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/gfp.h> 416f70fa51STao Ma #include <linux/slab.h> 426f70fa51STao Ma #include <linux/writeback.h> 436f70fa51STao Ma #include <linux/pagevec.h> 446f70fa51STao Ma #include <linux/swap.h> 45bd50873dSTao Ma #include <linux/security.h> 46bd50873dSTao Ma #include <linux/fsnotify.h> 47bd50873dSTao Ma #include <linux/quotaops.h> 48bd50873dSTao Ma #include <linux/namei.h> 49bd50873dSTao Ma #include <linux/mount.h> 506f70fa51STao Ma 516f70fa51STao Ma struct ocfs2_cow_context { 526f70fa51STao Ma struct inode *inode; 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 279374a263eSTao Ma 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 int ocfs2_lock_refcount_tree_by_inode(struct inode *inode, int rw, 528374a263eSTao Ma struct ocfs2_refcount_tree **ret_tree, 529374a263eSTao Ma struct buffer_head **ref_bh) 530374a263eSTao Ma { 531374a263eSTao Ma int ret; 532374a263eSTao Ma u64 ref_blkno; 533374a263eSTao Ma 534374a263eSTao Ma ret = ocfs2_get_refcount_block(inode, &ref_blkno); 535374a263eSTao Ma if (ret) { 536374a263eSTao Ma mlog_errno(ret); 537374a263eSTao Ma return ret; 538374a263eSTao Ma } 539374a263eSTao Ma 540374a263eSTao Ma return ocfs2_lock_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, 541374a263eSTao Ma rw, ret_tree, ref_bh); 542374a263eSTao Ma } 543374a263eSTao Ma 544374a263eSTao Ma void ocfs2_unlock_refcount_tree(struct ocfs2_super *osb, 545374a263eSTao Ma struct ocfs2_refcount_tree *tree, int rw) 546374a263eSTao Ma { 547374a263eSTao Ma if (rw) 548374a263eSTao Ma up_write(&tree->rf_sem); 549374a263eSTao Ma else 550374a263eSTao Ma up_read(&tree->rf_sem); 551374a263eSTao Ma 552374a263eSTao Ma ocfs2_refcount_unlock(tree, rw); 553374a263eSTao Ma ocfs2_refcount_tree_put(tree); 554374a263eSTao Ma } 555374a263eSTao Ma 556374a263eSTao Ma void ocfs2_purge_refcount_trees(struct ocfs2_super *osb) 557374a263eSTao Ma { 558374a263eSTao Ma struct rb_node *node; 559374a263eSTao Ma struct ocfs2_refcount_tree *tree; 560374a263eSTao Ma struct rb_root *root = &osb->osb_rf_lock_tree; 561374a263eSTao Ma 562374a263eSTao Ma while ((node = rb_last(root)) != NULL) { 563374a263eSTao Ma tree = rb_entry(node, struct ocfs2_refcount_tree, rf_node); 564374a263eSTao Ma 565374a263eSTao Ma mlog(0, "Purge tree %llu\n", 566374a263eSTao Ma (unsigned long long) tree->rf_blkno); 567374a263eSTao Ma 568374a263eSTao Ma rb_erase(&tree->rf_node, root); 569374a263eSTao Ma ocfs2_free_refcount_tree(tree); 570374a263eSTao Ma } 571374a263eSTao Ma } 5728bf396deSTao Ma 5738bf396deSTao Ma /* 5748bf396deSTao Ma * Create a refcount tree for an inode. 5758bf396deSTao Ma * We take for granted that the inode is already locked. 5768bf396deSTao Ma */ 5778bf396deSTao Ma static int ocfs2_create_refcount_tree(struct inode *inode, 5788bf396deSTao Ma struct buffer_head *di_bh) 5798bf396deSTao Ma { 5808bf396deSTao Ma int ret; 5818bf396deSTao Ma handle_t *handle = NULL; 5828bf396deSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 5838bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 5848bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 5858bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 5868bf396deSTao Ma struct buffer_head *new_bh = NULL; 5878bf396deSTao Ma struct ocfs2_refcount_block *rb; 5888bf396deSTao Ma struct ocfs2_refcount_tree *new_tree = NULL, *tree = NULL; 5898bf396deSTao Ma u16 suballoc_bit_start; 5908bf396deSTao Ma u32 num_got; 5918bf396deSTao Ma u64 first_blkno; 5928bf396deSTao Ma 5938bf396deSTao Ma BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL); 5948bf396deSTao Ma 5958bf396deSTao Ma mlog(0, "create tree for inode %lu\n", inode->i_ino); 5968bf396deSTao Ma 5978bf396deSTao Ma ret = ocfs2_reserve_new_metadata_blocks(osb, 1, &meta_ac); 5988bf396deSTao Ma if (ret) { 5998bf396deSTao Ma mlog_errno(ret); 6008bf396deSTao Ma goto out; 6018bf396deSTao Ma } 6028bf396deSTao Ma 6038bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_CREATE_CREDITS); 6048bf396deSTao Ma if (IS_ERR(handle)) { 6058bf396deSTao Ma ret = PTR_ERR(handle); 6068bf396deSTao Ma mlog_errno(ret); 6078bf396deSTao Ma goto out; 6088bf396deSTao Ma } 6098bf396deSTao Ma 6108bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 6118bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 6128bf396deSTao Ma if (ret) { 6138bf396deSTao Ma mlog_errno(ret); 6148bf396deSTao Ma goto out_commit; 6158bf396deSTao Ma } 6168bf396deSTao Ma 6178bf396deSTao Ma ret = ocfs2_claim_metadata(osb, handle, meta_ac, 1, 6188bf396deSTao Ma &suballoc_bit_start, &num_got, 6198bf396deSTao Ma &first_blkno); 6208bf396deSTao Ma if (ret) { 6218bf396deSTao Ma mlog_errno(ret); 6228bf396deSTao Ma goto out_commit; 6238bf396deSTao Ma } 6248bf396deSTao Ma 6258bf396deSTao Ma new_tree = ocfs2_allocate_refcount_tree(osb, first_blkno); 6268bf396deSTao Ma if (!new_tree) { 6278bf396deSTao Ma ret = -ENOMEM; 6288bf396deSTao Ma mlog_errno(ret); 6298bf396deSTao Ma goto out_commit; 6308bf396deSTao Ma } 6318bf396deSTao Ma 6328bf396deSTao Ma new_bh = sb_getblk(inode->i_sb, first_blkno); 6338bf396deSTao Ma ocfs2_set_new_buffer_uptodate(&new_tree->rf_ci, new_bh); 6348bf396deSTao Ma 6358bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &new_tree->rf_ci, new_bh, 6368bf396deSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 6378bf396deSTao Ma if (ret) { 6388bf396deSTao Ma mlog_errno(ret); 6398bf396deSTao Ma goto out_commit; 6408bf396deSTao Ma } 6418bf396deSTao Ma 6428bf396deSTao Ma /* Initialize ocfs2_refcount_block. */ 6438bf396deSTao Ma rb = (struct ocfs2_refcount_block *)new_bh->b_data; 6448bf396deSTao Ma memset(rb, 0, inode->i_sb->s_blocksize); 6458bf396deSTao Ma strcpy((void *)rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 6468bf396deSTao Ma rb->rf_suballoc_slot = cpu_to_le16(osb->slot_num); 6478bf396deSTao Ma rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 6488bf396deSTao Ma rb->rf_fs_generation = cpu_to_le32(osb->fs_generation); 6498bf396deSTao Ma rb->rf_blkno = cpu_to_le64(first_blkno); 6508bf396deSTao Ma rb->rf_count = cpu_to_le32(1); 6518bf396deSTao Ma rb->rf_records.rl_count = 6528bf396deSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(osb->sb)); 6538bf396deSTao Ma spin_lock(&osb->osb_lock); 6548bf396deSTao Ma rb->rf_generation = osb->s_next_generation++; 6558bf396deSTao Ma spin_unlock(&osb->osb_lock); 6568bf396deSTao Ma 6578bf396deSTao Ma ocfs2_journal_dirty(handle, new_bh); 6588bf396deSTao Ma 6598bf396deSTao Ma spin_lock(&oi->ip_lock); 6608bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 6618bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 6628bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(first_blkno); 6638bf396deSTao Ma spin_unlock(&oi->ip_lock); 6648bf396deSTao Ma 6658bf396deSTao Ma mlog(0, "created tree for inode %lu, refblock %llu\n", 6668bf396deSTao Ma inode->i_ino, (unsigned long long)first_blkno); 6678bf396deSTao Ma 6688bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 6698bf396deSTao Ma 6708bf396deSTao Ma /* 6718bf396deSTao Ma * We have to init the tree lock here since it will use 6728bf396deSTao Ma * the generation number to create it. 6738bf396deSTao Ma */ 6748bf396deSTao Ma new_tree->rf_generation = le32_to_cpu(rb->rf_generation); 6758bf396deSTao Ma ocfs2_init_refcount_tree_lock(osb, new_tree, first_blkno, 6768bf396deSTao Ma new_tree->rf_generation); 6778bf396deSTao Ma 6788bf396deSTao Ma spin_lock(&osb->osb_lock); 6798bf396deSTao Ma tree = ocfs2_find_refcount_tree(osb, first_blkno); 6808bf396deSTao Ma 6818bf396deSTao Ma /* 6828bf396deSTao Ma * We've just created a new refcount tree in this block. If 6838bf396deSTao Ma * we found a refcount tree on the ocfs2_super, it must be 6848bf396deSTao Ma * one we just deleted. We free the old tree before 6858bf396deSTao Ma * inserting the new tree. 6868bf396deSTao Ma */ 6878bf396deSTao Ma BUG_ON(tree && tree->rf_generation == new_tree->rf_generation); 6888bf396deSTao Ma if (tree) 6898bf396deSTao Ma ocfs2_erase_refcount_tree_from_list_no_lock(osb, tree); 6908bf396deSTao Ma ocfs2_insert_refcount_tree(osb, new_tree); 6918bf396deSTao Ma spin_unlock(&osb->osb_lock); 6928bf396deSTao Ma new_tree = NULL; 6938bf396deSTao Ma if (tree) 6948bf396deSTao Ma ocfs2_refcount_tree_put(tree); 6958bf396deSTao Ma 6968bf396deSTao Ma out_commit: 6978bf396deSTao Ma ocfs2_commit_trans(osb, handle); 6988bf396deSTao Ma 6998bf396deSTao Ma out: 7008bf396deSTao Ma if (new_tree) { 7018bf396deSTao Ma ocfs2_metadata_cache_exit(&new_tree->rf_ci); 7028bf396deSTao Ma kfree(new_tree); 7038bf396deSTao Ma } 7048bf396deSTao Ma 7058bf396deSTao Ma brelse(new_bh); 7068bf396deSTao Ma if (meta_ac) 7078bf396deSTao Ma ocfs2_free_alloc_context(meta_ac); 7088bf396deSTao Ma 7098bf396deSTao Ma return ret; 7108bf396deSTao Ma } 7118bf396deSTao Ma 7128bf396deSTao Ma static int ocfs2_set_refcount_tree(struct inode *inode, 7138bf396deSTao Ma struct buffer_head *di_bh, 7148bf396deSTao Ma u64 refcount_loc) 7158bf396deSTao Ma { 7168bf396deSTao Ma int ret; 7178bf396deSTao Ma handle_t *handle = NULL; 7188bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7198bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7208bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7218bf396deSTao Ma struct buffer_head *ref_root_bh = NULL; 7228bf396deSTao Ma struct ocfs2_refcount_block *rb; 7238bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7248bf396deSTao Ma 7258bf396deSTao Ma BUG_ON(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL); 7268bf396deSTao Ma 7278bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, refcount_loc, 1, 7288bf396deSTao Ma &ref_tree, &ref_root_bh); 7298bf396deSTao Ma if (ret) { 7308bf396deSTao Ma mlog_errno(ret); 7318bf396deSTao Ma return ret; 7328bf396deSTao Ma } 7338bf396deSTao Ma 7348bf396deSTao Ma handle = ocfs2_start_trans(osb, OCFS2_REFCOUNT_TREE_SET_CREDITS); 7358bf396deSTao Ma if (IS_ERR(handle)) { 7368bf396deSTao Ma ret = PTR_ERR(handle); 7378bf396deSTao Ma mlog_errno(ret); 7388bf396deSTao Ma goto out; 7398bf396deSTao Ma } 7408bf396deSTao Ma 7418bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 7428bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7438bf396deSTao Ma if (ret) { 7448bf396deSTao Ma mlog_errno(ret); 7458bf396deSTao Ma goto out_commit; 7468bf396deSTao Ma } 7478bf396deSTao Ma 7488bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, ref_root_bh, 7498bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 7508bf396deSTao Ma if (ret) { 7518bf396deSTao Ma mlog_errno(ret); 7528bf396deSTao Ma goto out_commit; 7538bf396deSTao Ma } 7548bf396deSTao Ma 7558bf396deSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 7568bf396deSTao Ma le32_add_cpu(&rb->rf_count, 1); 7578bf396deSTao Ma 7588bf396deSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 7598bf396deSTao Ma 7608bf396deSTao Ma spin_lock(&oi->ip_lock); 7618bf396deSTao Ma oi->ip_dyn_features |= OCFS2_HAS_REFCOUNT_FL; 7628bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 7638bf396deSTao Ma di->i_refcount_loc = cpu_to_le64(refcount_loc); 7648bf396deSTao Ma spin_unlock(&oi->ip_lock); 7658bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 7668bf396deSTao Ma 7678bf396deSTao Ma out_commit: 7688bf396deSTao Ma ocfs2_commit_trans(osb, handle); 7698bf396deSTao Ma out: 7708bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 7718bf396deSTao Ma brelse(ref_root_bh); 7728bf396deSTao Ma 7738bf396deSTao Ma return ret; 7748bf396deSTao Ma } 7758bf396deSTao Ma 7768bf396deSTao Ma int ocfs2_remove_refcount_tree(struct inode *inode, struct buffer_head *di_bh) 7778bf396deSTao Ma { 7788bf396deSTao Ma int ret, delete_tree = 0; 7798bf396deSTao Ma handle_t *handle = NULL; 7808bf396deSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7818bf396deSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 7828bf396deSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7838bf396deSTao Ma struct ocfs2_refcount_block *rb; 7848bf396deSTao Ma struct inode *alloc_inode = NULL; 7858bf396deSTao Ma struct buffer_head *alloc_bh = NULL; 7868bf396deSTao Ma struct buffer_head *blk_bh = NULL; 7878bf396deSTao Ma struct ocfs2_refcount_tree *ref_tree; 7888bf396deSTao Ma int credits = OCFS2_REFCOUNT_TREE_REMOVE_CREDITS; 7898bf396deSTao Ma u64 blk = 0, bg_blkno = 0, ref_blkno = le64_to_cpu(di->i_refcount_loc); 7908bf396deSTao Ma u16 bit = 0; 7918bf396deSTao Ma 7928bf396deSTao Ma if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) 7938bf396deSTao Ma return 0; 7948bf396deSTao Ma 7958bf396deSTao Ma BUG_ON(!ref_blkno); 7968bf396deSTao Ma ret = ocfs2_lock_refcount_tree(osb, ref_blkno, 1, &ref_tree, &blk_bh); 7978bf396deSTao Ma if (ret) { 7988bf396deSTao Ma mlog_errno(ret); 7998bf396deSTao Ma return ret; 8008bf396deSTao Ma } 8018bf396deSTao Ma 8028bf396deSTao Ma rb = (struct ocfs2_refcount_block *)blk_bh->b_data; 8038bf396deSTao Ma 8048bf396deSTao Ma /* 8058bf396deSTao Ma * If we are the last user, we need to free the block. 8068bf396deSTao Ma * So lock the allocator ahead. 