1f2c870e3STao Ma /* -*- mode: c; c-basic-offset: 8; -*- 2f2c870e3STao Ma * vim: noexpandtab sw=8 ts=8 sts=0: 3f2c870e3STao Ma * 4f2c870e3STao Ma * refcounttree.c 5f2c870e3STao Ma * 6f2c870e3STao Ma * Copyright (C) 2009 Oracle. All rights reserved. 7f2c870e3STao Ma * 8f2c870e3STao Ma * This program is free software; you can redistribute it and/or 9f2c870e3STao Ma * modify it under the terms of the GNU General Public 10f2c870e3STao Ma * License version 2 as published by the Free Software Foundation. 11f2c870e3STao Ma * 12f2c870e3STao Ma * This program is distributed in the hope that it will be useful, 13f2c870e3STao Ma * but WITHOUT ANY WARRANTY; without even the implied warranty of 14f2c870e3STao Ma * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 15f2c870e3STao Ma * General Public License for more details. 16f2c870e3STao Ma */ 17f2c870e3STao Ma 18e73a819dSTao Ma #include <linux/sort.h> 19f2c870e3STao Ma #define MLOG_MASK_PREFIX ML_REFCOUNT 20f2c870e3STao Ma #include <cluster/masklog.h> 21f2c870e3STao Ma #include "ocfs2.h" 22f2c870e3STao Ma #include "inode.h" 23f2c870e3STao Ma #include "alloc.h" 24f2c870e3STao Ma #include "suballoc.h" 25f2c870e3STao Ma #include "journal.h" 26f2c870e3STao Ma #include "uptodate.h" 27f2c870e3STao Ma #include "super.h" 28f2c870e3STao Ma #include "buffer_head_io.h" 29f2c870e3STao Ma #include "blockcheck.h" 30c732eb16STao Ma #include "refcounttree.h" 318bf396deSTao Ma #include "sysfile.h" 32374a263eSTao Ma #include "dlmglue.h" 33e73a819dSTao Ma #include "extent_map.h" 346f70fa51STao Ma #include "aops.h" 35492a8a33STao Ma #include "xattr.h" 3609bf27a0STao Ma #include "namei.h" 376f70fa51STao Ma 386f70fa51STao Ma #include <linux/bio.h> 396f70fa51STao Ma #include <linux/blkdev.h> 406f70fa51STao Ma #include <linux/slab.h> 416f70fa51STao Ma #include <linux/writeback.h> 426f70fa51STao Ma #include <linux/pagevec.h> 436f70fa51STao Ma #include <linux/swap.h> 44bd50873dSTao Ma #include <linux/security.h> 45bd50873dSTao Ma #include <linux/fsnotify.h> 46bd50873dSTao Ma #include <linux/quotaops.h> 47bd50873dSTao Ma #include <linux/namei.h> 48bd50873dSTao Ma #include <linux/mount.h> 496f70fa51STao Ma 506f70fa51STao Ma struct ocfs2_cow_context { 516f70fa51STao Ma struct inode *inode; 526ea4843fSTao Ma struct file *file; 536f70fa51STao Ma u32 cow_start; 546f70fa51STao Ma u32 cow_len; 55913580b4STao Ma struct ocfs2_extent_tree data_et; 56913580b4STao Ma struct ocfs2_refcount_tree *ref_tree; 576f70fa51STao Ma struct buffer_head *ref_root_bh; 586f70fa51STao Ma struct ocfs2_alloc_context *meta_ac; 596f70fa51STao Ma struct ocfs2_alloc_context *data_ac; 606f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 61492a8a33STao Ma void *cow_object; 62492a8a33STao Ma struct ocfs2_post_refcount *post_refcount; 63492a8a33STao Ma int extra_credits; 64913580b4STao Ma int (*get_clusters)(struct ocfs2_cow_context *context, 65913580b4STao Ma u32 v_cluster, u32 *p_cluster, 66913580b4STao Ma u32 *num_clusters, 67913580b4STao Ma unsigned int *extent_flags); 68913580b4STao Ma int (*cow_duplicate_clusters)(handle_t *handle, 69913580b4STao Ma struct ocfs2_cow_context *context, 70913580b4STao Ma u32 cpos, u32 old_cluster, 71913580b4STao Ma u32 new_cluster, u32 new_len); 726f70fa51STao Ma }; 73c732eb16STao Ma 74c732eb16STao Ma static inline struct ocfs2_refcount_tree * 75c732eb16STao Ma cache_info_to_refcount(struct ocfs2_caching_info *ci) 76c732eb16STao Ma { 77c732eb16STao Ma return container_of(ci, struct ocfs2_refcount_tree, rf_ci); 78c732eb16STao Ma } 79f2c870e3STao Ma 80f2c870e3STao Ma static int ocfs2_validate_refcount_block(struct super_block *sb, 81f2c870e3STao Ma struct buffer_head *bh) 82f2c870e3STao Ma { 83f2c870e3STao Ma int rc; 84f2c870e3STao Ma struct ocfs2_refcount_block *rb = 85f2c870e3STao Ma (struct ocfs2_refcount_block *)bh->b_data; 86f2c870e3STao Ma 87f2c870e3STao Ma mlog(0, "Validating refcount block %llu\n", 88f2c870e3STao Ma (unsigned long long)bh->b_blocknr); 89f2c870e3STao Ma 90f2c870e3STao Ma BUG_ON(!buffer_uptodate(bh)); 91f2c870e3STao Ma 92f2c870e3STao Ma /* 93f2c870e3STao Ma * If the ecc fails, we return the error but otherwise 94f2c870e3STao Ma * leave the filesystem running. We know any error is 95f2c870e3STao Ma * local to this block. 96f2c870e3STao Ma */ 97f2c870e3STao Ma rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &rb->rf_check); 98f2c870e3STao Ma if (rc) { 99f2c870e3STao Ma mlog(ML_ERROR, "Checksum failed for refcount block %llu\n", 100f2c870e3STao Ma (unsigned long long)bh->b_blocknr); 101f2c870e3STao Ma return rc; 102f2c870e3STao Ma } 103f2c870e3STao Ma 104f2c870e3STao Ma 105f2c870e3STao Ma if (!OCFS2_IS_VALID_REFCOUNT_BLOCK(rb)) { 106f2c870e3STao Ma ocfs2_error(sb, 107f2c870e3STao Ma "Refcount block #%llu has bad signature %.*s", 108f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 7, 109f2c870e3STao Ma rb->rf_signature); 110f2c870e3STao Ma return -EINVAL; 111f2c870e3STao Ma } 112f2c870e3STao Ma 113f2c870e3STao Ma if (le64_to_cpu(rb->rf_blkno) != bh->b_blocknr) { 114f2c870e3STao Ma ocfs2_error(sb, 115f2c870e3STao Ma "Refcount block #%llu has an invalid rf_blkno " 116f2c870e3STao Ma "of %llu", 117f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 118f2c870e3STao Ma (unsigned long long)le64_to_cpu(rb->rf_blkno)); 119f2c870e3STao Ma return -EINVAL; 120f2c870e3STao Ma } 121f2c870e3STao Ma 122f2c870e3STao Ma if (le32_to_cpu(rb->rf_fs_generation) != OCFS2_SB(sb)->fs_generation) { 123f2c870e3STao Ma ocfs2_error(sb, 124f2c870e3STao Ma "Refcount block #%llu has an invalid " 125f2c870e3STao Ma "rf_fs_generation of #%u", 126f2c870e3STao Ma (unsigned long long)bh->b_blocknr, 127f2c870e3STao Ma le32_to_cpu(rb->rf_fs_generation)); 128f2c870e3STao Ma return -EINVAL; 129f2c870e3STao Ma } 130f2c870e3STao Ma 131f2c870e3STao Ma return 0; 132f2c870e3STao Ma } 133f2c870e3STao Ma 134f2c870e3STao Ma static int ocfs2_read_refcount_block(struct ocfs2_caching_info *ci, 135f2c870e3STao Ma u64 rb_blkno, 136f2c870e3STao Ma struct buffer_head **bh) 137f2c870e3STao Ma { 138f2c870e3STao Ma int rc; 139f2c870e3STao Ma struct buffer_head *tmp = *bh; 140f2c870e3STao Ma 141f2c870e3STao Ma rc = ocfs2_read_block(ci, rb_blkno, &tmp, 142f2c870e3STao Ma ocfs2_validate_refcount_block); 143f2c870e3STao Ma 144f2c870e3STao Ma /* If ocfs2_read_block() got us a new bh, pass it up. */ 145f2c870e3STao Ma if (!rc && !*bh) 146f2c870e3STao Ma *bh = tmp; 147f2c870e3STao Ma 148f2c870e3STao Ma return rc; 149f2c870e3STao Ma } 150c732eb16STao Ma 151c732eb16STao Ma static u64 ocfs2_refcount_cache_owner(struct ocfs2_caching_info *ci) 152c732eb16STao Ma { 153c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 154c732eb16STao Ma 155c732eb16STao Ma return rf->rf_blkno; 156c732eb16STao Ma } 157c732eb16STao Ma 158c732eb16STao Ma static struct super_block * 159c732eb16STao Ma ocfs2_refcount_cache_get_super(struct ocfs2_caching_info *ci) 160c732eb16STao Ma { 161c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 162c732eb16STao Ma 163c732eb16STao Ma return rf->rf_sb; 164c732eb16STao Ma } 165c732eb16STao Ma 166c732eb16STao Ma static void ocfs2_refcount_cache_lock(struct ocfs2_caching_info *ci) 167c732eb16STao Ma { 168c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 169c732eb16STao Ma 170c732eb16STao Ma spin_lock(&rf->rf_lock); 171c732eb16STao Ma } 172c732eb16STao Ma 173c732eb16STao Ma static void ocfs2_refcount_cache_unlock(struct ocfs2_caching_info *ci) 174c732eb16STao Ma { 175c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 176c732eb16STao Ma 177c732eb16STao Ma spin_unlock(&rf->rf_lock); 178c732eb16STao Ma } 179c732eb16STao Ma 180c732eb16STao Ma static void ocfs2_refcount_cache_io_lock(struct ocfs2_caching_info *ci) 181c732eb16STao Ma { 182c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 183c732eb16STao Ma 184c732eb16STao Ma mutex_lock(&rf->rf_io_mutex); 185c732eb16STao Ma } 186c732eb16STao Ma 187c732eb16STao Ma static void ocfs2_refcount_cache_io_unlock(struct ocfs2_caching_info *ci) 188c732eb16STao Ma { 189c732eb16STao Ma struct ocfs2_refcount_tree *rf = cache_info_to_refcount(ci); 190c732eb16STao Ma 191c732eb16STao Ma mutex_unlock(&rf->rf_io_mutex); 192c732eb16STao Ma } 193c732eb16STao Ma 194c732eb16STao Ma static const struct ocfs2_caching_operations ocfs2_refcount_caching_ops = { 195c732eb16STao Ma .co_owner = ocfs2_refcount_cache_owner, 196c732eb16STao Ma .co_get_super = ocfs2_refcount_cache_get_super, 197c732eb16STao Ma .co_cache_lock = ocfs2_refcount_cache_lock, 198c732eb16STao Ma .co_cache_unlock = ocfs2_refcount_cache_unlock, 199c732eb16STao Ma .co_io_lock = ocfs2_refcount_cache_io_lock, 200c732eb16STao Ma .co_io_unlock = ocfs2_refcount_cache_io_unlock, 201c732eb16STao Ma }; 202374a263eSTao Ma 203374a263eSTao Ma static struct ocfs2_refcount_tree * 204374a263eSTao Ma ocfs2_find_refcount_tree(struct ocfs2_super *osb, u64 blkno) 205374a263eSTao Ma { 206374a263eSTao Ma struct rb_node *n = osb->osb_rf_lock_tree.rb_node; 207374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL; 208374a263eSTao Ma 209374a263eSTao Ma while (n) { 210374a263eSTao Ma tree = rb_entry(n, struct ocfs2_refcount_tree, rf_node); 211374a263eSTao Ma 212374a263eSTao Ma if (blkno < tree->rf_blkno) 213374a263eSTao Ma n = n->rb_left; 214374a263eSTao Ma else if (blkno > tree->rf_blkno) 215374a263eSTao Ma n = n->rb_right; 216374a263eSTao Ma else 217374a263eSTao Ma return tree; 218374a263eSTao Ma } 219374a263eSTao Ma 220374a263eSTao Ma return NULL; 221374a263eSTao Ma } 222374a263eSTao Ma 223374a263eSTao Ma /* osb_lock is already locked. */ 224374a263eSTao Ma static void ocfs2_insert_refcount_tree(struct ocfs2_super *osb, 225374a263eSTao Ma struct ocfs2_refcount_tree *new) 226374a263eSTao Ma { 227374a263eSTao Ma u64 rf_blkno = new->rf_blkno; 228374a263eSTao Ma struct rb_node *parent = NULL; 229374a263eSTao Ma struct rb_node **p = &osb->osb_rf_lock_tree.rb_node; 230374a263eSTao Ma struct ocfs2_refcount_tree *tmp; 231374a263eSTao Ma 232374a263eSTao Ma while (*p) { 233374a263eSTao Ma parent = *p; 234374a263eSTao Ma 235374a263eSTao Ma tmp = rb_entry(parent, struct ocfs2_refcount_tree, 236374a263eSTao Ma rf_node); 237374a263eSTao Ma 238374a263eSTao Ma if (rf_blkno < tmp->rf_blkno) 239374a263eSTao Ma p = &(*p)->rb_left; 240374a263eSTao Ma else if (rf_blkno > tmp->rf_blkno) 241374a263eSTao Ma p = &(*p)->rb_right; 242374a263eSTao Ma else { 243374a263eSTao Ma /* This should never happen! */ 244374a263eSTao Ma mlog(ML_ERROR, "Duplicate refcount block %llu found!\n", 245374a263eSTao Ma (unsigned long long)rf_blkno); 246374a263eSTao Ma BUG(); 247374a263eSTao Ma } 248374a263eSTao Ma } 249374a263eSTao Ma 250374a263eSTao Ma rb_link_node(&new->rf_node, parent, p); 251374a263eSTao Ma rb_insert_color(&new->rf_node, &osb->osb_rf_lock_tree); 252374a263eSTao Ma } 253374a263eSTao Ma 254374a263eSTao Ma static void ocfs2_free_refcount_tree(struct ocfs2_refcount_tree *tree) 255374a263eSTao Ma { 256374a263eSTao Ma ocfs2_metadata_cache_exit(&tree->rf_ci); 257374a263eSTao Ma ocfs2_simple_drop_lockres(OCFS2_SB(tree->rf_sb), &tree->rf_lockres); 258374a263eSTao Ma ocfs2_lock_res_free(&tree->rf_lockres); 259374a263eSTao Ma kfree(tree); 260374a263eSTao Ma } 261374a263eSTao Ma 262374a263eSTao Ma static inline void 263374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(struct ocfs2_super *osb, 264374a263eSTao Ma struct ocfs2_refcount_tree *tree) 265374a263eSTao Ma { 266374a263eSTao Ma rb_erase(&tree->rf_node, &osb->osb_rf_lock_tree); 267374a263eSTao Ma if (osb->osb_ref_tree_lru && osb->osb_ref_tree_lru == tree) 268374a263eSTao Ma osb->osb_ref_tree_lru = NULL; 269374a263eSTao Ma } 270374a263eSTao Ma 271374a263eSTao Ma static void ocfs2_erase_refcount_tree_from_list(struct ocfs2_super *osb, 272374a263eSTao Ma struct ocfs2_refcount_tree *tree) 273374a263eSTao Ma { 274374a263eSTao Ma spin_lock(&osb->osb_lock); 275374a263eSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree); 276374a263eSTao Ma spin_unlock(&osb->osb_lock); 277374a263eSTao Ma } 278374a263eSTao Ma 27912d4cec9STao Ma static void ocfs2_kref_remove_refcount_tree(struct kref *kref) 280374a263eSTao Ma { 281374a263eSTao Ma struct ocfs2_refcount_tree *tree = 282374a263eSTao Ma container_of(kref, struct ocfs2_refcount_tree, rf_getcnt); 283374a263eSTao Ma 284374a263eSTao Ma ocfs2_free_refcount_tree(tree); 285374a263eSTao Ma } 286374a263eSTao Ma 287374a263eSTao Ma static inline void 288374a263eSTao Ma ocfs2_refcount_tree_get(struct ocfs2_refcount_tree *tree) 289374a263eSTao Ma { 290374a263eSTao Ma kref_get(&tree->rf_getcnt); 291374a263eSTao Ma } 292374a263eSTao Ma 293374a263eSTao Ma static inline void 294374a263eSTao Ma ocfs2_refcount_tree_put(struct ocfs2_refcount_tree *tree) 295374a263eSTao Ma { 296374a263eSTao Ma kref_put(&tree->rf_getcnt, ocfs2_kref_remove_refcount_tree); 297374a263eSTao Ma } 298374a263eSTao Ma 299374a263eSTao Ma static inline void ocfs2_init_refcount_tree_ci(struct ocfs2_refcount_tree *new, 300374a263eSTao Ma struct super_block *sb) 301374a263eSTao Ma { 302374a263eSTao Ma ocfs2_metadata_cache_init(&new->rf_ci, &ocfs2_refcount_caching_ops); 303374a263eSTao Ma mutex_init(&new->rf_io_mutex); 304374a263eSTao Ma new->rf_sb = sb; 305374a263eSTao Ma spin_lock_init(&new->rf_lock); 306374a263eSTao Ma } 307374a263eSTao Ma 308374a263eSTao Ma static inline void ocfs2_init_refcount_tree_lock(struct ocfs2_super *osb, 309374a263eSTao Ma struct ocfs2_refcount_tree *new, 310374a263eSTao Ma u64 rf_blkno, u32 generation) 311374a263eSTao Ma { 312374a263eSTao Ma init_rwsem(&new->rf_sem); 313374a263eSTao Ma ocfs2_refcount_lock_res_init(&new->rf_lockres, osb, 314374a263eSTao Ma rf_blkno, generation); 315374a263eSTao Ma } 316374a263eSTao Ma 3178bf396deSTao Ma static struct ocfs2_refcount_tree* 3188bf396deSTao Ma ocfs2_allocate_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno) 3198bf396deSTao Ma { 3208bf396deSTao Ma struct ocfs2_refcount_tree *new; 3218bf396deSTao Ma 3228bf396deSTao Ma new = kzalloc(sizeof(struct ocfs2_refcount_tree), GFP_NOFS); 3238bf396deSTao Ma if (!new) 3248bf396deSTao Ma return NULL; 3258bf396deSTao Ma 3268bf396deSTao Ma new->rf_blkno = rf_blkno; 3278bf396deSTao Ma kref_init(&new->rf_getcnt); 3288bf396deSTao Ma ocfs2_init_refcount_tree_ci(new, osb->sb); 3298bf396deSTao Ma 3308bf396deSTao Ma return new; 3318bf396deSTao Ma } 3328bf396deSTao Ma 333374a263eSTao Ma static int ocfs2_get_refcount_tree(struct ocfs2_super *osb, u64 rf_blkno, 334374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree) 335374a263eSTao Ma { 336374a263eSTao Ma int ret = 0; 337374a263eSTao Ma struct ocfs2_refcount_tree *tree, *new = NULL; 338374a263eSTao Ma struct buffer_head *ref_root_bh = NULL; 339374a263eSTao Ma struct ocfs2_refcount_block *ref_rb; 340374a263eSTao Ma 341374a263eSTao Ma spin_lock(&osb->osb_lock); 342374a263eSTao Ma if (osb->osb_ref_tree_lru && 343374a263eSTao Ma osb->osb_ref_tree_lru->rf_blkno == rf_blkno) 344374a263eSTao Ma tree = osb->osb_ref_tree_lru; 345374a263eSTao Ma else 346374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno); 347374a263eSTao Ma if (tree) 348374a263eSTao Ma goto out; 349374a263eSTao Ma 350374a263eSTao Ma spin_unlock(&osb->osb_lock); 351374a263eSTao Ma 3528bf396deSTao Ma new = ocfs2_allocate_refcount_tree(osb, rf_blkno); 353374a263eSTao Ma if (!new) { 354374a263eSTao Ma ret = -ENOMEM; 3558bf396deSTao Ma mlog_errno(ret); 356374a263eSTao Ma return ret; 357374a263eSTao Ma } 358374a263eSTao Ma /* 359374a263eSTao Ma * We need the generation to create the refcount tree lock and since 360374a263eSTao Ma * it isn't changed during the tree modification, we are safe here to 361374a263eSTao Ma * read without protection. 362374a263eSTao Ma * We also have to purge the cache after we create the lock since the 363374a263eSTao Ma * refcount block may have the stale data. It can only be trusted when 364374a263eSTao Ma * we hold the refcount lock. 365374a263eSTao Ma */ 366374a263eSTao Ma ret = ocfs2_read_refcount_block(&new->rf_ci, rf_blkno, &ref_root_bh); 367374a263eSTao Ma if (ret) { 368374a263eSTao Ma mlog_errno(ret); 369374a263eSTao Ma ocfs2_metadata_cache_exit(&new->rf_ci); 370374a263eSTao Ma kfree(new); 371374a263eSTao Ma return ret; 372374a263eSTao Ma } 373374a263eSTao Ma 374374a263eSTao Ma ref_rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 375374a263eSTao Ma new->rf_generation = le32_to_cpu(ref_rb->rf_generation); 376374a263eSTao Ma ocfs2_init_refcount_tree_lock(osb, new, rf_blkno, 377374a263eSTao Ma new->rf_generation); 378374a263eSTao Ma ocfs2_metadata_cache_purge(&new->rf_ci); 379374a263eSTao Ma 380374a263eSTao Ma spin_lock(&osb->osb_lock); 381374a263eSTao Ma tree = ocfs2_find_refcount_tree(osb, rf_blkno); 382374a263eSTao Ma if (tree) 383374a263eSTao Ma goto out; 384374a263eSTao Ma 385374a263eSTao Ma ocfs2_insert_refcount_tree(osb, new); 386374a263eSTao Ma 387374a263eSTao Ma tree = new; 388374a263eSTao Ma new = NULL; 389374a263eSTao Ma 390374a263eSTao Ma out: 391374a263eSTao Ma *ret_tree = tree; 392374a263eSTao Ma 393374a263eSTao Ma osb->osb_ref_tree_lru = tree; 394374a263eSTao Ma 395374a263eSTao Ma spin_unlock(&osb->osb_lock); 396374a263eSTao Ma 397374a263eSTao Ma if (new) 398374a263eSTao Ma ocfs2_free_refcount_tree(new); 399374a263eSTao Ma 400374a263eSTao Ma brelse(ref_root_bh); 401374a263eSTao Ma return ret; 402374a263eSTao Ma } 403374a263eSTao Ma 404374a263eSTao Ma static int ocfs2_get_refcount_block(struct inode *inode, u64 *ref_blkno) 405374a263eSTao Ma { 406374a263eSTao Ma int ret; 407374a263eSTao Ma struct buffer_head *di_bh = NULL; 408374a263eSTao Ma struct ocfs2_dinode *di; 409374a263eSTao Ma 410374a263eSTao Ma ret = ocfs2_read_inode_block(inode, &di_bh); 411374a263eSTao Ma if (ret) { 412374a263eSTao Ma mlog_errno(ret); 413374a263eSTao Ma goto out; 414374a263eSTao Ma } 415374a263eSTao Ma 416374a263eSTao Ma BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 417374a263eSTao Ma 418374a263eSTao Ma di = (struct ocfs2_dinode *)di_bh->b_data; 419374a263eSTao Ma *ref_blkno = le64_to_cpu(di->i_refcount_loc); 420374a263eSTao Ma brelse(di_bh); 421374a263eSTao Ma out: 422374a263eSTao Ma return ret; 423374a263eSTao Ma } 424374a263eSTao Ma 425374a263eSTao Ma static int __ocfs2_lock_refcount_tree(struct ocfs2_super *osb, 426374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw) 427374a263eSTao Ma { 428374a263eSTao Ma int ret; 429374a263eSTao Ma 430374a263eSTao Ma ret = ocfs2_refcount_lock(tree, rw); 431374a263eSTao Ma if (ret) { 432374a263eSTao Ma mlog_errno(ret); 433374a263eSTao Ma goto out; 434374a263eSTao Ma } 435374a263eSTao Ma 436374a263eSTao Ma if (rw) 437374a263eSTao Ma down_write(&tree->rf_sem); 438374a263eSTao Ma else 439374a263eSTao Ma down_read(&tree->rf_sem); 440374a263eSTao Ma 441374a263eSTao Ma out: 442374a263eSTao Ma return ret; 443374a263eSTao Ma } 444374a263eSTao Ma 445374a263eSTao Ma /* 446374a263eSTao Ma * Lock the refcount tree pointed by ref_blkno and return the tree. 447374a263eSTao Ma * In most case, we lock the tree and read the refcount block. 448374a263eSTao Ma * So read it here if the caller really needs it. 449374a263eSTao Ma * 450374a263eSTao Ma * If the tree has been re-created by other node, it will free the 451374a263eSTao Ma * old one and re-create it. 452374a263eSTao Ma */ 453374a263eSTao Ma int ocfs2_lock_refcount_tree(struct ocfs2_super *osb, 454374a263eSTao Ma u64 ref_blkno, int rw, 455374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree, 456374a263eSTao Ma struct buffer_head **ref_bh) 457374a263eSTao Ma { 458374a263eSTao Ma int ret, delete_tree = 0; 459374a263eSTao Ma struct ocfs2_refcount_tree *tree = NULL; 460374a263eSTao Ma struct buffer_head *ref_root_bh = NULL; 461374a263eSTao Ma struct ocfs2_refcount_block *rb; 462374a263eSTao Ma 463374a263eSTao Ma again: 464374a263eSTao Ma ret = ocfs2_get_refcount_tree(osb, ref_blkno, &tree); 465374a263eSTao Ma if (ret) { 466374a263eSTao Ma mlog_errno(ret); 467374a263eSTao Ma return ret; 468374a263eSTao Ma } 469374a263eSTao Ma 470374a263eSTao Ma ocfs2_refcount_tree_get(tree); 471374a263eSTao Ma 472374a263eSTao Ma ret = __ocfs2_lock_refcount_tree(osb, tree, rw); 473374a263eSTao Ma if (ret) { 474374a263eSTao Ma mlog_errno(ret); 475374a263eSTao Ma ocfs2_refcount_tree_put(tree); 476374a263eSTao Ma goto out; 477374a263eSTao Ma } 478374a263eSTao Ma 479374a263eSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno, 480374a263eSTao Ma &ref_root_bh); 481374a263eSTao Ma if (ret) { 482374a263eSTao Ma mlog_errno(ret); 483374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw); 484374a263eSTao Ma ocfs2_refcount_tree_put(tree); 485374a263eSTao Ma goto out; 486374a263eSTao Ma } 487374a263eSTao Ma 488374a263eSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 489374a263eSTao Ma /* 490374a263eSTao Ma * If the refcount block has been freed and re-created, we may need 491374a263eSTao Ma * to recreate the refcount tree also. 492374a263eSTao Ma * 493374a263eSTao Ma * Here we just remove the tree from the rb-tree, and the last 494374a263eSTao Ma * kref holder will unlock and delete this refcount_tree. 495374a263eSTao Ma * Then we goto "again" and ocfs2_get_refcount_tree will create 496374a263eSTao Ma * the new refcount tree for us. 497374a263eSTao Ma */ 498374a263eSTao Ma if (tree->rf_generation != le32_to_cpu(rb->rf_generation)) { 499374a263eSTao Ma if (!tree->rf_removed) { 500374a263eSTao Ma ocfs2_erase_refcount_tree_from_list(osb, tree); 501374a263eSTao Ma tree->rf_removed = 1; 502374a263eSTao Ma delete_tree = 1; 503374a263eSTao Ma } 504374a263eSTao Ma 505374a263eSTao Ma ocfs2_unlock_refcount_tree(osb, tree, rw); 506374a263eSTao Ma /* 507374a263eSTao Ma * We get an extra reference when we create the refcount 508374a263eSTao Ma * tree, so another put will destroy it. 509374a263eSTao Ma */ 510374a263eSTao Ma if (delete_tree) 511374a263eSTao Ma ocfs2_refcount_tree_put(tree); 512374a263eSTao Ma brelse(ref_root_bh); 513374a263eSTao Ma ref_root_bh = NULL; 514374a263eSTao Ma goto again; 515374a263eSTao Ma } 516374a263eSTao Ma 517374a263eSTao Ma *ret_tree = tree; 518374a263eSTao Ma if (ref_bh) { 519374a263eSTao Ma *ref_bh = ref_root_bh; 520374a263eSTao Ma ref_root_bh = NULL; 521374a263eSTao Ma } 522374a263eSTao Ma out: 523374a263eSTao Ma brelse(ref_root_bh); 524374a263eSTao Ma return ret; 525374a263eSTao Ma } 526374a263eSTao Ma 527374a263eSTao Ma void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb, 528374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw) 529374a263eSTao Ma { 530374a263eSTao Ma if (rw) 531374a263eSTao Ma up_write(&tree->rf_sem); 532374a263eSTao Ma else 533374a263eSTao Ma up_read(&tree->rf_sem); 534374a263eSTao Ma 535374a263eSTao Ma ocfs2_refcount_unlock(tree, rw); 536374a263eSTao Ma ocfs2_refcount_tree_put(tree); 537374a263eSTao Ma } 538374a263eSTao Ma 539374a263eSTao Ma void ocfs2_purge_refcount_trees(struct ocfs2_super *osb) 540374a263eSTao Ma { 541374a263eSTao Ma struct rb_node *node; 542374a263eSTao Ma struct ocfs2_refcount_tree *tree; 543374a263eSTao Ma struct rb_root *root = &osb->osb_rf_lock_tree; 544374a263eSTao Ma 545374a263eSTao Ma while ((node = rb_last(root)) != NULL) { 546374a263eSTao Ma tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node); 547374a263eSTao Ma 548374a263eSTao Ma mlog(0, "Purge tree %llu\n", 549374a263eSTao Ma (unsigned long long) tree->rf_blkno); 550374a263eSTao Ma 551374a263eSTao Ma rb_erase(&tree->rf_node, root); 552374a263eSTao Ma ocfs2_free_refcount_tree(tree); 553374a263eSTao Ma } 554374a263eSTao Ma } 5558bf396deSTao Ma 5568bf396deSTao Ma /* 5578bf396deSTao Ma * Create a refcount tree for an inode. 5588bf396deSTao Ma * We take for granted that the inode is already locked. 