Only called when we have inserted a new refcount rec at index 0 1650e73a819dSTao Ma * which means ocfs2_extent_rec.e_cpos may need some change. 1651e73a819dSTao Ma */ 1652e73a819dSTao Ma static int ocfs2_adjust_refcount_rec(handle_t *handle, 1653e73a819dSTao Ma struct ocfs2_caching_info *ci, 1654e73a819dSTao Ma struct buffer_head *ref_root_bh, 1655e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1656e73a819dSTao Ma struct ocfs2_refcount_rec *rec) 1657e73a819dSTao Ma { 1658e73a819dSTao Ma int ret = 0, i; 1659e73a819dSTao Ma u32 new_cpos, old_cpos; 1660e73a819dSTao Ma struct ocfs2_path *path = NULL; 1661e73a819dSTao Ma struct ocfs2_extent_tree et; 1662e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1663e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1664e73a819dSTao Ma struct ocfs2_extent_list *el; 1665e73a819dSTao Ma 1666e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) 1667e73a819dSTao Ma goto out; 1668e73a819dSTao Ma 1669e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1670e73a819dSTao Ma old_cpos = le32_to_cpu(rb->rf_cpos); 1671e73a819dSTao Ma new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK; 1672e73a819dSTao Ma if (old_cpos <= new_cpos) 1673e73a819dSTao Ma goto out; 1674e73a819dSTao Ma 1675e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 1676e73a819dSTao Ma 1677e73a819dSTao Ma path = ocfs2_new_path_from_et(&et); 1678e73a819dSTao Ma if (!path) { 1679e73a819dSTao Ma ret = -ENOMEM; 1680e73a819dSTao Ma mlog_errno(ret); 1681e73a819dSTao Ma goto out; 1682e73a819dSTao Ma } 1683e73a819dSTao Ma 1684e73a819dSTao Ma ret = ocfs2_find_path(ci, path, old_cpos); 1685e73a819dSTao Ma if (ret) { 1686e73a819dSTao Ma mlog_errno(ret); 1687e73a819dSTao Ma goto out; 1688e73a819dSTao Ma } 1689e73a819dSTao Ma 1690e73a819dSTao Ma /* 1691e73a819dSTao Ma * 2 more credits, one for the leaf refcount block, one for 1692e73a819dSTao Ma * the extent block contains the extent rec. 1693e73a819dSTao Ma */ 1694c901fb00STao Ma ret = ocfs2_extend_trans(handle, 2); 1695e73a819dSTao Ma if (ret < 0) { 1696e73a819dSTao Ma mlog_errno(ret); 1697e73a819dSTao Ma goto out; 1698e73a819dSTao Ma } 1699e73a819dSTao Ma 1700e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1701e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1702e73a819dSTao Ma if (ret < 0) { 1703e73a819dSTao Ma mlog_errno(ret); 1704e73a819dSTao Ma goto out; 1705e73a819dSTao Ma } 1706e73a819dSTao Ma 1707e73a819dSTao Ma ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path), 1708e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1709e73a819dSTao Ma if (ret < 0) { 1710e73a819dSTao Ma mlog_errno(ret); 1711e73a819dSTao Ma goto out; 1712e73a819dSTao Ma } 1713e73a819dSTao Ma 1714e73a819dSTao Ma /* change the leaf extent block first. */ 1715e73a819dSTao Ma el = path_leaf_el(path); 1716e73a819dSTao Ma 1717e73a819dSTao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) 1718e73a819dSTao Ma if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos) 1719e73a819dSTao Ma break; 1720e73a819dSTao Ma 1721e73a819dSTao Ma BUG_ON(i == le16_to_cpu(el->l_next_free_rec)); 1722e73a819dSTao Ma 1723e73a819dSTao Ma el->l_recs[i].e_cpos = cpu_to_le32(new_cpos); 1724e73a819dSTao Ma 1725e73a819dSTao Ma /* change the r_cpos in the leaf block. */ 1726e73a819dSTao Ma rb->rf_cpos = cpu_to_le32(new_cpos); 1727e73a819dSTao Ma 1728e73a819dSTao Ma ocfs2_journal_dirty(handle, path_leaf_bh(path)); 1729e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1730e73a819dSTao Ma 1731e73a819dSTao Ma out: 1732e73a819dSTao Ma ocfs2_free_path(path); 1733e73a819dSTao Ma return ret; 1734e73a819dSTao Ma } 1735e73a819dSTao Ma 1736e73a819dSTao Ma static int ocfs2_insert_refcount_rec(handle_t *handle, 1737e73a819dSTao Ma struct ocfs2_caching_info *ci, 1738e73a819dSTao Ma struct buffer_head *ref_root_bh, 1739e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1740e73a819dSTao Ma struct ocfs2_refcount_rec *rec, 17417540c1a7STao Ma int index, int merge, 1742e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1743e73a819dSTao Ma { 1744e73a819dSTao Ma int ret; 1745e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1746e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1747e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1748e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1749e73a819dSTao Ma 1750e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1751e73a819dSTao Ma 1752e73a819dSTao Ma if (rf_list->rl_used == rf_list->rl_count) { 1753e73a819dSTao Ma u64 cpos = le64_to_cpu(rec->r_cpos); 1754e73a819dSTao Ma u32 len = le32_to_cpu(rec->r_clusters); 1755e73a819dSTao Ma 1756e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1757e73a819dSTao Ma ref_leaf_bh, meta_ac); 1758e73a819dSTao Ma if (ret) { 1759e73a819dSTao Ma mlog_errno(ret); 1760e73a819dSTao Ma goto out; 1761e73a819dSTao Ma } 1762e73a819dSTao Ma 1763e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1764e73a819dSTao Ma cpos, len, NULL, &index, 1765e73a819dSTao Ma &new_bh); 1766e73a819dSTao Ma if (ret) { 1767e73a819dSTao Ma mlog_errno(ret); 1768e73a819dSTao Ma goto out; 1769e73a819dSTao Ma } 1770e73a819dSTao Ma 1771e73a819dSTao Ma ref_leaf_bh = new_bh; 1772e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1773e73a819dSTao Ma rf_list = &rb->rf_records; 1774e73a819dSTao Ma } 1775e73a819dSTao Ma 1776e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1777e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1778e73a819dSTao Ma if (ret) { 1779e73a819dSTao Ma mlog_errno(ret); 1780e73a819dSTao Ma goto out; 1781e73a819dSTao Ma } 1782e73a819dSTao Ma 1783e73a819dSTao Ma if (index < le16_to_cpu(rf_list->rl_used)) 1784e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1], 1785e73a819dSTao Ma &rf_list->rl_recs[index], 1786e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index) * 1787e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1788e73a819dSTao Ma 1789198aac28STao Ma trace_ocfs2_insert_refcount_rec( 1790198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index, 1791e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos), 1792198aac28STao Ma le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount)); 1793e73a819dSTao Ma 1794e73a819dSTao Ma rf_list->rl_recs[index] = *rec; 1795e73a819dSTao Ma 1796e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, 1); 1797e73a819dSTao Ma 17987540c1a7STao Ma if (merge) 1799e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1800e73a819dSTao Ma 1801ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1802e73a819dSTao Ma 1803e73a819dSTao Ma if (index == 0) { 1804e73a819dSTao Ma ret = ocfs2_adjust_refcount_rec(handle, ci, 1805e73a819dSTao Ma ref_root_bh, 1806e73a819dSTao Ma ref_leaf_bh, rec); 1807e73a819dSTao Ma if (ret) 1808e73a819dSTao Ma mlog_errno(ret); 1809e73a819dSTao Ma } 1810e73a819dSTao Ma out: 1811e73a819dSTao Ma brelse(new_bh); 1812e73a819dSTao Ma return ret; 1813e73a819dSTao Ma } 1814e73a819dSTao Ma 1815e73a819dSTao Ma /* 1816e73a819dSTao Ma * Split the refcount_rec indexed by "index" in ref_leaf_bh. 1817e73a819dSTao Ma * This is much simple than our b-tree code. 1818e73a819dSTao Ma * split_rec is the new refcount rec we want to insert. 1819e73a819dSTao Ma * If split_rec->r_refcount > 0, we are changing the refcount(in case we 1820e73a819dSTao Ma * increase refcount or decrease a refcount to non-zero). 1821e73a819dSTao Ma * If split_rec->r_refcount == 0, we are punching a hole in current refcount 1822e73a819dSTao Ma * rec( in case we decrease a refcount to zero). 1823e73a819dSTao Ma */ 1824e73a819dSTao Ma static int ocfs2_split_refcount_rec(handle_t *handle, 1825e73a819dSTao Ma struct ocfs2_caching_info *ci, 1826e73a819dSTao Ma struct buffer_head *ref_root_bh, 1827e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1828e73a819dSTao Ma struct ocfs2_refcount_rec *split_rec, 18297540c1a7STao Ma int index, int merge, 1830e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1831e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1832e73a819dSTao Ma { 1833e73a819dSTao Ma int ret, recs_need; 1834e73a819dSTao Ma u32 len; 1835e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1836e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1837e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1838e73a819dSTao Ma struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index]; 1839e73a819dSTao Ma struct ocfs2_refcount_rec *tail_rec = NULL; 1840e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1841e73a819dSTao Ma 1842e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1843e73a819dSTao Ma 1844198aac28STao Ma trace_ocfs2_split_refcount_rec(le64_to_cpu(orig_rec->r_cpos), 1845198aac28STao Ma le32_to_cpu(orig_rec->r_clusters), 1846198aac28STao Ma le32_to_cpu(orig_rec->r_refcount), 1847e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos), 1848198aac28STao Ma le32_to_cpu(split_rec->r_clusters), 1849198aac28STao Ma le32_to_cpu(split_rec->r_refcount)); 1850e73a819dSTao Ma 1851e73a819dSTao Ma /* 1852e73a819dSTao Ma * If we just need to split the header or tail clusters, 1853e73a819dSTao Ma * no more recs are needed, just split is OK. 1854e73a819dSTao Ma * Otherwise we at least need one new recs. 1855e73a819dSTao Ma */ 1856e73a819dSTao Ma if (!split_rec->r_refcount && 1857e73a819dSTao Ma (split_rec->r_cpos == orig_rec->r_cpos || 1858e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1859e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) == 1860e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1861e73a819dSTao Ma recs_need = 0; 1862e73a819dSTao Ma else 1863e73a819dSTao Ma recs_need = 1; 1864e73a819dSTao Ma 1865e73a819dSTao Ma /* 1866e73a819dSTao Ma * We need one more rec if we split in the middle and the new rec have 1867e73a819dSTao Ma * some refcount in it. 1868e73a819dSTao Ma */ 1869e73a819dSTao Ma if (split_rec->r_refcount && 1870e73a819dSTao Ma (split_rec->r_cpos != orig_rec->r_cpos && 1871e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1872e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) != 1873e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1874e73a819dSTao Ma recs_need++; 1875e73a819dSTao Ma 1876e73a819dSTao Ma /* If the leaf block don't have enough record, expand it. */ 187712d4cec9STao Ma if (le16_to_cpu(rf_list->rl_used) + recs_need > 187812d4cec9STao Ma le16_to_cpu(rf_list->rl_count)) { 1879e73a819dSTao Ma struct ocfs2_refcount_rec tmp_rec; 1880e73a819dSTao Ma u64 cpos = le64_to_cpu(orig_rec->r_cpos); 1881e73a819dSTao Ma len = le32_to_cpu(orig_rec->r_clusters); 1882e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1883e73a819dSTao Ma ref_leaf_bh, meta_ac); 1884e73a819dSTao Ma if (ret) { 1885e73a819dSTao Ma mlog_errno(ret); 1886e73a819dSTao Ma goto out; 1887e73a819dSTao Ma } 1888e73a819dSTao Ma 1889e73a819dSTao Ma /* 1890e73a819dSTao Ma * We have to re-get it since now cpos may be moved to 1891e73a819dSTao Ma * another leaf block. 1892e73a819dSTao Ma */ 1893e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1894e73a819dSTao Ma cpos, len, &tmp_rec, &index, 1895e73a819dSTao Ma &new_bh); 1896e73a819dSTao Ma if (ret) { 1897e73a819dSTao Ma mlog_errno(ret); 1898e73a819dSTao Ma goto out; 1899e73a819dSTao Ma } 1900e73a819dSTao Ma 1901e73a819dSTao Ma ref_leaf_bh = new_bh; 1902e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1903e73a819dSTao Ma rf_list = &rb->rf_records; 1904e73a819dSTao Ma orig_rec = &rf_list->rl_recs[index]; 1905e73a819dSTao Ma } 1906e73a819dSTao Ma 1907e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1908e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1909e73a819dSTao Ma if (ret) { 1910e73a819dSTao Ma mlog_errno(ret); 1911e73a819dSTao Ma goto out; 1912e73a819dSTao Ma } 1913e73a819dSTao Ma 1914e73a819dSTao Ma /* 1915e73a819dSTao Ma * We have calculated out how many new records we need and store 1916e73a819dSTao Ma * in recs_need, so spare enough space first by moving the records 1917e73a819dSTao Ma * after "index" to the end. 1918e73a819dSTao Ma */ 1919e73a819dSTao Ma if (index != le16_to_cpu(rf_list->rl_used) - 1) 1920e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1 + recs_need], 1921e73a819dSTao Ma &rf_list->rl_recs[index + 1], 1922e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index - 1) * 1923e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1924e73a819dSTao Ma 1925e73a819dSTao Ma len = (le64_to_cpu(orig_rec->r_cpos) + 1926e73a819dSTao Ma le32_to_cpu(orig_rec->r_clusters)) - 1927e73a819dSTao Ma (le64_to_cpu(split_rec->r_cpos) + 1928e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters)); 1929e73a819dSTao Ma 1930e73a819dSTao Ma /* 1931e73a819dSTao Ma * If we have "len", the we will split in the tail and move it 1932e73a819dSTao Ma * to the end of the space we have just spared. 1933e73a819dSTao Ma */ 1934e73a819dSTao Ma if (len) { 1935e73a819dSTao Ma tail_rec = &rf_list->rl_recs[index + recs_need]; 1936e73a819dSTao Ma 1937e73a819dSTao Ma memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec)); 1938e73a819dSTao Ma le64_add_cpu(&tail_rec->r_cpos, 1939e73a819dSTao Ma le32_to_cpu(tail_rec->r_clusters) - len); 194012d4cec9STao Ma tail_rec->r_clusters = cpu_to_le32(len); 1941e73a819dSTao Ma } 1942e73a819dSTao Ma 1943e73a819dSTao Ma /* 1944e73a819dSTao Ma * If the split pos isn't the same as the original one, we need to 1945e73a819dSTao Ma * split in the head. 1946e73a819dSTao Ma * 1947e73a819dSTao Ma * Note: We have the chance that split_rec.r_refcount = 0, 1948e73a819dSTao Ma * recs_need = 0 and len > 0, which means we just cut the head from 1949e73a819dSTao Ma * the orig_rec and in that case we have done some modification in 1950e73a819dSTao Ma * orig_rec above, so the check for r_cpos is faked. 1951e73a819dSTao Ma */ 1952e73a819dSTao Ma if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) { 1953e73a819dSTao Ma len = le64_to_cpu(split_rec->r_cpos) - 1954e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos); 1955e73a819dSTao Ma orig_rec->r_clusters = cpu_to_le32(len); 1956e73a819dSTao Ma index++; 1957e73a819dSTao Ma } 1958e73a819dSTao Ma 1959e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, recs_need); 1960e73a819dSTao Ma 1961e73a819dSTao Ma if (split_rec->r_refcount) { 1962e73a819dSTao Ma rf_list->rl_recs[index] = *split_rec; 1963198aac28STao Ma trace_ocfs2_split_refcount_rec_insert( 1964198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index, 1965e73a819dSTao Ma (unsigned long long)le64_to_cpu(split_rec->r_cpos), 1966e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters), 1967198aac28STao Ma le32_to_cpu(split_rec->r_refcount)); 1968e73a819dSTao Ma 19697540c1a7STao Ma if (merge) 1970e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1971e73a819dSTao Ma } 1972e73a819dSTao Ma 1973ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1974e73a819dSTao Ma 1975e73a819dSTao Ma out: 1976e73a819dSTao Ma brelse(new_bh); 1977e73a819dSTao Ma return ret; 1978e73a819dSTao Ma } 1979e73a819dSTao Ma 19807540c1a7STao Ma static int __ocfs2_increase_refcount(handle_t *handle, 1981e73a819dSTao Ma struct ocfs2_caching_info *ci, 1982e73a819dSTao Ma struct buffer_head *ref_root_bh, 19837540c1a7STao Ma u64 cpos, u32 len, int merge, 1984e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1985e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1986e73a819dSTao Ma { 1987e73a819dSTao Ma int ret = 0, index; 1988e73a819dSTao Ma struct buffer_head *ref_leaf_bh = NULL; 1989e73a819dSTao Ma struct ocfs2_refcount_rec rec; 1990e73a819dSTao Ma unsigned int set_len = 0; 1991e73a819dSTao Ma 1992198aac28STao Ma trace_ocfs2_increase_refcount_begin( 1993e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1994e73a819dSTao Ma (unsigned long long)cpos, len); 1995e73a819dSTao Ma 1996e73a819dSTao Ma while (len) { 1997e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1998e73a819dSTao Ma cpos, len, &rec, &index, 1999e73a819dSTao Ma &ref_leaf_bh); 2000e73a819dSTao Ma if (ret) { 2001e73a819dSTao Ma mlog_errno(ret); 2002e73a819dSTao Ma goto out; 2003e73a819dSTao Ma } 2004e73a819dSTao Ma 2005e73a819dSTao Ma set_len = le32_to_cpu(rec.r_clusters); 2006e73a819dSTao Ma 2007e73a819dSTao Ma /* 2008e73a819dSTao Ma * Here we may meet with 3 situations: 2009e73a819dSTao Ma * 2010e73a819dSTao Ma * 1. If we find an already existing record, and the length 2011e73a819dSTao Ma * is the same, cool, we just need to increase the r_refcount 2012e73a819dSTao Ma * and it is OK. 2013e73a819dSTao Ma * 2. If we find a hole, just insert it with r_refcount = 1. 