8078bf396deSTao Ma */ 8088bf396deSTao Ma if (le32_to_cpu(rb->rf_count) == 1) { 8098bf396deSTao Ma blk = le64_to_cpu(rb->rf_blkno); 8108bf396deSTao Ma bit = le16_to_cpu(rb->rf_suballoc_bit); 8118bf396deSTao Ma bg_blkno = ocfs2_which_suballoc_group(blk, bit); 8128bf396deSTao Ma 8138bf396deSTao Ma alloc_inode = ocfs2_get_system_file_inode(osb, 8148bf396deSTao Ma EXTENT_ALLOC_SYSTEM_INODE, 8158bf396deSTao Ma le16_to_cpu(rb->rf_suballoc_slot)); 8168bf396deSTao Ma if (!alloc_inode) { 8178bf396deSTao Ma ret = -ENOMEM; 8188bf396deSTao Ma mlog_errno(ret); 8198bf396deSTao Ma goto out; 8208bf396deSTao Ma } 8218bf396deSTao Ma mutex_lock(&alloc_inode->i_mutex); 8228bf396deSTao Ma 8238bf396deSTao Ma ret = ocfs2_inode_lock(alloc_inode, &alloc_bh, 1); 8248bf396deSTao Ma if (ret) { 8258bf396deSTao Ma mlog_errno(ret); 8268bf396deSTao Ma goto out_mutex; 8278bf396deSTao Ma } 8288bf396deSTao Ma 8298bf396deSTao Ma credits += OCFS2_SUBALLOC_FREE; 8308bf396deSTao Ma } 8318bf396deSTao Ma 8328bf396deSTao Ma handle = ocfs2_start_trans(osb, credits); 8338bf396deSTao Ma if (IS_ERR(handle)) { 8348bf396deSTao Ma ret = PTR_ERR(handle); 8358bf396deSTao Ma mlog_errno(ret); 8368bf396deSTao Ma goto out_unlock; 8378bf396deSTao Ma } 8388bf396deSTao Ma 8398bf396deSTao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 8408bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8418bf396deSTao Ma if (ret) { 8428bf396deSTao Ma mlog_errno(ret); 8438bf396deSTao Ma goto out_commit; 8448bf396deSTao Ma } 8458bf396deSTao Ma 8468bf396deSTao Ma ret = ocfs2_journal_access_rb(handle, &ref_tree->rf_ci, blk_bh, 8478bf396deSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 8488bf396deSTao Ma if (ret) { 8498bf396deSTao Ma mlog_errno(ret); 8508bf396deSTao Ma goto out_commit; 8518bf396deSTao Ma } 8528bf396deSTao Ma 8538bf396deSTao Ma spin_lock(&oi->ip_lock); 8548bf396deSTao Ma oi->ip_dyn_features &= ~OCFS2_HAS_REFCOUNT_FL; 8558bf396deSTao Ma di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 8568bf396deSTao Ma di->i_refcount_loc = 0; 8578bf396deSTao Ma spin_unlock(&oi->ip_lock); 8588bf396deSTao Ma ocfs2_journal_dirty(handle, di_bh); 8598bf396deSTao Ma 8608bf396deSTao Ma le32_add_cpu(&rb->rf_count , -1); 8618bf396deSTao Ma ocfs2_journal_dirty(handle, blk_bh); 8628bf396deSTao Ma 8638bf396deSTao Ma if (!rb->rf_count) { 8648bf396deSTao Ma delete_tree = 1; 8658bf396deSTao Ma ocfs2_erase_refcount_tree_from_list(osb, ref_tree); 8668bf396deSTao Ma ret = ocfs2_free_suballoc_bits(handle, alloc_inode, 8678bf396deSTao Ma alloc_bh, bit, bg_blkno, 1); 8688bf396deSTao Ma if (ret) 8698bf396deSTao Ma mlog_errno(ret); 8708bf396deSTao Ma } 8718bf396deSTao Ma 8728bf396deSTao Ma out_commit: 8738bf396deSTao Ma ocfs2_commit_trans(osb, handle); 8748bf396deSTao Ma out_unlock: 8758bf396deSTao Ma if (alloc_inode) { 8768bf396deSTao Ma ocfs2_inode_unlock(alloc_inode, 1); 8778bf396deSTao Ma brelse(alloc_bh); 8788bf396deSTao Ma } 8798bf396deSTao Ma out_mutex: 8808bf396deSTao Ma if (alloc_inode) { 8818bf396deSTao Ma mutex_unlock(&alloc_inode->i_mutex); 8828bf396deSTao Ma iput(alloc_inode); 8838bf396deSTao Ma } 8848bf396deSTao Ma out: 8858bf396deSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 8868bf396deSTao Ma if (delete_tree) 8878bf396deSTao Ma ocfs2_refcount_tree_put(ref_tree); 8888bf396deSTao Ma brelse(blk_bh); 8898bf396deSTao Ma 8908bf396deSTao Ma return ret; 8918bf396deSTao Ma } 892e73a819dSTao Ma 893e73a819dSTao Ma static void ocfs2_find_refcount_rec_in_rl(struct ocfs2_caching_info *ci, 894e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 895e73a819dSTao Ma u64 cpos, unsigned int len, 896e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 897e73a819dSTao Ma int *index) 898e73a819dSTao Ma { 899e73a819dSTao Ma int i = 0; 900e73a819dSTao Ma struct ocfs2_refcount_block *rb = 901e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 902e73a819dSTao Ma struct ocfs2_refcount_rec *rec = NULL; 903e73a819dSTao Ma 904e73a819dSTao Ma for (; i < le16_to_cpu(rb->rf_records.rl_used); i++) { 905e73a819dSTao Ma rec = &rb->rf_records.rl_recs[i]; 906e73a819dSTao Ma 907e73a819dSTao Ma if (le64_to_cpu(rec->r_cpos) + 908e73a819dSTao Ma le32_to_cpu(rec->r_clusters) <= cpos) 909e73a819dSTao Ma continue; 910e73a819dSTao Ma else if (le64_to_cpu(rec->r_cpos) > cpos) 911e73a819dSTao Ma break; 912e73a819dSTao Ma 913e73a819dSTao Ma /* ok, cpos fail in this rec. Just return. */ 914e73a819dSTao Ma if (ret_rec) 915e73a819dSTao Ma *ret_rec = *rec; 916e73a819dSTao Ma goto out; 917e73a819dSTao Ma } 918e73a819dSTao Ma 919e73a819dSTao Ma if (ret_rec) { 920e73a819dSTao Ma /* We meet with a hole here, so fake the rec. */ 921e73a819dSTao Ma ret_rec->r_cpos = cpu_to_le64(cpos); 922e73a819dSTao Ma ret_rec->r_refcount = 0; 923e73a819dSTao Ma if (i < le16_to_cpu(rb->rf_records.rl_used) && 924e73a819dSTao Ma le64_to_cpu(rec->r_cpos) < cpos + len) 925e73a819dSTao Ma ret_rec->r_clusters = 926e73a819dSTao Ma cpu_to_le32(le64_to_cpu(rec->r_cpos) - cpos); 927e73a819dSTao Ma else 928e73a819dSTao Ma ret_rec->r_clusters = cpu_to_le32(len); 929e73a819dSTao Ma } 930e73a819dSTao Ma 931e73a819dSTao Ma out: 932e73a819dSTao Ma *index = i; 933e73a819dSTao Ma } 934e73a819dSTao Ma 935e73a819dSTao Ma /* 9368b2c0dbaSTao Ma * Try to remove refcount tree. The mechanism is: 9378b2c0dbaSTao Ma * 1) Check whether i_clusters == 0, if no, exit. 9388b2c0dbaSTao Ma * 2) check whether we have i_xattr_loc in dinode. if yes, exit. 9398b2c0dbaSTao Ma * 3) Check whether we have inline xattr stored outside, if yes, exit. 9408b2c0dbaSTao Ma * 4) Remove the tree. 9418b2c0dbaSTao Ma */ 9428b2c0dbaSTao Ma int ocfs2_try_remove_refcount_tree(struct inode *inode, 9438b2c0dbaSTao Ma struct buffer_head *di_bh) 9448b2c0dbaSTao Ma { 9458b2c0dbaSTao Ma int ret; 9468b2c0dbaSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 9478b2c0dbaSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 9488b2c0dbaSTao Ma 9498b2c0dbaSTao Ma down_write(&oi->ip_xattr_sem); 9508b2c0dbaSTao Ma down_write(&oi->ip_alloc_sem); 9518b2c0dbaSTao Ma 9528b2c0dbaSTao Ma if (oi->ip_clusters) 9538b2c0dbaSTao Ma goto out; 9548b2c0dbaSTao Ma 9558b2c0dbaSTao Ma if ((oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) && di->i_xattr_loc) 9568b2c0dbaSTao Ma goto out; 9578b2c0dbaSTao Ma 9588b2c0dbaSTao Ma if (oi->ip_dyn_features & OCFS2_INLINE_XATTR_FL && 9598b2c0dbaSTao Ma ocfs2_has_inline_xattr_value_outside(inode, di)) 9608b2c0dbaSTao Ma goto out; 9618b2c0dbaSTao Ma 9628b2c0dbaSTao Ma ret = ocfs2_remove_refcount_tree(inode, di_bh); 9638b2c0dbaSTao Ma if (ret) 9648b2c0dbaSTao Ma mlog_errno(ret); 9658b2c0dbaSTao Ma out: 9668b2c0dbaSTao Ma up_write(&oi->ip_alloc_sem); 9678b2c0dbaSTao Ma up_write(&oi->ip_xattr_sem); 9688b2c0dbaSTao Ma return 0; 9698b2c0dbaSTao Ma } 9708b2c0dbaSTao Ma 9718b2c0dbaSTao Ma /* 972e73a819dSTao Ma * Given a cpos and len, try to find the refcount record which contains cpos. 973e73a819dSTao Ma * 1. If cpos can be found in one refcount record, return the record. 974e73a819dSTao Ma * 2. If cpos can't be found, return a fake record which start from cpos 975e73a819dSTao Ma * and end at a small value between cpos+len and start of the next record. 976e73a819dSTao Ma * This fake record has r_refcount = 0. 977e73a819dSTao Ma */ 978e73a819dSTao Ma static int ocfs2_get_refcount_rec(struct ocfs2_caching_info *ci, 979e73a819dSTao Ma struct buffer_head *ref_root_bh, 980e73a819dSTao Ma u64 cpos, unsigned int len, 981e73a819dSTao Ma struct ocfs2_refcount_rec *ret_rec, 982e73a819dSTao Ma int *index, 983e73a819dSTao Ma struct buffer_head **ret_bh) 984e73a819dSTao Ma { 985e73a819dSTao Ma int ret = 0, i, found; 986e73a819dSTao Ma u32 low_cpos; 987e73a819dSTao Ma struct ocfs2_extent_list *el; 988e73a819dSTao Ma struct ocfs2_extent_rec *tmp, *rec = NULL; 989e73a819dSTao Ma struct ocfs2_extent_block *eb; 990e73a819dSTao Ma struct buffer_head *eb_bh = NULL, *ref_leaf_bh = NULL; 991e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 992e73a819dSTao Ma struct ocfs2_refcount_block *rb = 993e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 994e73a819dSTao Ma 995e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) { 996e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_root_bh, cpos, len, 997e73a819dSTao Ma ret_rec, index); 998e73a819dSTao Ma *ret_bh = ref_root_bh; 999e73a819dSTao Ma get_bh(ref_root_bh); 1000e73a819dSTao Ma return 0; 1001e73a819dSTao Ma } 1002e73a819dSTao Ma 1003e73a819dSTao Ma el = &rb->rf_list; 1004e73a819dSTao Ma low_cpos = cpos & OCFS2_32BIT_POS_MASK; 1005e73a819dSTao Ma 1006e73a819dSTao Ma if (el->l_tree_depth) { 1007e73a819dSTao Ma ret = ocfs2_find_leaf(ci, el, low_cpos, &eb_bh); 1008e73a819dSTao Ma if (ret) { 1009e73a819dSTao Ma mlog_errno(ret); 1010e73a819dSTao Ma goto out; 1011e73a819dSTao Ma } 1012e73a819dSTao Ma 1013e73a819dSTao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 1014e73a819dSTao Ma el = &eb->h_list; 1015e73a819dSTao Ma 1016e73a819dSTao Ma if (el->l_tree_depth) { 1017e73a819dSTao Ma ocfs2_error(sb, 1018e73a819dSTao Ma "refcount tree %llu has non zero tree " 1019e73a819dSTao Ma "depth in leaf btree tree block %llu\n", 1020e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1021e73a819dSTao Ma (unsigned long long)eb_bh->b_blocknr); 1022e73a819dSTao Ma ret = -EROFS; 1023e73a819dSTao Ma goto out; 1024e73a819dSTao Ma } 1025e73a819dSTao Ma } 1026e73a819dSTao Ma 1027e73a819dSTao Ma found = 0; 1028e73a819dSTao Ma for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) { 1029e73a819dSTao Ma rec = &el->l_recs[i]; 1030e73a819dSTao Ma 1031e73a819dSTao Ma if (le32_to_cpu(rec->e_cpos) <= low_cpos) { 1032e73a819dSTao Ma found = 1; 1033e73a819dSTao Ma break; 1034e73a819dSTao Ma } 1035e73a819dSTao Ma } 1036e73a819dSTao Ma 1037e73a819dSTao Ma /* adjust len when we have ocfs2_extent_rec after it. */ 1038e73a819dSTao Ma if (found && i < le16_to_cpu(el->l_next_free_rec) - 1) { 1039e73a819dSTao Ma tmp = &el->l_recs[i+1]; 1040e73a819dSTao Ma 1041e73a819dSTao Ma if (le32_to_cpu(tmp->e_cpos) < cpos + len) 1042e73a819dSTao Ma len = le32_to_cpu(tmp->e_cpos) - cpos; 1043e73a819dSTao Ma } 1044e73a819dSTao Ma 1045e73a819dSTao Ma ret = ocfs2_read_refcount_block(ci, le64_to_cpu(rec->e_blkno), 1046e73a819dSTao Ma &ref_leaf_bh); 1047e73a819dSTao Ma if (ret) { 1048e73a819dSTao Ma mlog_errno(ret); 1049e73a819dSTao Ma goto out; 1050e73a819dSTao Ma } 1051e73a819dSTao Ma 1052e73a819dSTao Ma ocfs2_find_refcount_rec_in_rl(ci, ref_leaf_bh, cpos, len, 1053e73a819dSTao Ma ret_rec, index); 1054e73a819dSTao Ma *ret_bh = ref_leaf_bh; 1055e73a819dSTao Ma out: 1056e73a819dSTao Ma brelse(eb_bh); 1057e73a819dSTao Ma return ret; 1058e73a819dSTao Ma } 1059e73a819dSTao Ma 1060e73a819dSTao Ma enum ocfs2_ref_rec_contig { 1061e73a819dSTao Ma REF_CONTIG_NONE = 0, 1062e73a819dSTao Ma REF_CONTIG_LEFT, 1063e73a819dSTao Ma REF_CONTIG_RIGHT, 1064e73a819dSTao Ma REF_CONTIG_LEFTRIGHT, 1065e73a819dSTao Ma }; 1066e73a819dSTao Ma 1067e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1068e73a819dSTao Ma ocfs2_refcount_rec_adjacent(struct ocfs2_refcount_block *rb, 1069e73a819dSTao Ma int index) 1070e73a819dSTao Ma { 1071e73a819dSTao Ma if ((rb->rf_records.rl_recs[index].r_refcount == 1072e73a819dSTao Ma rb->rf_records.rl_recs[index + 1].r_refcount) && 1073e73a819dSTao Ma (le64_to_cpu(rb->rf_records.rl_recs[index].r_cpos) + 1074e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index].r_clusters) == 1075e73a819dSTao Ma le64_to_cpu(rb->rf_records.rl_recs[index + 1].r_cpos))) 1076e73a819dSTao Ma return REF_CONTIG_RIGHT; 1077e73a819dSTao Ma 1078e73a819dSTao Ma return REF_CONTIG_NONE; 1079e73a819dSTao Ma } 1080e73a819dSTao Ma 1081e73a819dSTao Ma static enum ocfs2_ref_rec_contig 1082e73a819dSTao Ma ocfs2_refcount_rec_contig(struct ocfs2_refcount_block *rb, 1083e73a819dSTao Ma int index) 1084e73a819dSTao Ma { 1085e73a819dSTao Ma enum ocfs2_ref_rec_contig ret = REF_CONTIG_NONE; 1086e73a819dSTao Ma 1087e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 1) 1088e73a819dSTao Ma ret = ocfs2_refcount_rec_adjacent(rb, index); 1089e73a819dSTao Ma 1090e73a819dSTao Ma if (index > 0) { 1091e73a819dSTao Ma enum ocfs2_ref_rec_contig tmp; 1092e73a819dSTao Ma 1093e73a819dSTao Ma tmp = ocfs2_refcount_rec_adjacent(rb, index - 1); 1094e73a819dSTao Ma 1095e73a819dSTao Ma if (tmp == REF_CONTIG_RIGHT) { 1096e73a819dSTao Ma if (ret == REF_CONTIG_RIGHT) 1097e73a819dSTao Ma ret = REF_CONTIG_LEFTRIGHT; 1098e73a819dSTao Ma else 1099e73a819dSTao Ma ret = REF_CONTIG_LEFT; 1100e73a819dSTao Ma } 1101e73a819dSTao Ma } 1102e73a819dSTao Ma 1103e73a819dSTao Ma return ret; 1104e73a819dSTao Ma } 1105e73a819dSTao Ma 1106e73a819dSTao Ma static void ocfs2_rotate_refcount_rec_left(struct ocfs2_refcount_block *rb, 1107e73a819dSTao Ma int index) 1108e73a819dSTao Ma { 1109e73a819dSTao Ma BUG_ON(rb->rf_records.rl_recs[index].r_refcount != 1110e73a819dSTao Ma rb->rf_records.rl_recs[index+1].r_refcount); 1111e73a819dSTao Ma 1112e73a819dSTao Ma le32_add_cpu(&rb->rf_records.rl_recs[index].r_clusters, 1113e73a819dSTao Ma le32_to_cpu(rb->rf_records.rl_recs[index+1].r_clusters)); 1114e73a819dSTao Ma 1115e73a819dSTao Ma if (index < le16_to_cpu(rb->rf_records.rl_used) - 2) 1116e73a819dSTao Ma memmove(&rb->rf_records.rl_recs[index + 1], 1117e73a819dSTao Ma &rb->rf_records.rl_recs[index + 2], 1118e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec) * 1119e73a819dSTao Ma (le16_to_cpu(rb->rf_records.rl_used) - index - 2)); 1120e73a819dSTao Ma 1121e73a819dSTao Ma memset(&rb->rf_records.rl_recs[le16_to_cpu(rb->rf_records.rl_used) - 1], 1122e73a819dSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 1123e73a819dSTao Ma le16_add_cpu(&rb->rf_records.rl_used, -1); 1124e73a819dSTao Ma } 1125e73a819dSTao Ma 1126e73a819dSTao Ma /* 1127e73a819dSTao Ma * Merge the refcount rec if we are contiguous with the adjacent recs. 1128e73a819dSTao Ma */ 1129e73a819dSTao Ma static void ocfs2_refcount_rec_merge(struct ocfs2_refcount_block *rb, 1130e73a819dSTao Ma int index) 1131e73a819dSTao Ma { 1132e73a819dSTao Ma enum ocfs2_ref_rec_contig contig = 1133e73a819dSTao Ma ocfs2_refcount_rec_contig(rb, index); 1134e73a819dSTao Ma 1135e73a819dSTao Ma if (contig == REF_CONTIG_NONE) 1136e73a819dSTao Ma return; 1137e73a819dSTao Ma 1138e73a819dSTao Ma if (contig == REF_CONTIG_LEFT || contig == REF_CONTIG_LEFTRIGHT) { 1139e73a819dSTao Ma BUG_ON(index == 0); 1140e73a819dSTao Ma index--; 1141e73a819dSTao Ma } 1142e73a819dSTao Ma 1143e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1144e73a819dSTao Ma 1145e73a819dSTao Ma if (contig == REF_CONTIG_LEFTRIGHT) 1146e73a819dSTao Ma ocfs2_rotate_refcount_rec_left(rb, index); 1147e73a819dSTao Ma } 1148e73a819dSTao Ma 11491823cb0bSTao Ma /* 11501823cb0bSTao Ma * Change the refcount indexed by "index" in ref_bh. 11511823cb0bSTao Ma * If refcount reaches 0, remove it. 11521823cb0bSTao Ma */ 1153e73a819dSTao Ma static int ocfs2_change_refcount_rec(handle_t *handle, 1154e73a819dSTao Ma struct ocfs2_caching_info *ci, 1155e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 11567540c1a7STao Ma int index, int merge, int change) 1157e73a819dSTao Ma { 1158e73a819dSTao Ma int ret; 1159e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1160e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 11611823cb0bSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 11621823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rl->rl_recs[index]; 1163e73a819dSTao Ma 1164e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1165e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1166e73a819dSTao Ma if (ret) { 1167e73a819dSTao Ma mlog_errno(ret); 1168e73a819dSTao Ma goto out; 1169e73a819dSTao Ma } 1170e73a819dSTao Ma 1171e73a819dSTao Ma mlog(0, "change index %d, old count %u, change %d\n", index, 1172e73a819dSTao Ma le32_to_cpu(rec->r_refcount), change); 1173e73a819dSTao Ma le32_add_cpu(&rec->r_refcount, change); 1174e73a819dSTao Ma 11751823cb0bSTao Ma if (!rec->r_refcount) { 11761823cb0bSTao Ma if (index != le16_to_cpu(rl->rl_used) - 1) { 11771823cb0bSTao Ma memmove(rec, rec + 1, 11781823cb0bSTao Ma (le16_to_cpu(rl->rl_used) - index - 1) * 11791823cb0bSTao Ma sizeof(struct ocfs2_refcount_rec)); 11801823cb0bSTao Ma memset(&rl->rl_recs[le16_to_cpu(rl->rl_used) - 1], 11811823cb0bSTao Ma 0, sizeof(struct ocfs2_refcount_rec)); 11821823cb0bSTao Ma } 11831823cb0bSTao Ma 11841823cb0bSTao Ma le16_add_cpu(&rl->rl_used, -1); 11857540c1a7STao Ma } else if (merge) 1186e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1187e73a819dSTao Ma 1188e73a819dSTao Ma ret = ocfs2_journal_dirty(handle, ref_leaf_bh); 1189e73a819dSTao Ma if (ret) 1190e73a819dSTao Ma mlog_errno(ret); 1191e73a819dSTao Ma out: 1192e73a819dSTao Ma return ret; 1193e73a819dSTao Ma } 1194e73a819dSTao Ma 1195e73a819dSTao Ma static int ocfs2_expand_inline_ref_root(handle_t *handle, 1196e73a819dSTao Ma struct ocfs2_caching_info *ci, 1197e73a819dSTao Ma struct buffer_head *ref_root_bh, 1198e73a819dSTao Ma struct buffer_head **ref_leaf_bh, 1199e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1200e73a819dSTao Ma { 1201e73a819dSTao Ma int ret; 1202e73a819dSTao Ma u16 suballoc_bit_start; 1203e73a819dSTao Ma u32 num_got; 1204e73a819dSTao Ma u64 blkno; 1205e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1206e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1207e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1208e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1209e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1210e73a819dSTao Ma 1211e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1212e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1213e73a819dSTao Ma if (ret) { 1214e73a819dSTao Ma mlog_errno(ret); 1215e73a819dSTao Ma goto out; 1216e73a819dSTao Ma } 1217e73a819dSTao Ma 1218e73a819dSTao Ma ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1, 1219e73a819dSTao Ma &suballoc_bit_start, &num_got, 1220e73a819dSTao Ma &blkno); 1221e73a819dSTao Ma if (ret) { 1222e73a819dSTao Ma mlog_errno(ret); 1223e73a819dSTao Ma goto out; 1224e73a819dSTao Ma } 1225e73a819dSTao Ma 1226e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1227e73a819dSTao Ma if (new_bh == NULL) { 1228e73a819dSTao Ma ret = -EIO; 1229e73a819dSTao Ma mlog_errno(ret); 1230e73a819dSTao Ma goto out; 1231e73a819dSTao Ma } 1232e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1233e73a819dSTao Ma 1234e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1235e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1236e73a819dSTao Ma if (ret) { 1237e73a819dSTao Ma mlog_errno(ret); 1238e73a819dSTao Ma goto out; 1239e73a819dSTao Ma } 1240e73a819dSTao Ma 1241e73a819dSTao Ma /* 1242e73a819dSTao Ma * Initialize ocfs2_refcount_block. 