5598bf396deSTao Ma */ 5608bf396deSTao Ma static int ocfs2_create_refcount_tree(struct inode *inode, 5618bf396deSTao Ma struct buffer_head *di_bh) 5628bf396deSTao Ma { 5638bf396deSTao Ma int ret; 5648bf396deSTao Ma handle_t *handle = NULL; 5658bf396deSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 5668bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 5678bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 5688bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 5698bf396deSTao Ma struct buffer_head *new_bh = NULL; 5708bf396deSTao Ma struct ocfs2_refcount_block *rb; 5718bf396deSTao Ma struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL; 5728bf396deSTao Ma u16 suballoc_bit_start; 5738bf396deSTao Ma u32 num_got; 5742b6cb576SJoel Becker u64 suballoc_loc, first_blkno; 5758bf396deSTao Ma 5768bf396deSTao Ma BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL); 5778bf396deSTao Ma 5788bf396deSTao Ma mlog(0, "create tree for inode %lu\n", inode->i_ino); 5798bf396deSTao Ma 5808bf396deSTao Ma ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac); 5818bf396deSTao Ma if (ret) { 5828bf396deSTao Ma mlog_errno(ret); 5838bf396deSTao Ma goto out; 5848bf396deSTao Ma } 5858bf396deSTao Ma 5868bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS); 5878bf396deSTao Ma if (IS_ERR(handle)) { 5888bf396deSTao Ma ret = PTR_ERR(handle); 5898bf396deSTao Ma mlog_errno(ret); 5908bf396deSTao Ma goto out; 5918bf396deSTao Ma } 5928bf396deSTao Ma 5938bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 5948bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 5958bf396deSTao Ma if (ret) { 5968bf396deSTao Ma mlog_errno(ret); 5978bf396deSTao Ma goto out_commit; 5988bf396deSTao Ma } 5998bf396deSTao Ma 6002b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 6018bf396deSTao Ma &suballoc_bit_start, &num_got, 6028bf396deSTao Ma &first_blkno); 6038bf396deSTao Ma if (ret) { 6048bf396deSTao Ma mlog_errno(ret); 6058bf396deSTao Ma goto out_commit; 6068bf396deSTao Ma } 6078bf396deSTao Ma 6088bf396deSTao Ma new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno); 6098bf396deSTao Ma if (!new_tree) { 6108bf396deSTao Ma ret = -ENOMEM; 6118bf396deSTao Ma mlog_errno(ret); 6128bf396deSTao Ma goto out_commit; 6138bf396deSTao Ma } 6148bf396deSTao Ma 6158bf396deSTao Ma new_bh = sb_getblk(inode->i_sb, first_blkno); 6168bf396deSTao Ma ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh); 6178bf396deSTao Ma 6188bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh, 6198bf396deSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 6208bf396deSTao Ma if (ret) { 6218bf396deSTao Ma mlog_errno(ret); 6228bf396deSTao Ma goto out_commit; 6238bf396deSTao Ma } 6248bf396deSTao Ma 6258bf396deSTao Ma /* Initialize ocfs2_refcount_block. */ 6268bf396deSTao Ma rb = (struct ocfs2_refcount_block *)new_bh->b_data; 6278bf396deSTao Ma memset(rb, 0, inode->i_sb->s_blocksize); 6288bf396deSTao Ma strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 629b89c5428STiger Yang rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 6302b6cb576SJoel Becker rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 6318bf396deSTao Ma rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 6328bf396deSTao Ma rb->rf_fs_generation = cpu_to_le32(osb->fs_generation); 6338bf396deSTao Ma rb->rf_blkno = cpu_to_le64(first_blkno); 6348bf396deSTao Ma rb->rf_count = cpu_to_le32(1); 6358bf396deSTao Ma rb->rf_records.rl_count = 6368bf396deSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb)); 6378bf396deSTao Ma spin_lock(&osb->osb_lock); 6388bf396deSTao Ma rb->rf_generation = osb->s_next_generation++; 6398bf396deSTao Ma spin_unlock(&osb->osb_lock); 6408bf396deSTao Ma 6418bf396deSTao Ma ocfs2_journal_dirty(handle, new_bh); 6428bf396deSTao Ma 6438bf396deSTao Ma spin_lock(&oi->ip_lock); 6448bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 6458bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 6468bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(first_blkno); 6478bf396deSTao Ma spin_unlock(&oi->ip_lock); 6488bf396deSTao Ma 6498bf396deSTao Ma mlog(0, "created tree for inode %lu, refblock %llu\n", 6508bf396deSTao Ma inode->i_ino, (unsigned long long)first_blkno); 6518bf396deSTao Ma 6528bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 6538bf396deSTao Ma 6548bf396deSTao Ma /* 6558bf396deSTao Ma * We have to init the tree lock here since it will use 6568bf396deSTao Ma * the generation number to create it. 6578bf396deSTao Ma */ 6588bf396deSTao Ma new_tree->rf_generation = le32_to_cpu(rb->rf_generation); 6598bf396deSTao Ma ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno, 6608bf396deSTao Ma new_tree->rf_generation); 6618bf396deSTao Ma 6628bf396deSTao Ma spin_lock(&osb->osb_lock); 6638bf396deSTao Ma tree = ocfs2_find_refcount_tree(osb, first_blkno); 6648bf396deSTao Ma 6658bf396deSTao Ma /* 6668bf396deSTao Ma * We've just created a new refcount tree in this block. If 6678bf396deSTao Ma * we found a refcount tree on the ocfs2_super, it must be 6688bf396deSTao Ma * one we just deleted. We free the old tree before 6698bf396deSTao Ma * inserting the new tree. 6708bf396deSTao Ma */ 6718bf396deSTao Ma BUG_ON(tree && tree->rf_generation == new_tree->rf_generation); 6728bf396deSTao Ma if (tree) 6738bf396deSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree); 6748bf396deSTao Ma ocfs2_insert_refcount_tree(osb, new_tree); 6758bf396deSTao Ma spin_unlock(&osb->osb_lock); 6768bf396deSTao Ma new_tree = NULL; 6778bf396deSTao Ma if (tree) 6788bf396deSTao Ma ocfs2_refcount_tree_put(tree); 6798bf396deSTao Ma 6808bf396deSTao Ma out_commit: 6818bf396deSTao Ma ocfs2_commit_trans(osb, handle); 6828bf396deSTao Ma 6838bf396deSTao Ma out: 6848bf396deSTao Ma if (new_tree) { 6858bf396deSTao Ma ocfs2_metadata_cache_exit(&new_tree->rf_ci); 6868bf396deSTao Ma kfree(new_tree); 6878bf396deSTao Ma } 6888bf396deSTao Ma 6898bf396deSTao Ma brelse(new_bh); 6908bf396deSTao Ma if (meta_ac) 6918bf396deSTao Ma ocfs2_free_alloc_context(meta_ac); 6928bf396deSTao Ma 6938bf396deSTao Ma return ret; 6948bf396deSTao Ma } 6958bf396deSTao Ma 6968bf396deSTao Ma static int ocfs2_set_refcount_tree(struct inode *inode, 6978bf396deSTao Ma struct buffer_head *di_bh, 6988bf396deSTao Ma u64 refcount_loc) 6998bf396deSTao Ma { 7008bf396deSTao Ma int ret; 7018bf396deSTao Ma handle_t *handle = NULL; 7028bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7038bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7048bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7058bf396deSTao Ma struct buffer_head *ref_root_bh = NULL; 7068bf396deSTao Ma struct ocfs2_refcount_block *rb; 7078bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7088bf396deSTao Ma 7098bf396deSTao Ma BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL); 7108bf396deSTao Ma 7118bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1, 7128bf396deSTao Ma &ref_tree, &ref_root_bh); 7138bf396deSTao Ma if (ret) { 7148bf396deSTao Ma mlog_errno(ret); 7158bf396deSTao Ma return ret; 7168bf396deSTao Ma } 7178bf396deSTao Ma 7188bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS); 7198bf396deSTao Ma if (IS_ERR(handle)) { 7208bf396deSTao Ma ret = PTR_ERR(handle); 7218bf396deSTao Ma mlog_errno(ret); 7228bf396deSTao Ma goto out; 7238bf396deSTao Ma } 7248bf396deSTao Ma 7258bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 7268bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7278bf396deSTao Ma if (ret) { 7288bf396deSTao Ma mlog_errno(ret); 7298bf396deSTao Ma goto out_commit; 7308bf396deSTao Ma } 7318bf396deSTao Ma 7328bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh, 7338bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7348bf396deSTao Ma if (ret) { 7358bf396deSTao Ma mlog_errno(ret); 7368bf396deSTao Ma goto out_commit; 7378bf396deSTao Ma } 7388bf396deSTao Ma 7398bf396deSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 7408bf396deSTao Ma le32_add_cpu(&rb->rf_count, 1); 7418bf396deSTao Ma 7428bf396deSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 7438bf396deSTao Ma 7448bf396deSTao Ma spin_lock(&oi->ip_lock); 7458bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 7468bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 7478bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(refcount_loc); 7488bf396deSTao Ma spin_unlock(&oi->ip_lock); 7498bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 7508bf396deSTao Ma 7518bf396deSTao Ma out_commit: 7528bf396deSTao Ma ocfs2_commit_trans(osb, handle); 7538bf396deSTao Ma out: 7548bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 7558bf396deSTao Ma brelse(ref_root_bh); 7568bf396deSTao Ma 7578bf396deSTao Ma return ret; 7588bf396deSTao Ma } 7598bf396deSTao Ma 7608bf396deSTao Ma int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh) 7618bf396deSTao Ma { 7628bf396deSTao Ma int ret, delete_tree = 0; 7638bf396deSTao Ma handle_t *handle = NULL; 7648bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7658bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7668bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7678bf396deSTao Ma struct ocfs2_refcount_block *rb; 7688bf396deSTao Ma struct inode *alloc_inode = NULL; 7698bf396deSTao Ma struct buffer_head *alloc_bh = NULL; 7708bf396deSTao Ma struct buffer_head *blk_bh = NULL; 7718bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7728bf396deSTao Ma int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS; 7738bf396deSTao Ma u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc); 7748bf396deSTao Ma u16 bit = 0; 7758bf396deSTao Ma 7768bf396deSTao Ma if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) 7778bf396deSTao Ma return 0; 7788bf396deSTao Ma 7798bf396deSTao Ma BUG_ON(!ref_blkno); 7808bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh); 7818bf396deSTao Ma if (ret) { 7828bf396deSTao Ma mlog_errno(ret); 7838bf396deSTao Ma return ret; 7848bf396deSTao Ma } 7858bf396deSTao Ma 7868bf396deSTao Ma rb = (struct ocfs2_refcount_block *)blk_bh->b_data; 7878bf396deSTao Ma 7888bf396deSTao Ma /* 7898bf396deSTao Ma * If we are the last user, we need to free the block. 7908bf396deSTao Ma * So lock the allocator ahead. 7918bf396deSTao Ma */ 7928bf396deSTao Ma if (le32_to_cpu(rb->rf_count) == 1) { 7938bf396deSTao Ma blk = le64_to_cpu(rb->rf_blkno); 7948bf396deSTao Ma bit = le16_to_cpu(rb->rf_suballoc_bit); 79574380c47STao Ma if (rb->rf_suballoc_loc) 79674380c47STao Ma bg_blkno = le64_to_cpu(rb->rf_suballoc_loc); 79774380c47STao Ma else 7988bf396deSTao Ma bg_blkno = ocfs2_which_suballoc_group(blk, bit); 7998bf396deSTao Ma 8008bf396deSTao Ma alloc_inode = ocfs2_get_system_file_inode(osb, 8018bf396deSTao Ma EXTENT_ALLOC_SYSTEM_INODE, 8028bf396deSTao Ma le16_to_cpu(rb->rf_suballoc_slot)); 8038bf396deSTao Ma if (!alloc_inode) { 8048bf396deSTao Ma ret = -ENOMEM; 8058bf396deSTao Ma mlog_errno(ret); 8068bf396deSTao Ma goto out; 8078bf396deSTao Ma } 8088bf396deSTao Ma mutex_lock(&alloc_inode->i_mutex); 8098bf396deSTao Ma 8108bf396deSTao Ma ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1); 8118bf396deSTao Ma if (ret) { 8128bf396deSTao Ma mlog_errno(ret); 8138bf396deSTao Ma goto out_mutex; 8148bf396deSTao Ma } 8158bf396deSTao Ma 8168bf396deSTao Ma credits += OCFS2_SUBALLOC_FREE; 8178bf396deSTao Ma } 8188bf396deSTao Ma 8198bf396deSTao Ma handle = ocfs2_start_trans(osb, credits); 8208bf396deSTao Ma if (IS_ERR(handle)) { 8218bf396deSTao Ma ret = PTR_ERR(handle); 8228bf396deSTao Ma mlog_errno(ret); 8238bf396deSTao Ma goto out_unlock; 8248bf396deSTao Ma } 8258bf396deSTao Ma 8268bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 8278bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8288bf396deSTao Ma if (ret) { 8298bf396deSTao Ma mlog_errno(ret); 8308bf396deSTao Ma goto out_commit; 8318bf396deSTao Ma } 8328bf396deSTao Ma 8338bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh, 8348bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8358bf396deSTao Ma if (ret) { 8368bf396deSTao Ma mlog_errno(ret); 8378bf396deSTao Ma goto out_commit; 8388bf396deSTao Ma } 8398bf396deSTao Ma 8408bf396deSTao Ma spin_lock(&oi->ip_lock); 8418bf396deSTao Ma oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL; 8428bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 8438bf396deSTao Ma di->i_refcount_loc = 0; 8448bf396deSTao Ma spin_unlock(&oi->ip_lock); 8458bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 8468bf396deSTao Ma 8478bf396deSTao Ma le32_add_cpu(&rb->rf_count , -1); 8488bf396deSTao Ma ocfs2_journal_dirty(handle, blk_bh); 8498bf396deSTao Ma 8508bf396deSTao Ma if (!rb->rf_count) { 8518bf396deSTao Ma delete_tree = 1; 8528bf396deSTao Ma ocfs2_erase_refcount_tree_from_list(osb, ref_tree); 8538bf396deSTao Ma ret = ocfs2_free_suballoc_bits(handle, alloc_inode, 8548bf396deSTao Ma alloc_bh, bit, bg_blkno, 1); 8558bf396deSTao Ma if (ret) 8568bf396deSTao Ma mlog_errno(ret); 8578bf396deSTao Ma } 8588bf396deSTao Ma 8598bf396deSTao Ma out_commit: 8608bf396deSTao Ma ocfs2_commit_trans(osb, handle); 8618bf396deSTao Ma out_unlock: 8628bf396deSTao Ma if (alloc_inode) { 8638bf396deSTao Ma ocfs2_inode_unlock(alloc_inode, 1); 8648bf396deSTao Ma brelse(alloc_bh); 8658bf396deSTao Ma } 8668bf396deSTao Ma out_mutex: 8678bf396deSTao Ma if (alloc_inode) { 8688bf396deSTao Ma mutex_unlock(&alloc_inode->i_mutex); 8698bf396deSTao Ma iput(alloc_inode); 8708bf396deSTao Ma } 8718bf396deSTao Ma out: 8728bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 8738bf396deSTao Ma if (delete_tree) 8748bf396deSTao Ma ocfs2_refcount_tree_put(ref_tree); 8758bf396deSTao Ma brelse(blk_bh); 8768bf396deSTao Ma 8778bf396deSTao Ma return ret; 8788bf396deSTao Ma } 879e73a819dSTao Ma 880e73a819dSTao Ma static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci, 881e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 882e73a819dSTao Ma u64 cpos, unsigned int len, 883e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 884e73a819dSTao Ma int *index) 885e73a819dSTao Ma { 886e73a819dSTao Ma int i = 0; 887e73a819dSTao Ma struct ocfs2_refcount_block *rb = 888e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 889e73a819dSTao Ma struct ocfs2_refcount_rec *rec = NULL; 890e73a819dSTao Ma 891e73a819dSTao Ma for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) { 892e73a819dSTao Ma rec = &rb->rf_records.rl_recs[i]; 893e73a819dSTao Ma 894e73a819dSTao Ma if (le64_to_cpu(rec->r_cpos) + 895e73a819dSTao Ma le32_to_cpu(rec->r_clusters) <= cpos) 896e73a819dSTao Ma continue; 897e73a819dSTao Ma else if (le64_to_cpu(rec->r_cpos) > cpos) 898e73a819dSTao Ma break; 899e73a819dSTao Ma 900e73a819dSTao Ma /* ok, cpos fail in this rec. Just return. */ 901e73a819dSTao Ma if (ret_rec) 902e73a819dSTao Ma *ret_rec = *rec; 903e73a819dSTao Ma goto out; 904e73a819dSTao Ma } 905e73a819dSTao Ma 906e73a819dSTao Ma if (ret_rec) { 907e73a819dSTao Ma /* We meet with a hole here, so fake the rec. */ 908e73a819dSTao Ma ret_rec->r_cpos = cpu_to_le64(cpos); 909e73a819dSTao Ma ret_rec->r_refcount = 0; 910e73a819dSTao Ma if (i < le16_to_cpu(rb->rf_records.rl_used) && 911e73a819dSTao Ma le64_to_cpu(rec->r_cpos) < cpos + len) 912e73a819dSTao Ma ret_rec->r_clusters = 913e73a819dSTao Ma cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos); 914e73a819dSTao Ma else 915e73a819dSTao Ma ret_rec->r_clusters = cpu_to_le32(len); 916e73a819dSTao Ma } 917e73a819dSTao Ma 918e73a819dSTao Ma out: 919e73a819dSTao Ma *index = i; 920e73a819dSTao Ma } 921e73a819dSTao Ma 922e73a819dSTao Ma /* 9238b2c0dbaSTao Ma * Try to remove refcount tree. The mechanism is: 9248b2c0dbaSTao Ma * 1) Check whether i_clusters == 0, if no, exit. 9258b2c0dbaSTao Ma * 2) check whether we have i_xattr_loc in dinode. if yes, exit. 9268b2c0dbaSTao Ma * 3) Check whether we have inline xattr stored outside, if yes, exit. 9278b2c0dbaSTao Ma * 4) Remove the tree. 9288b2c0dbaSTao Ma */ 9298b2c0dbaSTao Ma int ocfs2_try_remove_refcount_tree(struct inode *inode, 9308b2c0dbaSTao Ma struct buffer_head *di_bh) 9318b2c0dbaSTao Ma { 9328b2c0dbaSTao Ma int ret; 9338b2c0dbaSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 9348b2c0dbaSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 9358b2c0dbaSTao Ma 9368b2c0dbaSTao Ma down_write(&oi->ip_xattr_sem); 9378b2c0dbaSTao Ma down_write(&oi->ip_alloc_sem); 9388b2c0dbaSTao Ma 9398b2c0dbaSTao Ma if (oi->ip_clusters) 9408b2c0dbaSTao Ma goto out; 9418b2c0dbaSTao Ma 9428b2c0dbaSTao Ma if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc) 9438b2c0dbaSTao Ma goto out; 9448b2c0dbaSTao Ma 9458b2c0dbaSTao Ma if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL && 9468b2c0dbaSTao Ma ocfs2_has_inline_xattr_value_outside(inode, di)) 9478b2c0dbaSTao Ma goto out; 9488b2c0dbaSTao Ma 9498b2c0dbaSTao Ma ret = ocfs2_remove_refcount_tree(inode, di_bh); 9508b2c0dbaSTao Ma if (ret) 9518b2c0dbaSTao Ma mlog_errno(ret); 9528b2c0dbaSTao Ma out: 9538b2c0dbaSTao Ma up_write(&oi->ip_alloc_sem); 9548b2c0dbaSTao Ma up_write(&oi->ip_xattr_sem); 9558b2c0dbaSTao Ma return 0; 9568b2c0dbaSTao Ma } 9578b2c0dbaSTao Ma 9588b2c0dbaSTao Ma /* 95938a04e43STao Ma * Find the end range for a leaf refcount block indicated by 96038a04e43STao Ma * el->l_recs[index].e_blkno. 96138a04e43STao Ma */ 96238a04e43STao Ma static int ocfs2_get_refcount_cpos_end(struct ocfs2_caching_info *ci, 96338a04e43STao Ma struct buffer_head *ref_root_bh, 96438a04e43STao Ma struct ocfs2_extent_block *eb, 96538a04e43STao Ma struct ocfs2_extent_list *el, 96638a04e43STao Ma int index, u32 *cpos_end) 96738a04e43STao Ma { 96838a04e43STao Ma int ret, i, subtree_root; 96938a04e43STao Ma u32 cpos; 97038a04e43STao Ma u64 blkno; 97138a04e43STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 97238a04e43STao Ma struct ocfs2_path *left_path = NULL, *right_path = NULL; 97338a04e43STao Ma struct ocfs2_extent_tree et; 97438a04e43STao Ma struct ocfs2_extent_list *tmp_el; 97538a04e43STao Ma 97638a04e43STao Ma if (index < le16_to_cpu(el->l_next_free_rec) - 1) { 97738a04e43STao Ma /* 97838a04e43STao Ma * We have a extent rec after index, so just use the e_cpos 97938a04e43STao Ma * of the next extent rec. 98038a04e43STao Ma */ 98138a04e43STao Ma *cpos_end = le32_to_cpu(el->l_recs[index+1].e_cpos); 98238a04e43STao Ma return 0; 98338a04e43STao Ma } 98438a04e43STao Ma 98538a04e43STao Ma if (!eb || (eb && !eb->h_next_leaf_blk)) { 98638a04e43STao Ma /* 98738a04e43STao Ma * We are the last extent rec, so any high cpos should 98838a04e43STao Ma * be stored in this leaf refcount block. 98938a04e43STao Ma */ 99038a04e43STao Ma *cpos_end = UINT_MAX; 99138a04e43STao Ma return 0; 99238a04e43STao Ma } 99338a04e43STao Ma 99438a04e43STao Ma /* 99538a04e43STao Ma * If the extent block isn't the last one, we have to find 99638a04e43STao Ma * the subtree root between this extent block and the next 99738a04e43STao Ma * leaf extent block and get the corresponding e_cpos from 99838a04e43STao Ma * the subroot. Otherwise we may corrupt the b-tree. 99938a04e43STao Ma */ 100038a04e43STao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 100138a04e43STao Ma 100238a04e43STao Ma left_path = ocfs2_new_path_from_et(&et); 100338a04e43STao Ma if (!left_path) { 100438a04e43STao Ma ret = -ENOMEM; 100538a04e43STao Ma mlog_errno(ret); 100638a04e43STao Ma goto out; 100738a04e43STao Ma } 100838a04e43STao Ma 100938a04e43STao Ma cpos = le32_to_cpu(eb->h_list.l_recs[index].e_cpos); 101038a04e43STao Ma ret = ocfs2_find_path(ci, left_path, cpos); 101138a04e43STao Ma if (ret) { 101238a04e43STao Ma mlog_errno(ret); 101338a04e43STao Ma goto out; 101438a04e43STao Ma } 101538a04e43STao Ma 101638a04e43STao Ma right_path = ocfs2_new_path_from_path(left_path); 101738a04e43STao Ma if (!right_path) { 101838a04e43STao Ma ret = -ENOMEM; 101938a04e43STao Ma mlog_errno(ret); 102038a04e43STao Ma goto out; 102138a04e43STao Ma } 102238a04e43STao Ma 102338a04e43STao Ma ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, &cpos); 102438a04e43STao Ma if (ret) { 102538a04e43STao Ma mlog_errno(ret); 102638a04e43STao Ma goto out; 102738a04e43STao Ma } 102838a04e43STao Ma 102938a04e43STao Ma ret = ocfs2_find_path(ci, right_path, cpos); 103038a04e43STao Ma if (ret) { 103138a04e43STao Ma mlog_errno(ret); 103238a04e43STao Ma goto out; 103338a04e43STao Ma } 103438a04e43STao Ma 103538a04e43STao Ma subtree_root = ocfs2_find_subtree_root(&et, left_path, 103638a04e43STao Ma right_path); 103738a04e43STao Ma 103838a04e43STao Ma tmp_el = left_path->p_node[subtree_root].el; 103938a04e43STao Ma blkno = left_path->p_node[subtree_root+1].bh->b_blocknr; 104038a04e43STao Ma for (i = 0; i < le32_to_cpu(tmp_el->l_next_free_rec); i++) { 104138a04e43STao Ma if (le64_to_cpu(tmp_el->l_recs[i].e_blkno) == blkno) { 104238a04e43STao Ma *cpos_end = le32_to_cpu(tmp_el->l_recs[i+1].e_cpos); 104338a04e43STao Ma break; 104438a04e43STao Ma } 104538a04e43STao Ma } 104638a04e43STao Ma 104738a04e43STao Ma BUG_ON(i == le32_to_cpu(tmp_el->l_next_free_rec)); 104838a04e43STao Ma 104938a04e43STao Ma out: 105038a04e43STao Ma ocfs2_free_path(left_path); 105138a04e43STao Ma ocfs2_free_path(right_path); 105238a04e43STao Ma return ret; 105338a04e43STao Ma } 105438a04e43STao Ma 105538a04e43STao Ma /* 1056e73a819dSTao Ma * Given a cpos and len, try to find the refcount record which contains cpos. 1057e73a819dSTao Ma * 1. If cpos can be found in one refcount record, return the record. 1058e73a819dSTao Ma * 2. If cpos can't be found, return a fake record which start from cpos 1059e73a819dSTao Ma * and end at a small value between cpos+len and start of the next record. 1060e73a819dSTao Ma * This fake record has r_refcount = 0. 