2014e73a819dSTao Ma * 3. If we are in the middle of one extent record, split 2015e73a819dSTao Ma * it. 2016e73a819dSTao Ma */ 2017e73a819dSTao Ma if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos && 2018e73a819dSTao Ma set_len <= len) { 2019198aac28STao Ma trace_ocfs2_increase_refcount_change( 2020198aac28STao Ma (unsigned long long)cpos, set_len, 2021e73a819dSTao Ma le32_to_cpu(rec.r_refcount)); 2022e73a819dSTao Ma ret = ocfs2_change_refcount_rec(handle, ci, 20237540c1a7STao Ma ref_leaf_bh, index, 20247540c1a7STao Ma merge, 1); 2025e73a819dSTao Ma if (ret) { 2026e73a819dSTao Ma mlog_errno(ret); 2027e73a819dSTao Ma goto out; 2028e73a819dSTao Ma } 2029e73a819dSTao Ma } else if (!rec.r_refcount) { 2030e73a819dSTao Ma rec.r_refcount = cpu_to_le32(1); 2031e73a819dSTao Ma 2032198aac28STao Ma trace_ocfs2_increase_refcount_insert( 2033e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2034e73a819dSTao Ma set_len); 2035e73a819dSTao Ma ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh, 2036e73a819dSTao Ma ref_leaf_bh, 20377540c1a7STao Ma &rec, index, 20387540c1a7STao Ma merge, meta_ac); 2039e73a819dSTao Ma if (ret) { 2040e73a819dSTao Ma mlog_errno(ret); 2041e73a819dSTao Ma goto out; 2042e73a819dSTao Ma } 2043e73a819dSTao Ma } else { 2044e73a819dSTao Ma set_len = min((u64)(cpos + len), 2045e73a819dSTao Ma le64_to_cpu(rec.r_cpos) + set_len) - cpos; 2046e73a819dSTao Ma rec.r_cpos = cpu_to_le64(cpos); 2047e73a819dSTao Ma rec.r_clusters = cpu_to_le32(set_len); 2048e73a819dSTao Ma le32_add_cpu(&rec.r_refcount, 1); 2049e73a819dSTao Ma 2050198aac28STao Ma trace_ocfs2_increase_refcount_split( 2051e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2052e73a819dSTao Ma set_len, le32_to_cpu(rec.r_refcount)); 2053e73a819dSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 2054e73a819dSTao Ma ref_root_bh, ref_leaf_bh, 20557540c1a7STao Ma &rec, index, merge, 2056e73a819dSTao Ma meta_ac, dealloc); 2057e73a819dSTao Ma if (ret) { 2058e73a819dSTao Ma mlog_errno(ret); 2059e73a819dSTao Ma goto out; 2060e73a819dSTao Ma } 2061e73a819dSTao Ma } 2062e73a819dSTao Ma 2063e73a819dSTao Ma cpos += set_len; 2064e73a819dSTao Ma len -= set_len; 2065e73a819dSTao Ma brelse(ref_leaf_bh); 2066e73a819dSTao Ma ref_leaf_bh = NULL; 2067e73a819dSTao Ma } 2068e73a819dSTao Ma 2069e73a819dSTao Ma out: 2070e73a819dSTao Ma brelse(ref_leaf_bh); 2071e73a819dSTao Ma return ret; 2072e73a819dSTao Ma } 20731823cb0bSTao Ma 20741823cb0bSTao Ma static int ocfs2_remove_refcount_extent(handle_t *handle, 20751823cb0bSTao Ma struct ocfs2_caching_info *ci, 20761823cb0bSTao Ma struct buffer_head *ref_root_bh, 20771823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 20781823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 20791823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 20801823cb0bSTao Ma { 20811823cb0bSTao Ma int ret; 20821823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 20831823cb0bSTao Ma struct ocfs2_refcount_block *rb = 20841823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 20851823cb0bSTao Ma struct ocfs2_extent_tree et; 20861823cb0bSTao Ma 20871823cb0bSTao Ma BUG_ON(rb->rf_records.rl_used); 20881823cb0bSTao Ma 2089198aac28STao Ma trace_ocfs2_remove_refcount_extent( 2090198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 2091198aac28STao Ma (unsigned long long)ref_leaf_bh->b_blocknr, 2092198aac28STao Ma le32_to_cpu(rb->rf_cpos)); 2093198aac28STao Ma 20941823cb0bSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 20951823cb0bSTao Ma ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos), 20961823cb0bSTao Ma 1, meta_ac, dealloc); 20971823cb0bSTao Ma if (ret) { 20981823cb0bSTao Ma mlog_errno(ret); 20991823cb0bSTao Ma goto out; 21001823cb0bSTao Ma } 21011823cb0bSTao Ma 21021823cb0bSTao Ma ocfs2_remove_from_cache(ci, ref_leaf_bh); 21031823cb0bSTao Ma 21041823cb0bSTao Ma /* 21051823cb0bSTao Ma * add the freed block to the dealloc so that it will be freed 21061823cb0bSTao Ma * when we run dealloc. 21071823cb0bSTao Ma */ 21081823cb0bSTao Ma ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE, 21091823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_slot), 211074380c47STao Ma le64_to_cpu(rb->rf_suballoc_loc), 21111823cb0bSTao Ma le64_to_cpu(rb->rf_blkno), 21121823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_bit)); 21131823cb0bSTao Ma if (ret) { 21141823cb0bSTao Ma mlog_errno(ret); 21151823cb0bSTao Ma goto out; 21161823cb0bSTao Ma } 21171823cb0bSTao Ma 21181823cb0bSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 21191823cb0bSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 21201823cb0bSTao Ma if (ret) { 21211823cb0bSTao Ma mlog_errno(ret); 21221823cb0bSTao Ma goto out; 21231823cb0bSTao Ma } 21241823cb0bSTao Ma 21251823cb0bSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 21261823cb0bSTao Ma 21271823cb0bSTao Ma le32_add_cpu(&rb->rf_clusters, -1); 21281823cb0bSTao Ma 21291823cb0bSTao Ma /* 21301823cb0bSTao Ma * check whether we need to restore the root refcount block if 21311823cb0bSTao Ma * there is no leaf extent block at atll. 21321823cb0bSTao Ma */ 21331823cb0bSTao Ma if (!rb->rf_list.l_next_free_rec) { 21341823cb0bSTao Ma BUG_ON(rb->rf_clusters); 21351823cb0bSTao Ma 2136198aac28STao Ma trace_ocfs2_restore_refcount_block( 21371823cb0bSTao Ma (unsigned long long)ref_root_bh->b_blocknr); 21381823cb0bSTao Ma 21391823cb0bSTao Ma rb->rf_flags = 0; 21401823cb0bSTao Ma rb->rf_parent = 0; 21411823cb0bSTao Ma rb->rf_cpos = 0; 21421823cb0bSTao Ma memset(&rb->rf_records, 0, sb->s_blocksize - 21431823cb0bSTao Ma offsetof(struct ocfs2_refcount_block, rf_records)); 21441823cb0bSTao Ma rb->rf_records.rl_count = 21451823cb0bSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 21461823cb0bSTao Ma } 21471823cb0bSTao Ma 21481823cb0bSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 21491823cb0bSTao Ma 21501823cb0bSTao Ma out: 21511823cb0bSTao Ma return ret; 21521823cb0bSTao Ma } 21531823cb0bSTao Ma 21547540c1a7STao Ma int ocfs2_increase_refcount(handle_t *handle, 21557540c1a7STao Ma struct ocfs2_caching_info *ci, 21567540c1a7STao Ma struct buffer_head *ref_root_bh, 21577540c1a7STao Ma u64 cpos, u32 len, 21587540c1a7STao Ma struct ocfs2_alloc_context *meta_ac, 21597540c1a7STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 21607540c1a7STao Ma { 21617540c1a7STao Ma return __ocfs2_increase_refcount(handle, ci, ref_root_bh, 21627540c1a7STao Ma cpos, len, 1, 21637540c1a7STao Ma meta_ac, dealloc); 21647540c1a7STao Ma } 21657540c1a7STao Ma 21661823cb0bSTao Ma static int ocfs2_decrease_refcount_rec(handle_t *handle, 21671823cb0bSTao Ma struct ocfs2_caching_info *ci, 21681823cb0bSTao Ma struct buffer_head *ref_root_bh, 21691823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 21701823cb0bSTao Ma int index, u64 cpos, unsigned int len, 21711823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 21721823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 21731823cb0bSTao Ma { 21741823cb0bSTao Ma int ret; 21751823cb0bSTao Ma struct ocfs2_refcount_block *rb = 21761823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 21771823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index]; 21781823cb0bSTao Ma 21791823cb0bSTao Ma BUG_ON(cpos < le64_to_cpu(rec->r_cpos)); 21801823cb0bSTao Ma BUG_ON(cpos + len > 21811823cb0bSTao Ma le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters)); 21821823cb0bSTao Ma 2183198aac28STao Ma trace_ocfs2_decrease_refcount_rec( 2184198aac28STao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 2185198aac28STao Ma (unsigned long long)cpos, len); 2186198aac28STao Ma 21871823cb0bSTao Ma if (cpos == le64_to_cpu(rec->r_cpos) && 21881823cb0bSTao Ma len == le32_to_cpu(rec->r_clusters)) 21891823cb0bSTao Ma ret = ocfs2_change_refcount_rec(handle, ci, 21907540c1a7STao Ma ref_leaf_bh, index, 1, -1); 21911823cb0bSTao Ma else { 21921823cb0bSTao Ma struct ocfs2_refcount_rec split = *rec; 21931823cb0bSTao Ma split.r_cpos = cpu_to_le64(cpos); 21941823cb0bSTao Ma split.r_clusters = cpu_to_le32(len); 21951823cb0bSTao Ma 21961823cb0bSTao Ma le32_add_cpu(&split.r_refcount, -1); 21971823cb0bSTao Ma 21981823cb0bSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 21991823cb0bSTao Ma ref_root_bh, ref_leaf_bh, 22007540c1a7STao Ma &split, index, 1, 22011823cb0bSTao Ma meta_ac, dealloc); 22021823cb0bSTao Ma } 22031823cb0bSTao Ma 22041823cb0bSTao Ma if (ret) { 22051823cb0bSTao Ma mlog_errno(ret); 22061823cb0bSTao Ma goto out; 22071823cb0bSTao Ma } 22081823cb0bSTao Ma 22091823cb0bSTao Ma /* Remove the leaf refcount block if it contains no refcount record. */ 22101823cb0bSTao Ma if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) { 22111823cb0bSTao Ma ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh, 22121823cb0bSTao Ma ref_leaf_bh, meta_ac, 22131823cb0bSTao Ma dealloc); 22141823cb0bSTao Ma if (ret) 22151823cb0bSTao Ma mlog_errno(ret); 22161823cb0bSTao Ma } 22171823cb0bSTao Ma 22181823cb0bSTao Ma out: 22191823cb0bSTao Ma return ret; 22201823cb0bSTao Ma } 22211823cb0bSTao Ma 22221823cb0bSTao Ma static int __ocfs2_decrease_refcount(handle_t *handle, 22231823cb0bSTao Ma struct ocfs2_caching_info *ci, 22241823cb0bSTao Ma struct buffer_head *ref_root_bh, 22251823cb0bSTao Ma u64 cpos, u32 len, 22261823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 22276ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 22286ae23c55STao Ma int delete) 22291823cb0bSTao Ma { 22301823cb0bSTao Ma int ret = 0, index = 0; 22311823cb0bSTao Ma struct ocfs2_refcount_rec rec; 22321823cb0bSTao Ma unsigned int r_count = 0, r_len; 22331823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 22341823cb0bSTao Ma struct buffer_head *ref_leaf_bh = NULL; 22351823cb0bSTao Ma 2236198aac28STao Ma trace_ocfs2_decrease_refcount( 22371823cb0bSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 22386ae23c55STao Ma (unsigned long long)cpos, len, delete); 22391823cb0bSTao Ma 22401823cb0bSTao Ma while (len) { 22411823cb0bSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 22421823cb0bSTao Ma cpos, len, &rec, &index, 22431823cb0bSTao Ma &ref_leaf_bh); 22441823cb0bSTao Ma if (ret) { 22451823cb0bSTao Ma mlog_errno(ret); 22461823cb0bSTao Ma goto out; 22471823cb0bSTao Ma } 22481823cb0bSTao Ma 22491823cb0bSTao Ma r_count = le32_to_cpu(rec.r_refcount); 22501823cb0bSTao Ma BUG_ON(r_count == 0); 22516ae23c55STao Ma if (!delete) 22526ae23c55STao Ma BUG_ON(r_count > 1); 22531823cb0bSTao Ma 22541823cb0bSTao Ma r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) + 22551823cb0bSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 22561823cb0bSTao Ma 22571823cb0bSTao Ma ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh, 22581823cb0bSTao Ma ref_leaf_bh, index, 22591823cb0bSTao Ma cpos, r_len, 22601823cb0bSTao Ma meta_ac, dealloc); 22611823cb0bSTao Ma if (ret) { 22621823cb0bSTao Ma mlog_errno(ret); 22631823cb0bSTao Ma goto out; 22641823cb0bSTao Ma } 22651823cb0bSTao Ma 22666ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1 && delete) { 22671823cb0bSTao Ma ret = ocfs2_cache_cluster_dealloc(dealloc, 22681823cb0bSTao Ma ocfs2_clusters_to_blocks(sb, cpos), 22691823cb0bSTao Ma r_len); 22701823cb0bSTao Ma if (ret) { 22711823cb0bSTao Ma mlog_errno(ret); 22721823cb0bSTao Ma goto out; 22731823cb0bSTao Ma } 22741823cb0bSTao Ma } 22751823cb0bSTao Ma 22761823cb0bSTao Ma cpos += r_len; 22771823cb0bSTao Ma len -= r_len; 22781823cb0bSTao Ma brelse(ref_leaf_bh); 22791823cb0bSTao Ma ref_leaf_bh = NULL; 22801823cb0bSTao Ma } 22811823cb0bSTao Ma 22821823cb0bSTao Ma out: 22831823cb0bSTao Ma brelse(ref_leaf_bh); 22841823cb0bSTao Ma return ret; 22851823cb0bSTao Ma } 22861823cb0bSTao Ma 22871823cb0bSTao Ma /* Caller must hold refcount tree lock. */ 22881823cb0bSTao Ma int ocfs2_decrease_refcount(struct inode *inode, 22891823cb0bSTao Ma handle_t *handle, u32 cpos, u32 len, 22901823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 22916ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 22926ae23c55STao Ma int delete) 22931823cb0bSTao Ma { 22941823cb0bSTao Ma int ret; 22951823cb0bSTao Ma u64 ref_blkno; 22961823cb0bSTao Ma struct buffer_head *ref_root_bh = NULL; 22971823cb0bSTao Ma struct ocfs2_refcount_tree *tree; 22981823cb0bSTao Ma 229984e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode)); 23001823cb0bSTao Ma 23011823cb0bSTao Ma ret = ocfs2_get_refcount_block(inode, &ref_blkno); 23021823cb0bSTao Ma if (ret) { 23031823cb0bSTao Ma mlog_errno(ret); 23041823cb0bSTao Ma goto out; 23051823cb0bSTao Ma } 23061823cb0bSTao Ma 23071823cb0bSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree); 23081823cb0bSTao Ma if (ret) { 23091823cb0bSTao Ma mlog_errno(ret); 23101823cb0bSTao Ma goto out; 23111823cb0bSTao Ma } 23121823cb0bSTao Ma 23131823cb0bSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno, 23141823cb0bSTao Ma &ref_root_bh); 23151823cb0bSTao Ma if (ret) { 23161823cb0bSTao Ma mlog_errno(ret); 23171823cb0bSTao Ma goto out; 23181823cb0bSTao Ma } 23191823cb0bSTao Ma 23201823cb0bSTao Ma ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh, 23216ae23c55STao Ma cpos, len, meta_ac, dealloc, delete); 23221823cb0bSTao Ma if (ret) 23231823cb0bSTao Ma mlog_errno(ret); 23241823cb0bSTao Ma out: 23251823cb0bSTao Ma brelse(ref_root_bh); 23261823cb0bSTao Ma return ret; 23271823cb0bSTao Ma } 23281aa75feaSTao Ma 23291aa75feaSTao Ma /* 23301aa75feaSTao Ma * Mark the already-existing extent at cpos as refcounted for len clusters. 23311aa75feaSTao Ma * This adds the refcount extent flag. 23321aa75feaSTao Ma * 23331aa75feaSTao Ma * If the existing extent is larger than the request, initiate a 23341aa75feaSTao Ma * split. An attempt will be made at merging with adjacent extents. 23351aa75feaSTao Ma * 23361aa75feaSTao Ma * The caller is responsible for passing down meta_ac if we'll need it. 23371aa75feaSTao Ma */ 23381aa75feaSTao Ma static int ocfs2_mark_extent_refcounted(struct inode *inode, 23391aa75feaSTao Ma struct ocfs2_extent_tree *et, 23401aa75feaSTao Ma handle_t *handle, u32 cpos, 23411aa75feaSTao Ma u32 len, u32 phys, 23421aa75feaSTao Ma struct ocfs2_alloc_context *meta_ac, 23431aa75feaSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 23441aa75feaSTao Ma { 23451aa75feaSTao Ma int ret; 23461aa75feaSTao Ma 2347198aac28STao Ma trace_ocfs2_mark_extent_refcounted(OCFS2_I(inode)->ip_blkno, 2348198aac28STao Ma cpos, len, phys); 23491aa75feaSTao Ma 23501aa75feaSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 23517ecef14aSJoe Perches ret = ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n", 23527ecef14aSJoe Perches inode->i_ino); 23531aa75feaSTao Ma goto out; 23541aa75feaSTao Ma } 23551aa75feaSTao Ma 23561aa75feaSTao Ma ret = ocfs2_change_extent_flag(handle, et, cpos, 23571aa75feaSTao Ma len, phys, meta_ac, dealloc, 23581aa75feaSTao Ma OCFS2_EXT_REFCOUNTED, 0); 23591aa75feaSTao Ma if (ret) 23601aa75feaSTao Ma mlog_errno(ret); 23611aa75feaSTao Ma 23621aa75feaSTao Ma out: 23631aa75feaSTao Ma return ret; 23641aa75feaSTao Ma } 2365bcbbb24aSTao Ma 2366bcbbb24aSTao Ma /* 2367bcbbb24aSTao Ma * Given some contiguous physical clusters, calculate what we need 2368bcbbb24aSTao Ma * for modifying their refcount. 2369bcbbb24aSTao Ma */ 2370bcbbb24aSTao Ma static int ocfs2_calc_refcount_meta_credits(struct super_block *sb, 2371bcbbb24aSTao Ma struct ocfs2_caching_info *ci, 2372bcbbb24aSTao Ma struct buffer_head *ref_root_bh, 2373bcbbb24aSTao Ma u64 start_cpos, 2374bcbbb24aSTao Ma u32 clusters, 2375bcbbb24aSTao Ma int *meta_add, 2376bcbbb24aSTao Ma int *credits) 2377bcbbb24aSTao Ma { 2378bcbbb24aSTao Ma int ret = 0, index, ref_blocks = 0, recs_add = 0; 2379bcbbb24aSTao Ma u64 cpos = start_cpos; 2380bcbbb24aSTao Ma struct ocfs2_refcount_block *rb; 2381bcbbb24aSTao Ma struct ocfs2_refcount_rec rec; 2382bcbbb24aSTao Ma struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL; 2383bcbbb24aSTao Ma u32 len; 2384bcbbb24aSTao Ma 2385bcbbb24aSTao Ma while (clusters) { 2386bcbbb24aSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 2387bcbbb24aSTao Ma cpos, clusters, &rec, 2388bcbbb24aSTao Ma &index, &ref_leaf_bh); 2389bcbbb24aSTao Ma if (ret) { 2390bcbbb24aSTao Ma mlog_errno(ret); 2391bcbbb24aSTao Ma goto out; 2392bcbbb24aSTao Ma } 2393bcbbb24aSTao Ma 2394bcbbb24aSTao Ma if (ref_leaf_bh != prev_bh) { 2395bcbbb24aSTao Ma /* 2396bcbbb24aSTao Ma * Now we encounter a new leaf block, so calculate 2397bcbbb24aSTao Ma * whether we need to extend the old leaf. 2398bcbbb24aSTao Ma */ 2399bcbbb24aSTao Ma if (prev_bh) { 2400bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *) 2401bcbbb24aSTao Ma prev_bh->b_data; 2402bcbbb24aSTao Ma 2403e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) + 2404bcbbb24aSTao Ma recs_add > 2405bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2406bcbbb24aSTao Ma ref_blocks++; 2407bcbbb24aSTao Ma } 2408bcbbb24aSTao Ma 2409bcbbb24aSTao Ma recs_add = 0; 2410bcbbb24aSTao Ma *credits += 1; 2411bcbbb24aSTao Ma brelse(prev_bh); 2412bcbbb24aSTao Ma prev_bh = ref_leaf_bh; 2413bcbbb24aSTao Ma get_bh(prev_bh); 2414bcbbb24aSTao Ma } 2415bcbbb24aSTao Ma 2416198aac28STao Ma trace_ocfs2_calc_refcount_meta_credits_iterate( 2417bcbbb24aSTao Ma recs_add, (unsigned long long)cpos, clusters, 2418bcbbb24aSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2419bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters), 2420bcbbb24aSTao Ma le32_to_cpu(rec.r_refcount), index); 2421bcbbb24aSTao Ma 2422bcbbb24aSTao Ma len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) + 2423bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 2424bcbbb24aSTao Ma /* 2425bcbbb24aSTao Ma * We record all the records which will be inserted to the 2426bcbbb24aSTao Ma * same refcount block, so that we can tell exactly whether 2427bcbbb24aSTao Ma * we need a new refcount block or not. 24288a2e70c4STao Ma * 24298a2e70c4STao Ma * If we will insert a new one, this is easy and only happens 24308a2e70c4STao Ma * during adding refcounted flag to the extent, so we don't 24318a2e70c4STao Ma * have a chance of spliting. We just need one record. 24328a2e70c4STao Ma * 24338a2e70c4STao Ma * If the refcount rec already exists, that would be a little 24348a2e70c4STao Ma * complicated. we may have to: 24358a2e70c4STao Ma * 1) split at the beginning if the start pos isn't aligned. 24368a2e70c4STao Ma * we need 1 more record in this case. 24378a2e70c4STao Ma * 2) split int the end if the end pos isn't aligned. 24388a2e70c4STao Ma * we need 1 more record in this case. 24398a2e70c4STao Ma * 3) split in the middle because of file system fragmentation. 