1243e73a819dSTao Ma * It should contain the same information as the old root. 1244e73a819dSTao Ma * so just memcpy it and change the corresponding field. 1245e73a819dSTao Ma */ 1246e73a819dSTao Ma memcpy(new_bh->b_data, ref_root_bh->b_data, sb->s_blocksize); 1247e73a819dSTao Ma 1248e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1249e73a819dSTao Ma new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num); 1250e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1251e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1252e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(0); 1253e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1254e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1255e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1256e73a819dSTao Ma 1257e73a819dSTao Ma /* Now change the root. */ 1258e73a819dSTao Ma memset(&root_rb->rf_list, 0, sb->s_blocksize - 1259e73a819dSTao Ma offsetof(struct ocfs2_refcount_block, rf_list)); 1260e73a819dSTao Ma root_rb->rf_list.l_count = cpu_to_le16(ocfs2_extent_recs_per_rb(sb)); 1261e73a819dSTao Ma root_rb->rf_clusters = cpu_to_le32(1); 1262e73a819dSTao Ma root_rb->rf_list.l_next_free_rec = cpu_to_le16(1); 1263e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_blkno = cpu_to_le64(blkno); 1264e73a819dSTao Ma root_rb->rf_list.l_recs[0].e_leaf_clusters = cpu_to_le16(1); 1265e73a819dSTao Ma root_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_TREE_FL); 1266e73a819dSTao Ma 1267e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 1268e73a819dSTao Ma 1269e73a819dSTao Ma mlog(0, "new leaf block %llu, used %u\n", (unsigned long long)blkno, 1270e73a819dSTao Ma le16_to_cpu(new_rb->rf_records.rl_used)); 1271e73a819dSTao Ma 1272e73a819dSTao Ma *ref_leaf_bh = new_bh; 1273e73a819dSTao Ma new_bh = NULL; 1274e73a819dSTao Ma out: 1275e73a819dSTao Ma brelse(new_bh); 1276e73a819dSTao Ma return ret; 1277e73a819dSTao Ma } 1278e73a819dSTao Ma 1279e73a819dSTao Ma static int ocfs2_refcount_rec_no_intersect(struct ocfs2_refcount_rec *prev, 1280e73a819dSTao Ma struct ocfs2_refcount_rec *next) 1281e73a819dSTao Ma { 1282e73a819dSTao Ma if (ocfs2_get_ref_rec_low_cpos(prev) + le32_to_cpu(prev->r_clusters) <= 1283e73a819dSTao Ma ocfs2_get_ref_rec_low_cpos(next)) 1284e73a819dSTao Ma return 1; 1285e73a819dSTao Ma 1286e73a819dSTao Ma return 0; 1287e73a819dSTao Ma } 1288e73a819dSTao Ma 1289e73a819dSTao Ma static int cmp_refcount_rec_by_low_cpos(const void *a, const void *b) 1290e73a819dSTao Ma { 1291e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1292e73a819dSTao Ma u32 l_cpos = ocfs2_get_ref_rec_low_cpos(l); 1293e73a819dSTao Ma u32 r_cpos = ocfs2_get_ref_rec_low_cpos(r); 1294e73a819dSTao Ma 1295e73a819dSTao Ma if (l_cpos > r_cpos) 1296e73a819dSTao Ma return 1; 1297e73a819dSTao Ma if (l_cpos < r_cpos) 1298e73a819dSTao Ma return -1; 1299e73a819dSTao Ma return 0; 1300e73a819dSTao Ma } 1301e73a819dSTao Ma 1302e73a819dSTao Ma static int cmp_refcount_rec_by_cpos(const void *a, const void *b) 1303e73a819dSTao Ma { 1304e73a819dSTao Ma const struct ocfs2_refcount_rec *l = a, *r = b; 1305e73a819dSTao Ma u64 l_cpos = le64_to_cpu(l->r_cpos); 1306e73a819dSTao Ma u64 r_cpos = le64_to_cpu(r->r_cpos); 1307e73a819dSTao Ma 1308e73a819dSTao Ma if (l_cpos > r_cpos) 1309e73a819dSTao Ma return 1; 1310e73a819dSTao Ma if (l_cpos < r_cpos) 1311e73a819dSTao Ma return -1; 1312e73a819dSTao Ma return 0; 1313e73a819dSTao Ma } 1314e73a819dSTao Ma 1315e73a819dSTao Ma static void swap_refcount_rec(void *a, void *b, int size) 1316e73a819dSTao Ma { 1317e73a819dSTao Ma struct ocfs2_refcount_rec *l = a, *r = b, tmp; 1318e73a819dSTao Ma 1319e73a819dSTao Ma tmp = *(struct ocfs2_refcount_rec *)l; 1320e73a819dSTao Ma *(struct ocfs2_refcount_rec *)l = 1321e73a819dSTao Ma *(struct ocfs2_refcount_rec *)r; 1322e73a819dSTao Ma *(struct ocfs2_refcount_rec *)r = tmp; 1323e73a819dSTao Ma } 1324e73a819dSTao Ma 1325e73a819dSTao Ma /* 1326e73a819dSTao Ma * The refcount cpos are ordered by their 64bit cpos, 1327e73a819dSTao Ma * But we will use the low 32 bit to be the e_cpos in the b-tree. 1328e73a819dSTao Ma * So we need to make sure that this pos isn't intersected with others. 1329e73a819dSTao Ma * 1330e73a819dSTao Ma * Note: The refcount block is already sorted by their low 32 bit cpos, 1331e73a819dSTao Ma * So just try the middle pos first, and we will exit when we find 1332e73a819dSTao Ma * the good position. 1333e73a819dSTao Ma */ 1334e73a819dSTao Ma static int ocfs2_find_refcount_split_pos(struct ocfs2_refcount_list *rl, 1335e73a819dSTao Ma u32 *split_pos, int *split_index) 1336e73a819dSTao Ma { 1337e73a819dSTao Ma int num_used = le16_to_cpu(rl->rl_used); 1338e73a819dSTao Ma int delta, middle = num_used / 2; 1339e73a819dSTao Ma 1340e73a819dSTao Ma for (delta = 0; delta < middle; delta++) { 1341e73a819dSTao Ma /* Let's check delta earlier than middle */ 1342e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1343e73a819dSTao Ma &rl->rl_recs[middle - delta - 1], 1344e73a819dSTao Ma &rl->rl_recs[middle - delta])) { 1345e73a819dSTao Ma *split_index = middle - delta; 1346e73a819dSTao Ma break; 1347e73a819dSTao Ma } 1348e73a819dSTao Ma 1349e73a819dSTao Ma /* For even counts, don't walk off the end */ 1350e73a819dSTao Ma if ((middle + delta + 1) == num_used) 1351e73a819dSTao Ma continue; 1352e73a819dSTao Ma 1353e73a819dSTao Ma /* Now try delta past middle */ 1354e73a819dSTao Ma if (ocfs2_refcount_rec_no_intersect( 1355e73a819dSTao Ma &rl->rl_recs[middle + delta], 1356e73a819dSTao Ma &rl->rl_recs[middle + delta + 1])) { 1357e73a819dSTao Ma *split_index = middle + delta + 1; 1358e73a819dSTao Ma break; 1359e73a819dSTao Ma } 1360e73a819dSTao Ma } 1361e73a819dSTao Ma 1362e73a819dSTao Ma if (delta >= middle) 1363e73a819dSTao Ma return -ENOSPC; 1364e73a819dSTao Ma 1365e73a819dSTao Ma *split_pos = ocfs2_get_ref_rec_low_cpos(&rl->rl_recs[*split_index]); 1366e73a819dSTao Ma return 0; 1367e73a819dSTao Ma } 1368e73a819dSTao Ma 1369e73a819dSTao Ma static int ocfs2_divide_leaf_refcount_block(struct buffer_head *ref_leaf_bh, 1370e73a819dSTao Ma struct buffer_head *new_bh, 1371e73a819dSTao Ma u32 *split_cpos) 1372e73a819dSTao Ma { 1373e73a819dSTao Ma int split_index = 0, num_moved, ret; 1374e73a819dSTao Ma u32 cpos = 0; 1375e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1376e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1377e73a819dSTao Ma struct ocfs2_refcount_list *rl = &rb->rf_records; 1378e73a819dSTao Ma struct ocfs2_refcount_block *new_rb = 1379e73a819dSTao Ma (struct ocfs2_refcount_block *)new_bh->b_data; 1380e73a819dSTao Ma struct ocfs2_refcount_list *new_rl = &new_rb->rf_records; 1381e73a819dSTao Ma 1382e73a819dSTao Ma mlog(0, "split old leaf refcount block %llu, count = %u, used = %u\n", 1383e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, 1384e73a819dSTao Ma le32_to_cpu(rl->rl_count), le32_to_cpu(rl->rl_used)); 1385e73a819dSTao Ma 1386e73a819dSTao Ma /* 1387e73a819dSTao Ma * XXX: Improvement later. 1388e73a819dSTao Ma * If we know all the high 32 bit cpos is the same, no need to sort. 1389e73a819dSTao Ma * 1390e73a819dSTao Ma * In order to make the whole process safe, we do: 1391e73a819dSTao Ma * 1. sort the entries by their low 32 bit cpos first so that we can 1392e73a819dSTao Ma * find the split cpos easily. 1393e73a819dSTao Ma * 2. call ocfs2_insert_extent to insert the new refcount block. 1394e73a819dSTao Ma * 3. move the refcount rec to the new block. 1395e73a819dSTao Ma * 4. sort the entries by their 64 bit cpos. 1396e73a819dSTao Ma * 5. dirty the new_rb and rb. 1397e73a819dSTao Ma */ 1398e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1399e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1400e73a819dSTao Ma cmp_refcount_rec_by_low_cpos, swap_refcount_rec); 1401e73a819dSTao Ma 1402e73a819dSTao Ma ret = ocfs2_find_refcount_split_pos(rl, &cpos, &split_index); 1403e73a819dSTao Ma if (ret) { 1404e73a819dSTao Ma mlog_errno(ret); 1405e73a819dSTao Ma return ret; 1406e73a819dSTao Ma } 1407e73a819dSTao Ma 1408e73a819dSTao Ma new_rb->rf_cpos = cpu_to_le32(cpos); 1409e73a819dSTao Ma 1410e73a819dSTao Ma /* move refcount records starting from split_index to the new block. */ 1411e73a819dSTao Ma num_moved = le16_to_cpu(rl->rl_used) - split_index; 1412e73a819dSTao Ma memcpy(new_rl->rl_recs, &rl->rl_recs[split_index], 1413e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1414e73a819dSTao Ma 1415e73a819dSTao Ma /*ok, remove the entries we just moved over to the other block. */ 1416e73a819dSTao Ma memset(&rl->rl_recs[split_index], 0, 1417e73a819dSTao Ma num_moved * sizeof(struct ocfs2_refcount_rec)); 1418e73a819dSTao Ma 1419e73a819dSTao Ma /* change old and new rl_used accordingly. */ 1420e73a819dSTao Ma le16_add_cpu(&rl->rl_used, -num_moved); 1421e73a819dSTao Ma new_rl->rl_used = cpu_to_le32(num_moved); 1422e73a819dSTao Ma 1423e73a819dSTao Ma sort(&rl->rl_recs, le16_to_cpu(rl->rl_used), 1424e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1425e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1426e73a819dSTao Ma 1427e73a819dSTao Ma sort(&new_rl->rl_recs, le16_to_cpu(new_rl->rl_used), 1428e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec), 1429e73a819dSTao Ma cmp_refcount_rec_by_cpos, swap_refcount_rec); 1430e73a819dSTao Ma 1431e73a819dSTao Ma *split_cpos = cpos; 1432e73a819dSTao Ma return 0; 1433e73a819dSTao Ma } 1434e73a819dSTao Ma 1435e73a819dSTao Ma static int ocfs2_new_leaf_refcount_block(handle_t *handle, 1436e73a819dSTao Ma struct ocfs2_caching_info *ci, 1437e73a819dSTao Ma struct buffer_head *ref_root_bh, 1438e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1439e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1440e73a819dSTao Ma { 1441e73a819dSTao Ma int ret; 1442e73a819dSTao Ma u16 suballoc_bit_start; 1443e73a819dSTao Ma u32 num_got, new_cpos; 1444e73a819dSTao Ma u64 blkno; 1445e73a819dSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 1446e73a819dSTao Ma struct ocfs2_refcount_block *root_rb = 1447e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1448e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1449e73a819dSTao Ma struct ocfs2_refcount_block *new_rb; 1450e73a819dSTao Ma struct ocfs2_extent_tree ref_et; 1451e73a819dSTao Ma 1452e73a819dSTao Ma BUG_ON(!(le32_to_cpu(root_rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)); 1453e73a819dSTao Ma 1454e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 1455e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1456e73a819dSTao Ma if (ret) { 1457e73a819dSTao Ma mlog_errno(ret); 1458e73a819dSTao Ma goto out; 1459e73a819dSTao Ma } 1460e73a819dSTao Ma 1461e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1462e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1463e73a819dSTao Ma if (ret) { 1464e73a819dSTao Ma mlog_errno(ret); 1465e73a819dSTao Ma goto out; 1466e73a819dSTao Ma } 1467e73a819dSTao Ma 1468e73a819dSTao Ma ret = ocfs2_claim_metadata(OCFS2_SB(sb), handle, meta_ac, 1, 1469e73a819dSTao Ma &suballoc_bit_start, &num_got, 1470e73a819dSTao Ma &blkno); 1471e73a819dSTao Ma if (ret) { 1472e73a819dSTao Ma mlog_errno(ret); 1473e73a819dSTao Ma goto out; 1474e73a819dSTao Ma } 1475e73a819dSTao Ma 1476e73a819dSTao Ma new_bh = sb_getblk(sb, blkno); 1477e73a819dSTao Ma if (new_bh == NULL) { 1478e73a819dSTao Ma ret = -EIO; 1479e73a819dSTao Ma mlog_errno(ret); 1480e73a819dSTao Ma goto out; 1481e73a819dSTao Ma } 1482e73a819dSTao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 1483e73a819dSTao Ma 1484e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, new_bh, 1485e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_CREATE); 1486e73a819dSTao Ma if (ret) { 1487e73a819dSTao Ma mlog_errno(ret); 1488e73a819dSTao Ma goto out; 1489e73a819dSTao Ma } 1490e73a819dSTao Ma 1491e73a819dSTao Ma /* Initialize ocfs2_refcount_block. */ 1492e73a819dSTao Ma new_rb = (struct ocfs2_refcount_block *)new_bh->b_data; 1493e73a819dSTao Ma memset(new_rb, 0, sb->s_blocksize); 1494e73a819dSTao Ma strcpy((void *)new_rb, OCFS2_REFCOUNT_BLOCK_SIGNATURE); 1495e73a819dSTao Ma new_rb->rf_suballoc_slot = cpu_to_le16(OCFS2_SB(sb)->slot_num); 1496e73a819dSTao Ma new_rb->rf_suballoc_bit = cpu_to_le16(suballoc_bit_start); 1497e73a819dSTao Ma new_rb->rf_fs_generation = cpu_to_le32(OCFS2_SB(sb)->fs_generation); 1498e73a819dSTao Ma new_rb->rf_blkno = cpu_to_le64(blkno); 1499e73a819dSTao Ma new_rb->rf_parent = cpu_to_le64(ref_root_bh->b_blocknr); 1500e73a819dSTao Ma new_rb->rf_flags = cpu_to_le32(OCFS2_REFCOUNT_LEAF_FL); 1501e73a819dSTao Ma new_rb->rf_records.rl_count = 1502e73a819dSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 1503e73a819dSTao Ma new_rb->rf_generation = root_rb->rf_generation; 1504e73a819dSTao Ma 1505e73a819dSTao Ma ret = ocfs2_divide_leaf_refcount_block(ref_leaf_bh, new_bh, &new_cpos); 1506e73a819dSTao Ma if (ret) { 1507e73a819dSTao Ma mlog_errno(ret); 1508e73a819dSTao Ma goto out; 1509e73a819dSTao Ma } 1510e73a819dSTao Ma 1511e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1512e73a819dSTao Ma ocfs2_journal_dirty(handle, new_bh); 1513e73a819dSTao Ma 1514e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&ref_et, ci, ref_root_bh); 1515e73a819dSTao Ma 1516e73a819dSTao Ma mlog(0, "insert new leaf block %llu at %u\n", 1517e73a819dSTao Ma (unsigned long long)new_bh->b_blocknr, new_cpos); 1518e73a819dSTao Ma 1519e73a819dSTao Ma /* Insert the new leaf block with the specific offset cpos. */ 1520e73a819dSTao Ma ret = ocfs2_insert_extent(handle, &ref_et, new_cpos, new_bh->b_blocknr, 1521e73a819dSTao Ma 1, 0, meta_ac); 1522e73a819dSTao Ma if (ret) 1523e73a819dSTao Ma mlog_errno(ret); 1524e73a819dSTao Ma 1525e73a819dSTao Ma out: 1526e73a819dSTao Ma brelse(new_bh); 1527e73a819dSTao Ma return ret; 1528e73a819dSTao Ma } 1529e73a819dSTao Ma 1530e73a819dSTao Ma static int ocfs2_expand_refcount_tree(handle_t *handle, 1531e73a819dSTao Ma struct ocfs2_caching_info *ci, 1532e73a819dSTao Ma struct buffer_head *ref_root_bh, 1533e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1534e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1535e73a819dSTao Ma { 1536e73a819dSTao Ma int ret; 1537e73a819dSTao Ma struct buffer_head *expand_bh = NULL; 1538e73a819dSTao Ma 1539e73a819dSTao Ma if (ref_root_bh == ref_leaf_bh) { 1540e73a819dSTao Ma /* 1541e73a819dSTao Ma * the old root bh hasn't been expanded to a b-tree, 1542e73a819dSTao Ma * so expand it first. 