1061e73a819dSTao Ma */ 1062e73a819dSTao Ma static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci, 1063e73a819dSTao Ma struct buffer_head *ref_root_bh, 1064e73a819dSTao Ma u64 cpos, unsigned int len, 1065e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 1066e73a819dSTao Ma int *index, 1067e73a819dSTao Ma struct buffer_head **ret_bh) 1068e73a819dSTao Ma { 1069e73a819dSTao Ma int ret = 0, i, found; 107038a04e43STao Ma u32 low_cpos, uninitialized_var(cpos_end); 1071e73a819dSTao Ma struct ocfs2_extent_list *el; 107238a04e43STao Ma struct ocfs2_extent_rec *rec = NULL; 107338a04e43STao Ma struct ocfs2_extent_block *eb = NULL; 1074e73a819dSTao Ma struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL; 1075e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1076e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1077e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1078e73a819dSTao Ma 1079e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) { 1080e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len, 1081e73a819dSTao Ma ret_rec, index); 1082e73a819dSTao Ma *ret_bh = ref_root_bh; 1083e73a819dSTao Ma get_bh(ref_root_bh); 1084e73a819dSTao Ma return 0; 1085e73a819dSTao Ma } 1086e73a819dSTao Ma 1087e73a819dSTao Ma el = &rb->rf_list; 1088e73a819dSTao Ma low_cpos = cpos & OCFS2_32BIT_POS_MASK; 1089e73a819dSTao Ma 1090e73a819dSTao Ma if (el->l_tree_depth) { 1091e73a819dSTao Ma ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh); 1092e73a819dSTao Ma if (ret) { 1093e73a819dSTao Ma mlog_errno(ret); 1094e73a819dSTao Ma goto out; 1095e73a819dSTao Ma } 1096e73a819dSTao Ma 1097e73a819dSTao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 1098e73a819dSTao Ma el = &eb->h_list; 1099e73a819dSTao Ma 1100e73a819dSTao Ma if (el->l_tree_depth) { 1101e73a819dSTao Ma ocfs2_error(sb, 1102e73a819dSTao Ma "refcount tree %llu has non zero tree " 1103e73a819dSTao Ma "depth in leaf btree tree block %llu\n", 1104e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1105e73a819dSTao Ma (unsigned long long)eb_bh->b_blocknr); 1106e73a819dSTao Ma ret = -EROFS; 1107e73a819dSTao Ma goto out; 1108e73a819dSTao Ma } 1109e73a819dSTao Ma } 1110e73a819dSTao Ma 1111e73a819dSTao Ma found = 0; 1112e73a819dSTao Ma for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) { 1113e73a819dSTao Ma rec = &el->l_recs[i]; 1114e73a819dSTao Ma 1115e73a819dSTao Ma if (le32_to_cpu(rec->e_cpos) <= low_cpos) { 1116e73a819dSTao Ma found = 1; 1117e73a819dSTao Ma break; 1118e73a819dSTao Ma } 1119e73a819dSTao Ma } 1120e73a819dSTao Ma 112138a04e43STao Ma if (found) { 112238a04e43STao Ma ret = ocfs2_get_refcount_cpos_end(ci, ref_root_bh, 112338a04e43STao Ma eb, el, i, &cpos_end); 112438a04e43STao Ma if (ret) { 112538a04e43STao Ma mlog_errno(ret); 112638a04e43STao Ma goto out; 112738a04e43STao Ma } 1128e73a819dSTao Ma 112938a04e43STao Ma if (cpos_end < low_cpos + len) 113038a04e43STao Ma len = cpos_end - low_cpos; 1131e73a819dSTao Ma } 1132e73a819dSTao Ma 1133e73a819dSTao Ma ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno), 1134e73a819dSTao Ma &ref_leaf_bh); 1135e73a819dSTao Ma if (ret) { 1136e73a819dSTao Ma mlog_errno(ret); 1137e73a819dSTao Ma goto out; 1138e73a819dSTao Ma } 1139e73a819dSTao Ma 1140e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len, 1141e73a819dSTao Ma ret_rec, index); 1142e73a819dSTao Ma *ret_bh = ref_leaf_bh; 1143e73a819dSTao Ma out: 1144e73a819dSTao Ma brelse(eb_bh); 1145e73a819dSTao Ma return ret; 1146e73a819dSTao Ma } 1147e73a819dSTao Ma 1148e73a819dSTao Ma enum ocfs2_ref_rec_contig { 1149e73a819dSTao Ma REF_CONTIG_NONE = 0, 1150e73a819dSTao Ma REF_CONTIG_LEFT, 1151e73a819dSTao Ma REF_CONTIG_RIGHT, 1152e73a819dSTao Ma REF_CONTIG_LEFTRIGHT, 1153e73a819dSTao Ma }; 1154e73a819dSTao Ma 1155e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1156e73a819dSTao Ma ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb, 1157e73a819dSTao Ma int index) 1158e73a819dSTao Ma { 1159e73a819dSTao Ma if ((rb->rf_records.rl_recs[index].r_refcount == 1160e73a819dSTao Ma rb->rf_records.rl_recs[index + 1].r_refcount) && 1161e73a819dSTao Ma (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) + 1162e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) == 1163e73a819dSTao Ma le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos))) 1164e73a819dSTao Ma return REF_CONTIG_RIGHT; 1165e73a819dSTao Ma 1166e73a819dSTao Ma return REF_CONTIG_NONE; 1167e73a819dSTao Ma } 1168e73a819dSTao Ma 1169e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1170e73a819dSTao Ma ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb, 1171e73a819dSTao Ma int index) 1172e73a819dSTao Ma { 1173e73a819dSTao Ma enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE; 1174e73a819dSTao Ma 1175e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 1) 1176e73a819dSTao Ma ret = ocfs2_refcount_rec_adjacent(rb, index); 1177e73a819dSTao Ma 1178e73a819dSTao Ma if (index > 0) { 1179e73a819dSTao Ma enum ocfs2_ref_rec_contig tmp; 1180e73a819dSTao Ma 1181e73a819dSTao Ma tmp = ocfs2_refcount_rec_adjacent(rb, index - 1); 1182e73a819dSTao Ma 1183e73a819dSTao Ma if (tmp == REF_CONTIG_RIGHT) { 1184e73a819dSTao Ma if (ret == REF_CONTIG_RIGHT) 1185e73a819dSTao Ma ret = REF_CONTIG_LEFTRIGHT; 1186e73a819dSTao Ma else 1187e73a819dSTao Ma ret = REF_CONTIG_LEFT; 1188e73a819dSTao Ma } 1189e73a819dSTao Ma } 1190e73a819dSTao Ma 1191e73a819dSTao Ma return ret; 1192e73a819dSTao Ma } 1193e73a819dSTao Ma 1194e73a819dSTao Ma static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb, 1195e73a819dSTao Ma int index) 1196e73a819dSTao Ma { 1197e73a819dSTao Ma BUG_ON(rb->rf_records.rl_recs[index].r_refcount != 1198e73a819dSTao Ma rb->rf_records.rl_recs[index+1].r_refcount); 1199e73a819dSTao Ma 1200e73a819dSTao Ma le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters, 1201e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters)); 1202e73a819dSTao Ma 1203e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 2) 1204e73a819dSTao Ma memmove(&rb->rf_records.rl_recs[index + 1], 1205e73a819dSTao Ma &rb->rf_records.rl_recs[index + 2], 1206e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec) * 1207e73a819dSTao Ma (le16_to_cpu(rb->rf_records.rl_used) - index - 2)); 1208e73a819dSTao Ma 1209e73a819dSTao Ma memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1], 1210e73a819dSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 1211e73a819dSTao Ma le16_add_cpu(&rb->rf_records.rl_used, -1); 1212e73a819dSTao Ma } 1213e73a819dSTao Ma 1214e73a819dSTao Ma /* 1215e73a819dSTao Ma * Merge the refcount rec if we are contiguous with the adjacent recs. 1216e73a819dSTao Ma */ 1217e73a819dSTao Ma static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb, 1218e73a819dSTao Ma int index) 1219e73a819dSTao Ma { 1220e73a819dSTao Ma enum ocfs2_ref_rec_contig contig = 1221e73a819dSTao Ma ocfs2_refcount_rec_contig(rb, index); 1222e73a819dSTao Ma 1223e73a819dSTao Ma if (contig == REF_CONTIG_NONE) 1224e73a819dSTao Ma return; 1225e73a819dSTao Ma 1226e73a819dSTao Ma if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) { 1227e73a819dSTao Ma BUG_ON(index == 0); 1228e73a819dSTao Ma index--; 1229e73a819dSTao Ma } 1230e73a819dSTao Ma 1231e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1232e73a819dSTao Ma 1233e73a819dSTao Ma if (contig == REF_CONTIG_LEFTRIGHT) 1234e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1235e73a819dSTao Ma } 1236e73a819dSTao Ma 12371823cb0bSTao Ma /* 12381823cb0bSTao Ma * Change the refcount indexed by "index" in ref_bh. 12391823cb0bSTao Ma * If refcount reaches 0, remove it. 12401823cb0bSTao Ma */ 1241e73a819dSTao Ma static int ocfs2_change_refcount_rec(handle_t *handle, 1242e73a819dSTao Ma struct ocfs2_caching_info *ci, 1243e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 12447540c1a7STao Ma int index, int merge, int change) 1245e73a819dSTao Ma { 1246e73a819dSTao Ma int ret; 1247e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1248e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 12491823cb0bSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 12501823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rl->rl_recs[index]; 1251e73a819dSTao Ma 1252e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1253e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1254e73a819dSTao Ma if (ret) { 1255e73a819dSTao Ma mlog_errno(ret); 1256e73a819dSTao Ma goto out; 1257e73a819dSTao Ma } 1258e73a819dSTao Ma 1259e73a819dSTao Ma mlog(0, "change index %d, old count %u, change %d\n", index, 1260e73a819dSTao Ma le32_to_cpu(rec->r_refcount), change); 1261e73a819dSTao Ma le32_add_cpu(&rec->r_refcount, change); 1262e73a819dSTao Ma 12631823cb0bSTao Ma if (!rec->r_refcount) { 12641823cb0bSTao Ma if (index != le16_to_cpu(rl->rl_used) - 1) { 12651823cb0bSTao Ma memmove(rec, rec + 1, 12661823cb0bSTao Ma (le16_to_cpu(rl->rl_used) - index - 1) * 12671823cb0bSTao Ma sizeof(struct ocfs2_refcount_rec)); 12681823cb0bSTao Ma memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1], 12691823cb0bSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 12701823cb0bSTao Ma } 12711823cb0bSTao Ma 12721823cb0bSTao Ma le16_add_cpu(&rl->rl_used, -1); 12737540c1a7STao Ma } else if (merge) 1274e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1275e73a819dSTao Ma 1276ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1277e73a819dSTao Ma out: 1278e73a819dSTao Ma return ret; 1279e73a819dSTao Ma } 1280e73a819dSTao Ma 1281e73a819dSTao Ma static int ocfs2_expand_inline_ref_root(handle_t *handle, 1282e73a819dSTao Ma struct ocfs2_caching_info *ci, 1283e73a819dSTao Ma struct buffer_head *ref_root_bh, 1284e73a819dSTao Ma struct buffer_head **ref_leaf_bh, 1285e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1286e73a819dSTao Ma { 1287e73a819dSTao Ma int ret; 1288e73a819dSTao Ma u16 suballoc_bit_start; 1289e73a819dSTao Ma u32 num_got; 12902b6cb576SJoel Becker u64 suballoc_loc, blkno; 1291e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1292e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1293e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1294e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1295e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1296e73a819dSTao Ma 1297e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1298e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1299e73a819dSTao Ma if (ret) { 1300e73a819dSTao Ma mlog_errno(ret); 1301e73a819dSTao Ma goto out; 1302e73a819dSTao Ma } 1303e73a819dSTao Ma 13042b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 1305e73a819dSTao Ma &suballoc_bit_start, &num_got, 1306e73a819dSTao Ma &blkno); 1307e73a819dSTao Ma if (ret) { 1308e73a819dSTao Ma mlog_errno(ret); 1309e73a819dSTao Ma goto out; 1310e73a819dSTao Ma } 1311e73a819dSTao Ma 1312e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1313e73a819dSTao Ma if (new_bh == NULL) { 1314e73a819dSTao Ma ret = -EIO; 1315e73a819dSTao Ma mlog_errno(ret); 1316e73a819dSTao Ma goto out; 1317e73a819dSTao Ma } 1318e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1319e73a819dSTao Ma 1320e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1321e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1322e73a819dSTao Ma if (ret) { 1323e73a819dSTao Ma mlog_errno(ret); 1324e73a819dSTao Ma goto out; 1325e73a819dSTao Ma } 1326e73a819dSTao Ma 1327e73a819dSTao Ma /* 1328e73a819dSTao Ma * Initialize ocfs2_refcount_block. 1329e73a819dSTao Ma * It should contain the same information as the old root. 1330e73a819dSTao Ma * so just memcpy it and change the corresponding field. 1331e73a819dSTao Ma */ 1332e73a819dSTao Ma memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize); 1333e73a819dSTao Ma 1334e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1335b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 13362b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 1337e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1338e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1339e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(0); 1340e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1341e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1342e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1343e73a819dSTao Ma 1344e73a819dSTao Ma /* Now change the root. */ 1345e73a819dSTao Ma memset(&root_rb->rf_list, 0, sb->s_blocksize - 1346e73a819dSTao Ma offsetof(struct ocfs2_refcount_block, rf_list)); 1347e73a819dSTao Ma root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb)); 1348e73a819dSTao Ma root_rb->rf_clusters = cpu_to_le32(1); 1349e73a819dSTao Ma root_rb->rf_list.l_next_free_rec = cpu_to_le16(1); 1350e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno); 1351e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1); 1352e73a819dSTao Ma root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL); 1353e73a819dSTao Ma 1354e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 1355e73a819dSTao Ma 1356e73a819dSTao Ma mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno, 1357e73a819dSTao Ma le16_to_cpu(new_rb->rf_records.rl_used)); 1358e73a819dSTao Ma 1359e73a819dSTao Ma *ref_leaf_bh = new_bh; 1360e73a819dSTao Ma new_bh = NULL; 1361e73a819dSTao Ma out: 1362e73a819dSTao Ma brelse(new_bh); 1363e73a819dSTao Ma return ret; 1364e73a819dSTao Ma } 1365e73a819dSTao Ma 1366e73a819dSTao Ma static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev, 1367e73a819dSTao Ma struct ocfs2_refcount_rec *next) 1368e73a819dSTao Ma { 1369e73a819dSTao Ma if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <= 1370e73a819dSTao Ma ocfs2_get_ref_rec_low_cpos(next)) 1371e73a819dSTao Ma return 1; 1372e73a819dSTao Ma 1373e73a819dSTao Ma return 0; 1374e73a819dSTao Ma } 1375e73a819dSTao Ma 1376e73a819dSTao Ma static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b) 1377e73a819dSTao Ma { 1378e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1379e73a819dSTao Ma u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l); 1380e73a819dSTao Ma u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r); 1381e73a819dSTao Ma 1382e73a819dSTao Ma if (l_cpos > r_cpos) 1383e73a819dSTao Ma return 1; 1384e73a819dSTao Ma if (l_cpos < r_cpos) 1385e73a819dSTao Ma return -1; 1386e73a819dSTao Ma return 0; 1387e73a819dSTao Ma } 1388e73a819dSTao Ma 1389e73a819dSTao Ma static int cmp_refcount_rec_by_cpos(const void *a, const void *b) 1390e73a819dSTao Ma { 1391e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1392e73a819dSTao Ma u64 l_cpos = le64_to_cpu(l->r_cpos); 1393e73a819dSTao Ma u64 r_cpos = le64_to_cpu(r->r_cpos); 1394e73a819dSTao Ma 1395e73a819dSTao Ma if (l_cpos > r_cpos) 1396e73a819dSTao Ma return 1; 1397e73a819dSTao Ma if (l_cpos < r_cpos) 1398e73a819dSTao Ma return -1; 1399e73a819dSTao Ma return 0; 1400e73a819dSTao Ma } 1401e73a819dSTao Ma 1402e73a819dSTao Ma static void swap_refcount_rec(void *a, void *b, int size) 1403e73a819dSTao Ma { 1404e73a819dSTao Ma struct ocfs2_refcount_rec *l = a, *r = b, tmp; 1405e73a819dSTao Ma 1406e73a819dSTao Ma tmp = *(struct ocfs2_refcount_rec *)l; 1407e73a819dSTao Ma *(struct ocfs2_refcount_rec *)l = 1408e73a819dSTao Ma *(struct ocfs2_refcount_rec *)r; 1409e73a819dSTao Ma *(struct ocfs2_refcount_rec *)r = tmp; 1410e73a819dSTao Ma } 1411e73a819dSTao Ma 1412e73a819dSTao Ma /* 1413e73a819dSTao Ma * The refcount cpos are ordered by their 64bit cpos, 1414e73a819dSTao Ma * But we will use the low 32 bit to be the e_cpos in the b-tree. 1415e73a819dSTao Ma * So we need to make sure that this pos isn't intersected with others. 1416e73a819dSTao Ma * 1417e73a819dSTao Ma * Note: The refcount block is already sorted by their low 32 bit cpos, 1418e73a819dSTao Ma * So just try the middle pos first, and we will exit when we find 1419e73a819dSTao Ma * the good position. 1420e73a819dSTao Ma */ 1421e73a819dSTao Ma static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl, 1422e73a819dSTao Ma u32 *split_pos, int *split_index) 1423e73a819dSTao Ma { 1424e73a819dSTao Ma int num_used = le16_to_cpu(rl->rl_used); 1425e73a819dSTao Ma int delta, middle = num_used / 2; 1426e73a819dSTao Ma 1427e73a819dSTao Ma for (delta = 0; delta < middle; delta++) { 1428e73a819dSTao Ma /* Let's check delta earlier than middle */ 1429e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1430e73a819dSTao Ma &rl->rl_recs[middle - delta - 1], 1431e73a819dSTao Ma &rl->rl_recs[middle - delta])) { 1432e73a819dSTao Ma *split_index = middle - delta; 1433e73a819dSTao Ma break; 1434e73a819dSTao Ma } 1435e73a819dSTao Ma 1436e73a819dSTao Ma /* For even counts, don't walk off the end */ 1437e73a819dSTao Ma if ((middle + delta + 1) == num_used) 1438e73a819dSTao Ma continue; 1439e73a819dSTao Ma 1440e73a819dSTao Ma /* Now try delta past middle */ 1441e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1442e73a819dSTao Ma &rl->rl_recs[middle + delta], 1443e73a819dSTao Ma &rl->rl_recs[middle + delta + 1])) { 1444e73a819dSTao Ma *split_index = middle + delta + 1; 1445e73a819dSTao Ma break; 1446e73a819dSTao Ma } 1447e73a819dSTao Ma } 1448e73a819dSTao Ma 1449e73a819dSTao Ma if (delta >= middle) 1450e73a819dSTao Ma return -ENOSPC; 1451e73a819dSTao Ma 1452e73a819dSTao Ma *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]); 1453e73a819dSTao Ma return 0; 1454e73a819dSTao Ma } 1455e73a819dSTao Ma 1456e73a819dSTao Ma static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh, 1457e73a819dSTao Ma struct buffer_head *new_bh, 1458e73a819dSTao Ma u32 *split_cpos) 1459e73a819dSTao Ma { 1460e73a819dSTao Ma int split_index = 0, num_moved, ret; 1461e73a819dSTao Ma u32 cpos = 0; 1462e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1463e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1464e73a819dSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 1465e73a819dSTao Ma struct ocfs2_refcount_block *new_rb = 1466e73a819dSTao Ma (struct ocfs2_refcount_block *)new_bh->b_data; 1467e73a819dSTao Ma struct ocfs2_refcount_list *new_rl = &new_rb->rf_records; 1468e73a819dSTao Ma 1469e73a819dSTao Ma mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n", 1470e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, 1471e73a819dSTao Ma le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used)); 1472e73a819dSTao Ma 1473e73a819dSTao Ma /* 1474e73a819dSTao Ma * XXX: Improvement later. 1475e73a819dSTao Ma * If we know all the high 32 bit cpos is the same, no need to sort. 1476e73a819dSTao Ma * 1477e73a819dSTao Ma * In order to make the whole process safe, we do: 1478e73a819dSTao Ma * 1. sort the entries by their low 32 bit cpos first so that we can 1479e73a819dSTao Ma * find the split cpos easily. 1480e73a819dSTao Ma * 2. call ocfs2_insert_extent to insert the new refcount block. 1481e73a819dSTao Ma * 3. move the refcount rec to the new block. 1482e73a819dSTao Ma * 4. sort the entries by their 64 bit cpos. 1483e73a819dSTao Ma * 5. dirty the new_rb and rb. 1484e73a819dSTao Ma */ 1485e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1486e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1487e73a819dSTao Ma cmp_refcount_rec_by_low_cpos, swap_refcount_rec); 1488e73a819dSTao Ma 1489e73a819dSTao Ma ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index); 1490e73a819dSTao Ma if (ret) { 1491e73a819dSTao Ma mlog_errno(ret); 1492e73a819dSTao Ma return ret; 1493e73a819dSTao Ma } 1494e73a819dSTao Ma 1495e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(cpos); 1496e73a819dSTao Ma 1497e73a819dSTao Ma /* move refcount records starting from split_index to the new block. */ 1498e73a819dSTao Ma num_moved = le16_to_cpu(rl->rl_used) - split_index; 1499e73a819dSTao Ma memcpy(new_rl->rl_recs, &rl->rl_recs[split_index], 1500e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1501e73a819dSTao Ma 1502e73a819dSTao Ma /*ok, remove the entries we just moved over to the other block. */ 1503e73a819dSTao Ma memset(&rl->rl_recs[split_index], 0, 1504e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1505e73a819dSTao Ma 1506e73a819dSTao Ma /* change old and new rl_used accordingly. */ 1507e73a819dSTao Ma le16_add_cpu(&rl->rl_used, -num_moved); 150812d4cec9STao Ma new_rl->rl_used = cpu_to_le16(num_moved); 1509e73a819dSTao Ma 1510e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1511e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1512e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1513e73a819dSTao Ma 1514e73a819dSTao Ma sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used), 1515e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1516e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1517e73a819dSTao Ma 1518e73a819dSTao Ma *split_cpos = cpos; 1519e73a819dSTao Ma return 0; 1520e73a819dSTao Ma } 1521e73a819dSTao Ma 1522e73a819dSTao Ma static int ocfs2_new_leaf_refcount_block(handle_t *handle, 1523e73a819dSTao Ma struct ocfs2_caching_info *ci, 1524e73a819dSTao Ma struct buffer_head *ref_root_bh, 1525e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1526e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1527e73a819dSTao Ma { 1528e73a819dSTao Ma int ret; 1529e73a819dSTao Ma u16 suballoc_bit_start; 1530e73a819dSTao Ma u32 num_got, new_cpos; 15312b6cb576SJoel Becker u64 suballoc_loc, blkno; 1532e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1533e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1534e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1535e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1536e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1537e73a819dSTao Ma struct ocfs2_extent_tree ref_et; 1538e73a819dSTao Ma 1539e73a819dSTao Ma BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)); 1540e73a819dSTao Ma 1541e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1542e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1543e73a819dSTao Ma if (ret) { 1544e73a819dSTao Ma mlog_errno(ret); 1545e73a819dSTao Ma goto out; 1546e73a819dSTao Ma } 1547e73a819dSTao Ma 1548e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1549e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1550e73a819dSTao Ma if (ret) { 1551e73a819dSTao Ma mlog_errno(ret); 1552e73a819dSTao Ma goto out; 1553e73a819dSTao Ma } 1554e73a819dSTao Ma 15552b6cb576SJoel Becker ret = ocfs2_claim_metadata(handle, meta_ac, 1, &suballoc_loc, 1556e73a819dSTao Ma &suballoc_bit_start, &num_got, 1557e73a819dSTao Ma &blkno); 1558e73a819dSTao Ma if (ret) { 1559e73a819dSTao Ma mlog_errno(ret); 1560e73a819dSTao Ma goto out; 1561e73a819dSTao Ma } 1562e73a819dSTao Ma 1563e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1564e73a819dSTao Ma if (new_bh == NULL) { 1565e73a819dSTao Ma ret = -EIO; 1566e73a819dSTao Ma mlog_errno(ret); 1567e73a819dSTao Ma goto out; 1568e73a819dSTao Ma } 1569e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1570e73a819dSTao Ma 1571e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1572e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1573e73a819dSTao Ma if (ret) { 1574e73a819dSTao Ma mlog_errno(ret); 1575e73a819dSTao Ma goto out; 1576e73a819dSTao Ma } 1577e73a819dSTao Ma 1578e73a819dSTao Ma /* Initialize ocfs2_refcount_block. */ 1579e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1580e73a819dSTao Ma memset(new_rb, 0, sb->s_blocksize); 1581e73a819dSTao Ma strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 1582b89c5428STiger Yang new_rb->rf_suballoc_slot = cpu_to_le16(meta_ac->ac_alloc_slot); 15832b6cb576SJoel Becker new_rb->rf_suballoc_loc = cpu_to_le64(suballoc_loc); 1584e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1585e73a819dSTao Ma new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation); 1586e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1587e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1588e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1589e73a819dSTao Ma new_rb->rf_records.rl_count = 1590e73a819dSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 1591e73a819dSTao Ma new_rb->rf_generation = root_rb->rf_generation; 1592e73a819dSTao Ma 1593e73a819dSTao Ma ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos); 1594e73a819dSTao Ma if (ret) { 1595e73a819dSTao Ma mlog_errno(ret); 1596e73a819dSTao Ma goto out; 1597e73a819dSTao Ma } 1598e73a819dSTao Ma 1599e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1600e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1601e73a819dSTao Ma 1602e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh); 1603e73a819dSTao Ma 1604e73a819dSTao Ma mlog(0, "insert new leaf block %llu at %u\n", 1605e73a819dSTao Ma (unsigned long long)new_bh->b_blocknr, new_cpos); 1606e73a819dSTao Ma 1607e73a819dSTao Ma /* Insert the new leaf block with the specific offset cpos. */ 1608e73a819dSTao Ma ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr, 1609e73a819dSTao Ma 1, 0, meta_ac); 1610e73a819dSTao Ma if (ret) 1611e73a819dSTao Ma mlog_errno(ret); 1612e73a819dSTao Ma 1613e73a819dSTao Ma out: 1614e73a819dSTao Ma brelse(new_bh); 1615e73a819dSTao Ma return ret; 1616e73a819dSTao Ma } 1617e73a819dSTao Ma 1618e73a819dSTao Ma static int ocfs2_expand_refcount_tree(handle_t *handle, 1619e73a819dSTao Ma struct ocfs2_caching_info *ci, 1620e73a819dSTao Ma struct buffer_head *ref_root_bh, 1621e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1622e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1623e73a819dSTao Ma { 1624e73a819dSTao Ma int ret; 1625e73a819dSTao Ma struct buffer_head *expand_bh = NULL; 1626e73a819dSTao Ma 1627e73a819dSTao Ma if (ref_root_bh == ref_leaf_bh) { 1628e73a819dSTao Ma /* 1629e73a819dSTao Ma * the old root bh hasn't been expanded to a b-tree, 1630e73a819dSTao Ma * so expand it first. 1631e73a819dSTao Ma */ 1632e73a819dSTao Ma ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh, 1633e73a819dSTao Ma &expand_bh, meta_ac); 1634e73a819dSTao Ma if (ret) { 1635e73a819dSTao Ma mlog_errno(ret); 1636e73a819dSTao Ma goto out; 1637e73a819dSTao Ma } 1638e73a819dSTao Ma } else { 1639e73a819dSTao Ma expand_bh = ref_leaf_bh; 1640e73a819dSTao Ma get_bh(expand_bh); 1641e73a819dSTao Ma } 1642e73a819dSTao Ma 1643e73a819dSTao Ma 1644e73a819dSTao Ma /* Now add a new refcount block into the tree.*/ 1645e73a819dSTao Ma ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh, 1646e73a819dSTao Ma expand_bh, meta_ac); 1647e73a819dSTao Ma if (ret) 1648e73a819dSTao Ma mlog_errno(ret); 1649e73a819dSTao Ma out: 1650e73a819dSTao Ma brelse(expand_bh); 1651e73a819dSTao Ma return ret; 1652e73a819dSTao Ma } 1653e73a819dSTao Ma 1654e73a819dSTao Ma /* 1655e73a819dSTao Ma * Adjust the extent rec in b-tree representing ref_leaf_bh. 1656e73a819dSTao Ma * 1657e73a819dSTao Ma * Only called when we have inserted a new refcount rec at index 0 1658e73a819dSTao Ma * which means ocfs2_extent_rec.e_cpos may need some change. 