24408a2e70c4STao Ma * we need 2 more records in this case(we can't detect this 24418a2e70c4STao Ma * beforehand, so always think of the worst case). 2442bcbbb24aSTao Ma */ 2443bcbbb24aSTao Ma if (rec.r_refcount) { 24448a2e70c4STao Ma recs_add += 2; 2445bcbbb24aSTao Ma /* Check whether we need a split at the beginning. */ 2446bcbbb24aSTao Ma if (cpos == start_cpos && 2447bcbbb24aSTao Ma cpos != le64_to_cpu(rec.r_cpos)) 2448bcbbb24aSTao Ma recs_add++; 2449bcbbb24aSTao Ma 2450bcbbb24aSTao Ma /* Check whether we need a split in the end. */ 2451bcbbb24aSTao Ma if (cpos + clusters < le64_to_cpu(rec.r_cpos) + 2452bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) 2453bcbbb24aSTao Ma recs_add++; 2454bcbbb24aSTao Ma } else 2455bcbbb24aSTao Ma recs_add++; 2456bcbbb24aSTao Ma 2457bcbbb24aSTao Ma brelse(ref_leaf_bh); 2458bcbbb24aSTao Ma ref_leaf_bh = NULL; 2459bcbbb24aSTao Ma clusters -= len; 2460bcbbb24aSTao Ma cpos += len; 2461bcbbb24aSTao Ma } 2462bcbbb24aSTao Ma 2463bcbbb24aSTao Ma if (prev_bh) { 2464bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)prev_bh->b_data; 2465bcbbb24aSTao Ma 2466e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) + recs_add > 2467bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2468bcbbb24aSTao Ma ref_blocks++; 2469bcbbb24aSTao Ma 2470bcbbb24aSTao Ma *credits += 1; 2471bcbbb24aSTao Ma } 2472bcbbb24aSTao Ma 2473bcbbb24aSTao Ma if (!ref_blocks) 2474bcbbb24aSTao Ma goto out; 2475bcbbb24aSTao Ma 2476bcbbb24aSTao Ma *meta_add += ref_blocks; 2477bcbbb24aSTao Ma *credits += ref_blocks; 2478bcbbb24aSTao Ma 2479bcbbb24aSTao Ma /* 2480bcbbb24aSTao Ma * So we may need ref_blocks to insert into the tree. 2481bcbbb24aSTao Ma * That also means we need to change the b-tree and add that number 2482bcbbb24aSTao Ma * of records since we never merge them. 2483bcbbb24aSTao Ma * We need one more block for expansion since the new created leaf 2484bcbbb24aSTao Ma * block is also full and needs split. 2485bcbbb24aSTao Ma */ 2486bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 2487bcbbb24aSTao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) { 2488bcbbb24aSTao Ma struct ocfs2_extent_tree et; 2489bcbbb24aSTao Ma 2490bcbbb24aSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 2491bcbbb24aSTao Ma *meta_add += ocfs2_extend_meta_needed(et.et_root_el); 2492bcbbb24aSTao Ma *credits += ocfs2_calc_extend_credits(sb, 249306f9da6eSGoldwyn Rodrigues et.et_root_el); 2494bcbbb24aSTao Ma } else { 2495bcbbb24aSTao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 2496bcbbb24aSTao Ma *meta_add += 1; 2497bcbbb24aSTao Ma } 2498bcbbb24aSTao Ma 2499bcbbb24aSTao Ma out: 2500198aac28STao Ma 2501198aac28STao Ma trace_ocfs2_calc_refcount_meta_credits( 2502198aac28STao Ma (unsigned long long)start_cpos, clusters, 2503198aac28STao Ma *meta_add, *credits); 2504bcbbb24aSTao Ma brelse(ref_leaf_bh); 2505bcbbb24aSTao Ma brelse(prev_bh); 2506bcbbb24aSTao Ma return ret; 2507bcbbb24aSTao Ma } 2508bcbbb24aSTao Ma 2509bcbbb24aSTao Ma /* 2510bcbbb24aSTao Ma * For refcount tree, we will decrease some contiguous clusters 2511bcbbb24aSTao Ma * refcount count, so just go through it to see how many blocks 2512bcbbb24aSTao Ma * we gonna touch and whether we need to create new blocks. 2513bcbbb24aSTao Ma * 2514bcbbb24aSTao Ma * Normally the refcount blocks store these refcount should be 2515af901ca1SAndré Goddard Rosa * contiguous also, so that we can get the number easily. 251678f94673STristan Ye * We will at most add split 2 refcount records and 2 more 251778f94673STristan Ye * refcount blocks, so just check it in a rough way. 2518bcbbb24aSTao Ma * 2519bcbbb24aSTao Ma * Caller must hold refcount tree lock. 2520bcbbb24aSTao Ma */ 2521bcbbb24aSTao Ma int ocfs2_prepare_refcount_change_for_del(struct inode *inode, 252278f94673STristan Ye u64 refcount_loc, 2523bcbbb24aSTao Ma u64 phys_blkno, 2524bcbbb24aSTao Ma u32 clusters, 2525bcbbb24aSTao Ma int *credits, 252678f94673STristan Ye int *ref_blocks) 2527bcbbb24aSTao Ma { 252878f94673STristan Ye int ret; 2529bcbbb24aSTao Ma struct buffer_head *ref_root_bh = NULL; 2530bcbbb24aSTao Ma struct ocfs2_refcount_tree *tree; 2531bcbbb24aSTao Ma u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno); 2532bcbbb24aSTao Ma 2533bcbbb24aSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 25347ecef14aSJoe Perches ret = ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n", 25357ecef14aSJoe Perches inode->i_ino); 2536bcbbb24aSTao Ma goto out; 2537bcbbb24aSTao Ma } 2538bcbbb24aSTao Ma 253984e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode)); 2540bcbbb24aSTao Ma 2541bcbbb24aSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), 254278f94673STristan Ye refcount_loc, &tree); 2543bcbbb24aSTao Ma if (ret) { 2544bcbbb24aSTao Ma mlog_errno(ret); 2545bcbbb24aSTao Ma goto out; 2546bcbbb24aSTao Ma } 2547bcbbb24aSTao Ma 254878f94673STristan Ye ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc, 2549bcbbb24aSTao Ma &ref_root_bh); 2550bcbbb24aSTao Ma if (ret) { 2551bcbbb24aSTao Ma mlog_errno(ret); 2552bcbbb24aSTao Ma goto out; 2553bcbbb24aSTao Ma } 2554bcbbb24aSTao Ma 2555bcbbb24aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 2556bcbbb24aSTao Ma &tree->rf_ci, 2557bcbbb24aSTao Ma ref_root_bh, 2558bcbbb24aSTao Ma start_cpos, clusters, 255978f94673STristan Ye ref_blocks, credits); 2560bcbbb24aSTao Ma if (ret) { 2561bcbbb24aSTao Ma mlog_errno(ret); 2562bcbbb24aSTao Ma goto out; 2563bcbbb24aSTao Ma } 2564bcbbb24aSTao Ma 2565198aac28STao Ma trace_ocfs2_prepare_refcount_change_for_del(*ref_blocks, *credits); 2566bcbbb24aSTao Ma 2567bcbbb24aSTao Ma out: 2568bcbbb24aSTao Ma brelse(ref_root_bh); 2569bcbbb24aSTao Ma return ret; 2570bcbbb24aSTao Ma } 25716f70fa51STao Ma 25726f70fa51STao Ma #define MAX_CONTIG_BYTES 1048576 25736f70fa51STao Ma 25746f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb) 25756f70fa51STao Ma { 25766f70fa51STao Ma return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES); 25776f70fa51STao Ma } 25786f70fa51STao Ma 25796f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb) 25806f70fa51STao Ma { 25816f70fa51STao Ma return ~(ocfs2_cow_contig_clusters(sb) - 1); 25826f70fa51STao Ma } 25836f70fa51STao Ma 25846f70fa51STao Ma /* 25856f70fa51STao Ma * Given an extent that starts at 'start' and an I/O that starts at 'cpos', 25866f70fa51STao Ma * find an offset (start + (n * contig_clusters)) that is closest to cpos 25876f70fa51STao Ma * while still being less than or equal to it. 25886f70fa51STao Ma * 25896f70fa51STao Ma * The goal is to break the extent at a multiple of contig_clusters. 25906f70fa51STao Ma */ 25916f70fa51STao Ma static inline unsigned int ocfs2_cow_align_start(struct super_block *sb, 25926f70fa51STao Ma unsigned int start, 25936f70fa51STao Ma unsigned int cpos) 25946f70fa51STao Ma { 25956f70fa51STao Ma BUG_ON(start > cpos); 25966f70fa51STao Ma 25976f70fa51STao Ma return start + ((cpos - start) & ocfs2_cow_contig_mask(sb)); 25986f70fa51STao Ma } 25996f70fa51STao Ma 26006f70fa51STao Ma /* 26016f70fa51STao Ma * Given a cluster count of len, pad it out so that it is a multiple 26026f70fa51STao Ma * of contig_clusters. 26036f70fa51STao Ma */ 26046f70fa51STao Ma static inline unsigned int ocfs2_cow_align_length(struct super_block *sb, 26056f70fa51STao Ma unsigned int len) 26066f70fa51STao Ma { 26076f70fa51STao Ma unsigned int padded = 26086f70fa51STao Ma (len + (ocfs2_cow_contig_clusters(sb) - 1)) & 26096f70fa51STao Ma ocfs2_cow_contig_mask(sb); 26106f70fa51STao Ma 26116f70fa51STao Ma /* Did we wrap? */ 26126f70fa51STao Ma if (padded < len) 26136f70fa51STao Ma padded = UINT_MAX; 26146f70fa51STao Ma 26156f70fa51STao Ma return padded; 26166f70fa51STao Ma } 26176f70fa51STao Ma 26186f70fa51STao Ma /* 26196f70fa51STao Ma * Calculate out the start and number of virtual clusters we need to to CoW. 26206f70fa51STao Ma * 26216f70fa51STao Ma * cpos is vitual start cluster position we want to do CoW in a 26226f70fa51STao Ma * file and write_len is the cluster length. 262337f8a2bfSTao Ma * max_cpos is the place where we want to stop CoW intentionally. 26246f70fa51STao Ma * 26256f70fa51STao Ma * Normal we will start CoW from the beginning of extent record cotaining cpos. 26266f70fa51STao Ma * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we 26276f70fa51STao Ma * get good I/O from the resulting extent tree. 26286f70fa51STao Ma */ 26296f70fa51STao Ma static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, 2630913580b4STao Ma struct ocfs2_extent_list *el, 26316f70fa51STao Ma u32 cpos, 26326f70fa51STao Ma u32 write_len, 263337f8a2bfSTao Ma u32 max_cpos, 26346f70fa51STao Ma u32 *cow_start, 26356f70fa51STao Ma u32 *cow_len) 26366f70fa51STao Ma { 26376f70fa51STao Ma int ret = 0; 26386f70fa51STao Ma int tree_height = le16_to_cpu(el->l_tree_depth), i; 26396f70fa51STao Ma struct buffer_head *eb_bh = NULL; 26406f70fa51STao Ma struct ocfs2_extent_block *eb = NULL; 26416f70fa51STao Ma struct ocfs2_extent_rec *rec; 26426f70fa51STao Ma unsigned int want_clusters, rec_end = 0; 26436f70fa51STao Ma int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb); 26446f70fa51STao Ma int leaf_clusters; 26456f70fa51STao Ma 264637f8a2bfSTao Ma BUG_ON(cpos + write_len > max_cpos); 264737f8a2bfSTao Ma 26486f70fa51STao Ma if (tree_height > 0) { 26496f70fa51STao Ma ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh); 26506f70fa51STao Ma if (ret) { 26516f70fa51STao Ma mlog_errno(ret); 26526f70fa51STao Ma goto out; 26536f70fa51STao Ma } 26546f70fa51STao Ma 26556f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 26566f70fa51STao Ma el = &eb->h_list; 26576f70fa51STao Ma 26586f70fa51STao Ma if (el->l_tree_depth) { 265917a5b9abSGoldwyn Rodrigues ret = ocfs2_error(inode->i_sb, 26607ecef14aSJoe Perches "Inode %lu has non zero tree depth in leaf block %llu\n", 26617ecef14aSJoe Perches inode->i_ino, 26626f70fa51STao Ma (unsigned long long)eb_bh->b_blocknr); 26636f70fa51STao Ma goto out; 26646f70fa51STao Ma } 26656f70fa51STao Ma } 26666f70fa51STao Ma 26676f70fa51STao Ma *cow_len = 0; 26686f70fa51STao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 26696f70fa51STao Ma rec = &el->l_recs[i]; 26706f70fa51STao Ma 26716f70fa51STao Ma if (ocfs2_is_empty_extent(rec)) { 26726f70fa51STao Ma mlog_bug_on_msg(i != 0, "Inode %lu has empty record in " 26736f70fa51STao Ma "index %d\n", inode->i_ino, i); 26746f70fa51STao Ma continue; 26756f70fa51STao Ma } 26766f70fa51STao Ma 26776f70fa51STao Ma if (le32_to_cpu(rec->e_cpos) + 26786f70fa51STao Ma le16_to_cpu(rec->e_leaf_clusters) <= cpos) 26796f70fa51STao Ma continue; 26806f70fa51STao Ma 26816f70fa51STao Ma if (*cow_len == 0) { 26826f70fa51STao Ma /* 26836f70fa51STao Ma * We should find a refcounted record in the 26846f70fa51STao Ma * first pass. 26856f70fa51STao Ma */ 26866f70fa51STao Ma BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED)); 26876f70fa51STao Ma *cow_start = le32_to_cpu(rec->e_cpos); 26886f70fa51STao Ma } 26896f70fa51STao Ma 26906f70fa51STao Ma /* 269137f8a2bfSTao Ma * If we encounter a hole, a non-refcounted record or 269237f8a2bfSTao Ma * pass the max_cpos, stop the search. 26936f70fa51STao Ma */ 26946f70fa51STao Ma if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) || 269537f8a2bfSTao Ma (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) || 269637f8a2bfSTao Ma (max_cpos <= le32_to_cpu(rec->e_cpos))) 26976f70fa51STao Ma break; 26986f70fa51STao Ma 26996f70fa51STao Ma leaf_clusters = le16_to_cpu(rec->e_leaf_clusters); 27006f70fa51STao Ma rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters; 270137f8a2bfSTao Ma if (rec_end > max_cpos) { 270237f8a2bfSTao Ma rec_end = max_cpos; 270337f8a2bfSTao Ma leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos); 270437f8a2bfSTao Ma } 27056f70fa51STao Ma 27066f70fa51STao Ma /* 27076f70fa51STao Ma * How many clusters do we actually need from 27086f70fa51STao Ma * this extent? First we see how many we actually 27096f70fa51STao Ma * need to complete the write. If that's smaller 27106f70fa51STao Ma * than contig_clusters, we try for contig_clusters. 27116f70fa51STao Ma */ 27126f70fa51STao Ma if (!*cow_len) 27136f70fa51STao Ma want_clusters = write_len; 27146f70fa51STao Ma else 27156f70fa51STao Ma want_clusters = (cpos + write_len) - 27166f70fa51STao Ma (*cow_start + *cow_len); 27176f70fa51STao Ma if (want_clusters < contig_clusters) 27186f70fa51STao Ma want_clusters = contig_clusters; 27196f70fa51STao Ma 27206f70fa51STao Ma /* 27216f70fa51STao Ma * If the write does not cover the whole extent, we 27226f70fa51STao Ma * need to calculate how we're going to split the extent. 27236f70fa51STao Ma * We try to do it on contig_clusters boundaries. 27246f70fa51STao Ma * 27256f70fa51STao Ma * Any extent smaller than contig_clusters will be 27266f70fa51STao Ma * CoWed in its entirety. 27276f70fa51STao Ma */ 27286f70fa51STao Ma if (leaf_clusters <= contig_clusters) 27296f70fa51STao Ma *cow_len += leaf_clusters; 27306f70fa51STao Ma else if (*cow_len || (*cow_start == cpos)) { 27316f70fa51STao Ma /* 27326f70fa51STao Ma * This extent needs to be CoW'd from its 27336f70fa51STao Ma * beginning, so all we have to do is compute 27346f70fa51STao Ma * how many clusters to grab. We align 27356f70fa51STao Ma * want_clusters to the edge of contig_clusters 27366f70fa51STao Ma * to get better I/O. 27376f70fa51STao Ma */ 27386f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 27396f70fa51STao Ma want_clusters); 27406f70fa51STao Ma 27416f70fa51STao Ma if (leaf_clusters < want_clusters) 27426f70fa51STao Ma *cow_len += leaf_clusters; 27436f70fa51STao Ma else 27446f70fa51STao Ma *cow_len += want_clusters; 27456f70fa51STao Ma } else if ((*cow_start + contig_clusters) >= 27466f70fa51STao Ma (cpos + write_len)) { 27476f70fa51STao Ma /* 27486f70fa51STao Ma * Breaking off contig_clusters at the front 27496f70fa51STao Ma * of the extent will cover our write. That's 27506f70fa51STao Ma * easy. 27516f70fa51STao Ma */ 27526f70fa51STao Ma *cow_len = contig_clusters; 27536f70fa51STao Ma } else if ((rec_end - cpos) <= contig_clusters) { 27546f70fa51STao Ma /* 27556f70fa51STao Ma * Breaking off contig_clusters at the tail of 27566f70fa51STao Ma * this extent will cover cpos. 27576f70fa51STao Ma */ 27586f70fa51STao Ma *cow_start = rec_end - contig_clusters; 27596f70fa51STao Ma *cow_len = contig_clusters; 27606f70fa51STao Ma } else if ((rec_end - cpos) <= want_clusters) { 27616f70fa51STao Ma /* 27626f70fa51STao Ma * While we can't fit the entire write in this 27636f70fa51STao Ma * extent, we know that the write goes from cpos 27646f70fa51STao Ma * to the end of the extent. Break that off. 27656f70fa51STao Ma * We try to break it at some multiple of 27666f70fa51STao Ma * contig_clusters from the front of the extent. 27676f70fa51STao Ma * Failing that (ie, cpos is within 27686f70fa51STao Ma * contig_clusters of the front), we'll CoW the 27696f70fa51STao Ma * entire extent. 27706f70fa51STao Ma */ 27716f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 27726f70fa51STao Ma *cow_start, cpos); 27736f70fa51STao Ma *cow_len = rec_end - *cow_start; 27746f70fa51STao Ma } else { 27756f70fa51STao Ma /* 27766f70fa51STao Ma * Ok, the entire write lives in the middle of 27776f70fa51STao Ma * this extent. Let's try to slice the extent up 27786f70fa51STao Ma * nicely. Optimally, our CoW region starts at 27796f70fa51STao Ma * m*contig_clusters from the beginning of the 27806f70fa51STao Ma * extent and goes for n*contig_clusters, 27816f70fa51STao Ma * covering the entire write. 27826f70fa51STao Ma */ 27836f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 27846f70fa51STao Ma *cow_start, cpos); 27856f70fa51STao Ma 27866f70fa51STao Ma want_clusters = (cpos + write_len) - *cow_start; 27876f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 27886f70fa51STao Ma want_clusters); 27896f70fa51STao Ma if (*cow_start + want_clusters <= rec_end) 27906f70fa51STao Ma *cow_len = want_clusters; 27916f70fa51STao Ma else 27926f70fa51STao Ma *cow_len = rec_end - *cow_start; 27936f70fa51STao Ma } 27946f70fa51STao Ma 27956f70fa51STao Ma /* Have we covered our entire write yet? */ 27966f70fa51STao Ma if ((*cow_start + *cow_len) >= (cpos + write_len)) 27976f70fa51STao Ma break; 27986f70fa51STao Ma 27996f70fa51STao Ma /* 28006f70fa51STao Ma * If we reach the end of the extent block and don't get enough 28016f70fa51STao Ma * clusters, continue with the next extent block if possible. 28026f70fa51STao Ma */ 28036f70fa51STao Ma if (i + 1 == le16_to_cpu(el->l_next_free_rec) && 28046f70fa51STao Ma eb && eb->h_next_leaf_blk) { 28056f70fa51STao Ma brelse(eb_bh); 28066f70fa51STao Ma eb_bh = NULL; 28076f70fa51STao Ma 28086f70fa51STao Ma ret = ocfs2_read_extent_block(INODE_CACHE(inode), 28096f70fa51STao Ma le64_to_cpu(eb->h_next_leaf_blk), 28106f70fa51STao Ma &eb_bh); 28116f70fa51STao Ma if (ret) { 28126f70fa51STao Ma mlog_errno(ret); 28136f70fa51STao Ma goto out; 28146f70fa51STao Ma } 28156f70fa51STao Ma 28166f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 28176f70fa51STao Ma el = &eb->h_list; 28186f70fa51STao Ma i = -1; 28196f70fa51STao Ma } 28206f70fa51STao Ma } 28216f70fa51STao Ma 28226f70fa51STao Ma out: 28236f70fa51STao Ma brelse(eb_bh); 28246f70fa51STao Ma return ret; 28256f70fa51STao Ma } 28266f70fa51STao Ma 28276f70fa51STao Ma /* 28286f70fa51STao Ma * Prepare meta_ac, data_ac and calculate credits when we want to add some 28296f70fa51STao Ma * num_clusters in data_tree "et" and change the refcount for the old 28306f70fa51STao Ma * clusters(starting form p_cluster) in the refcount tree. 28316f70fa51STao Ma * 28326f70fa51STao Ma * Note: 28336f70fa51STao Ma * 1. since we may split the old tree, so we at most will need num_clusters + 2 28346f70fa51STao Ma * more new leaf records. 