1543e73a819dSTao Ma */ 1544e73a819dSTao Ma ret = ocfs2_expand_inline_ref_root(handle, ci, ref_root_bh, 1545e73a819dSTao Ma &expand_bh, meta_ac); 1546e73a819dSTao Ma if (ret) { 1547e73a819dSTao Ma mlog_errno(ret); 1548e73a819dSTao Ma goto out; 1549e73a819dSTao Ma } 1550e73a819dSTao Ma } else { 1551e73a819dSTao Ma expand_bh = ref_leaf_bh; 1552e73a819dSTao Ma get_bh(expand_bh); 1553e73a819dSTao Ma } 1554e73a819dSTao Ma 1555e73a819dSTao Ma 1556e73a819dSTao Ma /* Now add a new refcount block into the tree.*/ 1557e73a819dSTao Ma ret = ocfs2_new_leaf_refcount_block(handle, ci, ref_root_bh, 1558e73a819dSTao Ma expand_bh, meta_ac); 1559e73a819dSTao Ma if (ret) 1560e73a819dSTao Ma mlog_errno(ret); 1561e73a819dSTao Ma out: 1562e73a819dSTao Ma brelse(expand_bh); 1563e73a819dSTao Ma return ret; 1564e73a819dSTao Ma } 1565e73a819dSTao Ma 1566e73a819dSTao Ma /* 1567e73a819dSTao Ma * Adjust the extent rec in b-tree representing ref_leaf_bh. 1568e73a819dSTao Ma * 1569e73a819dSTao Ma * Only called when we have inserted a new refcount rec at index 0 1570e73a819dSTao Ma * which means ocfs2_extent_rec.e_cpos may need some change. 1571e73a819dSTao Ma */ 1572e73a819dSTao Ma static int ocfs2_adjust_refcount_rec(handle_t *handle, 1573e73a819dSTao Ma struct ocfs2_caching_info *ci, 1574e73a819dSTao Ma struct buffer_head *ref_root_bh, 1575e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1576e73a819dSTao Ma struct ocfs2_refcount_rec *rec) 1577e73a819dSTao Ma { 1578e73a819dSTao Ma int ret = 0, i; 1579e73a819dSTao Ma u32 new_cpos, old_cpos; 1580e73a819dSTao Ma struct ocfs2_path *path = NULL; 1581e73a819dSTao Ma struct ocfs2_extent_tree et; 1582e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1583e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_root_bh->b_data; 1584e73a819dSTao Ma struct ocfs2_extent_list *el; 1585e73a819dSTao Ma 1586e73a819dSTao Ma if (!(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL)) 1587e73a819dSTao Ma goto out; 1588e73a819dSTao Ma 1589e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1590e73a819dSTao Ma old_cpos = le32_to_cpu(rb->rf_cpos); 1591e73a819dSTao Ma new_cpos = le64_to_cpu(rec->r_cpos) & OCFS2_32BIT_POS_MASK; 1592e73a819dSTao Ma if (old_cpos <= new_cpos) 1593e73a819dSTao Ma goto out; 1594e73a819dSTao Ma 1595e73a819dSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 1596e73a819dSTao Ma 1597e73a819dSTao Ma path = ocfs2_new_path_from_et(&et); 1598e73a819dSTao Ma if (!path) { 1599e73a819dSTao Ma ret = -ENOMEM; 1600e73a819dSTao Ma mlog_errno(ret); 1601e73a819dSTao Ma goto out; 1602e73a819dSTao Ma } 1603e73a819dSTao Ma 1604e73a819dSTao Ma ret = ocfs2_find_path(ci, path, old_cpos); 1605e73a819dSTao Ma if (ret) { 1606e73a819dSTao Ma mlog_errno(ret); 1607e73a819dSTao Ma goto out; 1608e73a819dSTao Ma } 1609e73a819dSTao Ma 1610e73a819dSTao Ma /* 1611e73a819dSTao Ma * 2 more credits, one for the leaf refcount block, one for 1612e73a819dSTao Ma * the extent block contains the extent rec. 1613e73a819dSTao Ma */ 1614e73a819dSTao Ma ret = ocfs2_extend_trans(handle, handle->h_buffer_credits + 2); 1615e73a819dSTao Ma if (ret < 0) { 1616e73a819dSTao Ma mlog_errno(ret); 1617e73a819dSTao Ma goto out; 1618e73a819dSTao Ma } 1619e73a819dSTao Ma 1620e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1621e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1622e73a819dSTao Ma if (ret < 0) { 1623e73a819dSTao Ma mlog_errno(ret); 1624e73a819dSTao Ma goto out; 1625e73a819dSTao Ma } 1626e73a819dSTao Ma 1627e73a819dSTao Ma ret = ocfs2_journal_access_eb(handle, ci, path_leaf_bh(path), 1628e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1629e73a819dSTao Ma if (ret < 0) { 1630e73a819dSTao Ma mlog_errno(ret); 1631e73a819dSTao Ma goto out; 1632e73a819dSTao Ma } 1633e73a819dSTao Ma 1634e73a819dSTao Ma /* change the leaf extent block first. */ 1635e73a819dSTao Ma el = path_leaf_el(path); 1636e73a819dSTao Ma 1637e73a819dSTao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) 1638e73a819dSTao Ma if (le32_to_cpu(el->l_recs[i].e_cpos) == old_cpos) 1639e73a819dSTao Ma break; 1640e73a819dSTao Ma 1641e73a819dSTao Ma BUG_ON(i == le16_to_cpu(el->l_next_free_rec)); 1642e73a819dSTao Ma 1643e73a819dSTao Ma el->l_recs[i].e_cpos = cpu_to_le32(new_cpos); 1644e73a819dSTao Ma 1645e73a819dSTao Ma /* change the r_cpos in the leaf block. */ 1646e73a819dSTao Ma rb->rf_cpos = cpu_to_le32(new_cpos); 1647e73a819dSTao Ma 1648e73a819dSTao Ma ocfs2_journal_dirty(handle, path_leaf_bh(path)); 1649e73a819dSTao Ma ocfs2_journal_dirty(handle, ref_leaf_bh); 1650e73a819dSTao Ma 1651e73a819dSTao Ma out: 1652e73a819dSTao Ma ocfs2_free_path(path); 1653e73a819dSTao Ma return ret; 1654e73a819dSTao Ma } 1655e73a819dSTao Ma 1656e73a819dSTao Ma static int ocfs2_insert_refcount_rec(handle_t *handle, 1657e73a819dSTao Ma struct ocfs2_caching_info *ci, 1658e73a819dSTao Ma struct buffer_head *ref_root_bh, 1659e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1660e73a819dSTao Ma struct ocfs2_refcount_rec *rec, 16617540c1a7STao Ma int index, int merge, 1662e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac) 1663e73a819dSTao Ma { 1664e73a819dSTao Ma int ret; 1665e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1666e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1667e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1668e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1669e73a819dSTao Ma 1670e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1671e73a819dSTao Ma 1672e73a819dSTao Ma if (rf_list->rl_used == rf_list->rl_count) { 1673e73a819dSTao Ma u64 cpos = le64_to_cpu(rec->r_cpos); 1674e73a819dSTao Ma u32 len = le32_to_cpu(rec->r_clusters); 1675e73a819dSTao Ma 1676e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1677e73a819dSTao Ma ref_leaf_bh, meta_ac); 1678e73a819dSTao Ma if (ret) { 1679e73a819dSTao Ma mlog_errno(ret); 1680e73a819dSTao Ma goto out; 1681e73a819dSTao Ma } 1682e73a819dSTao Ma 1683e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1684e73a819dSTao Ma cpos, len, NULL, &index, 1685e73a819dSTao Ma &new_bh); 1686e73a819dSTao Ma if (ret) { 1687e73a819dSTao Ma mlog_errno(ret); 1688e73a819dSTao Ma goto out; 1689e73a819dSTao Ma } 1690e73a819dSTao Ma 1691e73a819dSTao Ma ref_leaf_bh = new_bh; 1692e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1693e73a819dSTao Ma rf_list = &rb->rf_records; 1694e73a819dSTao Ma } 1695e73a819dSTao Ma 1696e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1697e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1698e73a819dSTao Ma if (ret) { 1699e73a819dSTao Ma mlog_errno(ret); 1700e73a819dSTao Ma goto out; 1701e73a819dSTao Ma } 1702e73a819dSTao Ma 1703e73a819dSTao Ma if (index < le16_to_cpu(rf_list->rl_used)) 1704e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1], 1705e73a819dSTao Ma &rf_list->rl_recs[index], 1706e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index) * 1707e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1708e73a819dSTao Ma 1709e73a819dSTao Ma mlog(0, "insert refcount record start %llu, len %u, count %u " 1710e73a819dSTao Ma "to leaf block %llu at index %d\n", 1711e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos), 1712e73a819dSTao Ma le32_to_cpu(rec->r_clusters), le32_to_cpu(rec->r_refcount), 1713e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index); 1714e73a819dSTao Ma 1715e73a819dSTao Ma rf_list->rl_recs[index] = *rec; 1716e73a819dSTao Ma 1717e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, 1); 1718e73a819dSTao Ma 17197540c1a7STao Ma if (merge) 1720e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1721e73a819dSTao Ma 1722e73a819dSTao Ma ret = ocfs2_journal_dirty(handle, ref_leaf_bh); 1723e73a819dSTao Ma if (ret) { 1724e73a819dSTao Ma mlog_errno(ret); 1725e73a819dSTao Ma goto out; 1726e73a819dSTao Ma } 1727e73a819dSTao Ma 1728e73a819dSTao Ma if (index == 0) { 1729e73a819dSTao Ma ret = ocfs2_adjust_refcount_rec(handle, ci, 1730e73a819dSTao Ma ref_root_bh, 1731e73a819dSTao Ma ref_leaf_bh, rec); 1732e73a819dSTao Ma if (ret) 1733e73a819dSTao Ma mlog_errno(ret); 1734e73a819dSTao Ma } 1735e73a819dSTao Ma out: 1736e73a819dSTao Ma brelse(new_bh); 1737e73a819dSTao Ma return ret; 1738e73a819dSTao Ma } 1739e73a819dSTao Ma 1740e73a819dSTao Ma /* 1741e73a819dSTao Ma * Split the refcount_rec indexed by "index" in ref_leaf_bh. 1742e73a819dSTao Ma * This is much simple than our b-tree code. 1743e73a819dSTao Ma * split_rec is the new refcount rec we want to insert. 1744e73a819dSTao Ma * If split_rec->r_refcount > 0, we are changing the refcount(in case we 1745e73a819dSTao Ma * increase refcount or decrease a refcount to non-zero). 1746e73a819dSTao Ma * If split_rec->r_refcount == 0, we are punching a hole in current refcount 1747e73a819dSTao Ma * rec( in case we decrease a refcount to zero). 1748e73a819dSTao Ma */ 1749e73a819dSTao Ma static int ocfs2_split_refcount_rec(handle_t *handle, 1750e73a819dSTao Ma struct ocfs2_caching_info *ci, 1751e73a819dSTao Ma struct buffer_head *ref_root_bh, 1752e73a819dSTao Ma struct buffer_head *ref_leaf_bh, 1753e73a819dSTao Ma struct ocfs2_refcount_rec *split_rec, 17547540c1a7STao Ma int index, int merge, 1755e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1756e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1757e73a819dSTao Ma { 1758e73a819dSTao Ma int ret, recs_need; 1759e73a819dSTao Ma u32 len; 1760e73a819dSTao Ma struct ocfs2_refcount_block *rb = 1761e73a819dSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1762e73a819dSTao Ma struct ocfs2_refcount_list *rf_list = &rb->rf_records; 1763e73a819dSTao Ma struct ocfs2_refcount_rec *orig_rec = &rf_list->rl_recs[index]; 1764e73a819dSTao Ma struct ocfs2_refcount_rec *tail_rec = NULL; 1765e73a819dSTao Ma struct buffer_head *new_bh = NULL; 1766e73a819dSTao Ma 1767e73a819dSTao Ma BUG_ON(le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL); 1768e73a819dSTao Ma 1769e73a819dSTao Ma mlog(0, "original r_pos %llu, cluster %u, split %llu, cluster %u\n", 1770e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos), le32_to_cpu(orig_rec->r_clusters), 1771e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos), 1772e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters)); 1773e73a819dSTao Ma 1774e73a819dSTao Ma /* 1775e73a819dSTao Ma * If we just need to split the header or tail clusters, 1776e73a819dSTao Ma * no more recs are needed, just split is OK. 1777e73a819dSTao Ma * Otherwise we at least need one new recs. 1778e73a819dSTao Ma */ 1779e73a819dSTao Ma if (!split_rec->r_refcount && 1780e73a819dSTao Ma (split_rec->r_cpos == orig_rec->r_cpos || 1781e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1782e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) == 1783e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1784e73a819dSTao Ma recs_need = 0; 1785e73a819dSTao Ma else 1786e73a819dSTao Ma recs_need = 1; 1787e73a819dSTao Ma 1788e73a819dSTao Ma /* 1789e73a819dSTao Ma * We need one more rec if we split in the middle and the new rec have 1790e73a819dSTao Ma * some refcount in it. 1791e73a819dSTao Ma */ 1792e73a819dSTao Ma if (split_rec->r_refcount && 1793e73a819dSTao Ma (split_rec->r_cpos != orig_rec->r_cpos && 1794e73a819dSTao Ma le64_to_cpu(split_rec->r_cpos) + 1795e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters) != 1796e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos) + le32_to_cpu(orig_rec->r_clusters))) 1797e73a819dSTao Ma recs_need++; 1798e73a819dSTao Ma 1799e73a819dSTao Ma /* If the leaf block don't have enough record, expand it. */ 1800e73a819dSTao Ma if (le16_to_cpu(rf_list->rl_used) + recs_need > rf_list->rl_count) { 1801e73a819dSTao Ma struct ocfs2_refcount_rec tmp_rec; 1802e73a819dSTao Ma u64 cpos = le64_to_cpu(orig_rec->r_cpos); 1803e73a819dSTao Ma len = le32_to_cpu(orig_rec->r_clusters); 1804e73a819dSTao Ma ret = ocfs2_expand_refcount_tree(handle, ci, ref_root_bh, 1805e73a819dSTao Ma ref_leaf_bh, meta_ac); 1806e73a819dSTao Ma if (ret) { 1807e73a819dSTao Ma mlog_errno(ret); 1808e73a819dSTao Ma goto out; 1809e73a819dSTao Ma } 1810e73a819dSTao Ma 1811e73a819dSTao Ma /* 1812e73a819dSTao Ma * We have to re-get it since now cpos may be moved to 1813e73a819dSTao Ma * another leaf block. 