1659e73a819dSTao Ma */ 1660e73a819dSTao Ma static int ocfs2_adjust_refcount_rec(handle_t *handle, 1661e73a819dSTao Ma struct ocfs2_caching_info *ci, 1662e73a819dSTao Ma struct buffer_head *ref_root_bh, 1663e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1664e73a819dSTao Ma struct ocfs2_refcount_rec *rec) 1665e73a819dSTao Ma { 1666e73a819dSTao Ma int ret = 0, i; 1667e73a819dSTao Ma u32 new_cpos, old_cpos; 1668e73a819dSTao Ma struct ocfs2_path *path = NULL; 1669e73a819dSTao Ma struct ocfs2_extent_tree et; 1670e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1671e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1672e73a819dSTao Ma struct ocfs2_extent_list *el; 1673e73a819dSTao Ma 1674e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) 1675e73a819dSTao Ma goto out; 1676e73a819dSTao Ma 1677e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1678e73a819dSTao Ma old_cpos = le32_to_cpu(rb->rf_cpos); 1679e73a819dSTao Ma new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK; 1680e73a819dSTao Ma if (old_cpos <= new_cpos) 1681e73a819dSTao Ma goto out; 1682e73a819dSTao Ma 1683e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 1684e73a819dSTao Ma 1685e73a819dSTao Ma path = ocfs2_new_path_from_et(&et); 1686e73a819dSTao Ma if (!path) { 1687e73a819dSTao Ma ret = -ENOMEM; 1688e73a819dSTao Ma mlog_errno(ret); 1689e73a819dSTao Ma goto out; 1690e73a819dSTao Ma } 1691e73a819dSTao Ma 1692e73a819dSTao Ma ret = ocfs2_find_path(ci, path, old_cpos); 1693e73a819dSTao Ma if (ret) { 1694e73a819dSTao Ma mlog_errno(ret); 1695e73a819dSTao Ma goto out; 1696e73a819dSTao Ma } 1697e73a819dSTao Ma 1698e73a819dSTao Ma /* 1699e73a819dSTao Ma * 2 more credits, one for the leaf refcount block, one for 1700e73a819dSTao Ma * the extent block contains the extent rec. 1701e73a819dSTao Ma */ 1702c901fb00STao Ma ret = ocfs2_extend_trans(handle, 2); 1703e73a819dSTao Ma if (ret < 0) { 1704e73a819dSTao Ma mlog_errno(ret); 1705e73a819dSTao Ma goto out; 1706e73a819dSTao Ma } 1707e73a819dSTao Ma 1708e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1709e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1710e73a819dSTao Ma if (ret < 0) { 1711e73a819dSTao Ma mlog_errno(ret); 1712e73a819dSTao Ma goto out; 1713e73a819dSTao Ma } 1714e73a819dSTao Ma 1715e73a819dSTao Ma ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path), 1716e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1717e73a819dSTao Ma if (ret < 0) { 1718e73a819dSTao Ma mlog_errno(ret); 1719e73a819dSTao Ma goto out; 1720e73a819dSTao Ma } 1721e73a819dSTao Ma 1722e73a819dSTao Ma /* change the leaf extent block first. */ 1723e73a819dSTao Ma el = path_leaf_el(path); 1724e73a819dSTao Ma 1725e73a819dSTao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) 1726e73a819dSTao Ma if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos) 1727e73a819dSTao Ma break; 1728e73a819dSTao Ma 1729e73a819dSTao Ma BUG_ON(i == le16_to_cpu(el->l_next_free_rec)); 1730e73a819dSTao Ma 1731e73a819dSTao Ma el->l_recs[i].e_cpos = cpu_to_le32(new_cpos); 1732e73a819dSTao Ma 1733e73a819dSTao Ma /* change the r_cpos in the leaf block. */ 1734e73a819dSTao Ma rb->rf_cpos = cpu_to_le32(new_cpos); 1735e73a819dSTao Ma 1736e73a819dSTao Ma ocfs2_journal_dirty(handle, path_leaf_bh(path)); 1737e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1738e73a819dSTao Ma 1739e73a819dSTao Ma out: 1740e73a819dSTao Ma ocfs2_free_path(path); 1741e73a819dSTao Ma return ret; 1742e73a819dSTao Ma } 1743e73a819dSTao Ma 1744e73a819dSTao Ma static int ocfs2_insert_refcount_rec(handle_t *handle, 1745e73a819dSTao Ma struct ocfs2_caching_info *ci, 1746e73a819dSTao Ma struct buffer_head *ref_root_bh, 1747e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1748e73a819dSTao Ma struct ocfs2_refcount_rec *rec, 17497540c1a7STao Ma int index, int merge, 1750e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1751e73a819dSTao Ma { 1752e73a819dSTao Ma int ret; 1753e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1754e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1755e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1756e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1757e73a819dSTao Ma 1758e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1759e73a819dSTao Ma 1760e73a819dSTao Ma if (rf_list->rl_used == rf_list->rl_count) { 1761e73a819dSTao Ma u64 cpos = le64_to_cpu(rec->r_cpos); 1762e73a819dSTao Ma u32 len = le32_to_cpu(rec->r_clusters); 1763e73a819dSTao Ma 1764e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1765e73a819dSTao Ma ref_leaf_bh, meta_ac); 1766e73a819dSTao Ma if (ret) { 1767e73a819dSTao Ma mlog_errno(ret); 1768e73a819dSTao Ma goto out; 1769e73a819dSTao Ma } 1770e73a819dSTao Ma 1771e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1772e73a819dSTao Ma cpos, len, NULL, &index, 1773e73a819dSTao Ma &new_bh); 1774e73a819dSTao Ma if (ret) { 1775e73a819dSTao Ma mlog_errno(ret); 1776e73a819dSTao Ma goto out; 1777e73a819dSTao Ma } 1778e73a819dSTao Ma 1779e73a819dSTao Ma ref_leaf_bh = new_bh; 1780e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1781e73a819dSTao Ma rf_list = &rb->rf_records; 1782e73a819dSTao Ma } 1783e73a819dSTao Ma 1784e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1785e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1786e73a819dSTao Ma if (ret) { 1787e73a819dSTao Ma mlog_errno(ret); 1788e73a819dSTao Ma goto out; 1789e73a819dSTao Ma } 1790e73a819dSTao Ma 1791e73a819dSTao Ma if (index < le16_to_cpu(rf_list->rl_used)) 1792e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1], 1793e73a819dSTao Ma &rf_list->rl_recs[index], 1794e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index) * 1795e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1796e73a819dSTao Ma 1797e73a819dSTao Ma mlog(0, "insert refcount record start %llu, len %u, count %u " 1798e73a819dSTao Ma "to leaf block %llu at index %d\n", 1799e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos), 1800e73a819dSTao Ma le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount), 1801e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index); 1802e73a819dSTao Ma 1803e73a819dSTao Ma rf_list->rl_recs[index] = *rec; 1804e73a819dSTao Ma 1805e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, 1); 1806e73a819dSTao Ma 18077540c1a7STao Ma if (merge) 1808e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1809e73a819dSTao Ma 1810ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1811e73a819dSTao Ma 1812e73a819dSTao Ma if (index == 0) { 1813e73a819dSTao Ma ret = ocfs2_adjust_refcount_rec(handle, ci, 1814e73a819dSTao Ma ref_root_bh, 1815e73a819dSTao Ma ref_leaf_bh, rec); 1816e73a819dSTao Ma if (ret) 1817e73a819dSTao Ma mlog_errno(ret); 1818e73a819dSTao Ma } 1819e73a819dSTao Ma out: 1820e73a819dSTao Ma brelse(new_bh); 1821e73a819dSTao Ma return ret; 1822e73a819dSTao Ma } 1823e73a819dSTao Ma 1824e73a819dSTao Ma /* 1825e73a819dSTao Ma * Split the refcount_rec indexed by "index" in ref_leaf_bh. 1826e73a819dSTao Ma * This is much simple than our b-tree code. 1827e73a819dSTao Ma * split_rec is the new refcount rec we want to insert. 1828e73a819dSTao Ma * If split_rec->r_refcount > 0, we are changing the refcount(in case we 1829e73a819dSTao Ma * increase refcount or decrease a refcount to non-zero). 1830e73a819dSTao Ma * If split_rec->r_refcount == 0, we are punching a hole in current refcount 1831e73a819dSTao Ma * rec( in case we decrease a refcount to zero). 1832e73a819dSTao Ma */ 1833e73a819dSTao Ma static int ocfs2_split_refcount_rec(handle_t *handle, 1834e73a819dSTao Ma struct ocfs2_caching_info *ci, 1835e73a819dSTao Ma struct buffer_head *ref_root_bh, 1836e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1837e73a819dSTao Ma struct ocfs2_refcount_rec *split_rec, 18387540c1a7STao Ma int index, int merge, 1839e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1840e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1841e73a819dSTao Ma { 1842e73a819dSTao Ma int ret, recs_need; 1843e73a819dSTao Ma u32 len; 1844e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1845e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1846e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1847e73a819dSTao Ma struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index]; 1848e73a819dSTao Ma struct ocfs2_refcount_rec *tail_rec = NULL; 1849e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1850e73a819dSTao Ma 1851e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1852e73a819dSTao Ma 1853e73a819dSTao Ma mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n", 1854e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters), 1855e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos), 1856e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters)); 1857e73a819dSTao Ma 1858e73a819dSTao Ma /* 1859e73a819dSTao Ma * If we just need to split the header or tail clusters, 1860e73a819dSTao Ma * no more recs are needed, just split is OK. 1861e73a819dSTao Ma * Otherwise we at least need one new recs. 1862e73a819dSTao Ma */ 1863e73a819dSTao Ma if (!split_rec->r_refcount && 1864e73a819dSTao Ma (split_rec->r_cpos == orig_rec->r_cpos || 1865e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1866e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) == 1867e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1868e73a819dSTao Ma recs_need = 0; 1869e73a819dSTao Ma else 1870e73a819dSTao Ma recs_need = 1; 1871e73a819dSTao Ma 1872e73a819dSTao Ma /* 1873e73a819dSTao Ma * We need one more rec if we split in the middle and the new rec have 1874e73a819dSTao Ma * some refcount in it. 1875e73a819dSTao Ma */ 1876e73a819dSTao Ma if (split_rec->r_refcount && 1877e73a819dSTao Ma (split_rec->r_cpos != orig_rec->r_cpos && 1878e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1879e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) != 1880e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1881e73a819dSTao Ma recs_need++; 1882e73a819dSTao Ma 1883e73a819dSTao Ma /* If the leaf block don't have enough record, expand it. */ 188412d4cec9STao Ma if (le16_to_cpu(rf_list->rl_used) + recs_need > 188512d4cec9STao Ma le16_to_cpu(rf_list->rl_count)) { 1886e73a819dSTao Ma struct ocfs2_refcount_rec tmp_rec; 1887e73a819dSTao Ma u64 cpos = le64_to_cpu(orig_rec->r_cpos); 1888e73a819dSTao Ma len = le32_to_cpu(orig_rec->r_clusters); 1889e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1890e73a819dSTao Ma ref_leaf_bh, meta_ac); 1891e73a819dSTao Ma if (ret) { 1892e73a819dSTao Ma mlog_errno(ret); 1893e73a819dSTao Ma goto out; 1894e73a819dSTao Ma } 1895e73a819dSTao Ma 1896e73a819dSTao Ma /* 1897e73a819dSTao Ma * We have to re-get it since now cpos may be moved to 1898e73a819dSTao Ma * another leaf block. 1899e73a819dSTao Ma */ 1900e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1901e73a819dSTao Ma cpos, len, &tmp_rec, &index, 1902e73a819dSTao Ma &new_bh); 1903e73a819dSTao Ma if (ret) { 1904e73a819dSTao Ma mlog_errno(ret); 1905e73a819dSTao Ma goto out; 1906e73a819dSTao Ma } 1907e73a819dSTao Ma 1908e73a819dSTao Ma ref_leaf_bh = new_bh; 1909e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1910e73a819dSTao Ma rf_list = &rb->rf_records; 1911e73a819dSTao Ma orig_rec = &rf_list->rl_recs[index]; 1912e73a819dSTao Ma } 1913e73a819dSTao Ma 1914e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1915e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1916e73a819dSTao Ma if (ret) { 1917e73a819dSTao Ma mlog_errno(ret); 1918e73a819dSTao Ma goto out; 1919e73a819dSTao Ma } 1920e73a819dSTao Ma 1921e73a819dSTao Ma /* 1922e73a819dSTao Ma * We have calculated out how many new records we need and store 1923e73a819dSTao Ma * in recs_need, so spare enough space first by moving the records 1924e73a819dSTao Ma * after "index" to the end. 1925e73a819dSTao Ma */ 1926e73a819dSTao Ma if (index != le16_to_cpu(rf_list->rl_used) - 1) 1927e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1 + recs_need], 1928e73a819dSTao Ma &rf_list->rl_recs[index + 1], 1929e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index - 1) * 1930e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1931e73a819dSTao Ma 1932e73a819dSTao Ma len = (le64_to_cpu(orig_rec->r_cpos) + 1933e73a819dSTao Ma le32_to_cpu(orig_rec->r_clusters)) - 1934e73a819dSTao Ma (le64_to_cpu(split_rec->r_cpos) + 1935e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters)); 1936e73a819dSTao Ma 1937e73a819dSTao Ma /* 1938e73a819dSTao Ma * If we have "len", the we will split in the tail and move it 1939e73a819dSTao Ma * to the end of the space we have just spared. 1940e73a819dSTao Ma */ 1941e73a819dSTao Ma if (len) { 1942e73a819dSTao Ma tail_rec = &rf_list->rl_recs[index + recs_need]; 1943e73a819dSTao Ma 1944e73a819dSTao Ma memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec)); 1945e73a819dSTao Ma le64_add_cpu(&tail_rec->r_cpos, 1946e73a819dSTao Ma le32_to_cpu(tail_rec->r_clusters) - len); 194712d4cec9STao Ma tail_rec->r_clusters = cpu_to_le32(len); 1948e73a819dSTao Ma } 1949e73a819dSTao Ma 1950e73a819dSTao Ma /* 1951e73a819dSTao Ma * If the split pos isn't the same as the original one, we need to 1952e73a819dSTao Ma * split in the head. 1953e73a819dSTao Ma * 1954e73a819dSTao Ma * Note: We have the chance that split_rec.r_refcount = 0, 1955e73a819dSTao Ma * recs_need = 0 and len > 0, which means we just cut the head from 1956e73a819dSTao Ma * the orig_rec and in that case we have done some modification in 1957e73a819dSTao Ma * orig_rec above, so the check for r_cpos is faked. 1958e73a819dSTao Ma */ 1959e73a819dSTao Ma if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) { 1960e73a819dSTao Ma len = le64_to_cpu(split_rec->r_cpos) - 1961e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos); 1962e73a819dSTao Ma orig_rec->r_clusters = cpu_to_le32(len); 1963e73a819dSTao Ma index++; 1964e73a819dSTao Ma } 1965e73a819dSTao Ma 1966e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, recs_need); 1967e73a819dSTao Ma 1968e73a819dSTao Ma if (split_rec->r_refcount) { 1969e73a819dSTao Ma rf_list->rl_recs[index] = *split_rec; 1970e73a819dSTao Ma mlog(0, "insert refcount record start %llu, len %u, count %u " 1971e73a819dSTao Ma "to leaf block %llu at index %d\n", 1972e73a819dSTao Ma (unsigned long long)le64_to_cpu(split_rec->r_cpos), 1973e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters), 1974e73a819dSTao Ma le32_to_cpu(split_rec->r_refcount), 1975e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index); 1976e73a819dSTao Ma 19777540c1a7STao Ma if (merge) 1978e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1979e73a819dSTao Ma } 1980e73a819dSTao Ma 1981ec20cec7SJoel Becker ocfs2_journal_dirty(handle, ref_leaf_bh); 1982e73a819dSTao Ma 1983e73a819dSTao Ma out: 1984e73a819dSTao Ma brelse(new_bh); 1985e73a819dSTao Ma return ret; 1986e73a819dSTao Ma } 1987e73a819dSTao Ma 19887540c1a7STao Ma static int __ocfs2_increase_refcount(handle_t *handle, 1989e73a819dSTao Ma struct ocfs2_caching_info *ci, 1990e73a819dSTao Ma struct buffer_head *ref_root_bh, 19917540c1a7STao Ma u64 cpos, u32 len, int merge, 1992e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1993e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1994e73a819dSTao Ma { 1995e73a819dSTao Ma int ret = 0, index; 1996e73a819dSTao Ma struct buffer_head *ref_leaf_bh = NULL; 1997e73a819dSTao Ma struct ocfs2_refcount_rec rec; 1998e73a819dSTao Ma unsigned int set_len = 0; 1999e73a819dSTao Ma 2000e73a819dSTao Ma mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n", 2001e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 2002e73a819dSTao Ma (unsigned long long)cpos, len); 2003e73a819dSTao Ma 2004e73a819dSTao Ma while (len) { 2005e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 2006e73a819dSTao Ma cpos, len, &rec, &index, 2007e73a819dSTao Ma &ref_leaf_bh); 2008e73a819dSTao Ma if (ret) { 2009e73a819dSTao Ma mlog_errno(ret); 2010e73a819dSTao Ma goto out; 2011e73a819dSTao Ma } 2012e73a819dSTao Ma 2013e73a819dSTao Ma set_len = le32_to_cpu(rec.r_clusters); 2014e73a819dSTao Ma 2015e73a819dSTao Ma /* 2016e73a819dSTao Ma * Here we may meet with 3 situations: 2017e73a819dSTao Ma * 2018e73a819dSTao Ma * 1. If we find an already existing record, and the length 2019e73a819dSTao Ma * is the same, cool, we just need to increase the r_refcount 2020e73a819dSTao Ma * and it is OK. 2021e73a819dSTao Ma * 2. If we find a hole, just insert it with r_refcount = 1. 2022e73a819dSTao Ma * 3. If we are in the middle of one extent record, split 2023e73a819dSTao Ma * it. 2024e73a819dSTao Ma */ 2025e73a819dSTao Ma if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos && 2026e73a819dSTao Ma set_len <= len) { 2027e73a819dSTao Ma mlog(0, "increase refcount rec, start %llu, len %u, " 2028e73a819dSTao Ma "count %u\n", (unsigned long long)cpos, set_len, 2029e73a819dSTao Ma le32_to_cpu(rec.r_refcount)); 2030e73a819dSTao Ma ret = ocfs2_change_refcount_rec(handle, ci, 20317540c1a7STao Ma ref_leaf_bh, index, 20327540c1a7STao Ma merge, 1); 2033e73a819dSTao Ma if (ret) { 2034e73a819dSTao Ma mlog_errno(ret); 2035e73a819dSTao Ma goto out; 2036e73a819dSTao Ma } 2037e73a819dSTao Ma } else if (!rec.r_refcount) { 2038e73a819dSTao Ma rec.r_refcount = cpu_to_le32(1); 2039e73a819dSTao Ma 2040e73a819dSTao Ma mlog(0, "insert refcount rec, start %llu, len %u\n", 2041e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2042e73a819dSTao Ma set_len); 2043e73a819dSTao Ma ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh, 2044e73a819dSTao Ma ref_leaf_bh, 20457540c1a7STao Ma &rec, index, 20467540c1a7STao Ma merge, meta_ac); 2047e73a819dSTao Ma if (ret) { 2048e73a819dSTao Ma mlog_errno(ret); 2049e73a819dSTao Ma goto out; 2050e73a819dSTao Ma } 2051e73a819dSTao Ma } else { 2052e73a819dSTao Ma set_len = min((u64)(cpos + len), 2053e73a819dSTao Ma le64_to_cpu(rec.r_cpos) + set_len) - cpos; 2054e73a819dSTao Ma rec.r_cpos = cpu_to_le64(cpos); 2055e73a819dSTao Ma rec.r_clusters = cpu_to_le32(set_len); 2056e73a819dSTao Ma le32_add_cpu(&rec.r_refcount, 1); 2057e73a819dSTao Ma 2058e73a819dSTao Ma mlog(0, "split refcount rec, start %llu, " 2059e73a819dSTao Ma "len %u, count %u\n", 2060e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2061e73a819dSTao Ma set_len, le32_to_cpu(rec.r_refcount)); 2062e73a819dSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 2063e73a819dSTao Ma ref_root_bh, ref_leaf_bh, 20647540c1a7STao Ma &rec, index, merge, 2065e73a819dSTao Ma meta_ac, dealloc); 2066e73a819dSTao Ma if (ret) { 2067e73a819dSTao Ma mlog_errno(ret); 2068e73a819dSTao Ma goto out; 2069e73a819dSTao Ma } 2070e73a819dSTao Ma } 2071e73a819dSTao Ma 2072e73a819dSTao Ma cpos += set_len; 2073e73a819dSTao Ma len -= set_len; 2074e73a819dSTao Ma brelse(ref_leaf_bh); 2075e73a819dSTao Ma ref_leaf_bh = NULL; 2076e73a819dSTao Ma } 2077e73a819dSTao Ma 2078e73a819dSTao Ma out: 2079e73a819dSTao Ma brelse(ref_leaf_bh); 2080e73a819dSTao Ma return ret; 2081e73a819dSTao Ma } 20821823cb0bSTao Ma 20831823cb0bSTao Ma static int ocfs2_remove_refcount_extent(handle_t *handle, 20841823cb0bSTao Ma struct ocfs2_caching_info *ci, 20851823cb0bSTao Ma struct buffer_head *ref_root_bh, 20861823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 20871823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 20881823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 20891823cb0bSTao Ma { 20901823cb0bSTao Ma int ret; 20911823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 20921823cb0bSTao Ma struct ocfs2_refcount_block *rb = 20931823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 20941823cb0bSTao Ma struct ocfs2_extent_tree et; 20951823cb0bSTao Ma 20961823cb0bSTao Ma BUG_ON(rb->rf_records.rl_used); 20971823cb0bSTao Ma 20981823cb0bSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 20991823cb0bSTao Ma ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos), 21001823cb0bSTao Ma 1, meta_ac, dealloc); 21011823cb0bSTao Ma if (ret) { 21021823cb0bSTao Ma mlog_errno(ret); 21031823cb0bSTao Ma goto out; 21041823cb0bSTao Ma } 21051823cb0bSTao Ma 21061823cb0bSTao Ma ocfs2_remove_from_cache(ci, ref_leaf_bh); 21071823cb0bSTao Ma 21081823cb0bSTao Ma /* 21091823cb0bSTao Ma * add the freed block to the dealloc so that it will be freed 21101823cb0bSTao Ma * when we run dealloc. 21111823cb0bSTao Ma */ 21121823cb0bSTao Ma ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE, 21131823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_slot), 211474380c47STao Ma le64_to_cpu(rb->rf_suballoc_loc), 21151823cb0bSTao Ma le64_to_cpu(rb->rf_blkno), 21161823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_bit)); 21171823cb0bSTao Ma if (ret) { 21181823cb0bSTao Ma mlog_errno(ret); 21191823cb0bSTao Ma goto out; 21201823cb0bSTao Ma } 21211823cb0bSTao Ma 21221823cb0bSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 21231823cb0bSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 21241823cb0bSTao Ma if (ret) { 21251823cb0bSTao Ma mlog_errno(ret); 21261823cb0bSTao Ma goto out; 21271823cb0bSTao Ma } 21281823cb0bSTao Ma 21291823cb0bSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 21301823cb0bSTao Ma 21311823cb0bSTao Ma le32_add_cpu(&rb->rf_clusters, -1); 21321823cb0bSTao Ma 21331823cb0bSTao Ma /* 21341823cb0bSTao Ma * check whether we need to restore the root refcount block if 21351823cb0bSTao Ma * there is no leaf extent block at atll. 21361823cb0bSTao Ma */ 21371823cb0bSTao Ma if (!rb->rf_list.l_next_free_rec) { 21381823cb0bSTao Ma BUG_ON(rb->rf_clusters); 21391823cb0bSTao Ma 21401823cb0bSTao Ma mlog(0, "reset refcount tree root %llu to be a record block.\n", 21411823cb0bSTao Ma (unsigned long long)ref_root_bh->b_blocknr); 21421823cb0bSTao Ma 21431823cb0bSTao Ma rb->rf_flags = 0; 21441823cb0bSTao Ma rb->rf_parent = 0; 21451823cb0bSTao Ma rb->rf_cpos = 0; 21461823cb0bSTao Ma memset(&rb->rf_records, 0, sb->s_blocksize - 21471823cb0bSTao Ma offsetof(struct ocfs2_refcount_block, rf_records)); 21481823cb0bSTao Ma rb->rf_records.rl_count = 21491823cb0bSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 21501823cb0bSTao Ma } 21511823cb0bSTao Ma 21521823cb0bSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 21531823cb0bSTao Ma 21541823cb0bSTao Ma out: 21551823cb0bSTao Ma return ret; 21561823cb0bSTao Ma } 21571823cb0bSTao Ma 21587540c1a7STao Ma int ocfs2_increase_refcount(handle_t *handle, 21597540c1a7STao Ma struct ocfs2_caching_info *ci, 21607540c1a7STao Ma struct buffer_head *ref_root_bh, 21617540c1a7STao Ma u64 cpos, u32 len, 21627540c1a7STao Ma struct ocfs2_alloc_context *meta_ac, 21637540c1a7STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 21647540c1a7STao Ma { 21657540c1a7STao Ma return __ocfs2_increase_refcount(handle, ci, ref_root_bh, 21667540c1a7STao Ma cpos, len, 1, 21677540c1a7STao Ma meta_ac, dealloc); 21687540c1a7STao Ma } 21697540c1a7STao Ma 21701823cb0bSTao Ma static int ocfs2_decrease_refcount_rec(handle_t *handle, 21711823cb0bSTao Ma struct ocfs2_caching_info *ci, 21721823cb0bSTao Ma struct buffer_head *ref_root_bh, 21731823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 21741823cb0bSTao Ma int index, u64 cpos, unsigned int len, 21751823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 21761823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 21771823cb0bSTao Ma { 21781823cb0bSTao Ma int ret; 21791823cb0bSTao Ma struct ocfs2_refcount_block *rb = 21801823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 21811823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index]; 21821823cb0bSTao Ma 21831823cb0bSTao Ma BUG_ON(cpos < le64_to_cpu(rec->r_cpos)); 21841823cb0bSTao Ma BUG_ON(cpos + len > 21851823cb0bSTao Ma le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters)); 21861823cb0bSTao 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 mlog(0, "split refcount rec, start %llu, " 21991823cb0bSTao Ma "len %u, count %u, original start %llu, len %u\n", 22001823cb0bSTao Ma (unsigned long long)le64_to_cpu(split.r_cpos), 22011823cb0bSTao Ma len, le32_to_cpu(split.r_refcount), 22021823cb0bSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos), 22031823cb0bSTao Ma le32_to_cpu(rec->r_clusters)); 22041823cb0bSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 22051823cb0bSTao Ma ref_root_bh, ref_leaf_bh, 22067540c1a7STao Ma &split, index, 1, 22071823cb0bSTao Ma meta_ac, dealloc); 22081823cb0bSTao Ma } 22091823cb0bSTao Ma 22101823cb0bSTao Ma if (ret) { 22111823cb0bSTao Ma mlog_errno(ret); 22121823cb0bSTao Ma goto out; 22131823cb0bSTao Ma } 22141823cb0bSTao Ma 22151823cb0bSTao Ma /* Remove the leaf refcount block if it contains no refcount record. */ 22161823cb0bSTao Ma if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) { 22171823cb0bSTao Ma ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh, 22181823cb0bSTao Ma ref_leaf_bh, meta_ac, 22191823cb0bSTao Ma dealloc); 22201823cb0bSTao Ma if (ret) 22211823cb0bSTao Ma mlog_errno(ret); 22221823cb0bSTao Ma } 22231823cb0bSTao Ma 22241823cb0bSTao Ma out: 22251823cb0bSTao Ma return ret; 22261823cb0bSTao Ma } 22271823cb0bSTao Ma 22281823cb0bSTao Ma static int __ocfs2_decrease_refcount(handle_t *handle, 22291823cb0bSTao Ma struct ocfs2_caching_info *ci, 22301823cb0bSTao Ma struct buffer_head *ref_root_bh, 22311823cb0bSTao Ma u64 cpos, u32 len, 22321823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 22336ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 22346ae23c55STao Ma int delete) 22351823cb0bSTao