28356f70fa51STao Ma * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so 28366f70fa51STao Ma * just give data_ac = NULL. 28376f70fa51STao Ma */ 28386f70fa51STao Ma static int ocfs2_lock_refcount_allocators(struct super_block *sb, 28396f70fa51STao Ma u32 p_cluster, u32 num_clusters, 28406f70fa51STao Ma struct ocfs2_extent_tree *et, 28416f70fa51STao Ma struct ocfs2_caching_info *ref_ci, 28426f70fa51STao Ma struct buffer_head *ref_root_bh, 28436f70fa51STao Ma struct ocfs2_alloc_context **meta_ac, 28446f70fa51STao Ma struct ocfs2_alloc_context **data_ac, 28456f70fa51STao Ma int *credits) 28466f70fa51STao Ma { 28476f70fa51STao Ma int ret = 0, meta_add = 0; 2848964f14a0SJun Piao int num_free_extents = ocfs2_num_free_extents(et); 28496f70fa51STao Ma 28506f70fa51STao Ma if (num_free_extents < 0) { 28516f70fa51STao Ma ret = num_free_extents; 28526f70fa51STao Ma mlog_errno(ret); 28536f70fa51STao Ma goto out; 28546f70fa51STao Ma } 28556f70fa51STao Ma 28566f70fa51STao Ma if (num_free_extents < num_clusters + 2) 28576f70fa51STao Ma meta_add = 28586f70fa51STao Ma ocfs2_extend_meta_needed(et->et_root_el); 28596f70fa51STao Ma 286006f9da6eSGoldwyn Rodrigues *credits += ocfs2_calc_extend_credits(sb, et->et_root_el); 28616f70fa51STao Ma 28626f70fa51STao Ma ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh, 28636f70fa51STao Ma p_cluster, num_clusters, 28646f70fa51STao Ma &meta_add, credits); 28656f70fa51STao Ma if (ret) { 28666f70fa51STao Ma mlog_errno(ret); 28676f70fa51STao Ma goto out; 28686f70fa51STao Ma } 28696f70fa51STao Ma 2870198aac28STao Ma trace_ocfs2_lock_refcount_allocators(meta_add, *credits); 28716f70fa51STao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add, 28726f70fa51STao Ma meta_ac); 28736f70fa51STao Ma if (ret) { 28746f70fa51STao Ma mlog_errno(ret); 28756f70fa51STao Ma goto out; 28766f70fa51STao Ma } 28776f70fa51STao Ma 28786f70fa51STao Ma if (data_ac) { 28796f70fa51STao Ma ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters, 28806f70fa51STao Ma data_ac); 28816f70fa51STao Ma if (ret) 28826f70fa51STao Ma mlog_errno(ret); 28836f70fa51STao Ma } 28846f70fa51STao Ma 28856f70fa51STao Ma out: 28866f70fa51STao Ma if (ret) { 28876f70fa51STao Ma if (*meta_ac) { 28886f70fa51STao Ma ocfs2_free_alloc_context(*meta_ac); 28896f70fa51STao Ma *meta_ac = NULL; 28906f70fa51STao Ma } 28916f70fa51STao Ma } 28926f70fa51STao Ma 28936f70fa51STao Ma return ret; 28946f70fa51STao Ma } 28956f70fa51STao Ma 28966f70fa51STao Ma static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh) 28976f70fa51STao Ma { 28986f70fa51STao Ma BUG_ON(buffer_dirty(bh)); 28996f70fa51STao Ma 29006f70fa51STao Ma clear_buffer_mapped(bh); 29016f70fa51STao Ma 29026f70fa51STao Ma return 0; 29036f70fa51STao Ma } 29046f70fa51STao Ma 29053e19a25eSTristan Ye int ocfs2_duplicate_clusters_by_page(handle_t *handle, 2906c7dd3392STiger Yang struct inode *inode, 29076f70fa51STao Ma u32 cpos, u32 old_cluster, 29086f70fa51STao Ma u32 new_cluster, u32 new_len) 29096f70fa51STao Ma { 29106f70fa51STao Ma int ret = 0, partial; 2911c7dd3392STiger Yang struct super_block *sb = inode->i_sb; 29126f70fa51STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 29136f70fa51STao Ma struct page *page; 29146f70fa51STao Ma pgoff_t page_index; 2915262d8a87SJoseph Qi unsigned int from, to; 29166f70fa51STao Ma loff_t offset, end, map_end; 29173e19a25eSTristan Ye struct address_space *mapping = inode->i_mapping; 29186f70fa51STao Ma 2919198aac28STao Ma trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster, 2920198aac28STao Ma new_cluster, new_len); 29216f70fa51STao Ma 29226f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 29236f70fa51STao Ma end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits); 2924f5e27b6dSTao Ma /* 2925f5e27b6dSTao Ma * We only duplicate pages until we reach the page contains i_size - 1. 2926f5e27b6dSTao Ma * So trim 'end' to i_size. 2927f5e27b6dSTao Ma */ 29283e19a25eSTristan Ye if (end > i_size_read(inode)) 29293e19a25eSTristan Ye end = i_size_read(inode); 29306f70fa51STao Ma 29316f70fa51STao Ma while (offset < end) { 293209cbfeafSKirill A. Shutemov page_index = offset >> PAGE_SHIFT; 293309cbfeafSKirill A. Shutemov map_end = ((loff_t)page_index + 1) << PAGE_SHIFT; 29346f70fa51STao Ma if (map_end > end) 29356f70fa51STao Ma map_end = end; 29366f70fa51STao Ma 29376f70fa51STao Ma /* from, to is the offset within the page. */ 293809cbfeafSKirill A. Shutemov from = offset & (PAGE_SIZE - 1); 293909cbfeafSKirill A. Shutemov to = PAGE_SIZE; 294009cbfeafSKirill A. Shutemov if (map_end & (PAGE_SIZE - 1)) 294109cbfeafSKirill A. Shutemov to = map_end & (PAGE_SIZE - 1); 29426f70fa51STao Ma 294369eb7765SLarry Chen retry: 29449b4c0ff3SJan Kara page = find_or_create_page(mapping, page_index, GFP_NOFS); 294562c61046SGu Zheng if (!page) { 294662c61046SGu Zheng ret = -ENOMEM; 294762c61046SGu Zheng mlog_errno(ret); 294862c61046SGu Zheng break; 294962c61046SGu Zheng } 29506f70fa51STao Ma 29510a1ea437STao Ma /* 295269eb7765SLarry Chen * In case PAGE_SIZE <= CLUSTER_SIZE, we do not expect a dirty 295369eb7765SLarry Chen * page, so write it back. 29540a1ea437STao Ma */ 295569eb7765SLarry Chen if (PAGE_SIZE <= OCFS2_SB(sb)->s_clustersize) { 295669eb7765SLarry Chen if (PageDirty(page)) { 295769eb7765SLarry Chen /* 295869eb7765SLarry Chen * write_on_page will unlock the page on return 295969eb7765SLarry Chen */ 296069eb7765SLarry Chen ret = write_one_page(page); 296169eb7765SLarry Chen goto retry; 296269eb7765SLarry Chen } 296369eb7765SLarry Chen } 29646f70fa51STao Ma 29656f70fa51STao Ma if (!PageUptodate(page)) { 29666f70fa51STao Ma ret = block_read_full_page(page, ocfs2_get_block); 29676f70fa51STao Ma if (ret) { 29686f70fa51STao Ma mlog_errno(ret); 29696f70fa51STao Ma goto unlock; 29706f70fa51STao Ma } 29716f70fa51STao Ma lock_page(page); 29726f70fa51STao Ma } 29736f70fa51STao Ma 29746f70fa51STao Ma if (page_has_buffers(page)) { 29756f70fa51STao Ma ret = walk_page_buffers(handle, page_buffers(page), 29766f70fa51STao Ma from, to, &partial, 29776f70fa51STao Ma ocfs2_clear_cow_buffer); 29786f70fa51STao Ma if (ret) { 29796f70fa51STao Ma mlog_errno(ret); 29806f70fa51STao Ma goto unlock; 29816f70fa51STao Ma } 29826f70fa51STao Ma } 29836f70fa51STao Ma 2984c7dd3392STiger Yang ocfs2_map_and_dirty_page(inode, 2985c7dd3392STiger Yang handle, from, to, 29866f70fa51STao Ma page, 0, &new_block); 29876f70fa51STao Ma mark_page_accessed(page); 29886f70fa51STao Ma unlock: 29896f70fa51STao Ma unlock_page(page); 299009cbfeafSKirill A. Shutemov put_page(page); 29916f70fa51STao Ma page = NULL; 29926f70fa51STao Ma offset = map_end; 29936f70fa51STao Ma if (ret) 29946f70fa51STao Ma break; 29956f70fa51STao Ma } 29966f70fa51STao Ma 29976f70fa51STao Ma return ret; 29986f70fa51STao Ma } 29996f70fa51STao Ma 30003e19a25eSTristan Ye int ocfs2_duplicate_clusters_by_jbd(handle_t *handle, 3001c7dd3392STiger Yang struct inode *inode, 3002492a8a33STao Ma u32 cpos, u32 old_cluster, 3003492a8a33STao Ma u32 new_cluster, u32 new_len) 3004492a8a33STao Ma { 3005492a8a33STao Ma int ret = 0; 30063e19a25eSTristan Ye struct super_block *sb = inode->i_sb; 30073e19a25eSTristan Ye struct ocfs2_caching_info *ci = INODE_CACHE(inode); 3008492a8a33STao Ma int i, blocks = ocfs2_clusters_to_blocks(sb, new_len); 3009492a8a33STao Ma u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster); 3010492a8a33STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 3011492a8a33STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 3012492a8a33STao Ma struct buffer_head *old_bh = NULL; 3013492a8a33STao Ma struct buffer_head *new_bh = NULL; 3014492a8a33STao Ma 3015198aac28STao Ma trace_ocfs2_duplicate_clusters_by_page(cpos, old_cluster, 3016492a8a33STao Ma new_cluster, new_len); 3017492a8a33STao Ma 3018492a8a33STao Ma for (i = 0; i < blocks; i++, old_block++, new_block++) { 3019492a8a33STao Ma new_bh = sb_getblk(osb->sb, new_block); 3020492a8a33STao Ma if (new_bh == NULL) { 30217391a294SRui Xiang ret = -ENOMEM; 3022492a8a33STao Ma mlog_errno(ret); 3023492a8a33STao Ma break; 3024492a8a33STao Ma } 3025492a8a33STao Ma 3026492a8a33STao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 3027492a8a33STao Ma 3028492a8a33STao Ma ret = ocfs2_read_block(ci, old_block, &old_bh, NULL); 3029492a8a33STao Ma if (ret) { 3030492a8a33STao Ma mlog_errno(ret); 3031492a8a33STao Ma break; 3032492a8a33STao Ma } 3033492a8a33STao Ma 3034492a8a33STao Ma ret = ocfs2_journal_access(handle, ci, new_bh, 3035492a8a33STao Ma OCFS2_JOURNAL_ACCESS_CREATE); 3036492a8a33STao Ma if (ret) { 3037492a8a33STao Ma mlog_errno(ret); 3038492a8a33STao Ma break; 3039492a8a33STao Ma } 3040492a8a33STao Ma 3041492a8a33STao Ma memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize); 3042ec20cec7SJoel Becker ocfs2_journal_dirty(handle, new_bh); 3043492a8a33STao Ma 3044492a8a33STao Ma brelse(new_bh); 3045492a8a33STao Ma brelse(old_bh); 3046492a8a33STao Ma new_bh = NULL; 3047492a8a33STao Ma old_bh = NULL; 3048492a8a33STao Ma } 3049492a8a33STao Ma 3050492a8a33STao Ma brelse(new_bh); 3051492a8a33STao Ma brelse(old_bh); 3052492a8a33STao Ma return ret; 3053492a8a33STao Ma } 3054492a8a33STao Ma 30556f70fa51STao Ma static int ocfs2_clear_ext_refcount(handle_t *handle, 30566f70fa51STao Ma struct ocfs2_extent_tree *et, 30576f70fa51STao Ma u32 cpos, u32 p_cluster, u32 len, 30586f70fa51STao Ma unsigned int ext_flags, 30596f70fa51STao Ma struct ocfs2_alloc_context *meta_ac, 30606f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 30616f70fa51STao Ma { 30626f70fa51STao Ma int ret, index; 30636f70fa51STao Ma struct ocfs2_extent_rec replace_rec; 30646f70fa51STao Ma struct ocfs2_path *path = NULL; 30656f70fa51STao Ma struct ocfs2_extent_list *el; 30666f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 30676f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(et->et_ci); 30686f70fa51STao Ma 3069198aac28STao Ma trace_ocfs2_clear_ext_refcount((unsigned long long)ino, 3070198aac28STao Ma cpos, len, p_cluster, ext_flags); 30716f70fa51STao Ma 30726f70fa51STao Ma memset(&replace_rec, 0, sizeof(replace_rec)); 30736f70fa51STao Ma replace_rec.e_cpos = cpu_to_le32(cpos); 30746f70fa51STao Ma replace_rec.e_leaf_clusters = cpu_to_le16(len); 30756f70fa51STao Ma replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb, 30766f70fa51STao Ma p_cluster)); 30776f70fa51STao Ma replace_rec.e_flags = ext_flags; 30786f70fa51STao Ma replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED; 30796f70fa51STao Ma 30806f70fa51STao Ma path = ocfs2_new_path_from_et(et); 30816f70fa51STao Ma if (!path) { 30826f70fa51STao Ma ret = -ENOMEM; 30836f70fa51STao Ma mlog_errno(ret); 30846f70fa51STao Ma goto out; 30856f70fa51STao Ma } 30866f70fa51STao Ma 30876f70fa51STao Ma ret = ocfs2_find_path(et->et_ci, path, cpos); 30886f70fa51STao Ma if (ret) { 30896f70fa51STao Ma mlog_errno(ret); 30906f70fa51STao Ma goto out; 30916f70fa51STao Ma } 30926f70fa51STao Ma 30936f70fa51STao Ma el = path_leaf_el(path); 30946f70fa51STao Ma 30956f70fa51STao Ma index = ocfs2_search_extent_list(el, cpos); 3096981035b4SYingtai Xie if (index == -1) { 309717a5b9abSGoldwyn Rodrigues ret = ocfs2_error(sb, 30987ecef14aSJoe Perches "Inode %llu has an extent at cpos %u which can no longer be found\n", 30996f70fa51STao Ma (unsigned long long)ino, cpos); 31006f70fa51STao Ma goto out; 31016f70fa51STao Ma } 31026f70fa51STao Ma 31036f70fa51STao Ma ret = ocfs2_split_extent(handle, et, path, index, 31046f70fa51STao Ma &replace_rec, meta_ac, dealloc); 31056f70fa51STao Ma if (ret) 31066f70fa51STao Ma mlog_errno(ret); 31076f70fa51STao Ma 31086f70fa51STao Ma out: 31096f70fa51STao Ma ocfs2_free_path(path); 31106f70fa51STao Ma return ret; 31116f70fa51STao Ma } 31126f70fa51STao Ma 31136f70fa51STao Ma static int ocfs2_replace_clusters(handle_t *handle, 31146f70fa51STao Ma struct ocfs2_cow_context *context, 31156f70fa51STao Ma u32 cpos, u32 old, 31166f70fa51STao Ma u32 new, u32 len, 31176f70fa51STao Ma unsigned int ext_flags) 31186f70fa51STao Ma { 31196f70fa51STao Ma int ret; 3120913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 31216f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(ci); 31226f70fa51STao Ma 3123198aac28STao Ma trace_ocfs2_replace_clusters((unsigned long long)ino, 3124198aac28STao Ma cpos, old, new, len, ext_flags); 31256f70fa51STao Ma 31266f70fa51STao Ma /*If the old clusters is unwritten, no need to duplicate. */ 31276f70fa51STao Ma if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) { 3128c7dd3392STiger Yang ret = context->cow_duplicate_clusters(handle, context->inode, 31293e19a25eSTristan Ye cpos, old, new, len); 31306f70fa51STao Ma if (ret) { 31316f70fa51STao Ma mlog_errno(ret); 31326f70fa51STao Ma goto out; 31336f70fa51STao Ma } 31346f70fa51STao Ma } 31356f70fa51STao Ma 3136913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, &context->data_et, 31376f70fa51STao Ma cpos, new, len, ext_flags, 31386f70fa51STao Ma context->meta_ac, &context->dealloc); 31396f70fa51STao Ma if (ret) 31406f70fa51STao Ma mlog_errno(ret); 31416f70fa51STao Ma out: 31426f70fa51STao Ma return ret; 31436f70fa51STao Ma } 31446f70fa51STao Ma 31453e19a25eSTristan Ye int ocfs2_cow_sync_writeback(struct super_block *sb, 31463e19a25eSTristan Ye struct inode *inode, 31476f70fa51STao Ma u32 cpos, u32 num_clusters) 31486f70fa51STao Ma { 31496f70fa51STao Ma int ret = 0; 31506f70fa51STao Ma loff_t offset, end, map_end; 31516f70fa51STao Ma pgoff_t page_index; 31526f70fa51STao Ma struct page *page; 31536f70fa51STao Ma 31543e19a25eSTristan Ye if (ocfs2_should_order_data(inode)) 31556f70fa51STao Ma return 0; 31566f70fa51STao Ma 31576f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 31586f70fa51STao Ma end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits); 31596f70fa51STao Ma 31603e19a25eSTristan Ye ret = filemap_fdatawrite_range(inode->i_mapping, 31616f70fa51STao Ma offset, end - 1); 31626f70fa51STao Ma if (ret < 0) { 31636f70fa51STao Ma mlog_errno(ret); 31646f70fa51STao Ma return ret; 31656f70fa51STao Ma } 31666f70fa51STao Ma 31676f70fa51STao Ma while (offset < end) { 316809cbfeafSKirill A. Shutemov page_index = offset >> PAGE_SHIFT; 316909cbfeafSKirill A. Shutemov map_end = ((loff_t)page_index + 1) << PAGE_SHIFT; 31706f70fa51STao Ma if (map_end > end) 31716f70fa51STao Ma map_end = end; 31726f70fa51STao Ma 31733e19a25eSTristan Ye page = find_or_create_page(inode->i_mapping, 31749b4c0ff3SJan Kara page_index, GFP_NOFS); 31756f70fa51STao Ma BUG_ON(!page); 31766f70fa51STao Ma 31776f70fa51STao Ma wait_on_page_writeback(page); 31786f70fa51STao Ma if (PageError(page)) { 31796f70fa51STao Ma ret = -EIO; 31806f70fa51STao Ma mlog_errno(ret); 31816f70fa51STao Ma } else 31826f70fa51STao Ma mark_page_accessed(page); 31836f70fa51STao Ma 31846f70fa51STao Ma unlock_page(page); 318509cbfeafSKirill A. Shutemov put_page(page); 31866f70fa51STao Ma page = NULL; 31876f70fa51STao Ma offset = map_end; 31886f70fa51STao Ma if (ret) 31896f70fa51STao Ma break; 31906f70fa51STao Ma } 31916f70fa51STao Ma 31926f70fa51STao Ma return ret; 31936f70fa51STao Ma } 31946f70fa51STao Ma 3195913580b4STao Ma static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context, 3196913580b4STao Ma u32 v_cluster, u32 *p_cluster, 3197913580b4STao Ma u32 *num_clusters, 3198913580b4STao Ma unsigned int *extent_flags) 3199913580b4STao Ma { 3200913580b4STao Ma return ocfs2_get_clusters(context->inode, v_cluster, p_cluster, 3201913580b4STao Ma num_clusters, extent_flags); 3202913580b4STao Ma } 3203913580b4STao Ma 32046f70fa51STao Ma static int ocfs2_make_clusters_writable(struct super_block *sb, 32056f70fa51STao Ma struct ocfs2_cow_context *context, 32066f70fa51STao Ma u32 cpos, u32 p_cluster, 32076f70fa51STao Ma u32 num_clusters, unsigned int e_flags) 32086f70fa51STao Ma { 32096ae23c55STao Ma int ret, delete, index, credits = 0; 3210acf3bb00STristan Ye u32 new_bit, new_len, orig_num_clusters; 32116ae23c55STao Ma unsigned int set_len; 32126f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 32136f70fa51STao Ma handle_t *handle; 32146ae23c55STao Ma struct buffer_head *ref_leaf_bh = NULL; 3215913580b4STao Ma struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci; 32166ae23c55STao Ma struct ocfs2_refcount_rec rec; 32176ae23c55STao Ma 3218198aac28STao Ma trace_ocfs2_make_clusters_writable(cpos, p_cluster, 3219198aac28STao Ma num_clusters, e_flags); 32206f70fa51STao Ma 32216f70fa51STao Ma ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters, 3222913580b4STao Ma &context->data_et, 3223913580b4STao Ma ref_ci, 32246f70fa51STao Ma context->ref_root_bh, 32256f70fa51STao Ma &context->meta_ac, 32266f70fa51STao Ma &context->data_ac, &credits); 32276f70fa51STao Ma if (ret) { 32286f70fa51STao Ma mlog_errno(ret); 32296f70fa51STao Ma return ret; 32306f70fa51STao Ma } 32316f70fa51STao Ma 3232492a8a33STao Ma if (context->post_refcount) 3233492a8a33STao Ma credits += context->post_refcount->credits; 3234492a8a33STao Ma 3235492a8a33STao Ma credits += context->extra_credits; 32366f70fa51STao Ma handle = ocfs2_start_trans(osb, credits); 32376f70fa51STao Ma if (IS_ERR(handle)) { 32386f70fa51STao Ma ret = PTR_ERR(handle); 32396f70fa51STao Ma mlog_errno(ret); 32406f70fa51STao Ma goto out; 32416f70fa51STao Ma } 32426f70fa51STao Ma 3243acf3bb00STristan Ye orig_num_clusters = num_clusters; 3244acf3bb00STristan Ye 32456f70fa51STao Ma while (num_clusters) { 3246913580b4STao Ma ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh, 32476ae23c55STao Ma p_cluster, num_clusters, 32486ae23c55STao Ma &rec, &index, &ref_leaf_bh); 32496ae23c55STao Ma if (ret) { 32506ae23c55STao Ma mlog_errno(ret); 32516ae23c55STao Ma goto out_commit; 32526ae23c55STao Ma } 32536ae23c55STao Ma 32546ae23c55STao Ma BUG_ON(!rec.r_refcount); 32556ae23c55STao Ma set_len = min((u64)p_cluster + num_clusters, 32566ae23c55STao Ma le64_to_cpu(rec.r_cpos) + 32576ae23c55STao Ma le32_to_cpu(rec.r_clusters)) - p_cluster; 32586ae23c55STao Ma 32596ae23c55STao Ma /* 32606ae23c55STao Ma * There are many different situation here. 32616ae23c55STao Ma * 1. If refcount == 1, remove the flag and don't COW. 32626ae23c55STao Ma * 2. If refcount > 1, allocate clusters. 