1814e73a819dSTao Ma */ 1815e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1816e73a819dSTao Ma cpos, len, &tmp_rec, &index, 1817e73a819dSTao Ma &new_bh); 1818e73a819dSTao Ma if (ret) { 1819e73a819dSTao Ma mlog_errno(ret); 1820e73a819dSTao Ma goto out; 1821e73a819dSTao Ma } 1822e73a819dSTao Ma 1823e73a819dSTao Ma ref_leaf_bh = new_bh; 1824e73a819dSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 1825e73a819dSTao Ma rf_list = &rb->rf_records; 1826e73a819dSTao Ma orig_rec = &rf_list->rl_recs[index]; 1827e73a819dSTao Ma } 1828e73a819dSTao Ma 1829e73a819dSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_leaf_bh, 1830e73a819dSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 1831e73a819dSTao Ma if (ret) { 1832e73a819dSTao Ma mlog_errno(ret); 1833e73a819dSTao Ma goto out; 1834e73a819dSTao Ma } 1835e73a819dSTao Ma 1836e73a819dSTao Ma /* 1837e73a819dSTao Ma * We have calculated out how many new records we need and store 1838e73a819dSTao Ma * in recs_need, so spare enough space first by moving the records 1839e73a819dSTao Ma * after "index" to the end. 1840e73a819dSTao Ma */ 1841e73a819dSTao Ma if (index != le16_to_cpu(rf_list->rl_used) - 1) 1842e73a819dSTao Ma memmove(&rf_list->rl_recs[index + 1 + recs_need], 1843e73a819dSTao Ma &rf_list->rl_recs[index + 1], 1844e73a819dSTao Ma (le16_to_cpu(rf_list->rl_used) - index - 1) * 1845e73a819dSTao Ma sizeof(struct ocfs2_refcount_rec)); 1846e73a819dSTao Ma 1847e73a819dSTao Ma len = (le64_to_cpu(orig_rec->r_cpos) + 1848e73a819dSTao Ma le32_to_cpu(orig_rec->r_clusters)) - 1849e73a819dSTao Ma (le64_to_cpu(split_rec->r_cpos) + 1850e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters)); 1851e73a819dSTao Ma 1852e73a819dSTao Ma /* 1853e73a819dSTao Ma * If we have "len", the we will split in the tail and move it 1854e73a819dSTao Ma * to the end of the space we have just spared. 1855e73a819dSTao Ma */ 1856e73a819dSTao Ma if (len) { 1857e73a819dSTao Ma tail_rec = &rf_list->rl_recs[index + recs_need]; 1858e73a819dSTao Ma 1859e73a819dSTao Ma memcpy(tail_rec, orig_rec, sizeof(struct ocfs2_refcount_rec)); 1860e73a819dSTao Ma le64_add_cpu(&tail_rec->r_cpos, 1861e73a819dSTao Ma le32_to_cpu(tail_rec->r_clusters) - len); 1862e73a819dSTao Ma tail_rec->r_clusters = le32_to_cpu(len); 1863e73a819dSTao Ma } 1864e73a819dSTao Ma 1865e73a819dSTao Ma /* 1866e73a819dSTao Ma * If the split pos isn't the same as the original one, we need to 1867e73a819dSTao Ma * split in the head. 1868e73a819dSTao Ma * 1869e73a819dSTao Ma * Note: We have the chance that split_rec.r_refcount = 0, 1870e73a819dSTao Ma * recs_need = 0 and len > 0, which means we just cut the head from 1871e73a819dSTao Ma * the orig_rec and in that case we have done some modification in 1872e73a819dSTao Ma * orig_rec above, so the check for r_cpos is faked. 1873e73a819dSTao Ma */ 1874e73a819dSTao Ma if (split_rec->r_cpos != orig_rec->r_cpos && tail_rec != orig_rec) { 1875e73a819dSTao Ma len = le64_to_cpu(split_rec->r_cpos) - 1876e73a819dSTao Ma le64_to_cpu(orig_rec->r_cpos); 1877e73a819dSTao Ma orig_rec->r_clusters = cpu_to_le32(len); 1878e73a819dSTao Ma index++; 1879e73a819dSTao Ma } 1880e73a819dSTao Ma 1881e73a819dSTao Ma le16_add_cpu(&rf_list->rl_used, recs_need); 1882e73a819dSTao Ma 1883e73a819dSTao Ma if (split_rec->r_refcount) { 1884e73a819dSTao Ma rf_list->rl_recs[index] = *split_rec; 1885e73a819dSTao Ma mlog(0, "insert refcount record start %llu, len %u, count %u " 1886e73a819dSTao Ma "to leaf block %llu at index %d\n", 1887e73a819dSTao Ma (unsigned long long)le64_to_cpu(split_rec->r_cpos), 1888e73a819dSTao Ma le32_to_cpu(split_rec->r_clusters), 1889e73a819dSTao Ma le32_to_cpu(split_rec->r_refcount), 1890e73a819dSTao Ma (unsigned long long)ref_leaf_bh->b_blocknr, index); 1891e73a819dSTao Ma 18927540c1a7STao Ma if (merge) 1893e73a819dSTao Ma ocfs2_refcount_rec_merge(rb, index); 1894e73a819dSTao Ma } 1895e73a819dSTao Ma 1896e73a819dSTao Ma ret = ocfs2_journal_dirty(handle, ref_leaf_bh); 1897e73a819dSTao Ma if (ret) 1898e73a819dSTao Ma mlog_errno(ret); 1899e73a819dSTao Ma 1900e73a819dSTao Ma out: 1901e73a819dSTao Ma brelse(new_bh); 1902e73a819dSTao Ma return ret; 1903e73a819dSTao Ma } 1904e73a819dSTao Ma 19057540c1a7STao Ma static int __ocfs2_increase_refcount(handle_t *handle, 1906e73a819dSTao Ma struct ocfs2_caching_info *ci, 1907e73a819dSTao Ma struct buffer_head *ref_root_bh, 19087540c1a7STao Ma u64 cpos, u32 len, int merge, 1909e73a819dSTao Ma struct ocfs2_alloc_context *meta_ac, 1910e73a819dSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 1911e73a819dSTao Ma { 1912e73a819dSTao Ma int ret = 0, index; 1913e73a819dSTao Ma struct buffer_head *ref_leaf_bh = NULL; 1914e73a819dSTao Ma struct ocfs2_refcount_rec rec; 1915e73a819dSTao Ma unsigned int set_len = 0; 1916e73a819dSTao Ma 1917e73a819dSTao Ma mlog(0, "Tree owner %llu, add refcount start %llu, len %u\n", 1918e73a819dSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 1919e73a819dSTao Ma (unsigned long long)cpos, len); 1920e73a819dSTao Ma 1921e73a819dSTao Ma while (len) { 1922e73a819dSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 1923e73a819dSTao Ma cpos, len, &rec, &index, 1924e73a819dSTao Ma &ref_leaf_bh); 1925e73a819dSTao Ma if (ret) { 1926e73a819dSTao Ma mlog_errno(ret); 1927e73a819dSTao Ma goto out; 1928e73a819dSTao Ma } 1929e73a819dSTao Ma 1930e73a819dSTao Ma set_len = le32_to_cpu(rec.r_clusters); 1931e73a819dSTao Ma 1932e73a819dSTao Ma /* 1933e73a819dSTao Ma * Here we may meet with 3 situations: 1934e73a819dSTao Ma * 1935e73a819dSTao Ma * 1. If we find an already existing record, and the length 1936e73a819dSTao Ma * is the same, cool, we just need to increase the r_refcount 1937e73a819dSTao Ma * and it is OK. 1938e73a819dSTao Ma * 2. If we find a hole, just insert it with r_refcount = 1. 1939e73a819dSTao Ma * 3. If we are in the middle of one extent record, split 1940e73a819dSTao Ma * it. 1941e73a819dSTao Ma */ 1942e73a819dSTao Ma if (rec.r_refcount && le64_to_cpu(rec.r_cpos) == cpos && 1943e73a819dSTao Ma set_len <= len) { 1944e73a819dSTao Ma mlog(0, "increase refcount rec, start %llu, len %u, " 1945e73a819dSTao Ma "count %u\n", (unsigned long long)cpos, set_len, 1946e73a819dSTao Ma le32_to_cpu(rec.r_refcount)); 1947e73a819dSTao Ma ret = ocfs2_change_refcount_rec(handle, ci, 19487540c1a7STao Ma ref_leaf_bh, index, 19497540c1a7STao Ma merge, 1); 1950e73a819dSTao Ma if (ret) { 1951e73a819dSTao Ma mlog_errno(ret); 1952e73a819dSTao Ma goto out; 1953e73a819dSTao Ma } 1954e73a819dSTao Ma } else if (!rec.r_refcount) { 1955e73a819dSTao Ma rec.r_refcount = cpu_to_le32(1); 1956e73a819dSTao Ma 1957e73a819dSTao Ma mlog(0, "insert refcount rec, start %llu, len %u\n", 1958e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 1959e73a819dSTao Ma set_len); 1960e73a819dSTao Ma ret = ocfs2_insert_refcount_rec(handle, ci, ref_root_bh, 1961e73a819dSTao Ma ref_leaf_bh, 19627540c1a7STao Ma &rec, index, 19637540c1a7STao Ma merge, meta_ac); 1964e73a819dSTao Ma if (ret) { 1965e73a819dSTao Ma mlog_errno(ret); 1966e73a819dSTao Ma goto out; 1967e73a819dSTao Ma } 1968e73a819dSTao Ma } else { 1969e73a819dSTao Ma set_len = min((u64)(cpos + len), 1970e73a819dSTao Ma le64_to_cpu(rec.r_cpos) + set_len) - cpos; 1971e73a819dSTao Ma rec.r_cpos = cpu_to_le64(cpos); 1972e73a819dSTao Ma rec.r_clusters = cpu_to_le32(set_len); 1973e73a819dSTao Ma le32_add_cpu(&rec.r_refcount, 1); 1974e73a819dSTao Ma 1975e73a819dSTao Ma mlog(0, "split refcount rec, start %llu, " 1976e73a819dSTao Ma "len %u, count %u\n", 1977e73a819dSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 1978e73a819dSTao Ma set_len, le32_to_cpu(rec.r_refcount)); 1979e73a819dSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 1980e73a819dSTao Ma ref_root_bh, ref_leaf_bh, 19817540c1a7STao Ma &rec, index, merge, 1982e73a819dSTao Ma meta_ac, dealloc); 1983e73a819dSTao Ma if (ret) { 1984e73a819dSTao Ma mlog_errno(ret); 1985e73a819dSTao Ma goto out; 1986e73a819dSTao Ma } 1987e73a819dSTao Ma } 1988e73a819dSTao Ma 1989e73a819dSTao Ma cpos += set_len; 1990e73a819dSTao Ma len -= set_len; 1991e73a819dSTao Ma brelse(ref_leaf_bh); 1992e73a819dSTao Ma ref_leaf_bh = NULL; 1993e73a819dSTao Ma } 1994e73a819dSTao Ma 1995e73a819dSTao Ma out: 1996e73a819dSTao Ma brelse(ref_leaf_bh); 1997e73a819dSTao Ma return ret; 1998e73a819dSTao Ma } 19991823cb0bSTao Ma 20001823cb0bSTao Ma static int ocfs2_remove_refcount_extent(handle_t *handle, 20011823cb0bSTao Ma struct ocfs2_caching_info *ci, 20021823cb0bSTao Ma struct buffer_head *ref_root_bh, 20031823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 20041823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 20051823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 20061823cb0bSTao Ma { 20071823cb0bSTao Ma int ret; 20081823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 20091823cb0bSTao Ma struct ocfs2_refcount_block *rb = 20101823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 20111823cb0bSTao Ma struct ocfs2_extent_tree et; 20121823cb0bSTao Ma 20131823cb0bSTao Ma BUG_ON(rb->rf_records.rl_used); 20141823cb0bSTao Ma 20151823cb0bSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 20161823cb0bSTao Ma ret = ocfs2_remove_extent(handle, &et, le32_to_cpu(rb->rf_cpos), 20171823cb0bSTao Ma 1, meta_ac, dealloc); 20181823cb0bSTao Ma if (ret) { 20191823cb0bSTao Ma mlog_errno(ret); 20201823cb0bSTao Ma goto out; 20211823cb0bSTao Ma } 20221823cb0bSTao Ma 20231823cb0bSTao Ma ocfs2_remove_from_cache(ci, ref_leaf_bh); 20241823cb0bSTao Ma 20251823cb0bSTao Ma /* 20261823cb0bSTao Ma * add the freed block to the dealloc so that it will be freed 20271823cb0bSTao Ma * when we run dealloc. 20281823cb0bSTao Ma */ 20291823cb0bSTao Ma ret = ocfs2_cache_block_dealloc(dealloc, EXTENT_ALLOC_SYSTEM_INODE, 20301823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_slot), 20311823cb0bSTao Ma le64_to_cpu(rb->rf_blkno), 20321823cb0bSTao Ma le16_to_cpu(rb->rf_suballoc_bit)); 20331823cb0bSTao Ma if (ret) { 20341823cb0bSTao Ma mlog_errno(ret); 20351823cb0bSTao Ma goto out; 20361823cb0bSTao Ma } 20371823cb0bSTao Ma 20381823cb0bSTao Ma ret = ocfs2_journal_access_rb(handle, ci, ref_root_bh, 20391823cb0bSTao Ma OCFS2_JOURNAL_ACCESS_WRITE); 20401823cb0bSTao Ma if (ret) { 20411823cb0bSTao Ma mlog_errno(ret); 20421823cb0bSTao Ma goto out; 20431823cb0bSTao Ma } 20441823cb0bSTao Ma 20451823cb0bSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 20461823cb0bSTao Ma 20471823cb0bSTao Ma le32_add_cpu(&rb->rf_clusters, -1); 20481823cb0bSTao Ma 20491823cb0bSTao Ma /* 20501823cb0bSTao Ma * check whether we need to restore the root refcount block if 20511823cb0bSTao Ma * there is no leaf extent block at atll. 20521823cb0bSTao Ma */ 20531823cb0bSTao Ma if (!rb->rf_list.l_next_free_rec) { 20541823cb0bSTao Ma BUG_ON(rb->rf_clusters); 20551823cb0bSTao Ma 20561823cb0bSTao Ma mlog(0, "reset refcount tree root %llu to be a record block.\n", 20571823cb0bSTao Ma (unsigned long long)ref_root_bh->b_blocknr); 20581823cb0bSTao Ma 20591823cb0bSTao Ma rb->rf_flags = 0; 20601823cb0bSTao Ma rb->rf_parent = 0; 20611823cb0bSTao Ma rb->rf_cpos = 0; 20621823cb0bSTao Ma memset(&rb->rf_records, 0, sb->s_blocksize - 20631823cb0bSTao Ma offsetof(struct ocfs2_refcount_block, rf_records)); 20641823cb0bSTao Ma rb->rf_records.rl_count = 20651823cb0bSTao Ma cpu_to_le16(ocfs2_refcount_recs_per_rb(sb)); 20661823cb0bSTao Ma } 20671823cb0bSTao Ma 20681823cb0bSTao Ma ocfs2_journal_dirty(handle, ref_root_bh); 20691823cb0bSTao Ma 20701823cb0bSTao Ma out: 20711823cb0bSTao Ma return ret; 20721823cb0bSTao Ma } 20731823cb0bSTao Ma 20747540c1a7STao Ma int ocfs2_increase_refcount(handle_t *handle, 20757540c1a7STao Ma struct ocfs2_caching_info *ci, 20767540c1a7STao Ma struct buffer_head *ref_root_bh, 20777540c1a7STao Ma u64 cpos, u32 len, 20787540c1a7STao Ma struct ocfs2_alloc_context *meta_ac, 20797540c1a7STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 20807540c1a7STao Ma { 20817540c1a7STao Ma return __ocfs2_increase_refcount(handle, ci, ref_root_bh, 20827540c1a7STao Ma cpos, len, 1, 20837540c1a7STao Ma meta_ac, dealloc); 20847540c1a7STao Ma } 20857540c1a7STao Ma 20861823cb0bSTao Ma static int ocfs2_decrease_refcount_rec(handle_t *handle, 20871823cb0bSTao Ma struct ocfs2_caching_info *ci, 20881823cb0bSTao Ma struct buffer_head *ref_root_bh, 20891823cb0bSTao Ma struct buffer_head *ref_leaf_bh, 20901823cb0bSTao Ma int index, u64 cpos, unsigned int len, 20911823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 20921823cb0bSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 20931823cb0bSTao Ma { 20941823cb0bSTao Ma int ret; 20951823cb0bSTao Ma struct ocfs2_refcount_block *rb = 20961823cb0bSTao Ma (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 20971823cb0bSTao Ma struct ocfs2_refcount_rec *rec = &rb->rf_records.rl_recs[index]; 20981823cb0bSTao Ma 20991823cb0bSTao Ma BUG_ON(cpos < le64_to_cpu(rec->r_cpos)); 21001823cb0bSTao Ma BUG_ON(cpos + len > 21011823cb0bSTao Ma le64_to_cpu(rec->r_cpos) + le32_to_cpu(rec->r_clusters)); 21021823cb0bSTao Ma 21031823cb0bSTao Ma if (cpos == le64_to_cpu(rec->r_cpos) && 21041823cb0bSTao Ma len == le32_to_cpu(rec->r_clusters)) 21051823cb0bSTao Ma ret = ocfs2_change_refcount_rec(handle, ci, 21067540c1a7STao Ma ref_leaf_bh, index, 1, -1); 21071823cb0bSTao Ma else { 21081823cb0bSTao Ma struct ocfs2_refcount_rec split = *rec; 21091823cb0bSTao Ma split.r_cpos = cpu_to_le64(cpos); 21101823cb0bSTao Ma split.r_clusters = cpu_to_le32(len); 21111823cb0bSTao Ma 21121823cb0bSTao Ma le32_add_cpu(&split.r_refcount, -1); 21131823cb0bSTao Ma 21141823cb0bSTao Ma mlog(0, "split refcount rec, start %llu, " 21151823cb0bSTao Ma "len %u, count %u, original start %llu, len %u\n", 21161823cb0bSTao Ma (unsigned long long)le64_to_cpu(split.r_cpos), 21171823cb0bSTao Ma len, le32_to_cpu(split.r_refcount), 21181823cb0bSTao Ma (unsigned long long)le64_to_cpu(rec->r_cpos), 21191823cb0bSTao Ma le32_to_cpu(rec->r_clusters)); 21201823cb0bSTao Ma ret = ocfs2_split_refcount_rec(handle, ci, 21211823cb0bSTao Ma ref_root_bh, ref_leaf_bh, 21227540c1a7STao Ma &split, index, 1, 21231823cb0bSTao Ma meta_ac, dealloc); 21241823cb0bSTao Ma } 21251823cb0bSTao Ma 21261823cb0bSTao Ma if (ret) { 21271823cb0bSTao Ma mlog_errno(ret); 21281823cb0bSTao Ma goto out; 21291823cb0bSTao Ma } 21301823cb0bSTao Ma 21311823cb0bSTao Ma /* Remove the leaf refcount block if it contains no refcount record. */ 21321823cb0bSTao Ma if (!rb->rf_records.rl_used && ref_leaf_bh != ref_root_bh) { 21331823cb0bSTao Ma ret = ocfs2_remove_refcount_extent(handle, ci, ref_root_bh, 21341823cb0bSTao Ma ref_leaf_bh, meta_ac, 21351823cb0bSTao Ma dealloc); 21361823cb0bSTao Ma if (ret) 21371823cb0bSTao Ma mlog_errno(ret); 21381823cb0bSTao Ma } 21391823cb0bSTao Ma 21401823cb0bSTao Ma out: 21411823cb0bSTao Ma return ret; 21421823cb0bSTao Ma } 21431823cb0bSTao Ma 21441823cb0bSTao Ma static int __ocfs2_decrease_refcount(handle_t *handle, 21451823cb0bSTao Ma struct ocfs2_caching_info *ci, 21461823cb0bSTao Ma struct buffer_head *ref_root_bh, 21471823cb0bSTao Ma u64 cpos, u32 len, 21481823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 21496ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 21506ae23c55STao Ma int delete) 21511823cb0bSTao Ma { 21521823cb0bSTao Ma int ret = 0, index = 0; 21531823cb0bSTao Ma struct ocfs2_refcount_rec rec; 21541823cb0bSTao Ma unsigned int r_count = 0, r_len; 21551823cb0bSTao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 21561823cb0bSTao Ma struct buffer_head *ref_leaf_bh = NULL; 21571823cb0bSTao Ma 21586ae23c55STao Ma mlog(0, "Tree owner %llu, decrease refcount start %llu, " 21596ae23c55STao Ma "len %u, delete %u\n", 21601823cb0bSTao Ma (unsigned long