Ma { 22361823cb0bSTao Ma int ret = 0, index = 0; 22371823cb0bSTao Ma struct ocfs2_refcount_rec rec; 22381823cb0bSTao Ma unsigned int r_count = 0, r_len; 22391823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 22401823cb0bSTao Ma struct buffer_head *ref_leaf_bh = NULL; 22411823cb0bSTao Ma 22426ae23c55STao Ma mlog(0, "Tree owner %llu, decrease refcount start %llu, " 22436ae23c55STao Ma "len %u, delete %u\n", 22441823cb0bSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 22456ae23c55STao Ma (unsigned long long)cpos, len, delete); 22461823cb0bSTao Ma 22471823cb0bSTao Ma while (len) { 22481823cb0bSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 22491823cb0bSTao Ma cpos, len, &rec, &index, 22501823cb0bSTao Ma &ref_leaf_bh); 22511823cb0bSTao Ma if (ret) { 22521823cb0bSTao Ma mlog_errno(ret); 22531823cb0bSTao Ma goto out; 22541823cb0bSTao Ma } 22551823cb0bSTao Ma 22561823cb0bSTao Ma r_count = le32_to_cpu(rec.r_refcount); 22571823cb0bSTao Ma BUG_ON(r_count == 0); 22586ae23c55STao Ma if (!delete) 22596ae23c55STao Ma BUG_ON(r_count > 1); 22601823cb0bSTao Ma 22611823cb0bSTao Ma r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) + 22621823cb0bSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 22631823cb0bSTao Ma 22641823cb0bSTao Ma ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh, 22651823cb0bSTao Ma ref_leaf_bh, index, 22661823cb0bSTao Ma cpos, r_len, 22671823cb0bSTao Ma meta_ac, dealloc); 22681823cb0bSTao Ma if (ret) { 22691823cb0bSTao Ma mlog_errno(ret); 22701823cb0bSTao Ma goto out; 22711823cb0bSTao Ma } 22721823cb0bSTao Ma 22736ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1 && delete) { 22741823cb0bSTao Ma ret = ocfs2_cache_cluster_dealloc(dealloc, 22751823cb0bSTao Ma ocfs2_clusters_to_blocks(sb, cpos), 22761823cb0bSTao Ma r_len); 22771823cb0bSTao Ma if (ret) { 22781823cb0bSTao Ma mlog_errno(ret); 22791823cb0bSTao Ma goto out; 22801823cb0bSTao Ma } 22811823cb0bSTao Ma } 22821823cb0bSTao Ma 22831823cb0bSTao Ma cpos += r_len; 22841823cb0bSTao Ma len -= r_len; 22851823cb0bSTao Ma brelse(ref_leaf_bh); 22861823cb0bSTao Ma ref_leaf_bh = NULL; 22871823cb0bSTao Ma } 22881823cb0bSTao Ma 22891823cb0bSTao Ma out: 22901823cb0bSTao Ma brelse(ref_leaf_bh); 22911823cb0bSTao Ma return ret; 22921823cb0bSTao Ma } 22931823cb0bSTao Ma 22941823cb0bSTao Ma /* Caller must hold refcount tree lock. */ 22951823cb0bSTao Ma int ocfs2_decrease_refcount(struct inode *inode, 22961823cb0bSTao Ma handle_t *handle, u32 cpos, u32 len, 22971823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 22986ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 22996ae23c55STao Ma int delete) 23001823cb0bSTao Ma { 23011823cb0bSTao Ma int ret; 23021823cb0bSTao Ma u64 ref_blkno; 23031823cb0bSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 23041823cb0bSTao Ma struct buffer_head *ref_root_bh = NULL; 23051823cb0bSTao Ma struct ocfs2_refcount_tree *tree; 23061823cb0bSTao Ma 23071823cb0bSTao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 23081823cb0bSTao Ma 23091823cb0bSTao Ma ret = ocfs2_get_refcount_block(inode, &ref_blkno); 23101823cb0bSTao Ma if (ret) { 23111823cb0bSTao Ma mlog_errno(ret); 23121823cb0bSTao Ma goto out; 23131823cb0bSTao Ma } 23141823cb0bSTao Ma 23151823cb0bSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree); 23161823cb0bSTao Ma if (ret) { 23171823cb0bSTao Ma mlog_errno(ret); 23181823cb0bSTao Ma goto out; 23191823cb0bSTao Ma } 23201823cb0bSTao Ma 23211823cb0bSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno, 23221823cb0bSTao Ma &ref_root_bh); 23231823cb0bSTao Ma if (ret) { 23241823cb0bSTao Ma mlog_errno(ret); 23251823cb0bSTao Ma goto out; 23261823cb0bSTao Ma } 23271823cb0bSTao Ma 23281823cb0bSTao Ma ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh, 23296ae23c55STao Ma cpos, len, meta_ac, dealloc, delete); 23301823cb0bSTao Ma if (ret) 23311823cb0bSTao Ma mlog_errno(ret); 23321823cb0bSTao Ma out: 23331823cb0bSTao Ma brelse(ref_root_bh); 23341823cb0bSTao Ma return ret; 23351823cb0bSTao Ma } 23361aa75feaSTao Ma 23371aa75feaSTao Ma /* 23381aa75feaSTao Ma * Mark the already-existing extent at cpos as refcounted for len clusters. 23391aa75feaSTao Ma * This adds the refcount extent flag. 23401aa75feaSTao Ma * 23411aa75feaSTao Ma * If the existing extent is larger than the request, initiate a 23421aa75feaSTao Ma * split. An attempt will be made at merging with adjacent extents. 23431aa75feaSTao Ma * 23441aa75feaSTao Ma * The caller is responsible for passing down meta_ac if we'll need it. 23451aa75feaSTao Ma */ 23461aa75feaSTao Ma static int ocfs2_mark_extent_refcounted(struct inode *inode, 23471aa75feaSTao Ma struct ocfs2_extent_tree *et, 23481aa75feaSTao Ma handle_t *handle, u32 cpos, 23491aa75feaSTao Ma u32 len, u32 phys, 23501aa75feaSTao Ma struct ocfs2_alloc_context *meta_ac, 23511aa75feaSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 23521aa75feaSTao Ma { 23531aa75feaSTao Ma int ret; 23541aa75feaSTao Ma 23551aa75feaSTao Ma mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n", 23561aa75feaSTao Ma inode->i_ino, cpos, len, phys); 23571aa75feaSTao Ma 23581aa75feaSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 23591aa75feaSTao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 23601aa75feaSTao Ma "tree, but the feature bit is not set in the " 23611aa75feaSTao Ma "super block.", inode->i_ino); 23621aa75feaSTao Ma ret = -EROFS; 23631aa75feaSTao Ma goto out; 23641aa75feaSTao Ma } 23651aa75feaSTao Ma 23661aa75feaSTao Ma ret = ocfs2_change_extent_flag(handle, et, cpos, 23671aa75feaSTao Ma len, phys, meta_ac, dealloc, 23681aa75feaSTao Ma OCFS2_EXT_REFCOUNTED, 0); 23691aa75feaSTao Ma if (ret) 23701aa75feaSTao Ma mlog_errno(ret); 23711aa75feaSTao Ma 23721aa75feaSTao Ma out: 23731aa75feaSTao Ma return ret; 23741aa75feaSTao Ma } 2375bcbbb24aSTao Ma 2376bcbbb24aSTao Ma /* 2377bcbbb24aSTao Ma * Given some contiguous physical clusters, calculate what we need 2378bcbbb24aSTao Ma * for modifying their refcount. 2379bcbbb24aSTao Ma */ 2380bcbbb24aSTao Ma static int ocfs2_calc_refcount_meta_credits(struct super_block *sb, 2381bcbbb24aSTao Ma struct ocfs2_caching_info *ci, 2382bcbbb24aSTao Ma struct buffer_head *ref_root_bh, 2383bcbbb24aSTao Ma u64 start_cpos, 2384bcbbb24aSTao Ma u32 clusters, 2385bcbbb24aSTao Ma int *meta_add, 2386bcbbb24aSTao Ma int *credits) 2387bcbbb24aSTao Ma { 2388bcbbb24aSTao Ma int ret = 0, index, ref_blocks = 0, recs_add = 0; 2389bcbbb24aSTao Ma u64 cpos = start_cpos; 2390bcbbb24aSTao Ma struct ocfs2_refcount_block *rb; 2391bcbbb24aSTao Ma struct ocfs2_refcount_rec rec; 2392bcbbb24aSTao Ma struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL; 2393bcbbb24aSTao Ma u32 len; 2394bcbbb24aSTao Ma 2395bcbbb24aSTao Ma mlog(0, "start_cpos %llu, clusters %u\n", 2396bcbbb24aSTao Ma (unsigned long long)start_cpos, clusters); 2397bcbbb24aSTao Ma while (clusters) { 2398bcbbb24aSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 2399bcbbb24aSTao Ma cpos, clusters, &rec, 2400bcbbb24aSTao Ma &index, &ref_leaf_bh); 2401bcbbb24aSTao Ma if (ret) { 2402bcbbb24aSTao Ma mlog_errno(ret); 2403bcbbb24aSTao Ma goto out; 2404bcbbb24aSTao Ma } 2405bcbbb24aSTao Ma 2406bcbbb24aSTao Ma if (ref_leaf_bh != prev_bh) { 2407bcbbb24aSTao Ma /* 2408bcbbb24aSTao Ma * Now we encounter a new leaf block, so calculate 2409bcbbb24aSTao Ma * whether we need to extend the old leaf. 2410bcbbb24aSTao Ma */ 2411bcbbb24aSTao Ma if (prev_bh) { 2412bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *) 2413bcbbb24aSTao Ma prev_bh->b_data; 2414bcbbb24aSTao Ma 2415bcbbb24aSTao Ma if (le64_to_cpu(rb->rf_records.rl_used) + 2416bcbbb24aSTao Ma recs_add > 2417bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2418bcbbb24aSTao Ma ref_blocks++; 2419bcbbb24aSTao Ma } 2420bcbbb24aSTao Ma 2421bcbbb24aSTao Ma recs_add = 0; 2422bcbbb24aSTao Ma *credits += 1; 2423bcbbb24aSTao Ma brelse(prev_bh); 2424bcbbb24aSTao Ma prev_bh = ref_leaf_bh; 2425bcbbb24aSTao Ma get_bh(prev_bh); 2426bcbbb24aSTao Ma } 2427bcbbb24aSTao Ma 2428bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 2429bcbbb24aSTao Ma 2430bcbbb24aSTao Ma mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu," 2431bcbbb24aSTao Ma "rec->r_clusters %u, rec->r_refcount %u, index %d\n", 2432bcbbb24aSTao Ma recs_add, (unsigned long long)cpos, clusters, 2433bcbbb24aSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2434bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters), 2435bcbbb24aSTao Ma le32_to_cpu(rec.r_refcount), index); 2436bcbbb24aSTao Ma 2437bcbbb24aSTao Ma len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) + 2438bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 2439bcbbb24aSTao Ma /* 2440bcbbb24aSTao Ma * We record all the records which will be inserted to the 2441bcbbb24aSTao Ma * same refcount block, so that we can tell exactly whether 2442bcbbb24aSTao Ma * we need a new refcount block or not. 24438a2e70c4STao Ma * 24448a2e70c4STao Ma * If we will insert a new one, this is easy and only happens 24458a2e70c4STao Ma * during adding refcounted flag to the extent, so we don't 24468a2e70c4STao Ma * have a chance of spliting. We just need one record. 24478a2e70c4STao Ma * 24488a2e70c4STao Ma * If the refcount rec already exists, that would be a little 24498a2e70c4STao Ma * complicated. we may have to: 24508a2e70c4STao Ma * 1) split at the beginning if the start pos isn't aligned. 24518a2e70c4STao Ma * we need 1 more record in this case. 24528a2e70c4STao Ma * 2) split int the end if the end pos isn't aligned. 24538a2e70c4STao Ma * we need 1 more record in this case. 24548a2e70c4STao Ma * 3) split in the middle because of file system fragmentation. 24558a2e70c4STao Ma * we need 2 more records in this case(we can't detect this 24568a2e70c4STao Ma * beforehand, so always think of the worst case). 2457bcbbb24aSTao Ma */ 2458bcbbb24aSTao Ma if (rec.r_refcount) { 24598a2e70c4STao Ma recs_add += 2; 2460bcbbb24aSTao Ma /* Check whether we need a split at the beginning. */ 2461bcbbb24aSTao Ma if (cpos == start_cpos && 2462bcbbb24aSTao Ma cpos != le64_to_cpu(rec.r_cpos)) 2463bcbbb24aSTao Ma recs_add++; 2464bcbbb24aSTao Ma 2465bcbbb24aSTao Ma /* Check whether we need a split in the end. */ 2466bcbbb24aSTao Ma if (cpos + clusters < le64_to_cpu(rec.r_cpos) + 2467bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) 2468bcbbb24aSTao Ma recs_add++; 2469bcbbb24aSTao Ma } else 2470bcbbb24aSTao Ma recs_add++; 2471bcbbb24aSTao Ma 2472bcbbb24aSTao Ma brelse(ref_leaf_bh); 2473bcbbb24aSTao Ma ref_leaf_bh = NULL; 2474bcbbb24aSTao Ma clusters -= len; 2475bcbbb24aSTao Ma cpos += len; 2476bcbbb24aSTao Ma } 2477bcbbb24aSTao Ma 2478bcbbb24aSTao Ma if (prev_bh) { 2479bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)prev_bh->b_data; 2480bcbbb24aSTao Ma 2481bcbbb24aSTao Ma if (le64_to_cpu(rb->rf_records.rl_used) + recs_add > 2482bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2483bcbbb24aSTao Ma ref_blocks++; 2484bcbbb24aSTao Ma 2485bcbbb24aSTao Ma *credits += 1; 2486bcbbb24aSTao Ma } 2487bcbbb24aSTao Ma 2488bcbbb24aSTao Ma if (!ref_blocks) 2489bcbbb24aSTao Ma goto out; 2490bcbbb24aSTao Ma 2491bcbbb24aSTao Ma mlog(0, "we need ref_blocks %d\n", ref_blocks); 2492bcbbb24aSTao Ma *meta_add += ref_blocks; 2493bcbbb24aSTao Ma *credits += ref_blocks; 2494bcbbb24aSTao Ma 2495bcbbb24aSTao Ma /* 2496bcbbb24aSTao Ma * So we may need ref_blocks to insert into the tree. 2497bcbbb24aSTao Ma * That also means we need to change the b-tree and add that number 2498bcbbb24aSTao Ma * of records since we never merge them. 2499bcbbb24aSTao Ma * We need one more block for expansion since the new created leaf 2500bcbbb24aSTao Ma * block is also full and needs split. 2501bcbbb24aSTao Ma */ 2502bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 2503bcbbb24aSTao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) { 2504bcbbb24aSTao Ma struct ocfs2_extent_tree et; 2505bcbbb24aSTao Ma 2506bcbbb24aSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 2507bcbbb24aSTao Ma *meta_add += ocfs2_extend_meta_needed(et.et_root_el); 2508bcbbb24aSTao Ma *credits += ocfs2_calc_extend_credits(sb, 2509bcbbb24aSTao Ma et.et_root_el, 2510bcbbb24aSTao Ma ref_blocks); 2511bcbbb24aSTao Ma } else { 2512bcbbb24aSTao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 2513bcbbb24aSTao Ma *meta_add += 1; 2514bcbbb24aSTao Ma } 2515bcbbb24aSTao Ma 2516bcbbb24aSTao Ma out: 2517bcbbb24aSTao Ma brelse(ref_leaf_bh); 2518bcbbb24aSTao Ma brelse(prev_bh); 2519bcbbb24aSTao Ma return ret; 2520bcbbb24aSTao Ma } 2521bcbbb24aSTao Ma 2522bcbbb24aSTao Ma /* 2523bcbbb24aSTao Ma * For refcount tree, we will decrease some contiguous clusters 2524bcbbb24aSTao Ma * refcount count, so just go through it to see how many blocks 2525bcbbb24aSTao Ma * we gonna touch and whether we need to create new blocks. 2526bcbbb24aSTao Ma * 2527bcbbb24aSTao Ma * Normally the refcount blocks store these refcount should be 2528af901ca1SAndré Goddard Rosa * contiguous also, so that we can get the number easily. 252978f94673STristan Ye * We will at most add split 2 refcount records and 2 more 253078f94673STristan Ye * refcount blocks, so just check it in a rough way. 2531bcbbb24aSTao Ma * 2532bcbbb24aSTao Ma * Caller must hold refcount tree lock. 2533bcbbb24aSTao Ma */ 2534bcbbb24aSTao Ma int ocfs2_prepare_refcount_change_for_del(struct inode *inode, 253578f94673STristan Ye u64 refcount_loc, 2536bcbbb24aSTao Ma u64 phys_blkno, 2537bcbbb24aSTao Ma u32 clusters, 2538bcbbb24aSTao Ma int *credits, 253978f94673STristan Ye int *ref_blocks) 2540bcbbb24aSTao Ma { 254178f94673STristan Ye int ret; 2542bcbbb24aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 2543bcbbb24aSTao Ma struct buffer_head *ref_root_bh = NULL; 2544bcbbb24aSTao Ma struct ocfs2_refcount_tree *tree; 2545bcbbb24aSTao Ma u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno); 2546bcbbb24aSTao Ma 2547bcbbb24aSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 2548bcbbb24aSTao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 2549bcbbb24aSTao Ma "tree, but the feature bit is not set in the " 2550bcbbb24aSTao Ma "super block.", inode->i_ino); 2551bcbbb24aSTao Ma ret = -EROFS; 2552bcbbb24aSTao Ma goto out; 2553bcbbb24aSTao Ma } 2554bcbbb24aSTao Ma 2555bcbbb24aSTao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 2556bcbbb24aSTao Ma 2557bcbbb24aSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), 255878f94673STristan Ye refcount_loc, &tree); 2559bcbbb24aSTao Ma if (ret) { 2560bcbbb24aSTao Ma mlog_errno(ret); 2561bcbbb24aSTao Ma goto out; 2562bcbbb24aSTao Ma } 2563bcbbb24aSTao Ma 256478f94673STristan Ye ret = ocfs2_read_refcount_block(&tree->rf_ci, refcount_loc, 2565bcbbb24aSTao Ma &ref_root_bh); 2566bcbbb24aSTao Ma if (ret) { 2567bcbbb24aSTao Ma mlog_errno(ret); 2568bcbbb24aSTao Ma goto out; 2569bcbbb24aSTao Ma } 2570bcbbb24aSTao Ma 2571bcbbb24aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 2572bcbbb24aSTao Ma &tree->rf_ci, 2573bcbbb24aSTao Ma ref_root_bh, 2574bcbbb24aSTao Ma start_cpos, clusters, 257578f94673STristan Ye ref_blocks, credits); 2576bcbbb24aSTao Ma if (ret) { 2577bcbbb24aSTao Ma mlog_errno(ret); 2578bcbbb24aSTao Ma goto out; 2579bcbbb24aSTao Ma } 2580bcbbb24aSTao Ma 258178f94673STristan Ye mlog(0, "reserve new metadata %d blocks, credits = %d\n", 258278f94673STristan Ye *ref_blocks, *credits); 2583bcbbb24aSTao Ma 2584bcbbb24aSTao Ma out: 2585bcbbb24aSTao Ma brelse(ref_root_bh); 2586bcbbb24aSTao Ma return ret; 2587bcbbb24aSTao Ma } 25886f70fa51STao Ma 25896f70fa51STao Ma #define MAX_CONTIG_BYTES 1048576 25906f70fa51STao Ma 25916f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb) 25926f70fa51STao Ma { 25936f70fa51STao Ma return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES); 25946f70fa51STao Ma } 25956f70fa51STao Ma 25966f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb) 25976f70fa51STao Ma { 25986f70fa51STao Ma return ~(ocfs2_cow_contig_clusters(sb) - 1); 25996f70fa51STao Ma } 26006f70fa51STao Ma 26016f70fa51STao Ma /* 26026f70fa51STao Ma * Given an extent that starts at 'start' and an I/O that starts at 'cpos', 26036f70fa51STao Ma * find an offset (start + (n * contig_clusters)) that is closest to cpos 26046f70fa51STao Ma * while still being less than or equal to it. 26056f70fa51STao Ma * 26066f70fa51STao Ma * The goal is to break the extent at a multiple of contig_clusters. 26076f70fa51STao Ma */ 26086f70fa51STao Ma static inline unsigned int ocfs2_cow_align_start(struct super_block *sb, 26096f70fa51STao Ma unsigned int start, 26106f70fa51STao Ma unsigned int cpos) 26116f70fa51STao Ma { 26126f70fa51STao Ma BUG_ON(start > cpos); 26136f70fa51STao Ma 26146f70fa51STao Ma return start + ((cpos - start) & ocfs2_cow_contig_mask(sb)); 26156f70fa51STao Ma } 26166f70fa51STao Ma 26176f70fa51STao Ma /* 26186f70fa51STao Ma * Given a cluster count of len, pad it out so that it is a multiple 26196f70fa51STao Ma * of contig_clusters. 26206f70fa51STao Ma */ 26216f70fa51STao Ma static inline unsigned int ocfs2_cow_align_length(struct super_block *sb, 26226f70fa51STao Ma unsigned int len) 26236f70fa51STao Ma { 26246f70fa51STao Ma unsigned int padded = 26256f70fa51STao Ma (len + (ocfs2_cow_contig_clusters(sb) - 1)) & 26266f70fa51STao Ma ocfs2_cow_contig_mask(sb); 26276f70fa51STao Ma 26286f70fa51STao Ma /* Did we wrap? */ 26296f70fa51STao Ma if (padded < len) 26306f70fa51STao Ma padded = UINT_MAX; 26316f70fa51STao Ma 26326f70fa51STao Ma return padded; 26336f70fa51STao Ma } 26346f70fa51STao Ma 26356f70fa51STao Ma /* 26366f70fa51STao Ma * Calculate out the start and number of virtual clusters we need to to CoW. 26376f70fa51STao Ma * 26386f70fa51STao Ma * cpos is vitual start cluster position we want to do CoW in a 26396f70fa51STao Ma * file and write_len is the cluster length. 264037f8a2bfSTao Ma * max_cpos is the place where we want to stop CoW intentionally. 26416f70fa51STao Ma * 26426f70fa51STao Ma * Normal we will start CoW from the beginning of extent record cotaining cpos. 26436f70fa51STao Ma * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we 26446f70fa51STao Ma * get good I/O from the resulting extent tree. 26456f70fa51STao Ma */ 26466f70fa51STao Ma static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, 2647913580b4STao Ma struct ocfs2_extent_list *el, 26486f70fa51STao Ma u32 cpos, 26496f70fa51STao Ma u32 write_len, 265037f8a2bfSTao Ma u32 max_cpos, 26516f70fa51STao Ma u32 *cow_start, 26526f70fa51STao Ma u32 *cow_len) 26536f70fa51STao Ma { 26546f70fa51STao Ma int ret = 0; 26556f70fa51STao Ma int tree_height = le16_to_cpu(el->l_tree_depth), i; 26566f70fa51STao Ma struct buffer_head *eb_bh = NULL; 26576f70fa51STao Ma struct ocfs2_extent_block *eb = NULL; 26586f70fa51STao Ma struct ocfs2_extent_rec *rec; 26596f70fa51STao Ma unsigned int want_clusters, rec_end = 0; 26606f70fa51STao Ma int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb); 26616f70fa51STao Ma int leaf_clusters; 26626f70fa51STao Ma 266337f8a2bfSTao Ma BUG_ON(cpos + write_len > max_cpos); 266437f8a2bfSTao Ma 26656f70fa51STao Ma if (tree_height > 0) { 26666f70fa51STao Ma ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh); 26676f70fa51STao Ma if (ret) { 26686f70fa51STao Ma mlog_errno(ret); 26696f70fa51STao Ma goto out; 26706f70fa51STao Ma } 26716f70fa51STao Ma 26726f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 26736f70fa51STao Ma el = &eb->h_list; 26746f70fa51STao Ma 26756f70fa51STao Ma if (el->l_tree_depth) { 26766f70fa51STao Ma ocfs2_error(inode->i_sb, 26776f70fa51STao Ma "Inode %lu has non zero tree depth in " 26786f70fa51STao Ma "leaf block %llu\n", inode->i_ino, 26796f70fa51STao Ma (unsigned long long)eb_bh->b_blocknr); 26806f70fa51STao Ma ret = -EROFS; 26816f70fa51STao Ma goto out; 26826f70fa51STao Ma } 26836f70fa51STao Ma } 26846f70fa51STao Ma 26856f70fa51STao Ma *cow_len = 0; 26866f70fa51STao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 26876f70fa51STao Ma rec = &el->l_recs[i]; 26886f70fa51STao Ma 26896f70fa51STao Ma if (ocfs2_is_empty_extent(rec)) { 26906f70fa51STao Ma mlog_bug_on_msg(i != 0, "Inode %lu has empty record in " 26916f70fa51STao Ma "index %d\n", inode->i_ino, i); 26926f70fa51STao Ma continue; 26936f70fa51STao Ma } 26946f70fa51STao Ma 26956f70fa51STao Ma if (le32_to_cpu(rec->e_cpos) + 26966f70fa51STao Ma le16_to_cpu(rec->e_leaf_clusters) <= cpos) 26976f70fa51STao Ma continue; 26986f70fa51STao Ma 26996f70fa51STao Ma if (*cow_len == 0) { 27006f70fa51STao Ma /* 27016f70fa51STao Ma * We should find a refcounted record in the 27026f70fa51STao Ma * first pass. 27036f70fa51STao Ma */ 27046f70fa51STao Ma BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED)); 27056f70fa51STao Ma *cow_start = le32_to_cpu(rec->e_cpos); 27066f70fa51STao Ma } 27076f70fa51STao Ma 27086f70fa51STao Ma /* 270937f8a2bfSTao Ma * If we encounter a hole, a non-refcounted record or 271037f8a2bfSTao Ma * pass the max_cpos, stop the search. 27116f70fa51STao Ma */ 27126f70fa51STao Ma if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) || 271337f8a2bfSTao Ma (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) || 271437f8a2bfSTao Ma (max_cpos <= le32_to_cpu(rec->e_cpos))) 27156f70fa51STao Ma break; 27166f70fa51STao Ma 27176f70fa51STao Ma leaf_clusters = le16_to_cpu(rec->e_leaf_clusters); 27186f70fa51STao Ma rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters; 271937f8a2bfSTao Ma if (rec_end > max_cpos) { 272037f8a2bfSTao Ma rec_end = max_cpos; 272137f8a2bfSTao Ma leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos); 272237f8a2bfSTao Ma } 27236f70fa51STao Ma 27246f70fa51STao Ma /* 27256f70fa51STao Ma * How many clusters do we actually need from 27266f70fa51STao Ma * this extent? First we see how many we actually 27276f70fa51STao Ma * need to complete the write. If that's smaller 27286f70fa51STao Ma * than contig_clusters, we try for contig_clusters. 27296f70fa51STao Ma */ 27306f70fa51STao Ma if (!*cow_len) 27316f70fa51STao Ma want_clusters = write_len; 27326f70fa51STao Ma else 27336f70fa51STao Ma want_clusters = (cpos + write_len) - 27346f70fa51STao Ma (*cow_start + *cow_len); 27356f70fa51STao Ma if (want_clusters < contig_clusters) 27366f70fa51STao Ma want_clusters = contig_clusters; 27376f70fa51STao Ma 27386f70fa51STao Ma /* 27396f70fa51STao Ma * If the write does not cover the whole extent, we 27406f70fa51STao Ma * need to calculate how we're going to split the extent. 27416f70fa51STao Ma * We try to do it on contig_clusters boundaries. 27426f70fa51STao Ma * 27436f70fa51STao Ma * Any extent smaller than contig_clusters will be 27446f70fa51STao Ma * CoWed in its entirety. 27456f70fa51STao Ma */ 27466f70fa51STao Ma if (leaf_clusters <= contig_clusters) 27476f70fa51STao Ma *cow_len += leaf_clusters; 27486f70fa51STao Ma else if (*cow_len || (*cow_start == cpos)) { 27496f70fa51STao Ma /* 27506f70fa51STao Ma * This extent needs to be CoW'd from its 27516f70fa51STao Ma * beginning, so all we have to do is compute 27526f70fa51STao Ma * how many clusters to grab. We align 27536f70fa51STao Ma * want_clusters to the edge of contig_clusters 27546f70fa51STao Ma * to get better I/O. 27556f70fa51STao Ma */ 27566f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 27576f70fa51STao Ma want_clusters); 27586f70fa51STao Ma 27596f70fa51STao Ma if (leaf_clusters < want_clusters) 27606f70fa51STao Ma *cow_len += leaf_clusters; 27616f70fa51STao Ma else 27626f70fa51STao Ma *cow_len += want_clusters; 27636f70fa51STao Ma } else if ((*cow_start + contig_clusters) >= 27646f70fa51STao Ma (cpos + write_len)) { 27656f70fa51STao Ma /* 27666f70fa51STao Ma * Breaking off contig_clusters at the front 27676f70fa51STao Ma * of the extent will cover our write. That's 27686f70fa51STao Ma * easy. 27696f70fa51STao Ma */ 27706f70fa51STao Ma *cow_len = contig_clusters; 27716f70fa51STao Ma } else if ((rec_end - cpos) <= contig_clusters) { 27726f70fa51STao Ma /* 27736f70fa51STao Ma * Breaking off contig_clusters at the tail of 27746f70fa51STao Ma * this extent will cover cpos. 27756f70fa51STao Ma */ 27766f70fa51STao Ma *cow_start = rec_end - contig_clusters; 27776f70fa51STao Ma *cow_len = contig_clusters; 27786f70fa51STao Ma } else if ((rec_end - cpos) <= want_clusters) { 27796f70fa51STao Ma /* 27806f70fa51STao Ma * While we can't fit the entire write in this 27816f70fa51STao Ma * extent, we know that the write goes from cpos 27826f70fa51STao Ma * to the end of the extent. Break that off. 27836f70fa51STao Ma * We try to break it at some multiple of 27846f70fa51STao Ma * contig_clusters from the front of the extent. 