32636ae23c55STao Ma * Here we may not allocate r_len once at a time, so continue 32646ae23c55STao Ma * until we reach num_clusters. 32656ae23c55STao Ma */ 32666ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1) { 32676ae23c55STao Ma delete = 0; 3268913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, 3269913580b4STao Ma &context->data_et, 32706ae23c55STao Ma cpos, p_cluster, 32716ae23c55STao Ma set_len, e_flags, 32726ae23c55STao Ma context->meta_ac, 32736ae23c55STao Ma &context->dealloc); 32746ae23c55STao Ma if (ret) { 32756ae23c55STao Ma mlog_errno(ret); 32766ae23c55STao Ma goto out_commit; 32776ae23c55STao Ma } 32786ae23c55STao Ma } else { 32796ae23c55STao Ma delete = 1; 32806ae23c55STao Ma 32811ed9b777SJoel Becker ret = __ocfs2_claim_clusters(handle, 32826ae23c55STao Ma context->data_ac, 32836ae23c55STao Ma 1, set_len, 32846f70fa51STao Ma &new_bit, &new_len); 32856f70fa51STao Ma if (ret) { 32866f70fa51STao Ma mlog_errno(ret); 32876f70fa51STao Ma goto out_commit; 32886f70fa51STao Ma } 32896f70fa51STao Ma 32906f70fa51STao Ma ret = ocfs2_replace_clusters(handle, context, 32916f70fa51STao Ma cpos, p_cluster, new_bit, 32926f70fa51STao Ma new_len, e_flags); 32936f70fa51STao Ma if (ret) { 32946f70fa51STao Ma mlog_errno(ret); 32956f70fa51STao Ma goto out_commit; 32966f70fa51STao Ma } 32976ae23c55STao Ma set_len = new_len; 32986f70fa51STao Ma } 32996f70fa51STao Ma 3300913580b4STao Ma ret = __ocfs2_decrease_refcount(handle, ref_ci, 33016f70fa51STao Ma context->ref_root_bh, 33026ae23c55STao Ma p_cluster, set_len, 33036f70fa51STao Ma context->meta_ac, 33046ae23c55STao Ma &context->dealloc, delete); 33056f70fa51STao Ma if (ret) { 33066f70fa51STao Ma mlog_errno(ret); 33076f70fa51STao Ma goto out_commit; 33086f70fa51STao Ma } 33096f70fa51STao Ma 33106ae23c55STao Ma cpos += set_len; 33116ae23c55STao Ma p_cluster += set_len; 33126ae23c55STao Ma num_clusters -= set_len; 33136ae23c55STao Ma brelse(ref_leaf_bh); 33146ae23c55STao Ma ref_leaf_bh = NULL; 33156ae23c55STao Ma } 33166ae23c55STao Ma 3317492a8a33STao Ma /* handle any post_cow action. */ 3318492a8a33STao Ma if (context->post_refcount && context->post_refcount->func) { 3319492a8a33STao Ma ret = context->post_refcount->func(context->inode, handle, 3320492a8a33STao Ma context->post_refcount->para); 3321492a8a33STao Ma if (ret) { 3322492a8a33STao Ma mlog_errno(ret); 3323492a8a33STao Ma goto out_commit; 3324492a8a33STao Ma } 3325492a8a33STao Ma } 3326492a8a33STao Ma 33276f70fa51STao Ma /* 33286f70fa51STao Ma * Here we should write the new page out first if we are 33296f70fa51STao Ma * in write-back mode. 33306f70fa51STao Ma */ 3331492a8a33STao Ma if (context->get_clusters == ocfs2_di_get_clusters) { 33323e19a25eSTristan Ye ret = ocfs2_cow_sync_writeback(sb, context->inode, cpos, 3333acf3bb00STristan Ye orig_num_clusters); 33346f70fa51STao Ma if (ret) 33356f70fa51STao Ma mlog_errno(ret); 3336492a8a33STao Ma } 33376f70fa51STao Ma 33386f70fa51STao Ma out_commit: 33396f70fa51STao Ma ocfs2_commit_trans(osb, handle); 33406f70fa51STao Ma 33416f70fa51STao Ma out: 33426f70fa51STao Ma if (context->data_ac) { 33436f70fa51STao Ma ocfs2_free_alloc_context(context->data_ac); 33446f70fa51STao Ma context->data_ac = NULL; 33456f70fa51STao Ma } 33466f70fa51STao Ma if (context->meta_ac) { 33476f70fa51STao Ma ocfs2_free_alloc_context(context->meta_ac); 33486f70fa51STao Ma context->meta_ac = NULL; 33496f70fa51STao Ma } 33506ae23c55STao Ma brelse(ref_leaf_bh); 33516f70fa51STao Ma 33526f70fa51STao Ma return ret; 33536f70fa51STao Ma } 33546f70fa51STao Ma 3355913580b4STao Ma static int ocfs2_replace_cow(struct ocfs2_cow_context *context) 33566f70fa51STao Ma { 33576f70fa51STao Ma int ret = 0; 3358913580b4STao Ma struct inode *inode = context->inode; 3359913580b4STao Ma u32 cow_start = context->cow_start, cow_len = context->cow_len; 3360913580b4STao Ma u32 p_cluster, num_clusters; 33616f70fa51STao Ma unsigned int ext_flags; 33626f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 33636f70fa51STao Ma 33641119d3c0Spiaojun if (!ocfs2_refcount_tree(osb)) { 33657ecef14aSJoe Perches return ocfs2_error(inode->i_sb, "Inode %lu want to use refcount tree, but the feature bit is not set in the super block\n", 33667ecef14aSJoe Perches inode->i_ino); 33676f70fa51STao Ma } 33686f70fa51STao Ma 33696f70fa51STao Ma ocfs2_init_dealloc_ctxt(&context->dealloc); 33706f70fa51STao Ma 33716f70fa51STao Ma while (cow_len) { 3372913580b4STao Ma ret = context->get_clusters(context, cow_start, &p_cluster, 33736f70fa51STao Ma &num_clusters, &ext_flags); 33746f70fa51STao Ma if (ret) { 33756f70fa51STao Ma mlog_errno(ret); 33766f70fa51STao Ma break; 33776f70fa51STao Ma } 33786f70fa51STao Ma 33796f70fa51STao Ma BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED)); 33806f70fa51STao Ma 33816f70fa51STao Ma if (cow_len < num_clusters) 33826f70fa51STao Ma num_clusters = cow_len; 33836f70fa51STao Ma 33846f70fa51STao Ma ret = ocfs2_make_clusters_writable(inode->i_sb, context, 33856f70fa51STao Ma cow_start, p_cluster, 33866f70fa51STao Ma num_clusters, ext_flags); 33876f70fa51STao Ma if (ret) { 33886f70fa51STao Ma mlog_errno(ret); 33896f70fa51STao Ma break; 33906f70fa51STao Ma } 33916f70fa51STao Ma 33926f70fa51STao Ma cow_len -= num_clusters; 33936f70fa51STao Ma cow_start += num_clusters; 33946f70fa51STao Ma } 33956f70fa51STao Ma 33966f70fa51STao Ma if (ocfs2_dealloc_has_cluster(&context->dealloc)) { 33976f70fa51STao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 33986f70fa51STao Ma ocfs2_run_deallocs(osb, &context->dealloc); 33996f70fa51STao Ma } 34006f70fa51STao Ma 34016f70fa51STao Ma return ret; 34026f70fa51STao Ma } 34036f70fa51STao Ma 34046f70fa51STao Ma /* 340537f8a2bfSTao Ma * Starting at cpos, try to CoW write_len clusters. Don't CoW 340637f8a2bfSTao Ma * past max_cpos. This will stop when it runs into a hole or an 340737f8a2bfSTao Ma * unrefcounted extent. 34086f70fa51STao Ma */ 34096f70fa51STao Ma static int ocfs2_refcount_cow_hunk(struct inode *inode, 34106f70fa51STao Ma struct buffer_head *di_bh, 341137f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 34126f70fa51STao Ma { 34136f70fa51STao Ma int ret; 34146f70fa51STao Ma u32 cow_start = 0, cow_len = 0; 34156f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 34166f70fa51STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 34176f70fa51STao Ma struct buffer_head *ref_root_bh = NULL; 34186f70fa51STao Ma struct ocfs2_refcount_tree *ref_tree; 3419913580b4STao Ma struct ocfs2_cow_context *context = NULL; 34206f70fa51STao Ma 342184e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode)); 34226f70fa51STao Ma 3423913580b4STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list, 342437f8a2bfSTao Ma cpos, write_len, max_cpos, 34256f70fa51STao Ma &cow_start, &cow_len); 34266f70fa51STao Ma if (ret) { 34276f70fa51STao Ma mlog_errno(ret); 34286f70fa51STao Ma goto out; 34296f70fa51STao Ma } 343037f8a2bfSTao Ma 3431198aac28STao Ma trace_ocfs2_refcount_cow_hunk(OCFS2_I(inode)->ip_blkno, 3432198aac28STao Ma cpos, write_len, max_cpos, 3433198aac28STao Ma cow_start, cow_len); 34346f70fa51STao Ma 34356f70fa51STao Ma BUG_ON(cow_len == 0); 34366f70fa51STao Ma 3437913580b4STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3438913580b4STao Ma if (!context) { 3439913580b4STao Ma ret = -ENOMEM; 3440913580b4STao Ma mlog_errno(ret); 3441913580b4STao Ma goto out; 3442913580b4STao Ma } 3443913580b4STao Ma 34446f70fa51STao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 34456f70fa51STao Ma 1, &ref_tree, &ref_root_bh); 34466f70fa51STao Ma if (ret) { 34476f70fa51STao Ma mlog_errno(ret); 34486f70fa51STao Ma goto out; 34496f70fa51STao Ma } 34506f70fa51STao Ma 3451913580b4STao Ma context->inode = inode; 3452913580b4STao Ma context->cow_start = cow_start; 3453913580b4STao Ma context->cow_len = cow_len; 3454913580b4STao Ma context->ref_tree = ref_tree; 3455913580b4STao Ma context->ref_root_bh = ref_root_bh; 3456913580b4STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page; 3457913580b4STao Ma context->get_clusters = ocfs2_di_get_clusters; 3458913580b4STao Ma 3459913580b4STao Ma ocfs2_init_dinode_extent_tree(&context->data_et, 3460913580b4STao Ma INODE_CACHE(inode), di_bh); 3461913580b4STao Ma 3462913580b4STao Ma ret = ocfs2_replace_cow(context); 34636f70fa51STao Ma if (ret) 34646f70fa51STao Ma mlog_errno(ret); 34656f70fa51STao Ma 3466913580b4STao Ma /* 3467913580b4STao Ma * truncate the extent map here since no matter whether we meet with 3468913580b4STao Ma * any error during the action, we shouldn't trust cached extent map 3469913580b4STao Ma * any more. 3470913580b4STao Ma */ 3471913580b4STao Ma ocfs2_extent_map_trunc(inode, cow_start); 3472913580b4STao Ma 34736f70fa51STao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 34746f70fa51STao Ma brelse(ref_root_bh); 34756f70fa51STao Ma out: 3476913580b4STao Ma kfree(context); 34776f70fa51STao Ma return ret; 34786f70fa51STao Ma } 34796f70fa51STao Ma 34806f70fa51STao Ma /* 34816f70fa51STao Ma * CoW any and all clusters between cpos and cpos+write_len. 348237f8a2bfSTao Ma * Don't CoW past max_cpos. If this returns successfully, all 348337f8a2bfSTao Ma * clusters between cpos and cpos+write_len are safe to modify. 34846f70fa51STao Ma */ 34856f70fa51STao Ma int ocfs2_refcount_cow(struct inode *inode, 34866f70fa51STao Ma struct buffer_head *di_bh, 348737f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 34886f70fa51STao Ma { 34896f70fa51STao Ma int ret = 0; 34906f70fa51STao Ma u32 p_cluster, num_clusters; 34916f70fa51STao Ma unsigned int ext_flags; 34926f70fa51STao Ma 34936f70fa51STao Ma while (write_len) { 34946f70fa51STao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 34956f70fa51STao Ma &num_clusters, &ext_flags); 34966f70fa51STao Ma if (ret) { 34976f70fa51STao Ma mlog_errno(ret); 34986f70fa51STao Ma break; 34996f70fa51STao Ma } 35006f70fa51STao Ma 35016f70fa51STao Ma if (write_len < num_clusters) 35026f70fa51STao Ma num_clusters = write_len; 35036f70fa51STao Ma 35046f70fa51STao Ma if (ext_flags & OCFS2_EXT_REFCOUNTED) { 3505c7dd3392STiger Yang ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos, 350637f8a2bfSTao Ma num_clusters, max_cpos); 35076f70fa51STao Ma if (ret) { 35086f70fa51STao Ma mlog_errno(ret); 35096f70fa51STao Ma break; 35106f70fa51STao Ma } 35116f70fa51STao Ma } 35126f70fa51STao Ma 35136f70fa51STao Ma write_len -= num_clusters; 35146f70fa51STao Ma cpos += num_clusters; 35156f70fa51STao Ma } 35166f70fa51STao Ma 35176f70fa51STao Ma return ret; 35186f70fa51STao Ma } 3519110a045aSTao Ma 3520492a8a33STao Ma static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context, 3521492a8a33STao Ma u32 v_cluster, u32 *p_cluster, 3522492a8a33STao Ma u32 *num_clusters, 3523492a8a33STao Ma unsigned int *extent_flags) 3524492a8a33STao Ma { 3525492a8a33STao Ma struct inode *inode = context->inode; 3526492a8a33STao Ma struct ocfs2_xattr_value_root *xv = context->cow_object; 3527492a8a33STao Ma 3528492a8a33STao Ma return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster, 3529492a8a33STao Ma num_clusters, &xv->xr_list, 3530492a8a33STao Ma extent_flags); 3531492a8a33STao Ma } 3532492a8a33STao Ma 3533492a8a33STao Ma /* 3534492a8a33STao Ma * Given a xattr value root, calculate the most meta/credits we need for 3535492a8a33STao Ma * refcount tree change if we truncate it to 0. 3536492a8a33STao Ma */ 3537492a8a33STao Ma int ocfs2_refcounted_xattr_delete_need(struct inode *inode, 3538492a8a33STao Ma struct ocfs2_caching_info *ref_ci, 3539492a8a33STao Ma struct buffer_head *ref_root_bh, 3540492a8a33STao Ma struct ocfs2_xattr_value_root *xv, 3541492a8a33STao Ma int *meta_add, int *credits) 3542492a8a33STao Ma { 3543492a8a33STao Ma int ret = 0, index, ref_blocks = 0; 3544492a8a33STao Ma u32 p_cluster, num_clusters; 3545492a8a33STao Ma u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters); 3546492a8a33STao Ma struct ocfs2_refcount_block *rb; 3547492a8a33STao Ma struct ocfs2_refcount_rec rec; 3548492a8a33STao Ma struct buffer_head *ref_leaf_bh = NULL; 3549492a8a33STao Ma 3550492a8a33STao Ma while (cpos < clusters) { 3551492a8a33STao Ma ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster, 3552492a8a33STao Ma &num_clusters, &xv->xr_list, 3553492a8a33STao Ma NULL); 3554492a8a33STao Ma if (ret) { 3555492a8a33STao Ma mlog_errno(ret); 3556492a8a33STao Ma goto out; 3557492a8a33STao Ma } 3558492a8a33STao Ma 3559492a8a33STao Ma cpos += num_clusters; 3560492a8a33STao Ma 3561492a8a33STao Ma while (num_clusters) { 3562492a8a33STao Ma ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh, 3563492a8a33STao Ma p_cluster, num_clusters, 3564492a8a33STao Ma &rec, &index, 3565492a8a33STao Ma &ref_leaf_bh); 3566492a8a33STao Ma if (ret) { 3567492a8a33STao Ma mlog_errno(ret); 3568492a8a33STao Ma goto out; 3569492a8a33STao Ma } 3570492a8a33STao Ma 3571492a8a33STao Ma BUG_ON(!rec.r_refcount); 3572492a8a33STao Ma 3573492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 3574492a8a33STao Ma 3575492a8a33STao Ma /* 3576492a8a33STao Ma * We really don't know whether the other clusters is in 3577492a8a33STao Ma * this refcount block or not, so just take the worst 3578492a8a33STao Ma * case that all the clusters are in this block and each 3579492a8a33STao Ma * one will split a refcount rec, so totally we need 3580492a8a33STao Ma * clusters * 2 new refcount rec. 3581492a8a33STao Ma */ 3582e1bf4cc6SAl Viro if (le16_to_cpu(rb->rf_records.rl_used) + clusters * 2 > 3583492a8a33STao Ma le16_to_cpu(rb->rf_records.rl_count)) 3584492a8a33STao Ma ref_blocks++; 3585492a8a33STao Ma 3586492a8a33STao Ma *credits += 1; 3587492a8a33STao Ma brelse(ref_leaf_bh); 3588492a8a33STao Ma ref_leaf_bh = NULL; 3589492a8a33STao Ma 3590492a8a33STao Ma if (num_clusters <= le32_to_cpu(rec.r_clusters)) 3591492a8a33STao Ma break; 3592492a8a33STao Ma else 3593492a8a33STao Ma num_clusters -= le32_to_cpu(rec.r_clusters); 3594492a8a33STao Ma p_cluster += num_clusters; 3595492a8a33STao Ma } 3596492a8a33STao Ma } 3597492a8a33STao Ma 3598492a8a33STao Ma *meta_add += ref_blocks; 3599492a8a33STao Ma if (!ref_blocks) 3600492a8a33STao Ma goto out; 3601492a8a33STao Ma 3602492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 3603492a8a33STao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) 3604492a8a33STao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 3605492a8a33STao Ma else { 3606492a8a33STao Ma struct ocfs2_extent_tree et; 3607492a8a33STao Ma 3608492a8a33STao Ma ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh); 3609492a8a33STao Ma *credits += ocfs2_calc_extend_credits(inode->i_sb, 361006f9da6eSGoldwyn Rodrigues et.et_root_el); 3611492a8a33STao Ma } 3612492a8a33STao Ma 3613492a8a33STao Ma out: 3614492a8a33STao Ma brelse(ref_leaf_bh); 3615492a8a33STao Ma return ret; 3616492a8a33STao Ma } 3617492a8a33STao Ma 3618492a8a33STao Ma /* 3619492a8a33STao Ma * Do CoW for xattr. 