long)ocfs2_metadata_cache_owner(ci), 21616ae23c55STao Ma (unsigned long long)cpos, len, delete); 21621823cb0bSTao Ma 21631823cb0bSTao Ma while (len) { 21641823cb0bSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 21651823cb0bSTao Ma cpos, len, &rec, &index, 21661823cb0bSTao Ma &ref_leaf_bh); 21671823cb0bSTao Ma if (ret) { 21681823cb0bSTao Ma mlog_errno(ret); 21691823cb0bSTao Ma goto out; 21701823cb0bSTao Ma } 21711823cb0bSTao Ma 21721823cb0bSTao Ma r_count = le32_to_cpu(rec.r_refcount); 21731823cb0bSTao Ma BUG_ON(r_count == 0); 21746ae23c55STao Ma if (!delete) 21756ae23c55STao Ma BUG_ON(r_count > 1); 21761823cb0bSTao Ma 21771823cb0bSTao Ma r_len = min((u64)(cpos + len), le64_to_cpu(rec.r_cpos) + 21781823cb0bSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 21791823cb0bSTao Ma 21801823cb0bSTao Ma ret = ocfs2_decrease_refcount_rec(handle, ci, ref_root_bh, 21811823cb0bSTao Ma ref_leaf_bh, index, 21821823cb0bSTao Ma cpos, r_len, 21831823cb0bSTao Ma meta_ac, dealloc); 21841823cb0bSTao Ma if (ret) { 21851823cb0bSTao Ma mlog_errno(ret); 21861823cb0bSTao Ma goto out; 21871823cb0bSTao Ma } 21881823cb0bSTao Ma 21896ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1 && delete) { 21901823cb0bSTao Ma ret = ocfs2_cache_cluster_dealloc(dealloc, 21911823cb0bSTao Ma ocfs2_clusters_to_blocks(sb, cpos), 21921823cb0bSTao Ma r_len); 21931823cb0bSTao Ma if (ret) { 21941823cb0bSTao Ma mlog_errno(ret); 21951823cb0bSTao Ma goto out; 21961823cb0bSTao Ma } 21971823cb0bSTao Ma } 21981823cb0bSTao Ma 21991823cb0bSTao Ma cpos += r_len; 22001823cb0bSTao Ma len -= r_len; 22011823cb0bSTao Ma brelse(ref_leaf_bh); 22021823cb0bSTao Ma ref_leaf_bh = NULL; 22031823cb0bSTao Ma } 22041823cb0bSTao Ma 22051823cb0bSTao Ma out: 22061823cb0bSTao Ma brelse(ref_leaf_bh); 22071823cb0bSTao Ma return ret; 22081823cb0bSTao Ma } 22091823cb0bSTao Ma 22101823cb0bSTao Ma /* Caller must hold refcount tree lock. */ 22111823cb0bSTao Ma int ocfs2_decrease_refcount(struct inode *inode, 22121823cb0bSTao Ma handle_t *handle, u32 cpos, u32 len, 22131823cb0bSTao Ma struct ocfs2_alloc_context *meta_ac, 22146ae23c55STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 22156ae23c55STao Ma int delete) 22161823cb0bSTao Ma { 22171823cb0bSTao Ma int ret; 22181823cb0bSTao Ma u64 ref_blkno; 22191823cb0bSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 22201823cb0bSTao Ma struct buffer_head *ref_root_bh = NULL; 22211823cb0bSTao Ma struct ocfs2_refcount_tree *tree; 22221823cb0bSTao Ma 22231823cb0bSTao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 22241823cb0bSTao Ma 22251823cb0bSTao Ma ret = ocfs2_get_refcount_block(inode, &ref_blkno); 22261823cb0bSTao Ma if (ret) { 22271823cb0bSTao Ma mlog_errno(ret); 22281823cb0bSTao Ma goto out; 22291823cb0bSTao Ma } 22301823cb0bSTao Ma 22311823cb0bSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), ref_blkno, &tree); 22321823cb0bSTao Ma if (ret) { 22331823cb0bSTao Ma mlog_errno(ret); 22341823cb0bSTao Ma goto out; 22351823cb0bSTao Ma } 22361823cb0bSTao Ma 22371823cb0bSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, tree->rf_blkno, 22381823cb0bSTao Ma &ref_root_bh); 22391823cb0bSTao Ma if (ret) { 22401823cb0bSTao Ma mlog_errno(ret); 22411823cb0bSTao Ma goto out; 22421823cb0bSTao Ma } 22431823cb0bSTao Ma 22441823cb0bSTao Ma ret = __ocfs2_decrease_refcount(handle, &tree->rf_ci, ref_root_bh, 22456ae23c55STao Ma cpos, len, meta_ac, dealloc, delete); 22461823cb0bSTao Ma if (ret) 22471823cb0bSTao Ma mlog_errno(ret); 22481823cb0bSTao Ma out: 22491823cb0bSTao Ma brelse(ref_root_bh); 22501823cb0bSTao Ma return ret; 22511823cb0bSTao Ma } 22521aa75feaSTao Ma 22531aa75feaSTao Ma /* 22541aa75feaSTao Ma * Mark the already-existing extent at cpos as refcounted for len clusters. 22551aa75feaSTao Ma * This adds the refcount extent flag. 22561aa75feaSTao Ma * 22571aa75feaSTao Ma * If the existing extent is larger than the request, initiate a 22581aa75feaSTao Ma * split. An attempt will be made at merging with adjacent extents. 22591aa75feaSTao Ma * 22601aa75feaSTao Ma * The caller is responsible for passing down meta_ac if we'll need it. 22611aa75feaSTao Ma */ 22621aa75feaSTao Ma static int ocfs2_mark_extent_refcounted(struct inode *inode, 22631aa75feaSTao Ma struct ocfs2_extent_tree *et, 22641aa75feaSTao Ma handle_t *handle, u32 cpos, 22651aa75feaSTao Ma u32 len, u32 phys, 22661aa75feaSTao Ma struct ocfs2_alloc_context *meta_ac, 22671aa75feaSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 22681aa75feaSTao Ma { 22691aa75feaSTao Ma int ret; 22701aa75feaSTao Ma 22711aa75feaSTao Ma mlog(0, "Inode %lu refcount tree cpos %u, len %u, phys cluster %u\n", 22721aa75feaSTao Ma inode->i_ino, cpos, len, phys); 22731aa75feaSTao Ma 22741aa75feaSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 22751aa75feaSTao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 22761aa75feaSTao Ma "tree, but the feature bit is not set in the " 22771aa75feaSTao Ma "super block.", inode->i_ino); 22781aa75feaSTao Ma ret = -EROFS; 22791aa75feaSTao Ma goto out; 22801aa75feaSTao Ma } 22811aa75feaSTao Ma 22821aa75feaSTao Ma ret = ocfs2_change_extent_flag(handle, et, cpos, 22831aa75feaSTao Ma len, phys, meta_ac, dealloc, 22841aa75feaSTao Ma OCFS2_EXT_REFCOUNTED, 0); 22851aa75feaSTao Ma if (ret) 22861aa75feaSTao Ma mlog_errno(ret); 22871aa75feaSTao Ma 22881aa75feaSTao Ma out: 22891aa75feaSTao Ma return ret; 22901aa75feaSTao Ma } 2291bcbbb24aSTao Ma 2292bcbbb24aSTao Ma /* 2293bcbbb24aSTao Ma * Given some contiguous physical clusters, calculate what we need 2294bcbbb24aSTao Ma * for modifying their refcount. 2295bcbbb24aSTao Ma */ 2296bcbbb24aSTao Ma static int ocfs2_calc_refcount_meta_credits(struct super_block *sb, 2297bcbbb24aSTao Ma struct ocfs2_caching_info *ci, 2298bcbbb24aSTao Ma struct buffer_head *ref_root_bh, 2299bcbbb24aSTao Ma u64 start_cpos, 2300bcbbb24aSTao Ma u32 clusters, 2301bcbbb24aSTao Ma int *meta_add, 2302bcbbb24aSTao Ma int *credits) 2303bcbbb24aSTao Ma { 2304bcbbb24aSTao Ma int ret = 0, index, ref_blocks = 0, recs_add = 0; 2305bcbbb24aSTao Ma u64 cpos = start_cpos; 2306bcbbb24aSTao Ma struct ocfs2_refcount_block *rb; 2307bcbbb24aSTao Ma struct ocfs2_refcount_rec rec; 2308bcbbb24aSTao Ma struct buffer_head *ref_leaf_bh = NULL, *prev_bh = NULL; 2309bcbbb24aSTao Ma u32 len; 2310bcbbb24aSTao Ma 2311bcbbb24aSTao Ma mlog(0, "start_cpos %llu, clusters %u\n", 2312bcbbb24aSTao Ma (unsigned long long)start_cpos, clusters); 2313bcbbb24aSTao Ma while (clusters) { 2314bcbbb24aSTao Ma ret = ocfs2_get_refcount_rec(ci, ref_root_bh, 2315bcbbb24aSTao Ma cpos, clusters, &rec, 2316bcbbb24aSTao Ma &index, &ref_leaf_bh); 2317bcbbb24aSTao Ma if (ret) { 2318bcbbb24aSTao Ma mlog_errno(ret); 2319bcbbb24aSTao Ma goto out; 2320bcbbb24aSTao Ma } 2321bcbbb24aSTao Ma 2322bcbbb24aSTao Ma if (ref_leaf_bh != prev_bh) { 2323bcbbb24aSTao Ma /* 2324bcbbb24aSTao Ma * Now we encounter a new leaf block, so calculate 2325bcbbb24aSTao Ma * whether we need to extend the old leaf. 2326bcbbb24aSTao Ma */ 2327bcbbb24aSTao Ma if (prev_bh) { 2328bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *) 2329bcbbb24aSTao Ma prev_bh->b_data; 2330bcbbb24aSTao Ma 2331bcbbb24aSTao Ma if (le64_to_cpu(rb->rf_records.rl_used) + 2332bcbbb24aSTao Ma recs_add > 2333bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2334bcbbb24aSTao Ma ref_blocks++; 2335bcbbb24aSTao Ma } 2336bcbbb24aSTao Ma 2337bcbbb24aSTao Ma recs_add = 0; 2338bcbbb24aSTao Ma *credits += 1; 2339bcbbb24aSTao Ma brelse(prev_bh); 2340bcbbb24aSTao Ma prev_bh = ref_leaf_bh; 2341bcbbb24aSTao Ma get_bh(prev_bh); 2342bcbbb24aSTao Ma } 2343bcbbb24aSTao Ma 2344bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 2345bcbbb24aSTao Ma 2346bcbbb24aSTao Ma mlog(0, "recs_add %d,cpos %llu, clusters %u, rec->r_cpos %llu," 2347bcbbb24aSTao Ma "rec->r_clusters %u, rec->r_refcount %u, index %d\n", 2348bcbbb24aSTao Ma recs_add, (unsigned long long)cpos, clusters, 2349bcbbb24aSTao Ma (unsigned long long)le64_to_cpu(rec.r_cpos), 2350bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters), 2351bcbbb24aSTao Ma le32_to_cpu(rec.r_refcount), index); 2352bcbbb24aSTao Ma 2353bcbbb24aSTao Ma len = min((u64)cpos + clusters, le64_to_cpu(rec.r_cpos) + 2354bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) - cpos; 2355bcbbb24aSTao Ma /* 2356bcbbb24aSTao Ma * If the refcount rec already exist, cool. We just need 2357bcbbb24aSTao Ma * to check whether there is a split. Otherwise we just need 2358bcbbb24aSTao Ma * to increase the refcount. 2359bcbbb24aSTao Ma * If we will insert one, increases recs_add. 2360bcbbb24aSTao Ma * 2361bcbbb24aSTao Ma * We record all the records which will be inserted to the 2362bcbbb24aSTao Ma * same refcount block, so that we can tell exactly whether 2363bcbbb24aSTao Ma * we need a new refcount block or not. 2364bcbbb24aSTao Ma */ 2365bcbbb24aSTao Ma if (rec.r_refcount) { 2366bcbbb24aSTao Ma /* Check whether we need a split at the beginning. */ 2367bcbbb24aSTao Ma if (cpos == start_cpos && 2368bcbbb24aSTao Ma cpos != le64_to_cpu(rec.r_cpos)) 2369bcbbb24aSTao Ma recs_add++; 2370bcbbb24aSTao Ma 2371bcbbb24aSTao Ma /* Check whether we need a split in the end. */ 2372bcbbb24aSTao Ma if (cpos + clusters < le64_to_cpu(rec.r_cpos) + 2373bcbbb24aSTao Ma le32_to_cpu(rec.r_clusters)) 2374bcbbb24aSTao Ma recs_add++; 2375bcbbb24aSTao Ma } else 2376bcbbb24aSTao Ma recs_add++; 2377bcbbb24aSTao Ma 2378bcbbb24aSTao Ma brelse(ref_leaf_bh); 2379bcbbb24aSTao Ma ref_leaf_bh = NULL; 2380bcbbb24aSTao Ma clusters -= len; 2381bcbbb24aSTao Ma cpos += len; 2382bcbbb24aSTao Ma } 2383bcbbb24aSTao Ma 2384bcbbb24aSTao Ma if (prev_bh) { 2385bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)prev_bh->b_data; 2386bcbbb24aSTao Ma 2387bcbbb24aSTao Ma if (le64_to_cpu(rb->rf_records.rl_used) + recs_add > 2388bcbbb24aSTao Ma le16_to_cpu(rb->rf_records.rl_count)) 2389bcbbb24aSTao Ma ref_blocks++; 2390bcbbb24aSTao Ma 2391bcbbb24aSTao Ma *credits += 1; 2392bcbbb24aSTao Ma } 2393bcbbb24aSTao Ma 2394bcbbb24aSTao Ma if (!ref_blocks) 2395bcbbb24aSTao Ma goto out; 2396bcbbb24aSTao Ma 2397bcbbb24aSTao Ma mlog(0, "we need ref_blocks %d\n", ref_blocks); 2398bcbbb24aSTao Ma *meta_add += ref_blocks; 2399bcbbb24aSTao Ma *credits += ref_blocks; 2400bcbbb24aSTao Ma 2401bcbbb24aSTao Ma /* 2402bcbbb24aSTao Ma * So we may need ref_blocks to insert into the tree. 2403bcbbb24aSTao Ma * That also means we need to change the b-tree and add that number 2404bcbbb24aSTao Ma * of records since we never merge them. 2405bcbbb24aSTao Ma * We need one more block for expansion since the new created leaf 2406bcbbb24aSTao Ma * block is also full and needs split. 2407bcbbb24aSTao Ma */ 2408bcbbb24aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 2409bcbbb24aSTao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) { 2410bcbbb24aSTao Ma struct ocfs2_extent_tree et; 2411bcbbb24aSTao Ma 2412bcbbb24aSTao Ma ocfs2_init_refcount_extent_tree(&et, ci, ref_root_bh); 2413bcbbb24aSTao Ma *meta_add += ocfs2_extend_meta_needed(et.et_root_el); 2414bcbbb24aSTao Ma *credits += ocfs2_calc_extend_credits(sb, 2415bcbbb24aSTao Ma et.et_root_el, 2416bcbbb24aSTao Ma ref_blocks); 2417bcbbb24aSTao Ma } else { 2418bcbbb24aSTao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 2419bcbbb24aSTao Ma *meta_add += 1; 2420bcbbb24aSTao Ma } 2421bcbbb24aSTao Ma 2422bcbbb24aSTao Ma out: 2423bcbbb24aSTao Ma brelse(ref_leaf_bh); 2424bcbbb24aSTao Ma brelse(prev_bh); 2425bcbbb24aSTao Ma return ret; 2426bcbbb24aSTao Ma } 2427bcbbb24aSTao Ma 2428bcbbb24aSTao Ma /* 2429bcbbb24aSTao Ma * For refcount tree, we will decrease some contiguous clusters 2430bcbbb24aSTao Ma * refcount count, so just go through it to see how many blocks 2431bcbbb24aSTao Ma * we gonna touch and whether we need to create new blocks. 2432bcbbb24aSTao Ma * 2433bcbbb24aSTao Ma * Normally the refcount blocks store these refcount should be 2434bcbbb24aSTao Ma * continguous also, so that we can get the number easily. 2435bcbbb24aSTao Ma * As for meta_ac, we will at most add split 2 refcount record and 2436bcbbb24aSTao Ma * 2 more refcount block, so just check it in a rough way. 2437bcbbb24aSTao Ma * 2438bcbbb24aSTao Ma * Caller must hold refcount tree lock. 2439bcbbb24aSTao Ma */ 2440bcbbb24aSTao Ma int ocfs2_prepare_refcount_change_for_del(struct inode *inode, 2441bcbbb24aSTao Ma struct buffer_head *di_bh, 2442bcbbb24aSTao Ma u64 phys_blkno, 2443bcbbb24aSTao Ma u32 clusters, 2444bcbbb24aSTao Ma int *credits, 2445bcbbb24aSTao Ma struct ocfs2_alloc_context **meta_ac) 2446bcbbb24aSTao Ma { 2447bcbbb24aSTao Ma int ret, ref_blocks = 0; 2448bcbbb24aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 2449bcbbb24aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 2450bcbbb24aSTao Ma struct buffer_head *ref_root_bh = NULL; 2451bcbbb24aSTao Ma struct ocfs2_refcount_tree *tree; 2452bcbbb24aSTao Ma u64 start_cpos = ocfs2_blocks_to_clusters(inode->i_sb, phys_blkno); 2453bcbbb24aSTao Ma 2454bcbbb24aSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 2455bcbbb24aSTao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 2456bcbbb24aSTao Ma "tree, but the feature bit is not set in the " 2457bcbbb24aSTao Ma "super block.", inode->i_ino); 2458bcbbb24aSTao Ma ret = -EROFS; 2459bcbbb24aSTao Ma goto out; 2460bcbbb24aSTao Ma } 2461bcbbb24aSTao Ma 2462bcbbb24aSTao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 2463bcbbb24aSTao Ma 2464bcbbb24aSTao Ma ret = ocfs2_get_refcount_tree(OCFS2_SB(inode->i_sb), 2465bcbbb24aSTao Ma le64_to_cpu(di->i_refcount_loc), &tree); 2466bcbbb24aSTao Ma if (ret) { 2467bcbbb24aSTao Ma mlog_errno(ret); 2468bcbbb24aSTao Ma goto out; 2469bcbbb24aSTao Ma } 2470bcbbb24aSTao Ma 2471bcbbb24aSTao Ma ret = ocfs2_read_refcount_block(&tree->rf_ci, 2472bcbbb24aSTao Ma le64_to_cpu(di->i_refcount_loc), 2473bcbbb24aSTao Ma &ref_root_bh); 2474bcbbb24aSTao Ma if (ret) { 2475bcbbb24aSTao Ma mlog_errno(ret); 2476bcbbb24aSTao Ma goto out; 2477bcbbb24aSTao Ma } 2478bcbbb24aSTao Ma 2479bcbbb24aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 2480bcbbb24aSTao Ma &tree->rf_ci, 2481bcbbb24aSTao Ma ref_root_bh, 2482bcbbb24aSTao Ma start_cpos, clusters, 2483bcbbb24aSTao Ma &ref_blocks, credits); 2484bcbbb24aSTao Ma if (ret) { 2485bcbbb24aSTao Ma mlog_errno(ret); 2486bcbbb24aSTao Ma goto out; 2487bcbbb24aSTao Ma } 2488bcbbb24aSTao Ma 2489bcbbb24aSTao Ma mlog(0, "reserve new metadata %d, credits = %d\n", 2490bcbbb24aSTao Ma ref_blocks, *credits); 2491bcbbb24aSTao Ma 2492bcbbb24aSTao Ma if (ref_blocks) { 2493bcbbb24aSTao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb), 2494bcbbb24aSTao Ma ref_blocks, meta_ac); 2495bcbbb24aSTao Ma if (ret) 2496bcbbb24aSTao Ma mlog_errno(ret); 2497bcbbb24aSTao Ma } 2498bcbbb24aSTao Ma 2499bcbbb24aSTao Ma out: 2500bcbbb24aSTao Ma brelse(ref_root_bh); 2501bcbbb24aSTao Ma return ret; 2502bcbbb24aSTao Ma } 25036f70fa51STao Ma 25046f70fa51STao Ma #define MAX_CONTIG_BYTES 1048576 25056f70fa51STao Ma 25066f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_clusters(struct super_block *sb) 25076f70fa51STao Ma { 25086f70fa51STao Ma return ocfs2_clusters_for_bytes(sb, MAX_CONTIG_BYTES); 25096f70fa51STao Ma } 25106f70fa51STao Ma 25116f70fa51STao Ma static inline unsigned int ocfs2_cow_contig_mask(struct super_block *sb) 25126f70fa51STao Ma { 25136f70fa51STao Ma return ~(ocfs2_cow_contig_clusters(sb) - 1); 25146f70fa51STao Ma } 25156f70fa51STao Ma 25166f70fa51STao Ma /* 25176f70fa51STao Ma * Given an extent that starts at 'start' and an I/O that starts at 'cpos', 25186f70fa51STao Ma * find an offset (start + (n * contig_clusters)) that is closest to cpos 25196f70fa51STao Ma * while still being less than or equal to it. 25206f70fa51STao Ma * 25216f70fa51STao Ma * The goal is to break the extent at a multiple of contig_clusters. 