27856f70fa51STao Ma * Failing that (ie, cpos is within 27866f70fa51STao Ma * contig_clusters of the front), we'll CoW the 27876f70fa51STao Ma * entire extent. 27886f70fa51STao Ma */ 27896f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 27906f70fa51STao Ma *cow_start, cpos); 27916f70fa51STao Ma *cow_len = rec_end - *cow_start; 27926f70fa51STao Ma } else { 27936f70fa51STao Ma /* 27946f70fa51STao Ma * Ok, the entire write lives in the middle of 27956f70fa51STao Ma * this extent. Let's try to slice the extent up 27966f70fa51STao Ma * nicely. Optimally, our CoW region starts at 27976f70fa51STao Ma * m*contig_clusters from the beginning of the 27986f70fa51STao Ma * extent and goes for n*contig_clusters, 27996f70fa51STao Ma * covering the entire write. 28006f70fa51STao Ma */ 28016f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 28026f70fa51STao Ma *cow_start, cpos); 28036f70fa51STao Ma 28046f70fa51STao Ma want_clusters = (cpos + write_len) - *cow_start; 28056f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 28066f70fa51STao Ma want_clusters); 28076f70fa51STao Ma if (*cow_start + want_clusters <= rec_end) 28086f70fa51STao Ma *cow_len = want_clusters; 28096f70fa51STao Ma else 28106f70fa51STao Ma *cow_len = rec_end - *cow_start; 28116f70fa51STao Ma } 28126f70fa51STao Ma 28136f70fa51STao Ma /* Have we covered our entire write yet? */ 28146f70fa51STao Ma if ((*cow_start + *cow_len) >= (cpos + write_len)) 28156f70fa51STao Ma break; 28166f70fa51STao Ma 28176f70fa51STao Ma /* 28186f70fa51STao Ma * If we reach the end of the extent block and don't get enough 28196f70fa51STao Ma * clusters, continue with the next extent block if possible. 28206f70fa51STao Ma */ 28216f70fa51STao Ma if (i + 1 == le16_to_cpu(el->l_next_free_rec) && 28226f70fa51STao Ma eb && eb->h_next_leaf_blk) { 28236f70fa51STao Ma brelse(eb_bh); 28246f70fa51STao Ma eb_bh = NULL; 28256f70fa51STao Ma 28266f70fa51STao Ma ret = ocfs2_read_extent_block(INODE_CACHE(inode), 28276f70fa51STao Ma le64_to_cpu(eb->h_next_leaf_blk), 28286f70fa51STao Ma &eb_bh); 28296f70fa51STao Ma if (ret) { 28306f70fa51STao Ma mlog_errno(ret); 28316f70fa51STao Ma goto out; 28326f70fa51STao Ma } 28336f70fa51STao Ma 28346f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 28356f70fa51STao Ma el = &eb->h_list; 28366f70fa51STao Ma i = -1; 28376f70fa51STao Ma } 28386f70fa51STao Ma } 28396f70fa51STao Ma 28406f70fa51STao Ma out: 28416f70fa51STao Ma brelse(eb_bh); 28426f70fa51STao Ma return ret; 28436f70fa51STao Ma } 28446f70fa51STao Ma 28456f70fa51STao Ma /* 28466f70fa51STao Ma * Prepare meta_ac, data_ac and calculate credits when we want to add some 28476f70fa51STao Ma * num_clusters in data_tree "et" and change the refcount for the old 28486f70fa51STao Ma * clusters(starting form p_cluster) in the refcount tree. 28496f70fa51STao Ma * 28506f70fa51STao Ma * Note: 28516f70fa51STao Ma * 1. since we may split the old tree, so we at most will need num_clusters + 2 28526f70fa51STao Ma * more new leaf records. 28536f70fa51STao Ma * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so 28546f70fa51STao Ma * just give data_ac = NULL. 28556f70fa51STao Ma */ 28566f70fa51STao Ma static int ocfs2_lock_refcount_allocators(struct super_block *sb, 28576f70fa51STao Ma u32 p_cluster, u32 num_clusters, 28586f70fa51STao Ma struct ocfs2_extent_tree *et, 28596f70fa51STao Ma struct ocfs2_caching_info *ref_ci, 28606f70fa51STao Ma struct buffer_head *ref_root_bh, 28616f70fa51STao Ma struct ocfs2_alloc_context **meta_ac, 28626f70fa51STao Ma struct ocfs2_alloc_context **data_ac, 28636f70fa51STao Ma int *credits) 28646f70fa51STao Ma { 28656f70fa51STao Ma int ret = 0, meta_add = 0; 28666f70fa51STao Ma int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et); 28676f70fa51STao Ma 28686f70fa51STao Ma if (num_free_extents < 0) { 28696f70fa51STao Ma ret = num_free_extents; 28706f70fa51STao Ma mlog_errno(ret); 28716f70fa51STao Ma goto out; 28726f70fa51STao Ma } 28736f70fa51STao Ma 28746f70fa51STao Ma if (num_free_extents < num_clusters + 2) 28756f70fa51STao Ma meta_add = 28766f70fa51STao Ma ocfs2_extend_meta_needed(et->et_root_el); 28776f70fa51STao Ma 28786f70fa51STao Ma *credits += ocfs2_calc_extend_credits(sb, et->et_root_el, 28796f70fa51STao Ma num_clusters + 2); 28806f70fa51STao Ma 28816f70fa51STao Ma ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh, 28826f70fa51STao Ma p_cluster, num_clusters, 28836f70fa51STao Ma &meta_add, credits); 28846f70fa51STao Ma if (ret) { 28856f70fa51STao Ma mlog_errno(ret); 28866f70fa51STao Ma goto out; 28876f70fa51STao Ma } 28886f70fa51STao Ma 28896f70fa51STao Ma mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n", 28906f70fa51STao Ma meta_add, num_clusters, *credits); 28916f70fa51STao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add, 28926f70fa51STao Ma meta_ac); 28936f70fa51STao Ma if (ret) { 28946f70fa51STao Ma mlog_errno(ret); 28956f70fa51STao Ma goto out; 28966f70fa51STao Ma } 28976f70fa51STao Ma 28986f70fa51STao Ma if (data_ac) { 28996f70fa51STao Ma ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters, 29006f70fa51STao Ma data_ac); 29016f70fa51STao Ma if (ret) 29026f70fa51STao Ma mlog_errno(ret); 29036f70fa51STao Ma } 29046f70fa51STao Ma 29056f70fa51STao Ma out: 29066f70fa51STao Ma if (ret) { 29076f70fa51STao Ma if (*meta_ac) { 29086f70fa51STao Ma ocfs2_free_alloc_context(*meta_ac); 29096f70fa51STao Ma *meta_ac = NULL; 29106f70fa51STao Ma } 29116f70fa51STao Ma } 29126f70fa51STao Ma 29136f70fa51STao Ma return ret; 29146f70fa51STao Ma } 29156f70fa51STao Ma 29166f70fa51STao Ma static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh) 29176f70fa51STao Ma { 29186f70fa51STao Ma BUG_ON(buffer_dirty(bh)); 29196f70fa51STao Ma 29206f70fa51STao Ma clear_buffer_mapped(bh); 29216f70fa51STao Ma 29226f70fa51STao Ma return 0; 29236f70fa51STao Ma } 29246f70fa51STao Ma 2925913580b4STao Ma static int ocfs2_duplicate_clusters_by_page(handle_t *handle, 29266f70fa51STao Ma struct ocfs2_cow_context *context, 29276f70fa51STao Ma u32 cpos, u32 old_cluster, 29286f70fa51STao Ma u32 new_cluster, u32 new_len) 29296f70fa51STao Ma { 29306f70fa51STao Ma int ret = 0, partial; 2931913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 29326f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 29336f70fa51STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 29346f70fa51STao Ma struct page *page; 29356f70fa51STao Ma pgoff_t page_index; 29366ea4843fSTao Ma unsigned int from, to, readahead_pages; 29376f70fa51STao Ma loff_t offset, end, map_end; 29386f70fa51STao Ma struct address_space *mapping = context->inode->i_mapping; 29396f70fa51STao Ma 29406f70fa51STao Ma mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster, 29416f70fa51STao Ma new_cluster, new_len, cpos); 29426f70fa51STao Ma 29436ea4843fSTao Ma readahead_pages = 29446ea4843fSTao Ma (ocfs2_cow_contig_clusters(sb) << 29456ea4843fSTao Ma OCFS2_SB(sb)->s_clustersize_bits) >> PAGE_CACHE_SHIFT; 29466f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 29476f70fa51STao Ma end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits); 2948f5e27b6dSTao Ma /* 2949f5e27b6dSTao Ma * We only duplicate pages until we reach the page contains i_size - 1. 2950f5e27b6dSTao Ma * So trim 'end' to i_size. 2951f5e27b6dSTao Ma */ 2952f5e27b6dSTao Ma if (end > i_size_read(context->inode)) 2953f5e27b6dSTao Ma end = i_size_read(context->inode); 29546f70fa51STao Ma 29556f70fa51STao Ma while (offset < end) { 29566f70fa51STao Ma page_index = offset >> PAGE_CACHE_SHIFT; 2957d622b89aSTao Ma map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT; 29586f70fa51STao Ma if (map_end > end) 29596f70fa51STao Ma map_end = end; 29606f70fa51STao Ma 29616f70fa51STao Ma /* from, to is the offset within the page. */ 29626f70fa51STao Ma from = offset & (PAGE_CACHE_SIZE - 1); 29636f70fa51STao Ma to = PAGE_CACHE_SIZE; 29646f70fa51STao Ma if (map_end & (PAGE_CACHE_SIZE - 1)) 29656f70fa51STao Ma to = map_end & (PAGE_CACHE_SIZE - 1); 29666f70fa51STao Ma 29679b4c0ff3SJan Kara page = find_or_create_page(mapping, page_index, GFP_NOFS); 29686f70fa51STao Ma 29690a1ea437STao Ma /* 29700a1ea437STao Ma * In case PAGE_CACHE_SIZE <= CLUSTER_SIZE, This page 29710a1ea437STao Ma * can't be dirtied before we CoW it out. 29720a1ea437STao Ma */ 29730a1ea437STao Ma if (PAGE_CACHE_SIZE <= OCFS2_SB(sb)->s_clustersize) 29746f70fa51STao Ma BUG_ON(PageDirty(page)); 29756f70fa51STao Ma 29766ea4843fSTao Ma if (PageReadahead(page) && context->file) { 29776ea4843fSTao Ma page_cache_async_readahead(mapping, 29786ea4843fSTao Ma &context->file->f_ra, 29796ea4843fSTao Ma context->file, 29806ea4843fSTao Ma page, page_index, 29816ea4843fSTao Ma readahead_pages); 29826ea4843fSTao Ma } 29836ea4843fSTao Ma 29846f70fa51STao Ma if (!PageUptodate(page)) { 29856f70fa51STao Ma ret = block_read_full_page(page, ocfs2_get_block); 29866f70fa51STao Ma if (ret) { 29876f70fa51STao Ma mlog_errno(ret); 29886f70fa51STao Ma goto unlock; 29896f70fa51STao Ma } 29906f70fa51STao Ma lock_page(page); 29916f70fa51STao Ma } 29926f70fa51STao Ma 29936f70fa51STao Ma if (page_has_buffers(page)) { 29946f70fa51STao Ma ret = walk_page_buffers(handle, page_buffers(page), 29956f70fa51STao Ma from, to, &partial, 29966f70fa51STao Ma ocfs2_clear_cow_buffer); 29976f70fa51STao Ma if (ret) { 29986f70fa51STao Ma mlog_errno(ret); 29996f70fa51STao Ma goto unlock; 30006f70fa51STao Ma } 30016f70fa51STao Ma } 30026f70fa51STao Ma 30036f70fa51STao Ma ocfs2_map_and_dirty_page(context->inode, 30046f70fa51STao Ma handle, from, to, 30056f70fa51STao Ma page, 0, &new_block); 30066f70fa51STao Ma mark_page_accessed(page); 30076f70fa51STao Ma unlock: 30086f70fa51STao Ma unlock_page(page); 30096f70fa51STao Ma page_cache_release(page); 30106f70fa51STao Ma page = NULL; 30116f70fa51STao Ma offset = map_end; 30126f70fa51STao Ma if (ret) 30136f70fa51STao Ma break; 30146f70fa51STao Ma } 30156f70fa51STao Ma 30166f70fa51STao Ma return ret; 30176f70fa51STao Ma } 30186f70fa51STao Ma 3019492a8a33STao Ma static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle, 3020492a8a33STao Ma struct ocfs2_cow_context *context, 3021492a8a33STao Ma u32 cpos, u32 old_cluster, 3022492a8a33STao Ma u32 new_cluster, u32 new_len) 3023492a8a33STao Ma { 3024492a8a33STao Ma int ret = 0; 3025492a8a33STao Ma struct super_block *sb = context->inode->i_sb; 3026492a8a33STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 3027492a8a33STao Ma int i, blocks = ocfs2_clusters_to_blocks(sb, new_len); 3028492a8a33STao Ma u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster); 3029492a8a33STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 3030492a8a33STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 3031492a8a33STao Ma struct buffer_head *old_bh = NULL; 3032492a8a33STao Ma struct buffer_head *new_bh = NULL; 3033492a8a33STao Ma 3034492a8a33STao Ma mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster, 3035492a8a33STao Ma new_cluster, new_len); 3036492a8a33STao Ma 3037492a8a33STao Ma for (i = 0; i < blocks; i++, old_block++, new_block++) { 3038492a8a33STao Ma new_bh = sb_getblk(osb->sb, new_block); 3039492a8a33STao Ma if (new_bh == NULL) { 3040492a8a33STao Ma ret = -EIO; 3041492a8a33STao Ma mlog_errno(ret); 3042492a8a33STao Ma break; 3043492a8a33STao Ma } 3044492a8a33STao Ma 3045492a8a33STao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 3046492a8a33STao Ma 3047492a8a33STao Ma ret = ocfs2_read_block(ci, old_block, &old_bh, NULL); 3048492a8a33STao Ma if (ret) { 3049492a8a33STao Ma mlog_errno(ret); 3050492a8a33STao Ma break; 3051492a8a33STao Ma } 3052492a8a33STao Ma 3053492a8a33STao Ma ret = ocfs2_journal_access(handle, ci, new_bh, 3054492a8a33STao Ma OCFS2_JOURNAL_ACCESS_CREATE); 3055492a8a33STao Ma if (ret) { 3056492a8a33STao Ma mlog_errno(ret); 3057492a8a33STao Ma break; 3058492a8a33STao Ma } 3059492a8a33STao Ma 3060492a8a33STao Ma memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize); 3061ec20cec7SJoel Becker ocfs2_journal_dirty(handle, new_bh); 3062492a8a33STao Ma 3063492a8a33STao Ma brelse(new_bh); 3064492a8a33STao Ma brelse(old_bh); 3065492a8a33STao Ma new_bh = NULL; 3066492a8a33STao Ma old_bh = NULL; 3067492a8a33STao Ma } 3068492a8a33STao Ma 3069492a8a33STao Ma brelse(new_bh); 3070492a8a33STao Ma brelse(old_bh); 3071492a8a33STao Ma return ret; 3072492a8a33STao Ma } 3073492a8a33STao Ma 30746f70fa51STao Ma static int ocfs2_clear_ext_refcount(handle_t *handle, 30756f70fa51STao Ma struct ocfs2_extent_tree *et, 30766f70fa51STao Ma u32 cpos, u32 p_cluster, u32 len, 30776f70fa51STao Ma unsigned int ext_flags, 30786f70fa51STao Ma struct ocfs2_alloc_context *meta_ac, 30796f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 30806f70fa51STao Ma { 30816f70fa51STao Ma int ret, index; 30826f70fa51STao Ma struct ocfs2_extent_rec replace_rec; 30836f70fa51STao Ma struct ocfs2_path *path = NULL; 30846f70fa51STao Ma struct ocfs2_extent_list *el; 30856f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 30866f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(et->et_ci); 30876f70fa51STao Ma 30886f70fa51STao Ma mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n", 30896f70fa51STao Ma (unsigned long long)ino, cpos, len, p_cluster, ext_flags); 30906f70fa51STao Ma 30916f70fa51STao Ma memset(&replace_rec, 0, sizeof(replace_rec)); 30926f70fa51STao Ma replace_rec.e_cpos = cpu_to_le32(cpos); 30936f70fa51STao Ma replace_rec.e_leaf_clusters = cpu_to_le16(len); 30946f70fa51STao Ma replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb, 30956f70fa51STao Ma p_cluster)); 30966f70fa51STao Ma replace_rec.e_flags = ext_flags; 30976f70fa51STao Ma replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED; 30986f70fa51STao Ma 30996f70fa51STao Ma path = ocfs2_new_path_from_et(et); 31006f70fa51STao Ma if (!path) { 31016f70fa51STao Ma ret = -ENOMEM; 31026f70fa51STao Ma mlog_errno(ret); 31036f70fa51STao Ma goto out; 31046f70fa51STao Ma } 31056f70fa51STao Ma 31066f70fa51STao Ma ret = ocfs2_find_path(et->et_ci, path, cpos); 31076f70fa51STao Ma if (ret) { 31086f70fa51STao Ma mlog_errno(ret); 31096f70fa51STao Ma goto out; 31106f70fa51STao Ma } 31116f70fa51STao Ma 31126f70fa51STao Ma el = path_leaf_el(path); 31136f70fa51STao Ma 31146f70fa51STao Ma index = ocfs2_search_extent_list(el, cpos); 31156f70fa51STao Ma if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) { 31166f70fa51STao Ma ocfs2_error(sb, 31176f70fa51STao Ma "Inode %llu has an extent at cpos %u which can no " 31186f70fa51STao Ma "longer be found.\n", 31196f70fa51STao Ma (unsigned long long)ino, cpos); 31206f70fa51STao Ma ret = -EROFS; 31216f70fa51STao Ma goto out; 31226f70fa51STao Ma } 31236f70fa51STao Ma 31246f70fa51STao Ma ret = ocfs2_split_extent(handle, et, path, index, 31256f70fa51STao Ma &replace_rec, meta_ac, dealloc); 31266f70fa51STao Ma if (ret) 31276f70fa51STao Ma mlog_errno(ret); 31286f70fa51STao Ma 31296f70fa51STao Ma out: 31306f70fa51STao Ma ocfs2_free_path(path); 31316f70fa51STao Ma return ret; 31326f70fa51STao Ma } 31336f70fa51STao Ma 31346f70fa51STao Ma static int ocfs2_replace_clusters(handle_t *handle, 31356f70fa51STao Ma struct ocfs2_cow_context *context, 31366f70fa51STao Ma u32 cpos, u32 old, 31376f70fa51STao Ma u32 new, u32 len, 31386f70fa51STao Ma unsigned int ext_flags) 31396f70fa51STao Ma { 31406f70fa51STao Ma int ret; 3141913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 31426f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(ci); 31436f70fa51STao Ma 31446f70fa51STao Ma mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n", 31456f70fa51STao Ma (unsigned long long)ino, cpos, old, new, len, ext_flags); 31466f70fa51STao Ma 31476f70fa51STao Ma /*If the old clusters is unwritten, no need to duplicate. */ 31486f70fa51STao Ma if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) { 3149913580b4STao Ma ret = context->cow_duplicate_clusters(handle, context, cpos, 31506f70fa51STao Ma old, new, len); 31516f70fa51STao Ma if (ret) { 31526f70fa51STao Ma mlog_errno(ret); 31536f70fa51STao Ma goto out; 31546f70fa51STao Ma } 31556f70fa51STao Ma } 31566f70fa51STao Ma 3157913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, &context->data_et, 31586f70fa51STao Ma cpos, new, len, ext_flags, 31596f70fa51STao Ma context->meta_ac, &context->dealloc); 31606f70fa51STao Ma if (ret) 31616f70fa51STao Ma mlog_errno(ret); 31626f70fa51STao Ma out: 31636f70fa51STao Ma return ret; 31646f70fa51STao Ma } 31656f70fa51STao Ma 31666f70fa51STao Ma static int ocfs2_cow_sync_writeback(struct super_block *sb, 31676f70fa51STao Ma struct ocfs2_cow_context *context, 31686f70fa51STao Ma u32 cpos, u32 num_clusters) 31696f70fa51STao Ma { 31706f70fa51STao Ma int ret = 0; 31716f70fa51STao Ma loff_t offset, end, map_end; 31726f70fa51STao Ma pgoff_t page_index; 31736f70fa51STao Ma struct page *page; 31746f70fa51STao Ma 31756f70fa51STao Ma if (ocfs2_should_order_data(context->inode)) 31766f70fa51STao Ma return 0; 31776f70fa51STao Ma 31786f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 31796f70fa51STao Ma end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits); 31806f70fa51STao Ma 31816f70fa51STao Ma ret = filemap_fdatawrite_range(context->inode->i_mapping, 31826f70fa51STao Ma offset, end - 1); 31836f70fa51STao Ma if (ret < 0) { 31846f70fa51STao Ma mlog_errno(ret); 31856f70fa51STao Ma return ret; 31866f70fa51STao Ma } 31876f70fa51STao Ma 31886f70fa51STao Ma while (offset < end) { 31896f70fa51STao Ma page_index = offset >> PAGE_CACHE_SHIFT; 3190d622b89aSTao Ma map_end = ((loff_t)page_index + 1) << PAGE_CACHE_SHIFT; 31916f70fa51STao Ma if (map_end > end) 31926f70fa51STao Ma map_end = end; 31936f70fa51STao Ma 31949b4c0ff3SJan Kara page = find_or_create_page(context->inode->i_mapping, 31959b4c0ff3SJan Kara page_index, GFP_NOFS); 31966f70fa51STao Ma BUG_ON(!page); 31976f70fa51STao Ma 31986f70fa51STao Ma wait_on_page_writeback(page); 31996f70fa51STao Ma if (PageError(page)) { 32006f70fa51STao Ma ret = -EIO; 32016f70fa51STao Ma mlog_errno(ret); 32026f70fa51STao Ma } else 32036f70fa51STao Ma mark_page_accessed(page); 32046f70fa51STao Ma 32056f70fa51STao Ma unlock_page(page); 32066f70fa51STao Ma page_cache_release(page); 32076f70fa51STao Ma page = NULL; 32086f70fa51STao Ma offset = map_end; 32096f70fa51STao Ma if (ret) 32106f70fa51STao Ma break; 32116f70fa51STao Ma } 32126f70fa51STao Ma 32136f70fa51STao Ma return ret; 32146f70fa51STao Ma } 32156f70fa51STao Ma 3216913580b4STao Ma static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context, 3217913580b4STao Ma u32 v_cluster, u32 *p_cluster, 3218913580b4STao Ma u32 *num_clusters, 3219913580b4STao Ma unsigned int *extent_flags) 3220913580b4STao Ma { 3221913580b4STao Ma return ocfs2_get_clusters(context->inode, v_cluster, p_cluster, 3222913580b4STao Ma num_clusters, extent_flags); 3223913580b4STao Ma } 3224913580b4STao Ma 32256f70fa51STao Ma static int ocfs2_make_clusters_writable(struct super_block *sb, 32266f70fa51STao Ma struct ocfs2_cow_context *context, 32276f70fa51STao Ma u32 cpos, u32 p_cluster, 32286f70fa51STao Ma u32 num_clusters, unsigned int e_flags) 32296f70fa51STao Ma { 32306ae23c55STao Ma int ret, delete, index, credits = 0; 3231acf3bb00STristan Ye u32 new_bit, new_len, orig_num_clusters; 32326ae23c55STao Ma unsigned int set_len; 32336f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 32346f70fa51STao Ma handle_t *handle; 32356ae23c55STao Ma struct buffer_head *ref_leaf_bh = NULL; 3236913580b4STao Ma struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci; 32376ae23c55STao Ma struct ocfs2_refcount_rec rec; 32386ae23c55STao Ma 32396ae23c55STao Ma mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n", 32406ae23c55STao Ma cpos, p_cluster, num_clusters, e_flags); 32416f70fa51STao Ma 32426f70fa51STao Ma ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters, 3243913580b4STao Ma &context->data_et, 3244913580b4STao Ma ref_ci, 32456f70fa51STao Ma context->ref_root_bh, 32466f70fa51STao Ma &context->meta_ac, 32476f70fa51STao Ma &context->data_ac, &credits); 32486f70fa51STao Ma if (ret) { 32496f70fa51STao Ma mlog_errno(ret); 32506f70fa51STao Ma return ret; 32516f70fa51STao Ma } 32526f70fa51STao Ma 3253492a8a33STao Ma if (context->post_refcount) 3254492a8a33STao Ma credits += context->post_refcount->credits; 3255492a8a33STao Ma 3256492a8a33STao Ma credits += context->extra_credits; 32576f70fa51STao Ma handle = ocfs2_start_trans(osb, credits); 32586f70fa51STao Ma if (IS_ERR(handle)) { 32596f70fa51STao Ma ret = PTR_ERR(handle); 32606f70fa51STao Ma mlog_errno(ret); 32616f70fa51STao Ma goto out; 32626f70fa51STao Ma } 32636f70fa51STao Ma 3264acf3bb00STristan Ye orig_num_clusters = num_clusters; 3265acf3bb00STristan Ye 32666f70fa51STao Ma while (num_clusters) { 3267913580b4STao Ma ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh, 32686ae23c55STao Ma p_cluster, num_clusters, 32696ae23c55STao Ma &rec, &index, &ref_leaf_bh); 32706ae23c55STao Ma if (ret) { 32716ae23c55STao Ma mlog_errno(ret); 32726ae23c55STao Ma goto out_commit; 32736ae23c55STao Ma } 32746ae23c55STao Ma 32756ae23c55STao Ma BUG_ON(!rec.r_refcount); 32766ae23c55STao Ma set_len = min((u64)p_cluster + num_clusters, 32776ae23c55STao Ma le64_to_cpu(rec.r_cpos) + 32786ae23c55STao Ma le32_to_cpu(rec.r_clusters)) - p_cluster; 32796ae23c55STao Ma 32806ae23c55STao Ma /* 32816ae23c55STao Ma * There are many different situation here. 32826ae23c55STao Ma * 1. If refcount == 1, remove the flag and don't COW. 32836ae23c55STao Ma * 2. If refcount > 1, allocate clusters. 32846ae23c55STao Ma * Here we may not allocate r_len once at a time, so continue 32856ae23c55STao Ma * until we reach num_clusters. 