3620492a8a33STao Ma */ 3621492a8a33STao Ma int ocfs2_refcount_cow_xattr(struct inode *inode, 3622492a8a33STao Ma struct ocfs2_dinode *di, 3623492a8a33STao Ma struct ocfs2_xattr_value_buf *vb, 3624492a8a33STao Ma struct ocfs2_refcount_tree *ref_tree, 3625492a8a33STao Ma struct buffer_head *ref_root_bh, 3626492a8a33STao Ma u32 cpos, u32 write_len, 3627492a8a33STao Ma struct ocfs2_post_refcount *post) 3628492a8a33STao Ma { 3629492a8a33STao Ma int ret; 3630492a8a33STao Ma struct ocfs2_xattr_value_root *xv = vb->vb_xv; 3631492a8a33STao Ma struct ocfs2_cow_context *context = NULL; 3632492a8a33STao Ma u32 cow_start, cow_len; 3633492a8a33STao Ma 363484e40080SDarrick J. Wong BUG_ON(!ocfs2_is_refcount_inode(inode)); 3635492a8a33STao Ma 3636492a8a33STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list, 3637492a8a33STao Ma cpos, write_len, UINT_MAX, 3638492a8a33STao Ma &cow_start, &cow_len); 3639492a8a33STao Ma if (ret) { 3640492a8a33STao Ma mlog_errno(ret); 3641492a8a33STao Ma goto out; 3642492a8a33STao Ma } 3643492a8a33STao Ma 3644492a8a33STao Ma BUG_ON(cow_len == 0); 3645492a8a33STao Ma 3646492a8a33STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3647492a8a33STao Ma if (!context) { 3648492a8a33STao Ma ret = -ENOMEM; 3649492a8a33STao Ma mlog_errno(ret); 3650492a8a33STao Ma goto out; 3651492a8a33STao Ma } 3652492a8a33STao Ma 3653492a8a33STao Ma context->inode = inode; 3654492a8a33STao Ma context->cow_start = cow_start; 3655492a8a33STao Ma context->cow_len = cow_len; 3656492a8a33STao Ma context->ref_tree = ref_tree; 36576eab04a8SJustin P. Mattock context->ref_root_bh = ref_root_bh; 3658492a8a33STao Ma context->cow_object = xv; 3659492a8a33STao Ma 3660492a8a33STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd; 3661492a8a33STao Ma /* We need the extra credits for duplicate_clusters by jbd. */ 3662492a8a33STao Ma context->extra_credits = 3663492a8a33STao Ma ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len; 3664492a8a33STao Ma context->get_clusters = ocfs2_xattr_value_get_clusters; 3665492a8a33STao Ma context->post_refcount = post; 3666492a8a33STao Ma 3667492a8a33STao Ma ocfs2_init_xattr_value_extent_tree(&context->data_et, 3668492a8a33STao Ma INODE_CACHE(inode), vb); 3669492a8a33STao Ma 3670492a8a33STao Ma ret = ocfs2_replace_cow(context); 3671492a8a33STao Ma if (ret) 3672492a8a33STao Ma mlog_errno(ret); 3673492a8a33STao Ma 3674492a8a33STao Ma out: 3675492a8a33STao Ma kfree(context); 3676492a8a33STao Ma return ret; 3677492a8a33STao Ma } 3678492a8a33STao Ma 3679110a045aSTao Ma /* 3680110a045aSTao Ma * Insert a new extent into refcount tree and mark a extent rec 3681110a045aSTao Ma * as refcounted in the dinode tree. 3682110a045aSTao Ma */ 3683110a045aSTao Ma int ocfs2_add_refcount_flag(struct inode *inode, 3684110a045aSTao Ma struct ocfs2_extent_tree *data_et, 3685110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3686110a045aSTao Ma struct buffer_head *ref_root_bh, 3687110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 36880129241eSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 36890129241eSTao Ma struct ocfs2_post_refcount *post) 3690110a045aSTao Ma { 3691110a045aSTao Ma int ret; 3692110a045aSTao Ma handle_t *handle; 3693110a045aSTao Ma int credits = 1, ref_blocks = 0; 3694110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3695110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3696110a045aSTao Ma 36973e10b793SDarrick J. Wong /* We need to be able to handle at least an extent tree split. */ 36983e10b793SDarrick J. Wong ref_blocks = ocfs2_extend_meta_needed(data_et->et_root_el); 36993e10b793SDarrick J. Wong 3700110a045aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 3701110a045aSTao Ma ref_ci, ref_root_bh, 3702110a045aSTao Ma p_cluster, num_clusters, 3703110a045aSTao Ma &ref_blocks, &credits); 3704110a045aSTao Ma if (ret) { 3705110a045aSTao Ma mlog_errno(ret); 3706110a045aSTao Ma goto out; 3707110a045aSTao Ma } 3708110a045aSTao Ma 3709198aac28STao Ma trace_ocfs2_add_refcount_flag(ref_blocks, credits); 3710110a045aSTao Ma 3711110a045aSTao Ma if (ref_blocks) { 37121119d3c0Spiaojun ret = ocfs2_reserve_new_metadata_blocks(osb, 3713110a045aSTao Ma ref_blocks, &meta_ac); 3714110a045aSTao Ma if (ret) { 3715110a045aSTao Ma mlog_errno(ret); 3716110a045aSTao Ma goto out; 3717110a045aSTao Ma } 3718110a045aSTao Ma } 3719110a045aSTao Ma 37200129241eSTao Ma if (post) 37210129241eSTao Ma credits += post->credits; 37220129241eSTao Ma 3723110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3724110a045aSTao Ma if (IS_ERR(handle)) { 3725110a045aSTao Ma ret = PTR_ERR(handle); 3726110a045aSTao Ma mlog_errno(ret); 3727110a045aSTao Ma goto out; 3728110a045aSTao Ma } 3729110a045aSTao Ma 3730110a045aSTao Ma ret = ocfs2_mark_extent_refcounted(inode, data_et, handle, 3731110a045aSTao Ma cpos, num_clusters, p_cluster, 3732110a045aSTao Ma meta_ac, dealloc); 3733110a045aSTao Ma if (ret) { 3734110a045aSTao Ma mlog_errno(ret); 3735110a045aSTao Ma goto out_commit; 3736110a045aSTao Ma } 3737110a045aSTao Ma 37387540c1a7STao Ma ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 37397540c1a7STao Ma p_cluster, num_clusters, 0, 3740110a045aSTao Ma meta_ac, dealloc); 37410129241eSTao Ma if (ret) { 37420129241eSTao Ma mlog_errno(ret); 37430129241eSTao Ma goto out_commit; 37440129241eSTao Ma } 37450129241eSTao Ma 37460129241eSTao Ma if (post && post->func) { 37470129241eSTao Ma ret = post->func(inode, handle, post->para); 3748110a045aSTao Ma if (ret) 3749110a045aSTao Ma mlog_errno(ret); 37500129241eSTao Ma } 3751110a045aSTao Ma 3752110a045aSTao Ma out_commit: 3753110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3754110a045aSTao Ma out: 3755110a045aSTao Ma if (meta_ac) 3756110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3757110a045aSTao Ma return ret; 3758110a045aSTao Ma } 3759110a045aSTao Ma 3760a9063ab9STao Ma static int ocfs2_change_ctime(struct inode *inode, 3761110a045aSTao Ma struct buffer_head *di_bh) 3762110a045aSTao Ma { 3763110a045aSTao Ma int ret; 3764a9063ab9STao Ma handle_t *handle; 3765a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3766a9063ab9STao Ma 3767a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb), 3768a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 3769a9063ab9STao Ma if (IS_ERR(handle)) { 3770a9063ab9STao Ma ret = PTR_ERR(handle); 3771a9063ab9STao Ma mlog_errno(ret); 3772a9063ab9STao Ma goto out; 3773a9063ab9STao Ma } 3774a9063ab9STao Ma 3775a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 3776a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 3777a9063ab9STao Ma if (ret) { 3778a9063ab9STao Ma mlog_errno(ret); 3779a9063ab9STao Ma goto out_commit; 3780a9063ab9STao Ma } 3781a9063ab9STao Ma 3782078cd827SDeepa Dinamani inode->i_ctime = current_time(inode); 3783a9063ab9STao Ma di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec); 3784a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec); 3785a9063ab9STao Ma 3786a9063ab9STao Ma ocfs2_journal_dirty(handle, di_bh); 3787a9063ab9STao Ma 3788a9063ab9STao Ma out_commit: 3789a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle); 3790a9063ab9STao Ma out: 3791a9063ab9STao Ma return ret; 3792a9063ab9STao Ma } 3793a9063ab9STao Ma 3794a9063ab9STao Ma static int ocfs2_attach_refcount_tree(struct inode *inode, 3795a9063ab9STao Ma struct buffer_head *di_bh) 3796a9063ab9STao Ma { 3797a9063ab9STao Ma int ret, data_changed = 0; 3798110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 3799110a045aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 3800110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3801110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3802110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 3803110a045aSTao Ma unsigned int ext_flags; 3804110a045aSTao Ma loff_t size; 3805110a045aSTao Ma u32 cpos, num_clusters, clusters, p_cluster; 3806110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 3807110a045aSTao Ma struct ocfs2_extent_tree di_et; 3808110a045aSTao Ma 3809110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 3810110a045aSTao Ma 381184e40080SDarrick J. Wong if (!ocfs2_is_refcount_inode(inode)) { 3812110a045aSTao Ma ret = ocfs2_create_refcount_tree(inode, di_bh); 3813110a045aSTao Ma if (ret) { 3814110a045aSTao Ma mlog_errno(ret); 3815110a045aSTao Ma goto out; 3816110a045aSTao Ma } 3817110a045aSTao Ma } 3818110a045aSTao Ma 3819110a045aSTao Ma BUG_ON(!di->i_refcount_loc); 3820110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, 3821110a045aSTao Ma le64_to_cpu(di->i_refcount_loc), 1, 3822110a045aSTao Ma &ref_tree, &ref_root_bh); 3823110a045aSTao Ma if (ret) { 3824110a045aSTao Ma mlog_errno(ret); 3825110a045aSTao Ma goto out; 3826110a045aSTao Ma } 3827110a045aSTao Ma 38282f48d593STao Ma if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 38292f48d593STao Ma goto attach_xattr; 38302f48d593STao Ma 3831110a045aSTao Ma ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh); 3832110a045aSTao Ma 3833110a045aSTao Ma size = i_size_read(inode); 3834110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(inode->i_sb, size); 3835110a045aSTao Ma 3836110a045aSTao Ma cpos = 0; 3837110a045aSTao Ma while (cpos < clusters) { 3838110a045aSTao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 3839110a045aSTao Ma &num_clusters, &ext_flags); 38402b0f6eaeSJoseph Qi if (ret) { 38412b0f6eaeSJoseph Qi mlog_errno(ret); 38422b0f6eaeSJoseph Qi goto unlock; 38432b0f6eaeSJoseph Qi } 3844110a045aSTao Ma if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) { 3845110a045aSTao Ma ret = ocfs2_add_refcount_flag(inode, &di_et, 3846110a045aSTao Ma &ref_tree->rf_ci, 3847110a045aSTao Ma ref_root_bh, cpos, 3848110a045aSTao Ma p_cluster, num_clusters, 38490129241eSTao Ma &dealloc, NULL); 3850110a045aSTao Ma if (ret) { 3851110a045aSTao Ma mlog_errno(ret); 3852a9063ab9STao Ma goto unlock; 3853110a045aSTao Ma } 3854a9063ab9STao Ma 3855a9063ab9STao Ma data_changed = 1; 3856110a045aSTao Ma } 3857110a045aSTao Ma cpos += num_clusters; 3858110a045aSTao Ma } 3859110a045aSTao Ma 38602f48d593STao Ma attach_xattr: 38610129241eSTao Ma if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 38620129241eSTao Ma ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh, 38630129241eSTao Ma &ref_tree->rf_ci, 38640129241eSTao Ma ref_root_bh, 38650129241eSTao Ma &dealloc); 38660129241eSTao Ma if (ret) { 38670129241eSTao Ma mlog_errno(ret); 38680129241eSTao Ma goto unlock; 38690129241eSTao Ma } 38700129241eSTao Ma } 38710129241eSTao Ma 3872a9063ab9STao Ma if (data_changed) { 3873a9063ab9STao Ma ret = ocfs2_change_ctime(inode, di_bh); 3874a9063ab9STao Ma if (ret) 3875a9063ab9STao Ma mlog_errno(ret); 3876a9063ab9STao Ma } 3877a9063ab9STao Ma 3878a9063ab9STao Ma unlock: 3879110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 3880110a045aSTao Ma brelse(ref_root_bh); 3881110a045aSTao Ma 3882110a045aSTao Ma if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) { 3883110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 3884110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 3885110a045aSTao Ma } 3886110a045aSTao Ma out: 3887110a045aSTao Ma /* 3888110a045aSTao Ma * Empty the extent map so that we may get the right extent 3889110a045aSTao Ma * record from the disk. 3890110a045aSTao Ma */ 3891110a045aSTao Ma ocfs2_extent_map_trunc(inode, 0); 3892110a045aSTao Ma 3893110a045aSTao Ma return ret; 3894110a045aSTao Ma } 3895110a045aSTao Ma 3896110a045aSTao Ma static int ocfs2_add_refcounted_extent(struct inode *inode, 3897110a045aSTao Ma struct ocfs2_extent_tree *et, 3898110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3899110a045aSTao Ma struct buffer_head *ref_root_bh, 3900110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 3901110a045aSTao Ma unsigned int ext_flags, 3902110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 3903110a045aSTao Ma { 3904110a045aSTao Ma int ret; 3905110a045aSTao Ma handle_t *handle; 3906110a045aSTao Ma int credits = 0; 3907110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3908110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3909110a045aSTao Ma 3910110a045aSTao Ma ret = ocfs2_lock_refcount_allocators(inode->i_sb, 3911110a045aSTao Ma p_cluster, num_clusters, 3912110a045aSTao Ma et, ref_ci, 3913110a045aSTao Ma ref_root_bh, &meta_ac, 3914110a045aSTao Ma NULL, &credits); 3915110a045aSTao Ma if (ret) { 3916110a045aSTao Ma mlog_errno(ret); 3917110a045aSTao Ma goto out; 3918110a045aSTao Ma } 3919110a045aSTao Ma 3920110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3921110a045aSTao Ma if (IS_ERR(handle)) { 3922110a045aSTao Ma ret = PTR_ERR(handle); 3923110a045aSTao Ma mlog_errno(ret); 3924110a045aSTao Ma goto out; 3925110a045aSTao Ma } 3926110a045aSTao Ma 3927110a045aSTao Ma ret = ocfs2_insert_extent(handle, et, cpos, 392812d4cec9STao Ma ocfs2_clusters_to_blocks(inode->i_sb, p_cluster), 3929110a045aSTao Ma num_clusters, ext_flags, meta_ac); 3930110a045aSTao Ma if (ret) { 3931110a045aSTao Ma mlog_errno(ret); 3932110a045aSTao Ma goto out_commit; 3933110a045aSTao Ma } 3934110a045aSTao Ma 39352999d12fSTao Ma ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 3936110a045aSTao Ma p_cluster, num_clusters, 3937110a045aSTao Ma meta_ac, dealloc); 393886e59436SDarrick J. Wong if (ret) { 393986e59436SDarrick J. Wong mlog_errno(ret); 394086e59436SDarrick J. Wong goto out_commit; 394186e59436SDarrick J. Wong } 394286e59436SDarrick J. Wong 394386e59436SDarrick J. Wong ret = dquot_alloc_space_nodirty(inode, 394486e59436SDarrick J. Wong ocfs2_clusters_to_bytes(osb->sb, num_clusters)); 3945110a045aSTao Ma if (ret) 3946110a045aSTao Ma mlog_errno(ret); 3947110a045aSTao Ma 3948110a045aSTao Ma out_commit: 3949110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3950110a045aSTao Ma out: 3951110a045aSTao Ma if (meta_ac) 3952110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3953110a045aSTao Ma return ret; 3954110a045aSTao Ma } 3955110a045aSTao Ma 39562f48d593STao Ma static int ocfs2_duplicate_inline_data(struct inode *s_inode, 39572f48d593STao Ma struct buffer_head *s_bh, 39582f48d593STao Ma struct inode *t_inode, 39592f48d593STao Ma struct buffer_head *t_bh) 39602f48d593STao Ma { 39612f48d593STao Ma int ret; 39622f48d593STao Ma handle_t *handle; 39632f48d593STao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb); 39642f48d593STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data; 39652f48d593STao Ma struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data; 39662f48d593STao Ma 39672f48d593STao Ma BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)); 39682f48d593STao Ma 39692f48d593STao Ma handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 39702f48d593STao Ma if (IS_ERR(handle)) { 39712f48d593STao Ma ret = PTR_ERR(handle); 39722f48d593STao Ma mlog_errno(ret); 39732f48d593STao Ma goto out; 39742f48d593STao Ma } 39752f48d593STao Ma 39762f48d593STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh, 39772f48d593STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 39782f48d593STao Ma if (ret) { 39792f48d593STao Ma mlog_errno(ret); 39802f48d593STao Ma goto out_commit; 39812f48d593STao Ma } 39822f48d593STao Ma 39832f48d593STao Ma t_di->id2.i_data.id_count = s_di->id2.i_data.id_count; 39842f48d593STao Ma memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data, 39852f48d593STao Ma le16_to_cpu(s_di->id2.i_data.id_count)); 39862f48d593STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock); 39872f48d593STao Ma OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL; 39882f48d593STao Ma t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features); 39892f48d593STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock); 39902f48d593STao Ma 39912f48d593STao Ma ocfs2_journal_dirty(handle, t_bh); 39922f48d593STao Ma 39932f48d593STao Ma out_commit: 39942f48d593STao Ma ocfs2_commit_trans(osb, handle); 39952f48d593STao Ma out: 39962f48d593STao Ma return ret; 39972f48d593STao Ma } 39982f48d593STao Ma 3999110a045aSTao Ma static int ocfs2_duplicate_extent_list(struct inode *s_inode, 4000110a045aSTao Ma struct inode *t_inode, 4001110a045aSTao Ma struct buffer_head *t_bh, 4002110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 4003110a045aSTao Ma struct buffer_head *ref_root_bh, 4004110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 4005110a045aSTao Ma { 4006110a045aSTao Ma int ret = 0; 4007110a045aSTao Ma u32 p_cluster, num_clusters, clusters, cpos; 4008110a045aSTao Ma loff_t size; 4009110a045aSTao Ma unsigned int ext_flags; 4010110a045aSTao Ma struct ocfs2_extent_tree et; 4011110a045aSTao Ma 4012110a045aSTao Ma ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh); 4013110a045aSTao Ma 4014110a045aSTao Ma size = i_size_read(s_inode); 4015110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size); 4016110a045aSTao Ma 4017110a045aSTao Ma cpos = 0; 4018110a045aSTao Ma while (cpos < clusters) { 4019110a045aSTao Ma ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster, 4020110a045aSTao Ma &num_clusters, &ext_flags); 40212b0f6eaeSJoseph Qi if (ret) { 40222b0f6eaeSJoseph Qi mlog_errno(ret); 40232b0f6eaeSJoseph Qi goto out; 40242b0f6eaeSJoseph Qi } 4025110a045aSTao Ma if (p_cluster) { 4026110a045aSTao Ma ret = ocfs2_add_refcounted_extent(t_inode, &et, 4027110a045aSTao Ma ref_ci, ref_root_bh, 4028110a045aSTao Ma cpos, p_cluster, 4029110a045aSTao Ma num_clusters, 4030110a045aSTao Ma ext_flags, 4031110a045aSTao Ma dealloc); 4032110a045aSTao Ma if (ret) { 4033110a045aSTao Ma mlog_errno(ret); 4034110a045aSTao Ma goto out; 4035110a045aSTao Ma } 4036110a045aSTao Ma } 4037110a045aSTao Ma 4038110a045aSTao Ma cpos += num_clusters; 4039110a045aSTao Ma } 4040110a045aSTao Ma 4041110a045aSTao Ma out: 4042110a045aSTao Ma return ret; 4043110a045aSTao Ma } 4044110a045aSTao Ma 4045a9063ab9STao Ma /* 4046a9063ab9STao Ma * change the new file's attributes to the src. 4047a9063ab9STao Ma * 4048a9063ab9STao Ma * reflink creates a snapshot of a file, that means the attributes 4049a9063ab9STao Ma * must be identical except for three exceptions - nlink, ino, and ctime. 