25226f70fa51STao Ma */ 25236f70fa51STao Ma static inline unsigned int ocfs2_cow_align_start(struct super_block *sb, 25246f70fa51STao Ma unsigned int start, 25256f70fa51STao Ma unsigned int cpos) 25266f70fa51STao Ma { 25276f70fa51STao Ma BUG_ON(start > cpos); 25286f70fa51STao Ma 25296f70fa51STao Ma return start + ((cpos - start) & ocfs2_cow_contig_mask(sb)); 25306f70fa51STao Ma } 25316f70fa51STao Ma 25326f70fa51STao Ma /* 25336f70fa51STao Ma * Given a cluster count of len, pad it out so that it is a multiple 25346f70fa51STao Ma * of contig_clusters. 25356f70fa51STao Ma */ 25366f70fa51STao Ma static inline unsigned int ocfs2_cow_align_length(struct super_block *sb, 25376f70fa51STao Ma unsigned int len) 25386f70fa51STao Ma { 25396f70fa51STao Ma unsigned int padded = 25406f70fa51STao Ma (len + (ocfs2_cow_contig_clusters(sb) - 1)) & 25416f70fa51STao Ma ocfs2_cow_contig_mask(sb); 25426f70fa51STao Ma 25436f70fa51STao Ma /* Did we wrap? */ 25446f70fa51STao Ma if (padded < len) 25456f70fa51STao Ma padded = UINT_MAX; 25466f70fa51STao Ma 25476f70fa51STao Ma return padded; 25486f70fa51STao Ma } 25496f70fa51STao Ma 25506f70fa51STao Ma /* 25516f70fa51STao Ma * Calculate out the start and number of virtual clusters we need to to CoW. 25526f70fa51STao Ma * 25536f70fa51STao Ma * cpos is vitual start cluster position we want to do CoW in a 25546f70fa51STao Ma * file and write_len is the cluster length. 255537f8a2bfSTao Ma * max_cpos is the place where we want to stop CoW intentionally. 25566f70fa51STao Ma * 25576f70fa51STao Ma * Normal we will start CoW from the beginning of extent record cotaining cpos. 25586f70fa51STao Ma * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we 25596f70fa51STao Ma * get good I/O from the resulting extent tree. 25606f70fa51STao Ma */ 25616f70fa51STao Ma static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, 2562913580b4STao Ma struct ocfs2_extent_list *el, 25636f70fa51STao Ma u32 cpos, 25646f70fa51STao Ma u32 write_len, 256537f8a2bfSTao Ma u32 max_cpos, 25666f70fa51STao Ma u32 *cow_start, 25676f70fa51STao Ma u32 *cow_len) 25686f70fa51STao Ma { 25696f70fa51STao Ma int ret = 0; 25706f70fa51STao Ma int tree_height = le16_to_cpu(el->l_tree_depth), i; 25716f70fa51STao Ma struct buffer_head *eb_bh = NULL; 25726f70fa51STao Ma struct ocfs2_extent_block *eb = NULL; 25736f70fa51STao Ma struct ocfs2_extent_rec *rec; 25746f70fa51STao Ma unsigned int want_clusters, rec_end = 0; 25756f70fa51STao Ma int contig_clusters = ocfs2_cow_contig_clusters(inode->i_sb); 25766f70fa51STao Ma int leaf_clusters; 25776f70fa51STao Ma 257837f8a2bfSTao Ma BUG_ON(cpos + write_len > max_cpos); 257937f8a2bfSTao Ma 25806f70fa51STao Ma if (tree_height > 0) { 25816f70fa51STao Ma ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh); 25826f70fa51STao Ma if (ret) { 25836f70fa51STao Ma mlog_errno(ret); 25846f70fa51STao Ma goto out; 25856f70fa51STao Ma } 25866f70fa51STao Ma 25876f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 25886f70fa51STao Ma el = &eb->h_list; 25896f70fa51STao Ma 25906f70fa51STao Ma if (el->l_tree_depth) { 25916f70fa51STao Ma ocfs2_error(inode->i_sb, 25926f70fa51STao Ma "Inode %lu has non zero tree depth in " 25936f70fa51STao Ma "leaf block %llu\n", inode->i_ino, 25946f70fa51STao Ma (unsigned long long)eb_bh->b_blocknr); 25956f70fa51STao Ma ret = -EROFS; 25966f70fa51STao Ma goto out; 25976f70fa51STao Ma } 25986f70fa51STao Ma } 25996f70fa51STao Ma 26006f70fa51STao Ma *cow_len = 0; 26016f70fa51STao Ma for (i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 26026f70fa51STao Ma rec = &el->l_recs[i]; 26036f70fa51STao Ma 26046f70fa51STao Ma if (ocfs2_is_empty_extent(rec)) { 26056f70fa51STao Ma mlog_bug_on_msg(i != 0, "Inode %lu has empty record in " 26066f70fa51STao Ma "index %d\n", inode->i_ino, i); 26076f70fa51STao Ma continue; 26086f70fa51STao Ma } 26096f70fa51STao Ma 26106f70fa51STao Ma if (le32_to_cpu(rec->e_cpos) + 26116f70fa51STao Ma le16_to_cpu(rec->e_leaf_clusters) <= cpos) 26126f70fa51STao Ma continue; 26136f70fa51STao Ma 26146f70fa51STao Ma if (*cow_len == 0) { 26156f70fa51STao Ma /* 26166f70fa51STao Ma * We should find a refcounted record in the 26176f70fa51STao Ma * first pass. 26186f70fa51STao Ma */ 26196f70fa51STao Ma BUG_ON(!(rec->e_flags & OCFS2_EXT_REFCOUNTED)); 26206f70fa51STao Ma *cow_start = le32_to_cpu(rec->e_cpos); 26216f70fa51STao Ma } 26226f70fa51STao Ma 26236f70fa51STao Ma /* 262437f8a2bfSTao Ma * If we encounter a hole, a non-refcounted record or 262537f8a2bfSTao Ma * pass the max_cpos, stop the search. 26266f70fa51STao Ma */ 26276f70fa51STao Ma if ((!(rec->e_flags & OCFS2_EXT_REFCOUNTED)) || 262837f8a2bfSTao Ma (*cow_len && rec_end != le32_to_cpu(rec->e_cpos)) || 262937f8a2bfSTao Ma (max_cpos <= le32_to_cpu(rec->e_cpos))) 26306f70fa51STao Ma break; 26316f70fa51STao Ma 26326f70fa51STao Ma leaf_clusters = le16_to_cpu(rec->e_leaf_clusters); 26336f70fa51STao Ma rec_end = le32_to_cpu(rec->e_cpos) + leaf_clusters; 263437f8a2bfSTao Ma if (rec_end > max_cpos) { 263537f8a2bfSTao Ma rec_end = max_cpos; 263637f8a2bfSTao Ma leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos); 263737f8a2bfSTao Ma } 26386f70fa51STao Ma 26396f70fa51STao Ma /* 26406f70fa51STao Ma * How many clusters do we actually need from 26416f70fa51STao Ma * this extent? First we see how many we actually 26426f70fa51STao Ma * need to complete the write. If that's smaller 26436f70fa51STao Ma * than contig_clusters, we try for contig_clusters. 26446f70fa51STao Ma */ 26456f70fa51STao Ma if (!*cow_len) 26466f70fa51STao Ma want_clusters = write_len; 26476f70fa51STao Ma else 26486f70fa51STao Ma want_clusters = (cpos + write_len) - 26496f70fa51STao Ma (*cow_start + *cow_len); 26506f70fa51STao Ma if (want_clusters < contig_clusters) 26516f70fa51STao Ma want_clusters = contig_clusters; 26526f70fa51STao Ma 26536f70fa51STao Ma /* 26546f70fa51STao Ma * If the write does not cover the whole extent, we 26556f70fa51STao Ma * need to calculate how we're going to split the extent. 26566f70fa51STao Ma * We try to do it on contig_clusters boundaries. 26576f70fa51STao Ma * 26586f70fa51STao Ma * Any extent smaller than contig_clusters will be 26596f70fa51STao Ma * CoWed in its entirety. 26606f70fa51STao Ma */ 26616f70fa51STao Ma if (leaf_clusters <= contig_clusters) 26626f70fa51STao Ma *cow_len += leaf_clusters; 26636f70fa51STao Ma else if (*cow_len || (*cow_start == cpos)) { 26646f70fa51STao Ma /* 26656f70fa51STao Ma * This extent needs to be CoW'd from its 26666f70fa51STao Ma * beginning, so all we have to do is compute 26676f70fa51STao Ma * how many clusters to grab. We align 26686f70fa51STao Ma * want_clusters to the edge of contig_clusters 26696f70fa51STao Ma * to get better I/O. 26706f70fa51STao Ma */ 26716f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 26726f70fa51STao Ma want_clusters); 26736f70fa51STao Ma 26746f70fa51STao Ma if (leaf_clusters < want_clusters) 26756f70fa51STao Ma *cow_len += leaf_clusters; 26766f70fa51STao Ma else 26776f70fa51STao Ma *cow_len += want_clusters; 26786f70fa51STao Ma } else if ((*cow_start + contig_clusters) >= 26796f70fa51STao Ma (cpos + write_len)) { 26806f70fa51STao Ma /* 26816f70fa51STao Ma * Breaking off contig_clusters at the front 26826f70fa51STao Ma * of the extent will cover our write. That's 26836f70fa51STao Ma * easy. 26846f70fa51STao Ma */ 26856f70fa51STao Ma *cow_len = contig_clusters; 26866f70fa51STao Ma } else if ((rec_end - cpos) <= contig_clusters) { 26876f70fa51STao Ma /* 26886f70fa51STao Ma * Breaking off contig_clusters at the tail of 26896f70fa51STao Ma * this extent will cover cpos. 26906f70fa51STao Ma */ 26916f70fa51STao Ma *cow_start = rec_end - contig_clusters; 26926f70fa51STao Ma *cow_len = contig_clusters; 26936f70fa51STao Ma } else if ((rec_end - cpos) <= want_clusters) { 26946f70fa51STao Ma /* 26956f70fa51STao Ma * While we can't fit the entire write in this 26966f70fa51STao Ma * extent, we know that the write goes from cpos 26976f70fa51STao Ma * to the end of the extent. Break that off. 26986f70fa51STao Ma * We try to break it at some multiple of 26996f70fa51STao Ma * contig_clusters from the front of the extent. 27006f70fa51STao Ma * Failing that (ie, cpos is within 27016f70fa51STao Ma * contig_clusters of the front), we'll CoW the 27026f70fa51STao Ma * entire extent. 27036f70fa51STao Ma */ 27046f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 27056f70fa51STao Ma *cow_start, cpos); 27066f70fa51STao Ma *cow_len = rec_end - *cow_start; 27076f70fa51STao Ma } else { 27086f70fa51STao Ma /* 27096f70fa51STao Ma * Ok, the entire write lives in the middle of 27106f70fa51STao Ma * this extent. Let's try to slice the extent up 27116f70fa51STao Ma * nicely. Optimally, our CoW region starts at 27126f70fa51STao Ma * m*contig_clusters from the beginning of the 27136f70fa51STao Ma * extent and goes for n*contig_clusters, 27146f70fa51STao Ma * covering the entire write. 27156f70fa51STao Ma */ 27166f70fa51STao Ma *cow_start = ocfs2_cow_align_start(inode->i_sb, 27176f70fa51STao Ma *cow_start, cpos); 27186f70fa51STao Ma 27196f70fa51STao Ma want_clusters = (cpos + write_len) - *cow_start; 27206f70fa51STao Ma want_clusters = ocfs2_cow_align_length(inode->i_sb, 27216f70fa51STao Ma want_clusters); 27226f70fa51STao Ma if (*cow_start + want_clusters <= rec_end) 27236f70fa51STao Ma *cow_len = want_clusters; 27246f70fa51STao Ma else 27256f70fa51STao Ma *cow_len = rec_end - *cow_start; 27266f70fa51STao Ma } 27276f70fa51STao Ma 27286f70fa51STao Ma /* Have we covered our entire write yet? */ 27296f70fa51STao Ma if ((*cow_start + *cow_len) >= (cpos + write_len)) 27306f70fa51STao Ma break; 27316f70fa51STao Ma 27326f70fa51STao Ma /* 27336f70fa51STao Ma * If we reach the end of the extent block and don't get enough 27346f70fa51STao Ma * clusters, continue with the next extent block if possible. 27356f70fa51STao Ma */ 27366f70fa51STao Ma if (i + 1 == le16_to_cpu(el->l_next_free_rec) && 27376f70fa51STao Ma eb && eb->h_next_leaf_blk) { 27386f70fa51STao Ma brelse(eb_bh); 27396f70fa51STao Ma eb_bh = NULL; 27406f70fa51STao Ma 27416f70fa51STao Ma ret = ocfs2_read_extent_block(INODE_CACHE(inode), 27426f70fa51STao Ma le64_to_cpu(eb->h_next_leaf_blk), 27436f70fa51STao Ma &eb_bh); 27446f70fa51STao Ma if (ret) { 27456f70fa51STao Ma mlog_errno(ret); 27466f70fa51STao Ma goto out; 27476f70fa51STao Ma } 27486f70fa51STao Ma 27496f70fa51STao Ma eb = (struct ocfs2_extent_block *) eb_bh->b_data; 27506f70fa51STao Ma el = &eb->h_list; 27516f70fa51STao Ma i = -1; 27526f70fa51STao Ma } 27536f70fa51STao Ma } 27546f70fa51STao Ma 27556f70fa51STao Ma out: 27566f70fa51STao Ma brelse(eb_bh); 27576f70fa51STao Ma return ret; 27586f70fa51STao Ma } 27596f70fa51STao Ma 27606f70fa51STao Ma /* 27616f70fa51STao Ma * Prepare meta_ac, data_ac and calculate credits when we want to add some 27626f70fa51STao Ma * num_clusters in data_tree "et" and change the refcount for the old 27636f70fa51STao Ma * clusters(starting form p_cluster) in the refcount tree. 27646f70fa51STao Ma * 27656f70fa51STao Ma * Note: 27666f70fa51STao Ma * 1. since we may split the old tree, so we at most will need num_clusters + 2 27676f70fa51STao Ma * more new leaf records. 27686f70fa51STao Ma * 2. In some case, we may not need to reserve new clusters(e.g, reflink), so 27696f70fa51STao Ma * just give data_ac = NULL. 27706f70fa51STao Ma */ 27716f70fa51STao Ma static int ocfs2_lock_refcount_allocators(struct super_block *sb, 27726f70fa51STao Ma u32 p_cluster, u32 num_clusters, 27736f70fa51STao Ma struct ocfs2_extent_tree *et, 27746f70fa51STao Ma struct ocfs2_caching_info *ref_ci, 27756f70fa51STao Ma struct buffer_head *ref_root_bh, 27766f70fa51STao Ma struct ocfs2_alloc_context **meta_ac, 27776f70fa51STao Ma struct ocfs2_alloc_context **data_ac, 27786f70fa51STao Ma int *credits) 27796f70fa51STao Ma { 27806f70fa51STao Ma int ret = 0, meta_add = 0; 27816f70fa51STao Ma int num_free_extents = ocfs2_num_free_extents(OCFS2_SB(sb), et); 27826f70fa51STao Ma 27836f70fa51STao Ma if (num_free_extents < 0) { 27846f70fa51STao Ma ret = num_free_extents; 27856f70fa51STao Ma mlog_errno(ret); 27866f70fa51STao Ma goto out; 27876f70fa51STao Ma } 27886f70fa51STao Ma 27896f70fa51STao Ma if (num_free_extents < num_clusters + 2) 27906f70fa51STao Ma meta_add = 27916f70fa51STao Ma ocfs2_extend_meta_needed(et->et_root_el); 27926f70fa51STao Ma 27936f70fa51STao Ma *credits += ocfs2_calc_extend_credits(sb, et->et_root_el, 27946f70fa51STao Ma num_clusters + 2); 27956f70fa51STao Ma 27966f70fa51STao Ma ret = ocfs2_calc_refcount_meta_credits(sb, ref_ci, ref_root_bh, 27976f70fa51STao Ma p_cluster, num_clusters, 27986f70fa51STao Ma &meta_add, credits); 27996f70fa51STao Ma if (ret) { 28006f70fa51STao Ma mlog_errno(ret); 28016f70fa51STao Ma goto out; 28026f70fa51STao Ma } 28036f70fa51STao Ma 28046f70fa51STao Ma mlog(0, "reserve new metadata %d, clusters %u, credits = %d\n", 28056f70fa51STao Ma meta_add, num_clusters, *credits); 28066f70fa51STao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(sb), meta_add, 28076f70fa51STao Ma meta_ac); 28086f70fa51STao Ma if (ret) { 28096f70fa51STao Ma mlog_errno(ret); 28106f70fa51STao Ma goto out; 28116f70fa51STao Ma } 28126f70fa51STao Ma 28136f70fa51STao Ma if (data_ac) { 28146f70fa51STao Ma ret = ocfs2_reserve_clusters(OCFS2_SB(sb), num_clusters, 28156f70fa51STao Ma data_ac); 28166f70fa51STao Ma if (ret) 28176f70fa51STao Ma mlog_errno(ret); 28186f70fa51STao Ma } 28196f70fa51STao Ma 28206f70fa51STao Ma out: 28216f70fa51STao Ma if (ret) { 28226f70fa51STao Ma if (*meta_ac) { 28236f70fa51STao Ma ocfs2_free_alloc_context(*meta_ac); 28246f70fa51STao Ma *meta_ac = NULL; 28256f70fa51STao Ma } 28266f70fa51STao Ma } 28276f70fa51STao Ma 28286f70fa51STao Ma return ret; 28296f70fa51STao Ma } 28306f70fa51STao Ma 28316f70fa51STao