32866ae23c55STao Ma */ 32876ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1) { 32886ae23c55STao Ma delete = 0; 3289913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, 3290913580b4STao Ma &context->data_et, 32916ae23c55STao Ma cpos, p_cluster, 32926ae23c55STao Ma set_len, e_flags, 32936ae23c55STao Ma context->meta_ac, 32946ae23c55STao Ma &context->dealloc); 32956ae23c55STao Ma if (ret) { 32966ae23c55STao Ma mlog_errno(ret); 32976ae23c55STao Ma goto out_commit; 32986ae23c55STao Ma } 32996ae23c55STao Ma } else { 33006ae23c55STao Ma delete = 1; 33016ae23c55STao Ma 33021ed9b777SJoel Becker ret = __ocfs2_claim_clusters(handle, 33036ae23c55STao Ma context->data_ac, 33046ae23c55STao Ma 1, set_len, 33056f70fa51STao Ma &new_bit, &new_len); 33066f70fa51STao Ma if (ret) { 33076f70fa51STao Ma mlog_errno(ret); 33086f70fa51STao Ma goto out_commit; 33096f70fa51STao Ma } 33106f70fa51STao Ma 33116f70fa51STao Ma ret = ocfs2_replace_clusters(handle, context, 33126f70fa51STao Ma cpos, p_cluster, new_bit, 33136f70fa51STao Ma new_len, e_flags); 33146f70fa51STao Ma if (ret) { 33156f70fa51STao Ma mlog_errno(ret); 33166f70fa51STao Ma goto out_commit; 33176f70fa51STao Ma } 33186ae23c55STao Ma set_len = new_len; 33196f70fa51STao Ma } 33206f70fa51STao Ma 3321913580b4STao Ma ret = __ocfs2_decrease_refcount(handle, ref_ci, 33226f70fa51STao Ma context->ref_root_bh, 33236ae23c55STao Ma p_cluster, set_len, 33246f70fa51STao Ma context->meta_ac, 33256ae23c55STao Ma &context->dealloc, delete); 33266f70fa51STao Ma if (ret) { 33276f70fa51STao Ma mlog_errno(ret); 33286f70fa51STao Ma goto out_commit; 33296f70fa51STao Ma } 33306f70fa51STao Ma 33316ae23c55STao Ma cpos += set_len; 33326ae23c55STao Ma p_cluster += set_len; 33336ae23c55STao Ma num_clusters -= set_len; 33346ae23c55STao Ma brelse(ref_leaf_bh); 33356ae23c55STao Ma ref_leaf_bh = NULL; 33366ae23c55STao Ma } 33376ae23c55STao Ma 3338492a8a33STao Ma /* handle any post_cow action. */ 3339492a8a33STao Ma if (context->post_refcount && context->post_refcount->func) { 3340492a8a33STao Ma ret = context->post_refcount->func(context->inode, handle, 3341492a8a33STao Ma context->post_refcount->para); 3342492a8a33STao Ma if (ret) { 3343492a8a33STao Ma mlog_errno(ret); 3344492a8a33STao Ma goto out_commit; 3345492a8a33STao Ma } 3346492a8a33STao Ma } 3347492a8a33STao Ma 33486f70fa51STao Ma /* 33496f70fa51STao Ma * Here we should write the new page out first if we are 33506f70fa51STao Ma * in write-back mode. 33516f70fa51STao Ma */ 3352492a8a33STao Ma if (context->get_clusters == ocfs2_di_get_clusters) { 3353acf3bb00STristan Ye ret = ocfs2_cow_sync_writeback(sb, context, cpos, 3354acf3bb00STristan Ye orig_num_clusters); 33556f70fa51STao Ma if (ret) 33566f70fa51STao Ma mlog_errno(ret); 3357492a8a33STao Ma } 33586f70fa51STao Ma 33596f70fa51STao Ma out_commit: 33606f70fa51STao Ma ocfs2_commit_trans(osb, handle); 33616f70fa51STao Ma 33626f70fa51STao Ma out: 33636f70fa51STao Ma if (context->data_ac) { 33646f70fa51STao Ma ocfs2_free_alloc_context(context->data_ac); 33656f70fa51STao Ma context->data_ac = NULL; 33666f70fa51STao Ma } 33676f70fa51STao Ma if (context->meta_ac) { 33686f70fa51STao Ma ocfs2_free_alloc_context(context->meta_ac); 33696f70fa51STao Ma context->meta_ac = NULL; 33706f70fa51STao Ma } 33716ae23c55STao Ma brelse(ref_leaf_bh); 33726f70fa51STao Ma 33736f70fa51STao Ma return ret; 33746f70fa51STao Ma } 33756f70fa51STao Ma 3376913580b4STao Ma static int ocfs2_replace_cow(struct ocfs2_cow_context *context) 33776f70fa51STao Ma { 33786f70fa51STao Ma int ret = 0; 3379913580b4STao Ma struct inode *inode = context->inode; 3380913580b4STao Ma u32 cow_start = context->cow_start, cow_len = context->cow_len; 3381913580b4STao Ma u32 p_cluster, num_clusters; 33826f70fa51STao Ma unsigned int ext_flags; 33836f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 33846f70fa51STao Ma 33856f70fa51STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 33866f70fa51STao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 33876f70fa51STao Ma "tree, but the feature bit is not set in the " 33886f70fa51STao Ma "super block.", inode->i_ino); 33896f70fa51STao Ma return -EROFS; 33906f70fa51STao Ma } 33916f70fa51STao Ma 33926f70fa51STao Ma ocfs2_init_dealloc_ctxt(&context->dealloc); 33936f70fa51STao Ma 33946f70fa51STao Ma while (cow_len) { 3395913580b4STao Ma ret = context->get_clusters(context, cow_start, &p_cluster, 33966f70fa51STao Ma &num_clusters, &ext_flags); 33976f70fa51STao Ma if (ret) { 33986f70fa51STao Ma mlog_errno(ret); 33996f70fa51STao Ma break; 34006f70fa51STao Ma } 34016f70fa51STao Ma 34026f70fa51STao Ma BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED)); 34036f70fa51STao Ma 34046f70fa51STao Ma if (cow_len < num_clusters) 34056f70fa51STao Ma num_clusters = cow_len; 34066f70fa51STao Ma 34076f70fa51STao Ma ret = ocfs2_make_clusters_writable(inode->i_sb, context, 34086f70fa51STao Ma cow_start, p_cluster, 34096f70fa51STao Ma num_clusters, ext_flags); 34106f70fa51STao Ma if (ret) { 34116f70fa51STao Ma mlog_errno(ret); 34126f70fa51STao Ma break; 34136f70fa51STao Ma } 34146f70fa51STao Ma 34156f70fa51STao Ma cow_len -= num_clusters; 34166f70fa51STao Ma cow_start += num_clusters; 34176f70fa51STao Ma } 34186f70fa51STao Ma 34196f70fa51STao Ma if (ocfs2_dealloc_has_cluster(&context->dealloc)) { 34206f70fa51STao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 34216f70fa51STao Ma ocfs2_run_deallocs(osb, &context->dealloc); 34226f70fa51STao Ma } 34236f70fa51STao Ma 34246f70fa51STao Ma return ret; 34256f70fa51STao Ma } 34266f70fa51STao Ma 34277b61cf54STao Ma static void ocfs2_readahead_for_cow(struct inode *inode, 34287b61cf54STao Ma struct file *file, 34297b61cf54STao Ma u32 start, u32 len) 34307b61cf54STao Ma { 34317b61cf54STao Ma struct address_space *mapping; 34327b61cf54STao Ma pgoff_t index; 34337b61cf54STao Ma unsigned long num_pages; 34347b61cf54STao Ma int cs_bits = OCFS2_SB(inode->i_sb)->s_clustersize_bits; 34357b61cf54STao Ma 34367b61cf54STao Ma if (!file) 34377b61cf54STao Ma return; 34387b61cf54STao Ma 34397b61cf54STao Ma mapping = file->f_mapping; 34407b61cf54STao Ma num_pages = (len << cs_bits) >> PAGE_CACHE_SHIFT; 34417b61cf54STao Ma if (!num_pages) 34427b61cf54STao Ma num_pages = 1; 34437b61cf54STao Ma 34447b61cf54STao Ma index = ((loff_t)start << cs_bits) >> PAGE_CACHE_SHIFT; 34457b61cf54STao Ma page_cache_sync_readahead(mapping, &file->f_ra, file, 34467b61cf54STao Ma index, num_pages); 34477b61cf54STao Ma } 34487b61cf54STao Ma 34496f70fa51STao Ma /* 345037f8a2bfSTao Ma * Starting at cpos, try to CoW write_len clusters. Don't CoW 345137f8a2bfSTao Ma * past max_cpos. This will stop when it runs into a hole or an 345237f8a2bfSTao Ma * unrefcounted extent. 34536f70fa51STao Ma */ 34546f70fa51STao Ma static int ocfs2_refcount_cow_hunk(struct inode *inode, 345515502712STao Ma struct file *file, 34566f70fa51STao Ma struct buffer_head *di_bh, 345737f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 34586f70fa51STao Ma { 34596f70fa51STao Ma int ret; 34606f70fa51STao Ma u32 cow_start = 0, cow_len = 0; 34616f70fa51STao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 34626f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 34636f70fa51STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 34646f70fa51STao Ma struct buffer_head *ref_root_bh = NULL; 34656f70fa51STao Ma struct ocfs2_refcount_tree *ref_tree; 3466913580b4STao Ma struct ocfs2_cow_context *context = NULL; 34676f70fa51STao Ma 34686f70fa51STao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 34696f70fa51STao Ma 3470913580b4STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list, 347137f8a2bfSTao Ma cpos, write_len, max_cpos, 34726f70fa51STao Ma &cow_start, &cow_len); 34736f70fa51STao Ma if (ret) { 34746f70fa51STao Ma mlog_errno(ret); 34756f70fa51STao Ma goto out; 34766f70fa51STao Ma } 347737f8a2bfSTao Ma 34786f70fa51STao Ma mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, " 34796f70fa51STao Ma "cow_len %u\n", inode->i_ino, 34806f70fa51STao Ma cpos, write_len, cow_start, cow_len); 34816f70fa51STao Ma 34826f70fa51STao Ma BUG_ON(cow_len == 0); 34836f70fa51STao Ma 34847b61cf54STao Ma ocfs2_readahead_for_cow(inode, file, cow_start, cow_len); 34857b61cf54STao Ma 3486913580b4STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3487913580b4STao Ma if (!context) { 3488913580b4STao Ma ret = -ENOMEM; 3489913580b4STao Ma mlog_errno(ret); 3490913580b4STao Ma goto out; 3491913580b4STao Ma } 3492913580b4STao Ma 34936f70fa51STao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 34946f70fa51STao Ma 1, &ref_tree, &ref_root_bh); 34956f70fa51STao Ma if (ret) { 34966f70fa51STao Ma mlog_errno(ret); 34976f70fa51STao Ma goto out; 34986f70fa51STao Ma } 34996f70fa51STao Ma 3500913580b4STao Ma context->inode = inode; 3501913580b4STao Ma context->cow_start = cow_start; 3502913580b4STao Ma context->cow_len = cow_len; 3503913580b4STao Ma context->ref_tree = ref_tree; 3504913580b4STao Ma context->ref_root_bh = ref_root_bh; 3505913580b4STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page; 3506913580b4STao Ma context->get_clusters = ocfs2_di_get_clusters; 35076ea4843fSTao Ma context->file = file; 3508913580b4STao Ma 3509913580b4STao Ma ocfs2_init_dinode_extent_tree(&context->data_et, 3510913580b4STao Ma INODE_CACHE(inode), di_bh); 3511913580b4STao Ma 3512913580b4STao Ma ret = ocfs2_replace_cow(context); 35136f70fa51STao Ma if (ret) 35146f70fa51STao Ma mlog_errno(ret); 35156f70fa51STao Ma 3516913580b4STao Ma /* 3517913580b4STao Ma * truncate the extent map here since no matter whether we meet with 3518913580b4STao Ma * any error during the action, we shouldn't trust cached extent map 3519913580b4STao Ma * any more. 3520913580b4STao Ma */ 3521913580b4STao Ma ocfs2_extent_map_trunc(inode, cow_start); 3522913580b4STao Ma 35236f70fa51STao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 35246f70fa51STao Ma brelse(ref_root_bh); 35256f70fa51STao Ma out: 3526913580b4STao Ma kfree(context); 35276f70fa51STao Ma return ret; 35286f70fa51STao Ma } 35296f70fa51STao Ma 35306f70fa51STao Ma /* 35316f70fa51STao Ma * CoW any and all clusters between cpos and cpos+write_len. 353237f8a2bfSTao Ma * Don't CoW past max_cpos. If this returns successfully, all 353337f8a2bfSTao Ma * clusters between cpos and cpos+write_len are safe to modify. 35346f70fa51STao Ma */ 35356f70fa51STao Ma int ocfs2_refcount_cow(struct inode *inode, 353615502712STao Ma struct file *file, 35376f70fa51STao Ma struct buffer_head *di_bh, 353837f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 35396f70fa51STao Ma { 35406f70fa51STao Ma int ret = 0; 35416f70fa51STao Ma u32 p_cluster, num_clusters; 35426f70fa51STao Ma unsigned int ext_flags; 35436f70fa51STao Ma 35446f70fa51STao Ma while (write_len) { 35456f70fa51STao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 35466f70fa51STao Ma &num_clusters, &ext_flags); 35476f70fa51STao Ma if (ret) { 35486f70fa51STao Ma mlog_errno(ret); 35496f70fa51STao Ma break; 35506f70fa51STao Ma } 35516f70fa51STao Ma 35526f70fa51STao Ma if (write_len < num_clusters) 35536f70fa51STao Ma num_clusters = write_len; 35546f70fa51STao Ma 35556f70fa51STao Ma if (ext_flags & OCFS2_EXT_REFCOUNTED) { 355615502712STao Ma ret = ocfs2_refcount_cow_hunk(inode, file, di_bh, cpos, 355737f8a2bfSTao Ma num_clusters, max_cpos); 35586f70fa51STao Ma if (ret) { 35596f70fa51STao Ma mlog_errno(ret); 35606f70fa51STao Ma break; 35616f70fa51STao Ma } 35626f70fa51STao Ma } 35636f70fa51STao Ma 35646f70fa51STao Ma write_len -= num_clusters; 35656f70fa51STao Ma cpos += num_clusters; 35666f70fa51STao Ma } 35676f70fa51STao Ma 35686f70fa51STao Ma return ret; 35696f70fa51STao Ma } 3570110a045aSTao Ma 3571492a8a33STao Ma static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context, 3572492a8a33STao Ma u32 v_cluster, u32 *p_cluster, 3573492a8a33STao Ma u32 *num_clusters, 3574492a8a33STao Ma unsigned int *extent_flags) 3575492a8a33STao Ma { 3576492a8a33STao Ma struct inode *inode = context->inode; 3577492a8a33STao Ma struct ocfs2_xattr_value_root *xv = context->cow_object; 3578492a8a33STao Ma 3579492a8a33STao Ma return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster, 3580492a8a33STao Ma num_clusters, &xv->xr_list, 3581492a8a33STao Ma extent_flags); 3582492a8a33STao Ma } 3583492a8a33STao Ma 3584492a8a33STao Ma /* 3585492a8a33STao Ma * Given a xattr value root, calculate the most meta/credits we need for 3586492a8a33STao Ma * refcount tree change if we truncate it to 0. 3587492a8a33STao Ma */ 3588492a8a33STao Ma int ocfs2_refcounted_xattr_delete_need(struct inode *inode, 3589492a8a33STao Ma struct ocfs2_caching_info *ref_ci, 3590492a8a33STao Ma struct buffer_head *ref_root_bh, 3591492a8a33STao Ma struct ocfs2_xattr_value_root *xv, 3592492a8a33STao Ma int *meta_add, int *credits) 3593492a8a33STao Ma { 3594492a8a33STao Ma int ret = 0, index, ref_blocks = 0; 3595492a8a33STao Ma u32 p_cluster, num_clusters; 3596492a8a33STao Ma u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters); 3597492a8a33STao Ma struct ocfs2_refcount_block *rb; 3598492a8a33STao Ma struct ocfs2_refcount_rec rec; 3599492a8a33STao Ma struct buffer_head *ref_leaf_bh = NULL; 3600492a8a33STao Ma 3601492a8a33STao Ma while (cpos < clusters) { 3602492a8a33STao Ma ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster, 3603492a8a33STao Ma &num_clusters, &xv->xr_list, 3604492a8a33STao Ma NULL); 3605492a8a33STao Ma if (ret) { 3606492a8a33STao Ma mlog_errno(ret); 3607492a8a33STao Ma goto out; 3608492a8a33STao Ma } 3609492a8a33STao Ma 3610492a8a33STao Ma cpos += num_clusters; 3611492a8a33STao Ma 3612492a8a33STao Ma while (num_clusters) { 3613492a8a33STao Ma ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh, 3614492a8a33STao Ma p_cluster, num_clusters, 3615492a8a33STao Ma &rec, &index, 3616492a8a33STao Ma &ref_leaf_bh); 3617492a8a33STao Ma if (ret) { 3618492a8a33STao Ma mlog_errno(ret); 3619492a8a33STao Ma goto out; 3620492a8a33STao Ma } 3621492a8a33STao Ma 3622492a8a33STao Ma BUG_ON(!rec.r_refcount); 3623492a8a33STao Ma 3624492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 3625492a8a33STao Ma 3626492a8a33STao Ma /* 3627492a8a33STao Ma * We really don't know whether the other clusters is in 3628492a8a33STao Ma * this refcount block or not, so just take the worst 3629492a8a33STao Ma * case that all the clusters are in this block and each 3630492a8a33STao Ma * one will split a refcount rec, so totally we need 3631492a8a33STao Ma * clusters * 2 new refcount rec. 3632492a8a33STao Ma */ 3633492a8a33STao Ma if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 > 3634492a8a33STao Ma le16_to_cpu(rb->rf_records.rl_count)) 3635492a8a33STao Ma ref_blocks++; 3636492a8a33STao Ma 3637492a8a33STao Ma *credits += 1; 3638492a8a33STao Ma brelse(ref_leaf_bh); 3639492a8a33STao Ma ref_leaf_bh = NULL; 3640492a8a33STao Ma 3641492a8a33STao Ma if (num_clusters <= le32_to_cpu(rec.r_clusters)) 3642492a8a33STao Ma break; 3643492a8a33STao Ma else 3644492a8a33STao Ma num_clusters -= le32_to_cpu(rec.r_clusters); 3645492a8a33STao Ma p_cluster += num_clusters; 3646492a8a33STao Ma } 3647492a8a33STao Ma } 3648492a8a33STao Ma 3649492a8a33STao Ma *meta_add += ref_blocks; 3650492a8a33STao Ma if (!ref_blocks) 3651492a8a33STao Ma goto out; 3652492a8a33STao Ma 3653492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 3654492a8a33STao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) 3655492a8a33STao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 3656492a8a33STao Ma else { 3657492a8a33STao Ma struct ocfs2_extent_tree et; 3658492a8a33STao Ma 3659492a8a33STao Ma ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh); 3660492a8a33STao Ma *credits += ocfs2_calc_extend_credits(inode->i_sb, 3661492a8a33STao Ma et.et_root_el, 3662492a8a33STao Ma ref_blocks); 3663492a8a33STao Ma } 3664492a8a33STao Ma 3665492a8a33STao Ma out: 3666492a8a33STao Ma brelse(ref_leaf_bh); 3667492a8a33STao Ma return ret; 3668492a8a33STao Ma } 3669492a8a33STao Ma 3670492a8a33STao Ma /* 3671492a8a33STao Ma * Do CoW for xattr. 3672492a8a33STao Ma */ 3673492a8a33STao Ma int ocfs2_refcount_cow_xattr(struct inode *inode, 3674492a8a33STao Ma struct ocfs2_dinode *di, 3675492a8a33STao Ma struct ocfs2_xattr_value_buf *vb, 3676492a8a33STao Ma struct ocfs2_refcount_tree *ref_tree, 3677492a8a33STao Ma struct buffer_head *ref_root_bh, 3678492a8a33STao Ma u32 cpos, u32 write_len, 3679492a8a33STao Ma struct ocfs2_post_refcount *post) 3680492a8a33STao Ma { 3681492a8a33STao Ma int ret; 3682492a8a33STao Ma struct ocfs2_xattr_value_root *xv = vb->vb_xv; 3683492a8a33STao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 3684492a8a33STao Ma struct ocfs2_cow_context *context = NULL; 3685492a8a33STao Ma u32 cow_start, cow_len; 3686492a8a33STao Ma 3687492a8a33STao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 3688492a8a33STao Ma 3689492a8a33STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list, 3690492a8a33STao Ma cpos, write_len, UINT_MAX, 3691492a8a33STao Ma &cow_start, &cow_len); 3692492a8a33STao Ma if (ret) { 3693492a8a33STao Ma mlog_errno(ret); 3694492a8a33STao Ma goto out; 3695492a8a33STao Ma } 3696492a8a33STao Ma 3697492a8a33STao Ma BUG_ON(cow_len == 0); 3698492a8a33STao Ma 3699492a8a33STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3700492a8a33STao Ma if (!context) { 3701492a8a33STao Ma ret = -ENOMEM; 3702492a8a33STao Ma mlog_errno(ret); 3703492a8a33STao Ma goto out; 3704492a8a33STao Ma } 3705492a8a33STao Ma 3706492a8a33STao Ma context->inode = inode; 3707492a8a33STao Ma context->cow_start = cow_start; 3708492a8a33STao Ma context->cow_len = cow_len; 3709492a8a33STao Ma context->ref_tree = ref_tree; 3710492a8a33STao Ma context->ref_root_bh = ref_root_bh;; 3711492a8a33STao Ma context->cow_object = xv; 3712492a8a33STao Ma 3713492a8a33STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd; 3714492a8a33STao Ma /* We need the extra credits for duplicate_clusters by jbd. */ 3715492a8a33STao Ma context->extra_credits = 3716492a8a33STao Ma ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len; 3717492a8a33STao Ma context->get_clusters = ocfs2_xattr_value_get_clusters; 3718492a8a33STao Ma context->post_refcount = post; 3719492a8a33STao Ma 3720492a8a33STao Ma ocfs2_init_xattr_value_extent_tree(&context->data_et, 3721492a8a33STao Ma INODE_CACHE(inode), vb); 3722492a8a33STao Ma 3723492a8a33STao Ma ret = ocfs2_replace_cow(context); 3724492a8a33STao Ma if (ret) 3725492a8a33STao Ma mlog_errno(ret); 3726492a8a33STao Ma 3727492a8a33STao Ma out: 3728492a8a33STao Ma kfree(context); 3729492a8a33STao Ma return ret; 3730492a8a33STao Ma } 3731492a8a33STao Ma 3732110a045aSTao Ma /* 3733110a045aSTao Ma * Insert a new extent into refcount tree and mark a extent rec 3734110a045aSTao Ma * as refcounted in the dinode tree. 3735110a045aSTao Ma */ 3736110a045aSTao Ma int ocfs2_add_refcount_flag(struct inode *inode, 3737110a045aSTao Ma struct ocfs2_extent_tree *data_et, 3738110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3739110a045aSTao Ma struct buffer_head *ref_root_bh, 3740110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 37410129241eSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 37420129241eSTao Ma struct ocfs2_post_refcount *post) 3743110a045aSTao Ma { 3744110a045aSTao Ma int ret; 3745110a045aSTao Ma handle_t *handle; 3746110a045aSTao Ma int credits = 1, ref_blocks = 0; 3747110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3748110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3749110a045aSTao Ma 3750110a045aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 3751110a045aSTao Ma ref_ci, ref_root_bh, 3752110a045aSTao Ma p_cluster, num_clusters, 3753110a045aSTao Ma &ref_blocks, &credits); 3754110a045aSTao Ma if (ret) { 3755110a045aSTao Ma mlog_errno(ret); 3756110a045aSTao Ma goto out; 3757110a045aSTao Ma } 3758110a045aSTao Ma 3759110a045aSTao Ma mlog(0, "reserve new metadata %d, credits = %d\n", 3760110a045aSTao Ma ref_blocks, credits); 3761110a045aSTao Ma 3762110a045aSTao Ma if (ref_blocks) { 3763110a045aSTao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb), 3764110a045aSTao Ma ref_blocks, &meta_ac); 3765110a045aSTao Ma if (ret) { 3766110a045aSTao Ma mlog_errno(ret); 3767110a045aSTao Ma goto out; 3768110a045aSTao Ma } 3769110a045aSTao Ma } 3770110a045aSTao Ma 37710129241eSTao Ma if (post) 37720129241eSTao Ma credits += post->credits; 37730129241eSTao Ma 3774110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3775110a045aSTao Ma if (IS_ERR(handle)) { 3776110a045aSTao Ma ret = PTR_ERR(handle); 3777110a045aSTao Ma mlog_errno(ret); 3778110a045aSTao Ma goto out; 3779110a045aSTao Ma } 3780110a045aSTao Ma 3781110a045aSTao Ma ret = ocfs2_mark_extent_refcounted(inode, data_et, handle, 3782110a045aSTao Ma cpos, num_clusters, p_cluster, 3783110a045aSTao Ma meta_ac, dealloc); 3784110a045aSTao Ma if (ret) { 3785110a045aSTao Ma mlog_errno(ret); 3786110a045aSTao Ma goto out_commit; 3787110a045aSTao Ma } 3788110a045aSTao Ma 37897540c1a7STao Ma ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 37907540c1a7STao Ma p_cluster, num_clusters, 0, 3791110a045aSTao Ma meta_ac, dealloc); 37920129241eSTao Ma if (ret) { 37930129241eSTao Ma mlog_errno(ret); 37940129241eSTao Ma goto out_commit; 37950129241eSTao Ma } 37960129241eSTao Ma 37970129241eSTao Ma if (post && post->func) { 37980129241eSTao Ma ret = post->func(inode, handle, post->para); 3799110a045aSTao Ma if (ret) 3800110a045aSTao Ma mlog_errno(ret); 38010129241eSTao Ma } 3802110a045aSTao Ma 3803110a045aSTao Ma out_commit: 3804110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3805110a045aSTao Ma out: 3806110a045aSTao Ma if (meta_ac) 3807110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3808110a045aSTao Ma return ret; 3809110a045aSTao Ma } 3810110a045aSTao Ma 3811a9063ab9STao Ma static int ocfs2_change_ctime(struct inode *inode, 3812110a045aSTao Ma struct buffer_head *di_bh) 3813110a045aSTao Ma { 3814110a045aSTao Ma int ret; 3815a9063ab9STao Ma handle_t *handle; 3816a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3817a9063ab9STao Ma 3818a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb), 3819a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 3820a9063ab9STao Ma if (IS_ERR(handle)) { 3821a9063ab9STao Ma ret = PTR_ERR(handle); 3822a9063ab9STao Ma mlog_errno(ret); 3823a9063ab9STao Ma goto out; 3824a9063ab9STao Ma } 3825a9063ab9STao Ma 3826a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 3827a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 3828a9063ab9STao Ma if (ret) { 3829a9063ab9STao Ma mlog_errno(ret); 3830a9063ab9STao Ma goto out_commit; 3831a9063ab9STao Ma } 3832a9063ab9STao Ma 3833a9063ab9STao Ma inode->i_ctime = CURRENT_TIME; 3834a9063ab9STao Ma di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec); 3835a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec); 3836a9063ab9STao Ma 3837a9063ab9STao Ma ocfs2_journal_dirty(handle, di_bh); 3838a9063ab9STao Ma 3839a9063ab9STao Ma out_commit: 3840a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle); 3841a9063ab9STao Ma out: 3842a9063ab9STao Ma return ret; 3843a9063ab9STao Ma } 3844a9063ab9STao Ma 3845a9063ab9STao Ma static int ocfs2_attach_refcount_tree(struct inode *inode, 3846a9063ab9STao Ma struct buffer_head *di_bh) 3847a9063ab9STao Ma { 3848a9063ab9STao Ma int ret, data_changed = 0; 3849110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 3850110a045aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 3851110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3852110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3853110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 3854110a045aSTao Ma unsigned int ext_flags; 3855110a045aSTao Ma loff_t size; 3856110a045aSTao Ma u32 cpos, num_clusters, clusters, p_cluster; 3857110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 3858110a045aSTao Ma struct ocfs2_extent_tree di_et; 3859110a045aSTao Ma 3860110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 3861110a045aSTao Ma 3862110a045aSTao Ma if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) { 3863110a045aSTao Ma ret = ocfs2_create_refcount_tree(inode, di_bh); 3864110a045aSTao Ma if (ret) { 3865110a045aSTao Ma mlog_errno(ret); 3866110a045aSTao Ma goto out; 3867110a045aSTao Ma } 3868110a045aSTao Ma } 3869110a045aSTao Ma 3870110a045aSTao Ma BUG_ON(!di->i_refcount_loc); 3871110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, 3872110a045aSTao Ma le64_to_cpu(di->i_refcount_loc), 1, 3873110a045aSTao Ma &ref_tree, &ref_root_bh); 3874110a045aSTao Ma if (ret) { 3875110a045aSTao Ma mlog_errno(ret); 3876110a045aSTao Ma goto out; 3877110a045aSTao Ma } 3878110a045aSTao Ma 38792f48d593STao Ma if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) 38802f48d593STao Ma goto attach_xattr; 38812f48d593STao Ma 3882110a045aSTao Ma ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh); 3883110a045aSTao Ma 3884110a045aSTao Ma size = i_size_read(inode); 3885110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(inode->i_sb, size); 3886110a045aSTao Ma 3887110a045aSTao Ma cpos = 0; 3888110a045aSTao Ma while (cpos < clusters) { 3889110a045aSTao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 3890110a045aSTao Ma &num_clusters, &ext_flags); 3891110a045aSTao Ma 3892110a045aSTao Ma if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) { 3893110a045aSTao Ma ret = ocfs2_add_refcount_flag(inode, &di_et, 3894110a045aSTao Ma &ref_tree->rf_ci, 3895110a045aSTao Ma ref_root_bh, cpos, 3896110a045aSTao Ma p_cluster, num_clusters, 38970129241eSTao Ma &dealloc, NULL); 3898110a045aSTao Ma if (ret) { 3899110a045aSTao Ma mlog_errno(ret); 3900a9063ab9STao Ma goto unlock; 3901110a045aSTao Ma } 3902a9063ab9STao Ma 3903a9063ab9STao Ma data_changed = 1; 3904110a045aSTao Ma } 3905110a045aSTao Ma cpos += num_clusters; 3906110a045aSTao Ma } 3907110a045aSTao Ma 39082f48d593STao Ma attach_xattr: 39090129241eSTao Ma if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 39100129241eSTao Ma ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh, 39110129241eSTao Ma &ref_tree->rf_ci, 39120129241eSTao Ma ref_root_bh, 39130129241eSTao Ma &dealloc); 39140129241eSTao Ma if (ret) { 39150129241eSTao Ma mlog_errno(ret); 39160129241eSTao Ma goto unlock; 39170129241eSTao Ma } 39180129241eSTao Ma } 39190129241eSTao Ma 3920a9063ab9STao Ma if (data_changed) { 3921a9063ab9STao Ma ret = ocfs2_change_ctime(inode, di_bh); 3922a9063ab9STao Ma if (ret) 3923a9063ab9STao Ma mlog_errno(ret); 3924a9063ab9STao Ma } 3925a9063ab9STao Ma 3926a9063ab9STao Ma unlock: 3927110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 3928110a045aSTao Ma brelse(ref_root_bh); 3929110a045aSTao Ma 3930110a045aSTao Ma if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) { 3931110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 3932110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 3933110a045aSTao Ma } 3934110a045aSTao Ma out: 3935110a045aSTao Ma /* 3936110a045aSTao Ma * Empty the extent map so that we may get the right extent 3937110a045aSTao Ma * record from the disk. 3938110a045aSTao Ma */ 3939110a045aSTao Ma ocfs2_extent_map_trunc(inode, 0); 3940110a045aSTao Ma 3941110a045aSTao Ma return ret; 3942110a045aSTao Ma } 3943110a045aSTao Ma 3944110a045aSTao Ma static int ocfs2_add_refcounted_extent(struct inode *inode, 3945110a045aSTao Ma struct ocfs2_extent_tree *et, 3946110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3947110a045aSTao Ma struct buffer_head *ref_root_bh, 3948110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 3949110a045aSTao Ma unsigned int ext_flags, 3950110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 3951110a045aSTao Ma { 3952110a045aSTao Ma int ret; 3953110a045aSTao Ma handle_t *handle; 3954110a045aSTao Ma int credits = 0; 3955110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3956110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3957110a045aSTao Ma 3958110a045aSTao Ma ret = ocfs2_lock_refcount_allocators(inode->i_sb, 3959110a045aSTao Ma p_cluster, num_clusters, 3960110a045aSTao Ma et, ref_ci, 3961110a045aSTao Ma ref_root_bh, &meta_ac, 3962110a045aSTao Ma NULL, &credits); 3963110a045aSTao Ma if (ret) { 3964110a045aSTao Ma mlog_errno(ret); 3965110a045aSTao Ma goto out; 3966110a045aSTao Ma } 3967110a045aSTao Ma 3968110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3969110a045aSTao Ma if (IS_ERR(handle)) { 3970110a045aSTao Ma ret = PTR_ERR(handle); 3971110a045aSTao Ma mlog_errno(ret); 3972110a045aSTao Ma goto out; 3973110a045aSTao Ma } 3974110a045aSTao Ma 3975110a045aSTao Ma ret = ocfs2_insert_extent(handle, et, cpos, 397612d4cec9STao Ma ocfs2_clusters_to_blocks(inode->i_sb, p_cluster), 3977110a045aSTao Ma num_clusters, ext_flags, meta_ac); 3978110a045aSTao Ma if (ret) { 3979110a045aSTao Ma mlog_errno(ret); 3980110a045aSTao Ma goto out_commit; 3981110a045aSTao Ma } 3982110a045aSTao Ma 39832999d12fSTao Ma ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 3984110a045aSTao Ma p_cluster, num_clusters, 3985110a045aSTao Ma meta_ac, dealloc); 3986110a045aSTao Ma if (ret) 3987110a045aSTao Ma mlog_errno(ret); 3988110a045aSTao Ma 3989110a045aSTao Ma out_commit: 3990110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3991110a045aSTao Ma out: 3992110a045aSTao Ma if (meta_ac) 3993110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3994110a045aSTao Ma return ret; 3995110a045aSTao Ma } 3996110a045aSTao Ma 39972f48d593STao Ma static int ocfs2_duplicate_inline_data(struct inode *s_inode, 39982f48d593STao Ma struct buffer_head *s_bh, 39992f48d593STao Ma struct inode *t_inode, 40002f48d593STao Ma struct buffer_head *t_bh) 40012f48d593STao Ma { 40022f48d593STao Ma int ret; 40032f48d593STao Ma handle_t *handle; 40042f48d593STao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb); 40052f48d593STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data; 40062f48d593STao Ma struct ocfs2_dinode *t_di = (struct ocfs2_dinode *)t_bh->b_data; 40072f48d593STao Ma 40082f48d593STao Ma BUG_ON(!(OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)); 40092f48d593STao Ma 40102f48d593STao Ma handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 40112f48d593STao Ma if (IS_ERR(handle)) { 40122f48d593STao Ma ret = PTR_ERR(handle); 40132f48d593STao Ma mlog_errno(ret); 40142f48d593STao Ma goto out; 40152f48d593STao Ma } 40162f48d593STao Ma 40172f48d593STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh, 40182f48d593STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 40192f48d593STao Ma if (ret) { 40202f48d593STao Ma mlog_errno(ret); 40212f48d593STao Ma goto out_commit; 40222f48d593STao Ma } 40232f48d593STao Ma 40242f48d593STao Ma t_di->id2.i_data.id_count = s_di->id2.i_data.id_count; 40252f48d593STao Ma memcpy(t_di->id2.i_data.id_data, s_di->id2.i_data.id_data, 40262f48d593STao Ma le16_to_cpu(s_di->id2.i_data.id_count)); 40272f48d593STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock); 40282f48d593STao Ma OCFS2_I(t_inode)->ip_dyn_features |= OCFS2_INLINE_DATA_FL; 40292f48d593STao Ma t_di->i_dyn_features = cpu_to_le16(OCFS2_I(t_inode)->ip_dyn_features); 40302f48d593STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock); 40312f48d593STao Ma 40322f48d593STao Ma ocfs2_journal_dirty(handle, t_bh); 40332f48d593STao Ma 40342f48d593STao Ma out_commit: 40352f48d593STao Ma ocfs2_commit_trans(osb, handle); 40362f48d593STao Ma out: 40372f48d593STao Ma return ret; 40382f48d593STao Ma } 40392f48d593STao Ma 4040110a045aSTao Ma static int ocfs2_duplicate_extent_list(struct inode *s_inode, 4041110a045aSTao Ma struct inode *t_inode, 4042110a045aSTao Ma struct buffer_head *t_bh, 4043110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 4044110a045aSTao Ma struct buffer_head *ref_root_bh, 4045110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 4046110a045aSTao Ma { 4047110a045aSTao Ma int ret = 0; 4048110a045aSTao Ma u32 p_cluster, num_clusters, clusters, cpos; 4049110a045aSTao Ma loff_t size; 4050110a045aSTao Ma unsigned int ext_flags; 4051110a045aSTao Ma struct ocfs2_extent_tree et; 4052110a045aSTao Ma 4053110a045aSTao Ma ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh); 4054110a045aSTao Ma 4055110a045aSTao Ma size = i_size_read(s_inode); 4056110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size); 4057110a045aSTao Ma 4058110a045aSTao Ma cpos = 0; 4059110a045aSTao Ma while (cpos < clusters) { 4060110a045aSTao Ma ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster, 4061110a045aSTao Ma &num_clusters, &ext_flags); 4062110a045aSTao Ma 4063110a045aSTao Ma if (p_cluster) { 4064110a045aSTao Ma ret = ocfs2_add_refcounted_extent(t_inode, &et, 4065110a045aSTao Ma ref_ci, ref_root_bh, 4066110a045aSTao Ma cpos, p_cluster, 4067110a045aSTao Ma num_clusters, 4068110a045aSTao Ma ext_flags, 4069110a045aSTao Ma dealloc); 4070110a045aSTao Ma if (ret) { 4071110a045aSTao Ma mlog_errno(ret); 4072110a045aSTao Ma goto out; 4073110a045aSTao Ma } 4074110a045aSTao Ma } 4075110a045aSTao Ma 4076110a045aSTao Ma cpos += num_clusters; 4077110a045aSTao Ma } 4078110a045aSTao Ma 4079110a045aSTao Ma out: 4080110a045aSTao Ma return ret; 4081110a045aSTao Ma } 4082110a045aSTao Ma 4083a9063ab9STao Ma /* 4084a9063ab9STao Ma * change the new file's attributes to the src. 4085a9063ab9STao Ma * 4086a9063ab9STao Ma * reflink creates a snapshot of a file, that means the attributes 4087a9063ab9STao Ma * must be identical except for three exceptions - nlink, ino, and ctime. 4088a9063ab9STao Ma */ 4089a9063ab9STao Ma static int ocfs2_complete_reflink(struct inode *s_inode, 4090a9063ab9STao Ma struct buffer_head *s_bh, 4091a9063ab9STao Ma struct inode *t_inode, 40920fe9b66cSTao Ma struct buffer_head *t_bh, 40930fe9b66cSTao Ma bool preserve) 4094a9063ab9STao Ma { 4095a9063ab9STao Ma int ret; 4096a9063ab9STao Ma handle_t *handle; 4097a9063ab9STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data; 4098a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data; 4099a9063ab9STao Ma loff_t size = i_size_read(s_inode); 4100a9063ab9STao Ma 4101a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb), 4102a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 4103a9063ab9STao Ma if (IS_ERR(handle)) { 4104a9063ab9STao Ma ret = PTR_ERR(handle); 4105a9063ab9STao Ma mlog_errno(ret); 4106a9063ab9STao Ma return ret; 4107a9063ab9STao Ma } 4108a9063ab9STao Ma 4109a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh, 4110a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 4111a9063ab9STao Ma if (ret) { 4112a9063ab9STao Ma mlog_errno(ret); 4113a9063ab9STao Ma goto out_commit; 4114a9063ab9STao Ma } 4115a9063ab9STao Ma 4116a9063ab9STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock); 4117a9063ab9STao Ma OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters; 4118a9063ab9STao Ma OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr; 4119a9063ab9STao Ma OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features; 4120a9063ab9STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock); 4121a9063ab9STao Ma i_size_write(t_inode, size); 41226527f8f8STao Ma t_inode->i_blocks = s_inode->i_blocks; 4123a9063ab9STao Ma 4124a9063ab9STao Ma di->i_xattr_inline_size = s_di->i_xattr_inline_size; 4125a9063ab9STao Ma di->i_clusters = s_di->i_clusters; 4126a9063ab9STao Ma di->i_size = s_di->i_size; 4127a9063ab9STao Ma di->i_dyn_features = s_di->i_dyn_features; 4128a9063ab9STao Ma di->i_attr = s_di->i_attr; 41290fe9b66cSTao Ma 41300fe9b66cSTao Ma if (preserve) { 4131c21a534eSTao Ma t_inode->i_uid = s_inode->i_uid; 4132c21a534eSTao Ma t_inode->i_gid = s_inode->i_gid; 4133c21a534eSTao Ma t_inode->i_mode = s_inode->i_mode; 4134a9063ab9STao Ma di->i_uid = s_di->i_uid; 4135a9063ab9STao Ma di->i_gid = s_di->i_gid; 4136a9063ab9STao Ma di->i_mode = s_di->i_mode; 4137a9063ab9STao Ma 4138a9063ab9STao Ma /* 4139a9063ab9STao Ma * update time. 41400fe9b66cSTao Ma * we want mtime to appear identical to the source and 41410fe9b66cSTao Ma * update ctime. 4142a9063ab9STao Ma */ 4143a9063ab9STao Ma t_inode->i_ctime = CURRENT_TIME; 4144a9063ab9STao Ma 4145a9063ab9STao Ma di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec); 4146a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec); 4147a9063ab9STao Ma 4148a9063ab9STao Ma t_inode->i_mtime = s_inode->i_mtime; 4149a9063ab9STao Ma di->i_mtime = s_di->i_mtime; 4150a9063ab9STao Ma di->i_mtime_nsec = s_di->i_mtime_nsec; 41510fe9b66cSTao Ma } 4152a9063ab9STao Ma 4153a9063ab9STao Ma ocfs2_journal_dirty(handle, t_bh); 4154a9063ab9STao Ma 4155a9063ab9STao Ma out_commit: 4156a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle); 4157a9063ab9STao Ma return ret; 4158a9063ab9STao Ma } 4159a9063ab9STao Ma 4160110a045aSTao Ma static int ocfs2_create_reflink_node(struct inode *s_inode, 4161110a045aSTao Ma struct buffer_head *s_bh, 4162110a045aSTao Ma struct inode *t_inode, 41630fe9b66cSTao Ma struct buffer_head *t_bh, 41640fe9b66cSTao Ma bool preserve) 4165110a045aSTao Ma { 4166110a045aSTao Ma int ret; 4167110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 4168110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 4169110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb); 4170110a045aSTao Ma struct ocfs2_refcount_block *rb; 4171110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data; 4172110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 4173110a045aSTao Ma 4174110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 4175110a045aSTao Ma 4176110a045aSTao Ma ret = ocfs2_set_refcount_tree(t_inode, t_bh, 4177110a045aSTao Ma le64_to_cpu(di->i_refcount_loc)); 4178110a045aSTao Ma if (ret) { 4179110a045aSTao Ma mlog_errno(ret); 4180110a045aSTao Ma goto out; 4181110a045aSTao Ma } 4182110a045aSTao Ma 41832f48d593STao Ma if (OCFS2_I(s_inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) { 41842f48d593STao Ma ret = ocfs2_duplicate_inline_data(s_inode, s_bh, 41852f48d593STao Ma t_inode, t_bh); 41862f48d593STao Ma if (ret) 41872f48d593STao Ma mlog_errno(ret); 41882f48d593STao Ma goto out; 41892f48d593STao Ma } 41902f48d593STao Ma 4191110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 4192110a045aSTao Ma 1, &ref_tree, &ref_root_bh); 4193110a045aSTao Ma if (ret) { 4194110a045aSTao Ma mlog_errno(ret); 4195110a045aSTao Ma goto out; 4196110a045aSTao Ma } 4197110a045aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 4198110a045aSTao Ma 4199110a045aSTao Ma ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh, 4200110a045aSTao Ma &ref_tree->rf_ci, ref_root_bh, 4201110a045aSTao Ma &dealloc); 4202a9063ab9STao Ma if (ret) { 4203a9063ab9STao Ma mlog_errno(ret); 4204a9063ab9STao Ma goto out_unlock_refcount; 4205a9063ab9STao Ma } 4206a9063ab9STao Ma 4207a9063ab9STao Ma out_unlock_refcount: 4208110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 4209110a045aSTao Ma brelse(ref_root_bh); 4210110a045aSTao Ma out: 4211110a045aSTao Ma if (ocfs2_dealloc_has_cluster(&dealloc)) { 4212110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 4213110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 4214110a045aSTao Ma } 4215110a045aSTao Ma 4216110a045aSTao Ma return ret; 4217110a045aSTao Ma } 421809bf27a0STao Ma 421909bf27a0STao Ma static int __ocfs2_reflink(struct dentry *old_dentry, 422009bf27a0STao Ma struct buffer_head *old_bh, 422109bf27a0STao Ma struct inode *new_inode, 422209bf27a0STao Ma bool preserve) 422309bf27a0STao Ma { 422409bf27a0STao Ma int ret; 422509bf27a0STao Ma struct inode *inode = old_dentry->d_inode; 422609bf27a0STao Ma struct buffer_head *new_bh = NULL; 422709bf27a0STao Ma 42285693486bSJoel Becker if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) { 42295693486bSJoel Becker ret = -EINVAL; 42305693486bSJoel Becker mlog_errno(ret); 42315693486bSJoel Becker goto out; 42325693486bSJoel Becker } 42335693486bSJoel Becker 423409bf27a0STao Ma ret = filemap_fdatawrite(inode->i_mapping); 423509bf27a0STao Ma if (ret) { 423609bf27a0STao Ma mlog_errno(ret); 423709bf27a0STao Ma goto out; 423809bf27a0STao Ma } 423909bf27a0STao Ma 424009bf27a0STao Ma ret = ocfs2_attach_refcount_tree(inode, old_bh); 424109bf27a0STao Ma if (ret) { 424209bf27a0STao Ma mlog_errno(ret); 424309bf27a0STao Ma goto out; 424409bf27a0STao Ma } 424509bf27a0STao Ma 424607eaac94STao Ma mutex_lock_nested(&new_inode->i_mutex, I_MUTEX_CHILD); 424707eaac94STao Ma ret = ocfs2_inode_lock_nested(new_inode, &new_bh, 1, 424807eaac94STao Ma OI_LS_REFLINK_TARGET); 424909bf27a0STao Ma if (ret) { 425009bf27a0STao Ma mlog_errno(ret); 425109bf27a0STao Ma goto out_unlock; 425209bf27a0STao Ma } 425309bf27a0STao Ma 425409bf27a0STao Ma ret = ocfs2_create_reflink_node(inode, old_bh, 425509bf27a0STao Ma new_inode, new_bh, preserve); 425609bf27a0STao Ma if (ret) { 425709bf27a0STao Ma mlog_errno(ret); 425809bf27a0STao Ma goto inode_unlock; 425909bf27a0STao Ma } 426009bf27a0STao Ma 426109bf27a0STao Ma if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 426209bf27a0STao Ma ret = ocfs2_reflink_xattrs(inode, old_bh, 426309bf27a0STao Ma new_inode, new_bh, 426409bf27a0STao Ma preserve); 426587f4b1bbSTao Ma if (ret) { 426687f4b1bbSTao Ma mlog_errno(ret); 426787f4b1bbSTao Ma goto inode_unlock; 426887f4b1bbSTao Ma } 426987f4b1bbSTao Ma } 427087f4b1bbSTao Ma 427187f4b1bbSTao Ma ret = ocfs2_complete_reflink(inode, old_bh, 427287f4b1bbSTao Ma new_inode, new_bh, preserve); 427309bf27a0STao Ma if (ret) 427409bf27a0STao Ma mlog_errno(ret); 427587f4b1bbSTao Ma 427609bf27a0STao Ma inode_unlock: 427709bf27a0STao Ma ocfs2_inode_unlock(new_inode, 1); 427809bf27a0STao Ma brelse(new_bh); 427909bf27a0STao Ma out_unlock: 428009bf27a0STao Ma mutex_unlock(&new_inode->i_mutex); 428109bf27a0STao Ma out: 428209bf27a0STao Ma if (!ret) { 428309bf27a0STao Ma ret = filemap_fdatawait(inode->i_mapping); 428409bf27a0STao Ma if (ret) 428509bf27a0STao Ma mlog_errno(ret); 428609bf27a0STao Ma } 428709bf27a0STao Ma return ret; 428809bf27a0STao Ma } 428909bf27a0STao Ma 429009bf27a0STao Ma static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, 429109bf27a0STao Ma struct dentry *new_dentry, bool preserve) 429209bf27a0STao Ma { 429309bf27a0STao Ma int error; 429409bf27a0STao Ma struct inode *inode = old_dentry->d_inode; 429509bf27a0STao Ma struct buffer_head *old_bh = NULL; 429609bf27a0STao Ma struct inode *new_orphan_inode = NULL; 429709bf27a0STao Ma 429809bf27a0STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 429909bf27a0STao Ma return -EOPNOTSUPP; 430009bf27a0STao Ma 430109bf27a0STao Ma error = ocfs2_create_inode_in_orphan(dir, inode->i_mode, 430209bf27a0STao Ma &new_orphan_inode); 430309bf27a0STao Ma if (error) { 430409bf27a0STao Ma mlog_errno(error); 430509bf27a0STao Ma goto out; 430609bf27a0STao Ma } 430709bf27a0STao Ma 430809bf27a0STao Ma error = ocfs2_inode_lock(inode, &old_bh, 1); 430909bf27a0STao Ma if (error) { 431009bf27a0STao Ma mlog_errno(error); 431109bf27a0STao Ma goto out; 431209bf27a0STao Ma } 431309bf27a0STao Ma 431409bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_xattr_sem); 431509bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_alloc_sem); 431609bf27a0STao Ma error = __ocfs2_reflink(old_dentry, old_bh, 431709bf27a0STao Ma new_orphan_inode, preserve); 431809bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_alloc_sem); 431909bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_xattr_sem); 432009bf27a0STao Ma 432109bf27a0STao Ma ocfs2_inode_unlock(inode, 1); 432209bf27a0STao Ma brelse(old_bh); 432309bf27a0STao Ma 432409bf27a0STao Ma if (error) { 432509bf27a0STao Ma mlog_errno(error); 432609bf27a0STao Ma goto out; 432709bf27a0STao Ma } 432809bf27a0STao Ma 432909bf27a0STao Ma /* If the security isn't preserved, we need to re-initialize them. */ 433009bf27a0STao Ma if (!preserve) { 433109bf27a0STao Ma error = ocfs2_init_security_and_acl(dir, new_orphan_inode); 433209bf27a0STao Ma if (error) 433309bf27a0STao Ma mlog_errno(error); 433409bf27a0STao Ma } 433509bf27a0STao Ma out: 433609bf27a0STao Ma if (!error) { 433709bf27a0STao Ma error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, 433809bf27a0STao Ma new_dentry); 433909bf27a0STao Ma if (error) 434009bf27a0STao Ma mlog_errno(error); 434109bf27a0STao Ma } 434209bf27a0STao Ma 434309bf27a0STao Ma if (new_orphan_inode) { 434409bf27a0STao Ma /* 434509bf27a0STao Ma * We need to open_unlock the inode no matter whether we 434609bf27a0STao Ma * succeed or not, so that other nodes can delete it later. 434709bf27a0STao Ma */ 434809bf27a0STao Ma ocfs2_open_unlock(new_orphan_inode); 434909bf27a0STao Ma if (error) 435009bf27a0STao Ma iput(new_orphan_inode); 435109bf27a0STao Ma } 435209bf27a0STao Ma 435309bf27a0STao Ma return error; 435409bf27a0STao Ma } 4355bd50873dSTao Ma 4356bd50873dSTao Ma /* 4357bd50873dSTao Ma * Below here are the bits used by OCFS2_IOC_REFLINK() to fake 4358bd50873dSTao Ma * sys_reflink(). This will go away when vfs_reflink() exists in 4359bd50873dSTao Ma * fs/namei.c. 4360bd50873dSTao Ma */ 4361bd50873dSTao Ma 4362bd50873dSTao Ma /* copied from may_create in VFS. */ 4363bd50873dSTao Ma static inline int ocfs2_may_create(struct inode *dir, struct dentry *child) 4364bd50873dSTao Ma { 4365bd50873dSTao Ma if (child->d_inode) 4366bd50873dSTao Ma return -EEXIST; 4367bd50873dSTao Ma if (IS_DEADDIR(dir)) 4368bd50873dSTao Ma return -ENOENT; 4369bd50873dSTao Ma return inode_permission(dir, MAY_WRITE | MAY_EXEC); 4370bd50873dSTao Ma } 4371bd50873dSTao Ma 4372bd50873dSTao Ma /* copied from user_path_parent. */ 4373bd50873dSTao Ma static int ocfs2_user_path_parent(const char __user *path, 4374bd50873dSTao Ma struct nameidata *nd, char **name) 4375bd50873dSTao Ma { 4376bd50873dSTao Ma char *s = getname(path); 4377bd50873dSTao Ma int error; 4378bd50873dSTao Ma 4379bd50873dSTao Ma if (IS_ERR(s)) 4380bd50873dSTao Ma return PTR_ERR(s); 4381bd50873dSTao Ma 4382c9c6cac0SAl Viro error = kern_path_parent(s, nd); 4383bd50873dSTao Ma if (error) 4384bd50873dSTao Ma putname(s); 4385bd50873dSTao Ma else 4386bd50873dSTao Ma *name = s; 4387bd50873dSTao Ma 4388bd50873dSTao Ma return error; 4389bd50873dSTao Ma } 4390bd50873dSTao Ma 4391bd50873dSTao Ma /** 4392bd50873dSTao Ma * ocfs2_vfs_reflink - Create a reference-counted link 4393bd50873dSTao Ma * 4394bd50873dSTao Ma * @old_dentry: source dentry + inode 4395bd50873dSTao Ma * @dir: directory to create the target 4396bd50873dSTao Ma * @new_dentry: target dentry 4397bd50873dSTao Ma * @preserve: if true, preserve all file attributes 4398bd50873dSTao Ma */ 439912d4cec9STao Ma static int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir, 4400bd50873dSTao Ma struct dentry *new_dentry, bool preserve) 4401bd50873dSTao Ma { 4402bd50873dSTao Ma struct inode *inode = old_dentry->d_inode; 4403bd50873dSTao Ma int error; 4404bd50873dSTao Ma 4405bd50873dSTao Ma if (!inode) 4406bd50873dSTao Ma return -ENOENT; 4407bd50873dSTao Ma 4408bd50873dSTao Ma error = ocfs2_may_create(dir, new_dentry); 4409bd50873dSTao Ma if (error) 4410bd50873dSTao Ma return error; 4411bd50873dSTao Ma 4412bd50873dSTao Ma if (dir->i_sb != inode->i_sb) 4413bd50873dSTao Ma return -EXDEV; 4414bd50873dSTao Ma 4415bd50873dSTao Ma /* 4416bd50873dSTao Ma * A reflink to an append-only or immutable file cannot be created. 4417bd50873dSTao Ma */ 4418bd50873dSTao Ma if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) 4419bd50873dSTao Ma return -EPERM; 4420bd50873dSTao Ma 4421bd50873dSTao Ma /* Only regular files can be reflinked. */ 4422bd50873dSTao Ma if (!S_ISREG(inode->i_mode)) 4423bd50873dSTao Ma return -EPERM; 4424bd50873dSTao Ma 4425bd50873dSTao Ma /* 4426bd50873dSTao Ma * If the caller wants to preserve ownership, they require the 4427bd50873dSTao Ma * rights to do so. 4428bd50873dSTao Ma */ 4429bd50873dSTao Ma if (preserve) { 4430bd50873dSTao Ma if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN)) 4431bd50873dSTao Ma return -EPERM; 4432bd50873dSTao Ma if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN)) 4433bd50873dSTao Ma return -EPERM; 4434bd50873dSTao Ma } 4435bd50873dSTao Ma 4436bd50873dSTao Ma /* 4437bd50873dSTao Ma * If the caller is modifying any aspect of the attributes, they 4438bd50873dSTao Ma * are not creating a snapshot. They need read permission on the 4439bd50873dSTao Ma * file. 4440bd50873dSTao Ma */ 4441bd50873dSTao Ma if (!preserve) { 4442bd50873dSTao Ma error = inode_permission(inode, MAY_READ); 4443bd50873dSTao Ma if (error) 4444bd50873dSTao Ma return error; 4445bd50873dSTao Ma } 4446bd50873dSTao Ma 4447bd50873dSTao Ma mutex_lock(&inode->i_mutex); 4448871a2931SChristoph Hellwig dquot_initialize(dir); 4449bd50873dSTao Ma error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve); 4450bd50873dSTao Ma mutex_unlock(&inode->i_mutex); 4451bd50873dSTao Ma if (!error) 4452bd50873dSTao Ma fsnotify_create(dir, new_dentry); 4453bd50873dSTao Ma return error; 4454bd50873dSTao Ma } 4455bd50873dSTao Ma /* 4456bd50873dSTao Ma * Most codes are copied from sys_linkat. 4457bd50873dSTao Ma */ 4458bd50873dSTao Ma int ocfs2_reflink_ioctl(struct inode *inode, 4459bd50873dSTao Ma const char __user *oldname, 4460bd50873dSTao Ma const char __user *newname, 4461bd50873dSTao Ma bool preserve) 4462bd50873dSTao Ma { 4463bd50873dSTao Ma struct dentry *new_dentry; 4464bd50873dSTao Ma struct nameidata nd; 4465bd50873dSTao Ma struct path old_path; 4466bd50873dSTao Ma int error; 4467bd50873dSTao Ma char *to = NULL; 4468bd50873dSTao Ma 4469bd50873dSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 4470bd50873dSTao Ma return -EOPNOTSUPP; 4471bd50873dSTao Ma 4472bd50873dSTao Ma error = user_path_at(AT_FDCWD, oldname, 0, &old_path); 4473bd50873dSTao Ma if (error) { 4474bd50873dSTao Ma mlog_errno(error); 4475bd50873dSTao Ma return error; 4476bd50873dSTao Ma } 4477bd50873dSTao Ma 4478bd50873dSTao Ma error = ocfs2_user_path_parent(newname, &nd, &to); 4479bd50873dSTao Ma if (error) { 4480bd50873dSTao Ma mlog_errno(error); 4481bd50873dSTao Ma goto out; 4482bd50873dSTao Ma } 4483bd50873dSTao Ma 4484bd50873dSTao Ma error = -EXDEV; 4485bd50873dSTao Ma if (old_path.mnt != nd.path.mnt) 4486bd50873dSTao Ma goto out_release; 4487bd50873dSTao Ma new_dentry = lookup_create(&nd, 0); 4488bd50873dSTao Ma error = PTR_ERR(new_dentry); 4489bd50873dSTao Ma if (IS_ERR(new_dentry)) { 4490bd50873dSTao Ma mlog_errno(error); 4491bd50873dSTao Ma goto out_unlock; 4492bd50873dSTao Ma } 4493bd50873dSTao Ma 4494bd50873dSTao Ma error = mnt_want_write(nd.path.mnt); 4495bd50873dSTao Ma if (error) { 4496bd50873dSTao Ma mlog_errno(error); 4497bd50873dSTao Ma goto out_dput; 4498bd50873dSTao Ma } 4499bd50873dSTao Ma 4500bd50873dSTao Ma error = ocfs2_vfs_reflink(old_path.dentry, 4501bd50873dSTao Ma nd.path.dentry->d_inode, 4502bd50873dSTao Ma new_dentry, preserve); 4503bd50873dSTao Ma mnt_drop_write(nd.path.mnt); 4504bd50873dSTao Ma out_dput: 4505bd50873dSTao Ma dput(new_dentry); 4506bd50873dSTao Ma out_unlock: 4507bd50873dSTao Ma mutex_unlock(&nd.path.dentry->d_inode->i_mutex); 4508bd50873dSTao Ma out_release: 4509bd50873dSTao Ma path_put(&nd.path); 4510bd50873dSTao Ma putname(to); 4511bd50873dSTao Ma out: 4512bd50873dSTao Ma path_put(&old_path); 4513bd50873dSTao Ma 4514bd50873dSTao Ma return error; 4515bd50873dSTao Ma } 4516