4050a9063ab9STao Ma */ 4051a9063ab9STao Ma static int ocfs2_complete_reflink(struct inode *s_inode, 4052a9063ab9STao Ma struct buffer_head *s_bh, 4053a9063ab9STao Ma struct inode *t_inode, 40540fe9b66cSTao Ma struct buffer_head *t_bh, 40550fe9b66cSTao Ma bool preserve) 4056a9063ab9STao Ma { 4057a9063ab9STao Ma int ret; 4058a9063ab9STao Ma handle_t *handle; 4059a9063ab9STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data; 4060a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data; 4061a9063ab9STao Ma loff_t size = i_size_read(s_inode); 4062a9063ab9STao Ma 4063a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb), 4064a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 4065a9063ab9STao Ma if (IS_ERR(handle)) { 4066a9063ab9STao Ma ret = PTR_ERR(handle); 4067a9063ab9STao Ma mlog_errno(ret); 4068a9063ab9STao Ma return ret; 4069a9063ab9STao Ma } 4070a9063ab9STao Ma 4071a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh, 4072a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 4073a9063ab9STao Ma if (ret) { 4074a9063ab9STao Ma mlog_errno(ret); 4075a9063ab9STao Ma goto out_commit; 4076a9063ab9STao Ma } 4077a9063ab9STao Ma 4078a9063ab9STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock); 4079a9063ab9STao Ma OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters; 4080a9063ab9STao Ma OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr; 4081a9063ab9STao Ma OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features; 4082a9063ab9STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock); 4083a9063ab9STao Ma i_size_write(t_inode, size); 40846527f8f8STao Ma t_inode->i_blocks = s_inode->i_blocks; 4085a9063ab9STao Ma 4086a9063ab9STao Ma di->i_xattr_inline_size = s_di->i_xattr_inline_size; 4087a9063ab9STao Ma di->i_clusters = s_di->i_clusters; 4088a9063ab9STao Ma di->i_size = s_di->i_size; 4089a9063ab9STao Ma di->i_dyn_features = s_di->i_dyn_features; 4090a9063ab9STao Ma di->i_attr = s_di->i_attr; 40910fe9b66cSTao Ma 40920fe9b66cSTao Ma if (preserve) { 4093c21a534eSTao Ma t_inode->i_uid = s_inode->i_uid; 4094c21a534eSTao Ma t_inode->i_gid = s_inode->i_gid; 4095c21a534eSTao Ma t_inode->i_mode = s_inode->i_mode; 4096a9063ab9STao Ma di->i_uid = s_di->i_uid; 4097a9063ab9STao Ma di->i_gid = s_di->i_gid; 4098a9063ab9STao Ma di->i_mode = s_di->i_mode; 4099a9063ab9STao Ma 4100a9063ab9STao Ma /* 4101a9063ab9STao Ma * update time. 41020fe9b66cSTao Ma * we want mtime to appear identical to the source and 41030fe9b66cSTao Ma * update ctime. 4104a9063ab9STao Ma */ 4105078cd827SDeepa Dinamani t_inode->i_ctime = current_time(t_inode); 4106a9063ab9STao Ma 4107a9063ab9STao Ma di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec); 4108a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec); 4109a9063ab9STao Ma 4110a9063ab9STao Ma t_inode->i_mtime = s_inode->i_mtime; 4111a9063ab9STao Ma di->i_mtime = s_di->i_mtime; 4112a9063ab9STao Ma di->i_mtime_nsec = s_di->i_mtime_nsec; 41130fe9b66cSTao Ma } 4114a9063ab9STao Ma 4115a9063ab9STao Ma ocfs2_journal_dirty(handle, t_bh); 4116a9063ab9STao Ma 4117a9063ab9STao Ma out_commit: 4118a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle); 4119a9063ab9STao Ma return ret; 4120a9063ab9STao Ma } 4121a9063ab9STao Ma 4122110a045aSTao Ma static int ocfs2_create_reflink_node(struct inode *s_inode, 4123110a045aSTao Ma struct buffer_head *s_bh, 4124110a045aSTao Ma struct inode *t_inode, 41250fe9b66cSTao Ma struct buffer_head *t_bh, 41260fe9b66cSTao Ma bool preserve) 4127110a045aSTao Ma { 4128110a045aSTao Ma int ret; 4129110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 4130110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 4131110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb); 4132110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data; 4133110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 4134110a045aSTao Ma 4135110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 4136110a045aSTao Ma 4137110a045aSTao Ma ret = ocfs2_set_refcount_tree(t_inode, t_bh, 4138110a045aSTao Ma le64_to_cpu(di->i_refcount_loc)); 4139110a045aSTao Ma if (ret) { 4140110a045aSTao Ma mlog_errno(ret); 4141110a045aSTao Ma goto out; 4142110a045aSTao Ma } 4143110a045aSTao Ma 41442f48d593STao Ma if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 41452f48d593STao Ma ret = ocfs2_duplicate_inline_data(s_inode, s_bh, 41462f48d593STao Ma t_inode, t_bh); 41472f48d593STao Ma if (ret) 41482f48d593STao Ma mlog_errno(ret); 41492f48d593STao Ma goto out; 41502f48d593STao Ma } 41512f48d593STao Ma 4152110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 4153110a045aSTao Ma 1, &ref_tree, &ref_root_bh); 4154110a045aSTao Ma if (ret) { 4155110a045aSTao Ma mlog_errno(ret); 4156110a045aSTao Ma goto out; 4157110a045aSTao Ma } 4158110a045aSTao Ma 4159110a045aSTao Ma ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh, 4160110a045aSTao Ma &ref_tree->rf_ci, ref_root_bh, 4161110a045aSTao Ma &dealloc); 4162a9063ab9STao Ma if (ret) { 4163a9063ab9STao Ma mlog_errno(ret); 4164a9063ab9STao Ma goto out_unlock_refcount; 4165a9063ab9STao Ma } 4166a9063ab9STao Ma 4167a9063ab9STao Ma out_unlock_refcount: 4168110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 4169110a045aSTao Ma brelse(ref_root_bh); 4170110a045aSTao Ma out: 4171110a045aSTao Ma if (ocfs2_dealloc_has_cluster(&dealloc)) { 4172110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 4173110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 4174110a045aSTao Ma } 4175110a045aSTao Ma 4176110a045aSTao Ma return ret; 4177110a045aSTao Ma } 417809bf27a0STao Ma 417909bf27a0STao Ma static int __ocfs2_reflink(struct dentry *old_dentry, 418009bf27a0STao Ma struct buffer_head *old_bh, 418109bf27a0STao Ma struct inode *new_inode, 418209bf27a0STao Ma bool preserve) 418309bf27a0STao Ma { 418409bf27a0STao Ma int ret; 41852b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry); 418609bf27a0STao Ma struct buffer_head *new_bh = NULL; 418709bf27a0STao Ma 41885693486bSJoel Becker if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) { 41895693486bSJoel Becker ret = -EINVAL; 41905693486bSJoel Becker mlog_errno(ret); 41915693486bSJoel Becker goto out; 41925693486bSJoel Becker } 41935693486bSJoel Becker 419409bf27a0STao Ma ret = filemap_fdatawrite(inode->i_mapping); 419509bf27a0STao Ma if (ret) { 419609bf27a0STao Ma mlog_errno(ret); 419709bf27a0STao Ma goto out; 419809bf27a0STao Ma } 419909bf27a0STao Ma 420009bf27a0STao Ma ret = ocfs2_attach_refcount_tree(inode, old_bh); 420109bf27a0STao Ma if (ret) { 420209bf27a0STao Ma mlog_errno(ret); 420309bf27a0STao Ma goto out; 420409bf27a0STao Ma } 420509bf27a0STao Ma 42065955102cSAl Viro inode_lock_nested(new_inode, I_MUTEX_CHILD); 420707eaac94STao Ma ret = ocfs2_inode_lock_nested(new_inode, &new_bh, 1, 420807eaac94STao Ma OI_LS_REFLINK_TARGET); 420909bf27a0STao Ma if (ret) { 421009bf27a0STao Ma mlog_errno(ret); 421109bf27a0STao Ma goto out_unlock; 421209bf27a0STao Ma } 421309bf27a0STao Ma 421409bf27a0STao Ma ret = ocfs2_create_reflink_node(inode, old_bh, 421509bf27a0STao Ma new_inode, new_bh, preserve); 421609bf27a0STao Ma if (ret) { 421709bf27a0STao Ma mlog_errno(ret); 421809bf27a0STao Ma goto inode_unlock; 421909bf27a0STao Ma } 422009bf27a0STao Ma 422109bf27a0STao Ma if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 422209bf27a0STao Ma ret = ocfs2_reflink_xattrs(inode, old_bh, 422309bf27a0STao Ma new_inode, new_bh, 422409bf27a0STao Ma preserve); 422587f4b1bbSTao Ma if (ret) { 422687f4b1bbSTao Ma mlog_errno(ret); 422787f4b1bbSTao Ma goto inode_unlock; 422887f4b1bbSTao Ma } 422987f4b1bbSTao Ma } 423087f4b1bbSTao Ma 423187f4b1bbSTao Ma ret = ocfs2_complete_reflink(inode, old_bh, 423287f4b1bbSTao Ma new_inode, new_bh, preserve); 423309bf27a0STao Ma if (ret) 423409bf27a0STao Ma mlog_errno(ret); 423587f4b1bbSTao Ma 423609bf27a0STao Ma inode_unlock: 423709bf27a0STao Ma ocfs2_inode_unlock(new_inode, 1); 423809bf27a0STao Ma brelse(new_bh); 423909bf27a0STao Ma out_unlock: 42405955102cSAl Viro inode_unlock(new_inode); 424109bf27a0STao Ma out: 424209bf27a0STao Ma if (!ret) { 424309bf27a0STao Ma ret = filemap_fdatawait(inode->i_mapping); 424409bf27a0STao Ma if (ret) 424509bf27a0STao Ma mlog_errno(ret); 424609bf27a0STao Ma } 424709bf27a0STao Ma return ret; 424809bf27a0STao Ma } 424909bf27a0STao Ma 425009bf27a0STao Ma static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, 425109bf27a0STao Ma struct dentry *new_dentry, bool preserve) 425209bf27a0STao Ma { 4253e4383029SAshish Samant int error, had_lock; 42542b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry); 425509bf27a0STao Ma struct buffer_head *old_bh = NULL; 425609bf27a0STao Ma struct inode *new_orphan_inode = NULL; 4257e4383029SAshish Samant struct ocfs2_lock_holder oh; 425809bf27a0STao Ma 425909bf27a0STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 426009bf27a0STao Ma return -EOPNOTSUPP; 426109bf27a0STao Ma 4262702e5bc6SChristoph Hellwig 4263c25a1e06SJunxiao Bi error = ocfs2_create_inode_in_orphan(dir, inode->i_mode, 426409bf27a0STao Ma &new_orphan_inode); 426509bf27a0STao Ma if (error) { 426609bf27a0STao Ma mlog_errno(error); 426709bf27a0STao Ma goto out; 426809bf27a0STao Ma } 426909bf27a0STao Ma 42708a8ad1c2SWengang Wang error = ocfs2_rw_lock(inode, 1); 42718a8ad1c2SWengang Wang if (error) { 42728a8ad1c2SWengang Wang mlog_errno(error); 42738a8ad1c2SWengang Wang goto out; 42748a8ad1c2SWengang Wang } 42758a8ad1c2SWengang Wang 427609bf27a0STao Ma error = ocfs2_inode_lock(inode, &old_bh, 1); 427709bf27a0STao Ma if (error) { 427809bf27a0STao Ma mlog_errno(error); 42798a8ad1c2SWengang Wang ocfs2_rw_unlock(inode, 1); 428009bf27a0STao Ma goto out; 428109bf27a0STao Ma } 428209bf27a0STao Ma 428309bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_xattr_sem); 428409bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_alloc_sem); 428509bf27a0STao Ma error = __ocfs2_reflink(old_dentry, old_bh, 428609bf27a0STao Ma new_orphan_inode, preserve); 428709bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_alloc_sem); 428809bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_xattr_sem); 428909bf27a0STao Ma 429009bf27a0STao Ma ocfs2_inode_unlock(inode, 1); 42918a8ad1c2SWengang Wang ocfs2_rw_unlock(inode, 1); 429209bf27a0STao Ma brelse(old_bh); 429309bf27a0STao Ma 429409bf27a0STao Ma if (error) { 429509bf27a0STao Ma mlog_errno(error); 429609bf27a0STao Ma goto out; 429709bf27a0STao Ma } 429809bf27a0STao Ma 4299e4383029SAshish Samant had_lock = ocfs2_inode_lock_tracker(new_orphan_inode, NULL, 1, 4300e4383029SAshish Samant &oh); 4301e4383029SAshish Samant if (had_lock < 0) { 4302e4383029SAshish Samant error = had_lock; 4303e4383029SAshish Samant mlog_errno(error); 4304e4383029SAshish Samant goto out; 4305e4383029SAshish Samant } 4306e4383029SAshish Samant 430709bf27a0STao Ma /* If the security isn't preserved, we need to re-initialize them. */ 430809bf27a0STao Ma if (!preserve) { 43092a7dba39SEric Paris error = ocfs2_init_security_and_acl(dir, new_orphan_inode, 4310c25a1e06SJunxiao Bi &new_dentry->d_name); 431109bf27a0STao Ma if (error) 431209bf27a0STao Ma mlog_errno(error); 431309bf27a0STao Ma } 431409bf27a0STao Ma if (!error) { 431509bf27a0STao Ma error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, 431609bf27a0STao Ma new_dentry); 431709bf27a0STao Ma if (error) 431809bf27a0STao Ma mlog_errno(error); 431909bf27a0STao Ma } 4320e4383029SAshish Samant ocfs2_inode_unlock_tracker(new_orphan_inode, 1, &oh, had_lock); 432109bf27a0STao Ma 4322e4383029SAshish Samant out: 432309bf27a0STao Ma if (new_orphan_inode) { 432409bf27a0STao Ma /* 432509bf27a0STao Ma * We need to open_unlock the inode no matter whether we 432609bf27a0STao Ma * succeed or not, so that other nodes can delete it later. 432709bf27a0STao Ma */ 432809bf27a0STao Ma ocfs2_open_unlock(new_orphan_inode); 432909bf27a0STao Ma if (error) 433009bf27a0STao Ma iput(new_orphan_inode); 433109bf27a0STao Ma } 433209bf27a0STao Ma 433309bf27a0STao Ma return error; 433409bf27a0STao Ma } 4335bd50873dSTao Ma 4336bd50873dSTao Ma /* 4337bd50873dSTao Ma * Below here are the bits used by OCFS2_IOC_REFLINK() to fake 4338bd50873dSTao Ma * sys_reflink(). This will go away when vfs_reflink() exists in 4339bd50873dSTao Ma * fs/namei.c. 4340bd50873dSTao Ma */ 4341bd50873dSTao Ma 4342bd50873dSTao Ma /* copied from may_create in VFS. */ 4343bd50873dSTao Ma static inline int ocfs2_may_create(struct inode *dir, struct dentry *child) 4344bd50873dSTao Ma { 43452b0143b5SDavid Howells if (d_really_is_positive(child)) 4346bd50873dSTao Ma return -EEXIST; 4347bd50873dSTao Ma if (IS_DEADDIR(dir)) 4348bd50873dSTao Ma return -ENOENT; 4349bd50873dSTao Ma return inode_permission(dir, MAY_WRITE | MAY_EXEC); 4350bd50873dSTao Ma } 4351bd50873dSTao Ma 4352bd50873dSTao Ma /** 4353bd50873dSTao Ma * ocfs2_vfs_reflink - Create a reference-counted link 4354bd50873dSTao Ma * 4355bd50873dSTao Ma * @old_dentry: source dentry + inode 4356bd50873dSTao Ma * @dir: directory to create the target 4357bd50873dSTao Ma * @new_dentry: target dentry 4358bd50873dSTao Ma * @preserve: if true, preserve all file attributes 4359bd50873dSTao Ma */ 436012d4cec9STao Ma static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir, 4361bd50873dSTao Ma struct dentry *new_dentry, bool preserve) 4362bd50873dSTao Ma { 43632b0143b5SDavid Howells struct inode *inode = d_inode(old_dentry); 4364bd50873dSTao Ma int error; 4365bd50873dSTao Ma 4366bd50873dSTao Ma if (!inode) 4367bd50873dSTao Ma return -ENOENT; 4368bd50873dSTao Ma 4369bd50873dSTao Ma error = ocfs2_may_create(dir, new_dentry); 4370bd50873dSTao Ma if (error) 4371bd50873dSTao Ma return error; 4372bd50873dSTao Ma 4373bd50873dSTao Ma if (dir->i_sb != inode->i_sb) 4374bd50873dSTao Ma return -EXDEV; 4375bd50873dSTao Ma 4376bd50873dSTao Ma /* 4377bd50873dSTao Ma * A reflink to an append-only or immutable file cannot be created. 4378bd50873dSTao Ma */ 4379bd50873dSTao Ma if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) 4380bd50873dSTao Ma return -EPERM; 4381bd50873dSTao Ma 4382bd50873dSTao Ma /* Only regular files can be reflinked. */ 4383bd50873dSTao Ma if (!S_ISREG(inode->i_mode)) 4384bd50873dSTao Ma return -EPERM; 4385bd50873dSTao Ma 4386bd50873dSTao Ma /* 4387bd50873dSTao Ma * If the caller wants to preserve ownership, they require the 4388bd50873dSTao Ma * rights to do so. 4389bd50873dSTao Ma */ 4390bd50873dSTao Ma if (preserve) { 4391488c8ef0SEric W. Biederman if (!uid_eq(current_fsuid(), inode->i_uid) && !capable(CAP_CHOWN)) 4392bd50873dSTao Ma return -EPERM; 4393bd50873dSTao Ma if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN)) 4394bd50873dSTao Ma return -EPERM; 4395bd50873dSTao Ma } 4396bd50873dSTao Ma 4397bd50873dSTao Ma /* 4398bd50873dSTao Ma * If the caller is modifying any aspect of the attributes, they 4399bd50873dSTao Ma * are not creating a snapshot. They need read permission on the 4400bd50873dSTao Ma * file. 4401bd50873dSTao Ma */ 4402bd50873dSTao Ma if (!preserve) { 4403bd50873dSTao Ma error = inode_permission(inode, MAY_READ); 4404bd50873dSTao Ma if (error) 4405bd50873dSTao Ma return error; 4406bd50873dSTao Ma } 4407bd50873dSTao Ma 44085955102cSAl Viro inode_lock(inode); 44099c89fe0aSJan Kara error = dquot_initialize(dir); 44109c89fe0aSJan Kara if (!error) 4411bd50873dSTao Ma error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve); 44125955102cSAl Viro inode_unlock(inode); 4413bd50873dSTao Ma if (!error) 4414bd50873dSTao Ma fsnotify_create(dir, new_dentry); 4415bd50873dSTao Ma return error; 4416bd50873dSTao Ma } 4417bd50873dSTao Ma /* 4418bd50873dSTao Ma * Most codes are copied from sys_linkat. 4419bd50873dSTao Ma */ 4420bd50873dSTao Ma int ocfs2_reflink_ioctl(struct inode *inode, 4421bd50873dSTao Ma const char __user *oldname, 4422bd50873dSTao Ma const char __user *newname, 4423bd50873dSTao Ma bool preserve) 4424bd50873dSTao Ma { 4425bd50873dSTao Ma struct dentry *new_dentry; 4426dae6ad8fSAl Viro struct path old_path, new_path; 4427bd50873dSTao Ma int error; 4428bd50873dSTao Ma 4429bd50873dSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 4430bd50873dSTao Ma return -EOPNOTSUPP; 4431bd50873dSTao Ma 4432bd50873dSTao Ma error = user_path_at(AT_FDCWD, oldname, 0, &old_path); 4433bd50873dSTao Ma if (error) { 4434bd50873dSTao Ma mlog_errno(error); 4435bd50873dSTao Ma return error; 4436bd50873dSTao Ma } 4437bd50873dSTao Ma 4438dae6ad8fSAl Viro new_dentry = user_path_create(AT_FDCWD, newname, &new_path, 0); 4439dae6ad8fSAl Viro error = PTR_ERR(new_dentry); 4440dae6ad8fSAl Viro if (IS_ERR(new_dentry)) { 4441bd50873dSTao Ma mlog_errno(error); 4442bd50873dSTao Ma goto out; 4443bd50873dSTao Ma } 4444bd50873dSTao Ma 4445bd50873dSTao Ma error = -EXDEV; 4446dae6ad8fSAl Viro if (old_path.mnt != new_path.mnt) { 4447bd50873dSTao Ma mlog_errno(error); 4448dae6ad8fSAl Viro goto out_dput; 4449bd50873dSTao Ma } 4450bd50873dSTao Ma 4451bd50873dSTao Ma error = ocfs2_vfs_reflink(old_path.dentry, 44522b0143b5SDavid Howells d_inode(new_path.dentry), 4453bd50873dSTao Ma new_dentry, preserve); 4454bd50873dSTao Ma out_dput: 4455921a1650SAl Viro done_path_create(&new_path, new_dentry); 4456bd50873dSTao Ma out: 4457bd50873dSTao Ma path_put(&old_path); 4458bd50873dSTao Ma 4459bd50873dSTao Ma return error; 4460bd50873dSTao Ma } 446129ac8e85SDarrick J. Wong 446229ac8e85SDarrick J. Wong /* Update destination inode size, if necessary. */ 446365f098e9SDarrick J. Wong int ocfs2_reflink_update_dest(struct inode *dest, 446429ac8e85SDarrick J. Wong struct buffer_head *d_bh, 446529ac8e85SDarrick J. Wong loff_t newlen) 446629ac8e85SDarrick J. Wong { 446729ac8e85SDarrick J. Wong handle_t *handle; 446829ac8e85SDarrick J. Wong int ret; 446929ac8e85SDarrick J. Wong 447029ac8e85SDarrick J. Wong dest->i_blocks = ocfs2_inode_sector_count(dest); 447129ac8e85SDarrick J. Wong 447229ac8e85SDarrick J. Wong if (newlen <= i_size_read(dest)) 447329ac8e85SDarrick J. Wong return 0; 447429ac8e85SDarrick J. Wong 447529ac8e85SDarrick J. Wong handle = ocfs2_start_trans(OCFS2_SB(dest->i_sb), 447629ac8e85SDarrick J. Wong OCFS2_INODE_UPDATE_CREDITS); 447729ac8e85SDarrick J. Wong if (IS_ERR(handle)) { 447829ac8e85SDarrick J. Wong ret = PTR_ERR(handle); 447929ac8e85SDarrick J. Wong mlog_errno(ret); 448029ac8e85SDarrick J. Wong return ret; 448129ac8e85SDarrick J. Wong } 448229ac8e85SDarrick J. Wong 448329ac8e85SDarrick J. Wong /* Extend i_size if needed. */ 448429ac8e85SDarrick J. Wong spin_lock(&OCFS2_I(dest)->ip_lock); 448529ac8e85SDarrick J. Wong if (newlen > i_size_read(dest)) 448629ac8e85SDarrick J. Wong i_size_write(dest, newlen); 448729ac8e85SDarrick J. Wong spin_unlock(&OCFS2_I(dest)->ip_lock); 448829ac8e85SDarrick J. Wong dest->i_ctime = dest->i_mtime = current_time(dest); 448929ac8e85SDarrick J. Wong 449029ac8e85SDarrick J. Wong ret = ocfs2_mark_inode_dirty(handle, dest, d_bh); 449129ac8e85SDarrick J. Wong if (ret) { 449229ac8e85SDarrick J. Wong mlog_errno(ret); 449329ac8e85SDarrick J. Wong goto out_commit; 449429ac8e85SDarrick J. Wong } 449529ac8e85SDarrick J. Wong 449629ac8e85SDarrick J. Wong out_commit: 449729ac8e85SDarrick J. Wong ocfs2_commit_trans(OCFS2_SB(dest->i_sb), handle); 449829ac8e85SDarrick J. Wong return ret; 449929ac8e85SDarrick J. Wong } 450029ac8e85SDarrick J. Wong 450129ac8e85SDarrick J. Wong /* Remap the range pos_in:len in s_inode to pos_out:len in t_inode. */ 4502900611a1SDarrick J. Wong static loff_t ocfs2_reflink_remap_extent(struct inode *s_inode, 450329ac8e85SDarrick J. Wong struct buffer_head *s_bh, 450429ac8e85SDarrick J. Wong loff_t pos_in, 450529ac8e85SDarrick J. Wong struct inode *t_inode, 450629ac8e85SDarrick J. Wong struct buffer_head *t_bh, 450729ac8e85SDarrick J. Wong loff_t pos_out, 450829ac8e85SDarrick J. Wong loff_t len, 450929ac8e85SDarrick J. Wong struct ocfs2_cached_dealloc_ctxt *dealloc) 451029ac8e85SDarrick J. Wong { 451129ac8e85SDarrick J. Wong struct ocfs2_extent_tree s_et; 451229ac8e85SDarrick J. Wong struct ocfs2_extent_tree t_et; 451329ac8e85SDarrick J. Wong struct ocfs2_dinode *dis; 451429ac8e85SDarrick J. Wong struct buffer_head *ref_root_bh = NULL; 451529ac8e85SDarrick J. Wong struct ocfs2_refcount_tree *ref_tree; 451629ac8e85SDarrick J. Wong struct ocfs2_super *osb; 4517900611a1SDarrick J. Wong loff_t remapped_bytes = 0; 451829ac8e85SDarrick J. Wong loff_t pstart, plen; 4519900611a1SDarrick J. Wong u32 p_cluster, num_clusters, slast, spos, tpos, remapped_clus = 0; 452029ac8e85SDarrick J. Wong unsigned int ext_flags; 452129ac8e85SDarrick J. Wong int ret = 0; 452229ac8e85SDarrick J. Wong 452329ac8e85SDarrick J. Wong osb = OCFS2_SB(s_inode->i_sb); 452429ac8e85SDarrick J. Wong dis = (struct ocfs2_dinode *)s_bh->b_data; 452529ac8e85SDarrick J. Wong ocfs2_init_dinode_extent_tree(&s_et, INODE_CACHE(s_inode), s_bh); 452629ac8e85SDarrick J. Wong ocfs2_init_dinode_extent_tree(&t_et, INODE_CACHE(t_inode), t_bh); 452729ac8e85SDarrick J. Wong 452829ac8e85SDarrick J. Wong spos = ocfs2_bytes_to_clusters(s_inode->i_sb, pos_in); 452929ac8e85SDarrick J. Wong tpos = ocfs2_bytes_to_clusters(t_inode->i_sb, pos_out); 453029ac8e85SDarrick J. Wong slast = ocfs2_clusters_for_bytes(s_inode->i_sb, pos_in + len); 453129ac8e85SDarrick J. Wong 453229ac8e85SDarrick J. Wong while (spos < slast) { 453329ac8e85SDarrick J. Wong if (fatal_signal_pending(current)) { 453429ac8e85SDarrick J. Wong ret = -EINTR; 453529ac8e85SDarrick J. Wong goto out; 453629ac8e85SDarrick J. Wong } 453729ac8e85SDarrick J. Wong 453829ac8e85SDarrick J. Wong /* Look up the extent. */ 453929ac8e85SDarrick J. Wong ret = ocfs2_get_clusters(s_inode, spos, &p_cluster, 454029ac8e85SDarrick J. Wong &num_clusters, &ext_flags); 454129ac8e85SDarrick J. Wong if (ret) { 454229ac8e85SDarrick J. Wong mlog_errno(ret); 454329ac8e85SDarrick J. Wong goto out; 454429ac8e85SDarrick J. Wong } 454529ac8e85SDarrick J. Wong 454629ac8e85SDarrick J. Wong num_clusters = min_t(u32, num_clusters, slast - spos); 454729ac8e85SDarrick J. Wong 454829ac8e85SDarrick J. Wong /* Punch out the dest range. */ 454929ac8e85SDarrick J. Wong pstart = ocfs2_clusters_to_bytes(t_inode->i_sb, tpos); 455029ac8e85SDarrick J. Wong plen = ocfs2_clusters_to_bytes(t_inode->i_sb, num_clusters); 455129ac8e85SDarrick J. Wong ret = ocfs2_remove_inode_range(t_inode, t_bh, pstart, plen); 455229ac8e85SDarrick J. Wong if (ret) { 455329ac8e85SDarrick J. Wong mlog_errno(ret); 455429ac8e85SDarrick J. Wong goto out; 455529ac8e85SDarrick J. Wong } 455629ac8e85SDarrick J. Wong 455729ac8e85SDarrick J. Wong if (p_cluster == 0) 455829ac8e85SDarrick J. Wong goto next_loop; 455929ac8e85SDarrick J. Wong 456029ac8e85SDarrick J. Wong /* Lock the refcount btree... */ 456129ac8e85SDarrick J. Wong ret = ocfs2_lock_refcount_tree(osb, 456229ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc), 456329ac8e85SDarrick J. Wong 1, &ref_tree, &ref_root_bh); 456429ac8e85SDarrick J. Wong if (ret) { 456529ac8e85SDarrick J. Wong mlog_errno(ret); 456629ac8e85SDarrick J. Wong goto out; 456729ac8e85SDarrick J. Wong } 456829ac8e85SDarrick J. Wong 456929ac8e85SDarrick J. Wong /* Mark s_inode's extent as refcounted. */ 457029ac8e85SDarrick J. Wong if (!(ext_flags & OCFS2_EXT_REFCOUNTED)) { 457129ac8e85SDarrick J. Wong ret = ocfs2_add_refcount_flag(s_inode, &s_et, 457229ac8e85SDarrick J. Wong &ref_tree->rf_ci, 457329ac8e85SDarrick J. Wong ref_root_bh, spos, 457429ac8e85SDarrick J. Wong p_cluster, num_clusters, 457529ac8e85SDarrick J. Wong dealloc, NULL); 457629ac8e85SDarrick J. Wong if (ret) { 457729ac8e85SDarrick J. Wong mlog_errno(ret); 457829ac8e85SDarrick J. Wong goto out_unlock_refcount; 457929ac8e85SDarrick J. Wong } 458029ac8e85SDarrick J. Wong } 458129ac8e85SDarrick J. Wong 458229ac8e85SDarrick J. Wong /* Map in the new extent. */ 458329ac8e85SDarrick J. Wong ext_flags |= OCFS2_EXT_REFCOUNTED; 458429ac8e85SDarrick J. Wong ret = ocfs2_add_refcounted_extent(t_inode, &t_et, 458529ac8e85SDarrick J. Wong &ref_tree->rf_ci, 458629ac8e85SDarrick J. Wong ref_root_bh, 458729ac8e85SDarrick J. Wong tpos, p_cluster, 458829ac8e85SDarrick J. Wong num_clusters, 458929ac8e85SDarrick J. Wong ext_flags, 459029ac8e85SDarrick J. Wong dealloc); 459129ac8e85SDarrick J. Wong if (ret) { 459229ac8e85SDarrick J. Wong mlog_errno(ret); 459329ac8e85SDarrick J. Wong goto out_unlock_refcount; 459429ac8e85SDarrick J. Wong } 459529ac8e85SDarrick J. Wong 459629ac8e85SDarrick J. Wong ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 459729ac8e85SDarrick J. Wong brelse(ref_root_bh); 459829ac8e85SDarrick J. Wong next_loop: 459929ac8e85SDarrick J. Wong spos += num_clusters; 460029ac8e85SDarrick J. Wong tpos += num_clusters; 4601900611a1SDarrick J. Wong remapped_clus += num_clusters; 460229ac8e85SDarrick J. Wong } 460329ac8e85SDarrick J. Wong 4604900611a1SDarrick J. Wong goto out; 460529ac8e85SDarrick J. Wong out_unlock_refcount: 460629ac8e85SDarrick J. Wong ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 460729ac8e85SDarrick J. Wong brelse(ref_root_bh); 4608900611a1SDarrick J. Wong out: 4609900611a1SDarrick J. Wong remapped_bytes = ocfs2_clusters_to_bytes(t_inode->i_sb, remapped_clus); 4610900611a1SDarrick J. Wong remapped_bytes = min_t(loff_t, len, remapped_bytes); 4611900611a1SDarrick J. Wong 4612900611a1SDarrick J. Wong return remapped_bytes > 0 ? remapped_bytes : ret; 461329ac8e85SDarrick J. Wong } 461429ac8e85SDarrick J. Wong 461529ac8e85SDarrick J. Wong /* Set up refcount tree and remap s_inode to t_inode. */ 461665f098e9SDarrick J. Wong loff_t ocfs2_reflink_remap_blocks(struct inode *s_inode, 461729ac8e85SDarrick J. Wong struct buffer_head *s_bh, 461829ac8e85SDarrick J. Wong loff_t pos_in, 461929ac8e85SDarrick J. Wong struct inode *t_inode, 462029ac8e85SDarrick J. Wong struct buffer_head *t_bh, 462129ac8e85SDarrick J. Wong loff_t pos_out, 462229ac8e85SDarrick J. Wong loff_t len) 462329ac8e85SDarrick J. Wong { 462429ac8e85SDarrick J. Wong struct ocfs2_cached_dealloc_ctxt dealloc; 462529ac8e85SDarrick J. Wong struct ocfs2_super *osb; 462629ac8e85SDarrick J. Wong struct ocfs2_dinode *dis; 462729ac8e85SDarrick J. Wong struct ocfs2_dinode *dit; 4628900611a1SDarrick J. Wong loff_t ret; 462929ac8e85SDarrick J. Wong 463029ac8e85SDarrick J. Wong osb = OCFS2_SB(s_inode->i_sb); 463129ac8e85SDarrick J. Wong dis = (struct ocfs2_dinode *)s_bh->b_data; 463229ac8e85SDarrick J. Wong dit = (struct ocfs2_dinode *)t_bh->b_data; 463329ac8e85SDarrick J. Wong ocfs2_init_dealloc_ctxt(&dealloc); 463429ac8e85SDarrick J. Wong 463529ac8e85SDarrick J. Wong /* 463629ac8e85SDarrick J. Wong * If we're reflinking the entire file and the source is inline 463729ac8e85SDarrick J. Wong * data, just copy the contents. 463829ac8e85SDarrick J. Wong */ 463929ac8e85SDarrick J. Wong if (pos_in == pos_out && pos_in == 0 && len == i_size_read(s_inode) && 464029ac8e85SDarrick J. Wong i_size_read(t_inode) <= len && 464129ac8e85SDarrick J. Wong (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)) { 464229ac8e85SDarrick J. Wong ret = ocfs2_duplicate_inline_data(s_inode, s_bh, t_inode, t_bh); 464329ac8e85SDarrick J. Wong if (ret) 464429ac8e85SDarrick J. Wong mlog_errno(ret); 464529ac8e85SDarrick J. Wong goto out; 464629ac8e85SDarrick J. Wong } 464729ac8e85SDarrick J. Wong 464829ac8e85SDarrick J. Wong /* 464929ac8e85SDarrick J. Wong * If both inodes belong to two different refcount groups then 465029ac8e85SDarrick J. Wong * forget it because we don't know how (or want) to go merging 465129ac8e85SDarrick J. Wong * refcount trees. 465229ac8e85SDarrick J. Wong */ 465329ac8e85SDarrick J. Wong ret = -EOPNOTSUPP; 465429ac8e85SDarrick J. Wong if (ocfs2_is_refcount_inode(s_inode) && 465529ac8e85SDarrick J. Wong ocfs2_is_refcount_inode(t_inode) && 465629ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc) != 465729ac8e85SDarrick J. Wong le64_to_cpu(dit->i_refcount_loc)) 465829ac8e85SDarrick J. Wong goto out; 465929ac8e85SDarrick J. Wong 466029ac8e85SDarrick J. Wong /* Neither inode has a refcount tree. Add one to s_inode. */ 466129ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(s_inode) && 466229ac8e85SDarrick J. Wong !ocfs2_is_refcount_inode(t_inode)) { 466329ac8e85SDarrick J. Wong ret = ocfs2_create_refcount_tree(s_inode, s_bh); 466429ac8e85SDarrick J. Wong if (ret) { 466529ac8e85SDarrick J. Wong mlog_errno(ret); 466629ac8e85SDarrick J. Wong goto out; 466729ac8e85SDarrick J. Wong } 466829ac8e85SDarrick J. Wong } 466929ac8e85SDarrick J. Wong 467029ac8e85SDarrick J. Wong /* Ensure that both inodes end up with the same refcount tree. */ 467129ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(s_inode)) { 467229ac8e85SDarrick J. Wong ret = ocfs2_set_refcount_tree(s_inode, s_bh, 467329ac8e85SDarrick J. Wong le64_to_cpu(dit->i_refcount_loc)); 467429ac8e85SDarrick J. Wong if (ret) { 467529ac8e85SDarrick J. Wong mlog_errno(ret); 467629ac8e85SDarrick J. Wong goto out; 467729ac8e85SDarrick J. Wong } 467829ac8e85SDarrick J. Wong } 467929ac8e85SDarrick J. Wong if (!ocfs2_is_refcount_inode(t_inode)) { 468029ac8e85SDarrick J. Wong ret = ocfs2_set_refcount_tree(t_inode, t_bh, 468129ac8e85SDarrick J. Wong le64_to_cpu(dis->i_refcount_loc)); 468229ac8e85SDarrick J. Wong if (ret) { 468329ac8e85SDarrick J. Wong mlog_errno(ret); 468429ac8e85SDarrick J. Wong goto out; 468529ac8e85SDarrick J. Wong } 468629ac8e85SDarrick J. Wong } 468729ac8e85SDarrick J. Wong 468829ac8e85SDarrick J. Wong /* Turn off inline data in the dest file. */ 468929ac8e85SDarrick J. Wong if (OCFS2_I(t_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 469029ac8e85SDarrick J. Wong ret = ocfs2_convert_inline_data_to_extents(t_inode, t_bh); 469129ac8e85SDarrick J. Wong if (ret) { 469229ac8e85SDarrick J. Wong mlog_errno(ret); 469329ac8e85SDarrick J. Wong goto out; 469429ac8e85SDarrick J. Wong } 469529ac8e85SDarrick J. Wong } 469629ac8e85SDarrick J. Wong 469729ac8e85SDarrick J. Wong /* Actually remap extents now. */ 469829ac8e85SDarrick J. Wong ret = ocfs2_reflink_remap_extent(s_inode, s_bh, pos_in, t_inode, t_bh, 469929ac8e85SDarrick J. Wong pos_out, len, &dealloc); 4700900611a1SDarrick J. Wong if (ret < 0) { 470129ac8e85SDarrick J. Wong mlog_errno(ret); 470229ac8e85SDarrick J. Wong goto out; 470329ac8e85SDarrick J. Wong } 470429ac8e85SDarrick J. Wong 470529ac8e85SDarrick J. Wong out: 470629ac8e85SDarrick J. Wong if (ocfs2_dealloc_has_cluster(&dealloc)) { 470729ac8e85SDarrick J. Wong ocfs2_schedule_truncate_log_flush(osb, 1); 470829ac8e85SDarrick J. Wong ocfs2_run_deallocs(osb, &dealloc); 470929ac8e85SDarrick J. Wong } 471029ac8e85SDarrick J. Wong 471129ac8e85SDarrick J. Wong return ret; 471229ac8e85SDarrick J. Wong } 471329ac8e85SDarrick J. Wong 471429ac8e85SDarrick J. Wong /* Lock an inode and grab a bh pointing to the inode. */ 471565f098e9SDarrick J. Wong int ocfs2_reflink_inodes_lock(struct inode *s_inode, 4716e6a9467eSDarrick J. Wong struct buffer_head **bh_s, 471729ac8e85SDarrick J. Wong struct inode *t_inode, 4718e6a9467eSDarrick J. Wong struct buffer_head **bh_t) 471929ac8e85SDarrick J. Wong { 4720e6a9467eSDarrick J. Wong struct inode *inode1 = s_inode; 4721e6a9467eSDarrick J. Wong struct inode *inode2 = t_inode; 472229ac8e85SDarrick J. Wong struct ocfs2_inode_info *oi1; 472329ac8e85SDarrick J. Wong struct ocfs2_inode_info *oi2; 4724e6a9467eSDarrick J. Wong struct buffer_head *bh1 = NULL; 4725e6a9467eSDarrick J. Wong struct buffer_head *bh2 = NULL; 472629ac8e85SDarrick J. Wong bool same_inode = (s_inode == t_inode); 4727e6a9467eSDarrick J. Wong bool need_swap = (inode1->i_ino > inode2->i_ino); 472829ac8e85SDarrick J. Wong int status; 472929ac8e85SDarrick J. Wong 473029ac8e85SDarrick J. Wong /* First grab the VFS and rw locks. */ 473129ac8e85SDarrick J. Wong lock_two_nondirectories(s_inode, t_inode); 4732e6a9467eSDarrick J. Wong if (need_swap) 473329ac8e85SDarrick J. Wong swap(inode1, inode2); 473429ac8e85SDarrick J. Wong 473529ac8e85SDarrick J. Wong status = ocfs2_rw_lock(inode1, 1); 473629ac8e85SDarrick J. Wong if (status) { 473729ac8e85SDarrick J. Wong mlog_errno(status); 473829ac8e85SDarrick J. Wong goto out_i1; 473929ac8e85SDarrick J. Wong } 474029ac8e85SDarrick J. Wong if (!same_inode) { 474129ac8e85SDarrick J. Wong status = ocfs2_rw_lock(inode2, 1); 474229ac8e85SDarrick J. Wong if (status) { 474329ac8e85SDarrick J. Wong mlog_errno(status); 474429ac8e85SDarrick J. Wong goto out_i2; 474529ac8e85SDarrick J. Wong } 474629ac8e85SDarrick J. Wong } 474729ac8e85SDarrick J. Wong 474829ac8e85SDarrick J. Wong /* Now go for the cluster locks */ 474929ac8e85SDarrick J. Wong oi1 = OCFS2_I(inode1); 475029ac8e85SDarrick J. Wong oi2 = OCFS2_I(inode2); 475129ac8e85SDarrick J. Wong 475229ac8e85SDarrick J. Wong trace_ocfs2_double_lock((unsigned long long)oi1->ip_blkno, 475329ac8e85SDarrick J. Wong (unsigned long long)oi2->ip_blkno); 475429ac8e85SDarrick J. Wong 475529ac8e85SDarrick J. Wong /* We always want to lock the one with the lower lockid first. */ 475629ac8e85SDarrick J. Wong if (oi1->ip_blkno > oi2->ip_blkno) 475729ac8e85SDarrick J. Wong mlog_errno(-ENOLCK); 475829ac8e85SDarrick J. Wong 475929ac8e85SDarrick J. Wong /* lock id1 */ 4760e6a9467eSDarrick J. Wong status = ocfs2_inode_lock_nested(inode1, &bh1, 1, 4761e6a9467eSDarrick J. Wong OI_LS_REFLINK_TARGET); 476229ac8e85SDarrick J. Wong if (status < 0) { 476329ac8e85SDarrick J. Wong if (status != -ENOENT) 476429ac8e85SDarrick J. Wong mlog_errno(status); 476529ac8e85SDarrick J. Wong goto out_rw2; 476629ac8e85SDarrick J. Wong } 476729ac8e85SDarrick J. Wong 476829ac8e85SDarrick J. Wong /* lock id2 */ 476929ac8e85SDarrick J. Wong if (!same_inode) { 4770e6a9467eSDarrick J. Wong status = ocfs2_inode_lock_nested(inode2, &bh2, 1, 477129ac8e85SDarrick J. Wong OI_LS_REFLINK_TARGET); 477229ac8e85SDarrick J. Wong if (status < 0) { 477329ac8e85SDarrick J. Wong if (status != -ENOENT) 477429ac8e85SDarrick J. Wong mlog_errno(status); 477529ac8e85SDarrick J. Wong goto out_cl1; 477629ac8e85SDarrick J. Wong } 4777e6a9467eSDarrick J. Wong } else { 4778e6a9467eSDarrick J. Wong bh2 = bh1; 4779e6a9467eSDarrick J. Wong } 4780e6a9467eSDarrick J. Wong 4781e6a9467eSDarrick J. Wong /* 4782e6a9467eSDarrick J. Wong * If we swapped inode order above, we have to swap the buffer heads 4783e6a9467eSDarrick J. Wong * before passing them back to the caller. 4784e6a9467eSDarrick J. Wong */ 4785e6a9467eSDarrick J. Wong if (need_swap) 4786e6a9467eSDarrick J. Wong swap(bh1, bh2); 4787e6a9467eSDarrick J. Wong *bh_s = bh1; 4788e6a9467eSDarrick J. Wong *bh_t = bh2; 478929ac8e85SDarrick J. Wong 479029ac8e85SDarrick J. Wong trace_ocfs2_double_lock_end( 4791d324cd4cSpiaojun (unsigned long long)oi1->ip_blkno, 4792d324cd4cSpiaojun (unsigned long long)oi2->ip_blkno); 479329ac8e85SDarrick J. Wong 479429ac8e85SDarrick J. Wong return 0; 479529ac8e85SDarrick J. Wong 479629ac8e85SDarrick J. Wong out_cl1: 479729ac8e85SDarrick J. Wong ocfs2_inode_unlock(inode1, 1); 4798e6a9467eSDarrick J. Wong brelse(bh1); 479929ac8e85SDarrick J. Wong out_rw2: 480029ac8e85SDarrick J. Wong ocfs2_rw_unlock(inode2, 1); 480129ac8e85SDarrick J. Wong out_i2: 480229ac8e85SDarrick J. Wong ocfs2_rw_unlock(inode1, 1); 480329ac8e85SDarrick J. Wong out_i1: 480429ac8e85SDarrick J. Wong unlock_two_nondirectories(s_inode, t_inode); 480529ac8e85SDarrick J. Wong return status; 480629ac8e85SDarrick J. Wong } 480729ac8e85SDarrick J. Wong 480829ac8e85SDarrick J. Wong /* Unlock both inodes and release buffers. */ 480965f098e9SDarrick J. Wong void ocfs2_reflink_inodes_unlock(struct inode *s_inode, 481029ac8e85SDarrick J. Wong struct buffer_head *s_bh, 481129ac8e85SDarrick J. Wong struct inode *t_inode, 481229ac8e85SDarrick J. Wong struct buffer_head *t_bh) 481329ac8e85SDarrick J. Wong { 481429ac8e85SDarrick J. Wong ocfs2_inode_unlock(s_inode, 1); 481529ac8e85SDarrick J. Wong ocfs2_rw_unlock(s_inode, 1); 481629ac8e85SDarrick J. Wong brelse(s_bh); 481729ac8e85SDarrick J. Wong if (s_inode != t_inode) { 481829ac8e85SDarrick J. Wong ocfs2_inode_unlock(t_inode, 1); 481929ac8e85SDarrick J. Wong ocfs2_rw_unlock(t_inode, 1); 482029ac8e85SDarrick J. Wong brelse(t_bh); 482129ac8e85SDarrick J. Wong } 482229ac8e85SDarrick J. Wong unlock_two_nondirectories(s_inode, t_inode); 482329ac8e85SDarrick J. Wong } 4824