Ma static int ocfs2_clear_cow_buffer(handle_t *handle, struct buffer_head *bh) 28326f70fa51STao Ma { 28336f70fa51STao Ma BUG_ON(buffer_dirty(bh)); 28346f70fa51STao Ma 28356f70fa51STao Ma clear_buffer_mapped(bh); 28366f70fa51STao Ma 28376f70fa51STao Ma return 0; 28386f70fa51STao Ma } 28396f70fa51STao Ma 2840913580b4STao Ma static int ocfs2_duplicate_clusters_by_page(handle_t *handle, 28416f70fa51STao Ma struct ocfs2_cow_context *context, 28426f70fa51STao Ma u32 cpos, u32 old_cluster, 28436f70fa51STao Ma u32 new_cluster, u32 new_len) 28446f70fa51STao Ma { 28456f70fa51STao Ma int ret = 0, partial; 2846913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 28476f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(ci); 28486f70fa51STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 28496f70fa51STao Ma struct page *page; 28506f70fa51STao Ma pgoff_t page_index; 28516f70fa51STao Ma unsigned int from, to; 28526f70fa51STao Ma loff_t offset, end, map_end; 28536f70fa51STao Ma struct address_space *mapping = context->inode->i_mapping; 28546f70fa51STao Ma 28556f70fa51STao Ma mlog(0, "old_cluster %u, new %u, len %u at offset %u\n", old_cluster, 28566f70fa51STao Ma new_cluster, new_len, cpos); 28576f70fa51STao Ma 28586f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 28596f70fa51STao Ma end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits); 28606f70fa51STao Ma 28616f70fa51STao Ma while (offset < end) { 28626f70fa51STao Ma page_index = offset >> PAGE_CACHE_SHIFT; 28636f70fa51STao Ma map_end = (page_index + 1) << PAGE_CACHE_SHIFT; 28646f70fa51STao Ma if (map_end > end) 28656f70fa51STao Ma map_end = end; 28666f70fa51STao Ma 28676f70fa51STao Ma /* from, to is the offset within the page. */ 28686f70fa51STao Ma from = offset & (PAGE_CACHE_SIZE - 1); 28696f70fa51STao Ma to = PAGE_CACHE_SIZE; 28706f70fa51STao Ma if (map_end & (PAGE_CACHE_SIZE - 1)) 28716f70fa51STao Ma to = map_end & (PAGE_CACHE_SIZE - 1); 28726f70fa51STao Ma 28736f70fa51STao Ma page = grab_cache_page(mapping, page_index); 28746f70fa51STao Ma 28756f70fa51STao Ma /* This page can't be dirtied before we CoW it out. */ 28766f70fa51STao Ma BUG_ON(PageDirty(page)); 28776f70fa51STao Ma 28786f70fa51STao Ma if (!PageUptodate(page)) { 28796f70fa51STao Ma ret = block_read_full_page(page, ocfs2_get_block); 28806f70fa51STao Ma if (ret) { 28816f70fa51STao Ma mlog_errno(ret); 28826f70fa51STao Ma goto unlock; 28836f70fa51STao Ma } 28846f70fa51STao Ma lock_page(page); 28856f70fa51STao Ma } 28866f70fa51STao Ma 28876f70fa51STao Ma if (page_has_buffers(page)) { 28886f70fa51STao Ma ret = walk_page_buffers(handle, page_buffers(page), 28896f70fa51STao Ma from, to, &partial, 28906f70fa51STao Ma ocfs2_clear_cow_buffer); 28916f70fa51STao Ma if (ret) { 28926f70fa51STao Ma mlog_errno(ret); 28936f70fa51STao Ma goto unlock; 28946f70fa51STao Ma } 28956f70fa51STao Ma } 28966f70fa51STao Ma 28976f70fa51STao Ma ocfs2_map_and_dirty_page(context->inode, 28986f70fa51STao Ma handle, from, to, 28996f70fa51STao Ma page, 0, &new_block); 29006f70fa51STao Ma mark_page_accessed(page); 29016f70fa51STao Ma unlock: 29026f70fa51STao Ma unlock_page(page); 29036f70fa51STao Ma page_cache_release(page); 29046f70fa51STao Ma page = NULL; 29056f70fa51STao Ma offset = map_end; 29066f70fa51STao Ma if (ret) 29076f70fa51STao Ma break; 29086f70fa51STao Ma } 29096f70fa51STao Ma 29106f70fa51STao Ma return ret; 29116f70fa51STao Ma } 29126f70fa51STao Ma 2913492a8a33STao Ma static int ocfs2_duplicate_clusters_by_jbd(handle_t *handle, 2914492a8a33STao Ma struct ocfs2_cow_context *context, 2915492a8a33STao Ma u32 cpos, u32 old_cluster, 2916492a8a33STao Ma u32 new_cluster, u32 new_len) 2917492a8a33STao Ma { 2918492a8a33STao Ma int ret = 0; 2919492a8a33STao Ma struct super_block *sb = context->inode->i_sb; 2920492a8a33STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 2921492a8a33STao Ma int i, blocks = ocfs2_clusters_to_blocks(sb, new_len); 2922492a8a33STao Ma u64 old_block = ocfs2_clusters_to_blocks(sb, old_cluster); 2923492a8a33STao Ma u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster); 2924492a8a33STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 2925492a8a33STao Ma struct buffer_head *old_bh = NULL; 2926492a8a33STao Ma struct buffer_head *new_bh = NULL; 2927492a8a33STao Ma 2928492a8a33STao Ma mlog(0, "old_cluster %u, new %u, len %u\n", old_cluster, 2929492a8a33STao Ma new_cluster, new_len); 2930492a8a33STao Ma 2931492a8a33STao Ma for (i = 0; i < blocks; i++, old_block++, new_block++) { 2932492a8a33STao Ma new_bh = sb_getblk(osb->sb, new_block); 2933492a8a33STao Ma if (new_bh == NULL) { 2934492a8a33STao Ma ret = -EIO; 2935492a8a33STao Ma mlog_errno(ret); 2936492a8a33STao Ma break; 2937492a8a33STao Ma } 2938492a8a33STao Ma 2939492a8a33STao Ma ocfs2_set_new_buffer_uptodate(ci, new_bh); 2940492a8a33STao Ma 2941492a8a33STao Ma ret = ocfs2_read_block(ci, old_block, &old_bh, NULL); 2942492a8a33STao Ma if (ret) { 2943492a8a33STao Ma mlog_errno(ret); 2944492a8a33STao Ma break; 2945492a8a33STao Ma } 2946492a8a33STao Ma 2947492a8a33STao Ma ret = ocfs2_journal_access(handle, ci, new_bh, 2948492a8a33STao Ma OCFS2_JOURNAL_ACCESS_CREATE); 2949492a8a33STao Ma if (ret) { 2950492a8a33STao Ma mlog_errno(ret); 2951492a8a33STao Ma break; 2952492a8a33STao Ma } 2953492a8a33STao Ma 2954492a8a33STao Ma memcpy(new_bh->b_data, old_bh->b_data, sb->s_blocksize); 2955492a8a33STao Ma ret = ocfs2_journal_dirty(handle, new_bh); 2956492a8a33STao Ma if (ret) { 2957492a8a33STao Ma mlog_errno(ret); 2958492a8a33STao Ma break; 2959492a8a33STao Ma } 2960492a8a33STao Ma 2961492a8a33STao Ma brelse(new_bh); 2962492a8a33STao Ma brelse(old_bh); 2963492a8a33STao Ma new_bh = NULL; 2964492a8a33STao Ma old_bh = NULL; 2965492a8a33STao Ma } 2966492a8a33STao Ma 2967492a8a33STao Ma brelse(new_bh); 2968492a8a33STao Ma brelse(old_bh); 2969492a8a33STao Ma return ret; 2970492a8a33STao Ma } 2971492a8a33STao Ma 29726f70fa51STao Ma static int ocfs2_clear_ext_refcount(handle_t *handle, 29736f70fa51STao Ma struct ocfs2_extent_tree *et, 29746f70fa51STao Ma u32 cpos, u32 p_cluster, u32 len, 29756f70fa51STao Ma unsigned int ext_flags, 29766f70fa51STao Ma struct ocfs2_alloc_context *meta_ac, 29776f70fa51STao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 29786f70fa51STao Ma { 29796f70fa51STao Ma int ret, index; 29806f70fa51STao Ma struct ocfs2_extent_rec replace_rec; 29816f70fa51STao Ma struct ocfs2_path *path = NULL; 29826f70fa51STao Ma struct ocfs2_extent_list *el; 29836f70fa51STao Ma struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 29846f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(et->et_ci); 29856f70fa51STao Ma 29866f70fa51STao Ma mlog(0, "inode %llu cpos %u, len %u, p_cluster %u, ext_flags %u\n", 29876f70fa51STao Ma (unsigned long long)ino, cpos, len, p_cluster, ext_flags); 29886f70fa51STao Ma 29896f70fa51STao Ma memset(&replace_rec, 0, sizeof(replace_rec)); 29906f70fa51STao Ma replace_rec.e_cpos = cpu_to_le32(cpos); 29916f70fa51STao Ma replace_rec.e_leaf_clusters = cpu_to_le16(len); 29926f70fa51STao Ma replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(sb, 29936f70fa51STao Ma p_cluster)); 29946f70fa51STao Ma replace_rec.e_flags = ext_flags; 29956f70fa51STao Ma replace_rec.e_flags &= ~OCFS2_EXT_REFCOUNTED; 29966f70fa51STao Ma 29976f70fa51STao Ma path = ocfs2_new_path_from_et(et); 29986f70fa51STao Ma if (!path) { 29996f70fa51STao Ma ret = -ENOMEM; 30006f70fa51STao Ma mlog_errno(ret); 30016f70fa51STao Ma goto out; 30026f70fa51STao Ma } 30036f70fa51STao Ma 30046f70fa51STao Ma ret = ocfs2_find_path(et->et_ci, path, cpos); 30056f70fa51STao Ma if (ret) { 30066f70fa51STao Ma mlog_errno(ret); 30076f70fa51STao Ma goto out; 30086f70fa51STao Ma } 30096f70fa51STao Ma 30106f70fa51STao Ma el = path_leaf_el(path); 30116f70fa51STao Ma 30126f70fa51STao Ma index = ocfs2_search_extent_list(el, cpos); 30136f70fa51STao Ma if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) { 30146f70fa51STao Ma ocfs2_error(sb, 30156f70fa51STao Ma "Inode %llu has an extent at cpos %u which can no " 30166f70fa51STao Ma "longer be found.\n", 30176f70fa51STao Ma (unsigned long long)ino, cpos); 30186f70fa51STao Ma ret = -EROFS; 30196f70fa51STao Ma goto out; 30206f70fa51STao Ma } 30216f70fa51STao Ma 30226f70fa51STao Ma ret = ocfs2_split_extent(handle, et, path, index, 30236f70fa51STao Ma &replace_rec, meta_ac, dealloc); 30246f70fa51STao Ma if (ret) 30256f70fa51STao Ma mlog_errno(ret); 30266f70fa51STao Ma 30276f70fa51STao Ma out: 30286f70fa51STao Ma ocfs2_free_path(path); 30296f70fa51STao Ma return ret; 30306f70fa51STao Ma } 30316f70fa51STao Ma 30326f70fa51STao Ma static int ocfs2_replace_clusters(handle_t *handle, 30336f70fa51STao Ma struct ocfs2_cow_context *context, 30346f70fa51STao Ma u32 cpos, u32 old, 30356f70fa51STao Ma u32 new, u32 len, 30366f70fa51STao Ma unsigned int ext_flags) 30376f70fa51STao Ma { 30386f70fa51STao Ma int ret; 3039913580b4STao Ma struct ocfs2_caching_info *ci = context->data_et.et_ci; 30406f70fa51STao Ma u64 ino = ocfs2_metadata_cache_owner(ci); 30416f70fa51STao Ma 30426f70fa51STao Ma mlog(0, "inode %llu, cpos %u, old %u, new %u, len %u, ext_flags %u\n", 30436f70fa51STao Ma (unsigned long long)ino, cpos, old, new, len, ext_flags); 30446f70fa51STao Ma 30456f70fa51STao Ma /*If the old clusters is unwritten, no need to duplicate. */ 30466f70fa51STao Ma if (!(ext_flags & OCFS2_EXT_UNWRITTEN)) { 3047913580b4STao Ma ret = context->cow_duplicate_clusters(handle, context, cpos, 30486f70fa51STao Ma old, new, len); 30496f70fa51STao Ma if (ret) { 30506f70fa51STao Ma mlog_errno(ret); 30516f70fa51STao Ma goto out; 30526f70fa51STao Ma } 30536f70fa51STao Ma } 30546f70fa51STao Ma 3055913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, &context->data_et, 30566f70fa51STao Ma cpos, new, len, ext_flags, 30576f70fa51STao Ma context->meta_ac, &context->dealloc); 30586f70fa51STao Ma if (ret) 30596f70fa51STao Ma mlog_errno(ret); 30606f70fa51STao Ma out: 30616f70fa51STao Ma return ret; 30626f70fa51STao Ma } 30636f70fa51STao Ma 30646f70fa51STao Ma static int ocfs2_cow_sync_writeback(struct super_block *sb, 30656f70fa51STao Ma struct ocfs2_cow_context *context, 30666f70fa51STao Ma u32 cpos, u32 num_clusters) 30676f70fa51STao Ma { 30686f70fa51STao Ma int ret = 0; 30696f70fa51STao Ma loff_t offset, end, map_end; 30706f70fa51STao Ma pgoff_t page_index; 30716f70fa51STao Ma struct page *page; 30726f70fa51STao Ma 30736f70fa51STao Ma if (ocfs2_should_order_data(context->inode)) 30746f70fa51STao Ma return 0; 30756f70fa51STao Ma 30766f70fa51STao Ma offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits; 30776f70fa51STao Ma end = offset + (num_clusters << OCFS2_SB(sb)->s_clustersize_bits); 30786f70fa51STao Ma 30796f70fa51STao Ma ret = filemap_fdatawrite_range(context->inode->i_mapping, 30806f70fa51STao Ma offset, end - 1); 30816f70fa51STao Ma if (ret < 0) { 30826f70fa51STao Ma mlog_errno(ret); 30836f70fa51STao Ma return ret; 30846f70fa51STao Ma } 30856f70fa51STao Ma 30866f70fa51STao Ma while (offset < end) { 30876f70fa51STao Ma page_index = offset >> PAGE_CACHE_SHIFT; 30886f70fa51STao Ma map_end = (page_index + 1) << PAGE_CACHE_SHIFT; 30896f70fa51STao Ma if (map_end > end) 30906f70fa51STao Ma map_end = end; 30916f70fa51STao Ma 30926f70fa51STao Ma page = grab_cache_page(context->inode->i_mapping, page_index); 30936f70fa51STao Ma BUG_ON(!page); 30946f70fa51STao Ma 30956f70fa51STao Ma wait_on_page_writeback(page); 30966f70fa51STao Ma if (PageError(page)) { 30976f70fa51STao Ma ret = -EIO; 30986f70fa51STao Ma mlog_errno(ret); 30996f70fa51STao Ma } else 31006f70fa51STao Ma mark_page_accessed(page); 31016f70fa51STao Ma 31026f70fa51STao Ma unlock_page(page); 31036f70fa51STao Ma page_cache_release(page); 31046f70fa51STao Ma page = NULL; 31056f70fa51STao Ma offset = map_end; 31066f70fa51STao Ma if (ret) 31076f70fa51STao Ma break; 31086f70fa51STao Ma } 31096f70fa51STao Ma 31106f70fa51STao Ma return ret; 31116f70fa51STao Ma } 31126f70fa51STao Ma 3113913580b4STao Ma static int ocfs2_di_get_clusters(struct ocfs2_cow_context *context, 3114913580b4STao Ma u32 v_cluster, u32 *p_cluster, 3115913580b4STao Ma u32 *num_clusters, 3116913580b4STao Ma unsigned int *extent_flags) 3117913580b4STao Ma { 3118913580b4STao Ma return ocfs2_get_clusters(context->inode, v_cluster, p_cluster, 3119913580b4STao Ma num_clusters, extent_flags); 3120913580b4STao Ma } 3121913580b4STao Ma 31226f70fa51STao Ma static int ocfs2_make_clusters_writable(struct super_block *sb, 31236f70fa51STao Ma struct ocfs2_cow_context *context, 31246f70fa51STao Ma u32 cpos, u32 p_cluster, 31256f70fa51STao Ma u32 num_clusters, unsigned int e_flags) 31266f70fa51STao Ma { 31276ae23c55STao Ma int ret, delete, index, credits = 0; 31286f70fa51STao Ma u32 new_bit, new_len; 31296ae23c55STao Ma unsigned int set_len; 31306f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(sb); 31316f70fa51STao Ma handle_t *handle; 31326ae23c55STao Ma struct buffer_head *ref_leaf_bh = NULL; 3133913580b4STao Ma struct ocfs2_caching_info *ref_ci = &context->ref_tree->rf_ci; 31346ae23c55STao Ma struct ocfs2_refcount_rec rec; 31356ae23c55STao Ma 31366ae23c55STao Ma mlog(0, "cpos %u, p_cluster %u, num_clusters %u, e_flags %u\n", 31376ae23c55STao Ma cpos, p_cluster, num_clusters, e_flags); 31386f70fa51STao Ma 31396f70fa51STao Ma ret = ocfs2_lock_refcount_allocators(sb, p_cluster, num_clusters, 3140913580b4STao Ma &context->data_et, 3141913580b4STao Ma ref_ci, 31426f70fa51STao Ma context->ref_root_bh, 31436f70fa51STao Ma &context->meta_ac, 31446f70fa51STao Ma &context->data_ac, &credits); 31456f70fa51STao Ma if (ret) { 31466f70fa51STao Ma mlog_errno(ret); 31476f70fa51STao Ma return ret; 31486f70fa51STao Ma } 31496f70fa51STao Ma 3150492a8a33STao Ma if (context->post_refcount) 3151492a8a33STao Ma credits += context->post_refcount->credits; 3152492a8a33STao Ma 3153492a8a33STao Ma credits += context->extra_credits; 31546f70fa51STao Ma handle = ocfs2_start_trans(osb, credits); 31556f70fa51STao Ma if (IS_ERR(handle)) { 31566f70fa51STao Ma ret = PTR_ERR(handle); 31576f70fa51STao Ma mlog_errno(ret); 31586f70fa51STao Ma goto out; 31596f70fa51STao Ma } 31606f70fa51STao Ma 31616f70fa51STao Ma while (num_clusters) { 3162913580b4STao Ma ret = ocfs2_get_refcount_rec(ref_ci, context->ref_root_bh, 31636ae23c55STao Ma p_cluster, num_clusters, 31646ae23c55STao Ma &rec, &index, &ref_leaf_bh); 31656ae23c55STao Ma if (ret) { 31666ae23c55STao Ma mlog_errno(ret); 31676ae23c55STao Ma goto out_commit; 31686ae23c55STao Ma } 31696ae23c55STao Ma 31706ae23c55STao Ma BUG_ON(!rec.r_refcount); 31716ae23c55STao Ma set_len = min((u64)p_cluster + num_clusters, 31726ae23c55STao Ma le64_to_cpu(rec.r_cpos) + 31736ae23c55STao Ma le32_to_cpu(rec.r_clusters)) - p_cluster; 31746ae23c55STao Ma 31756ae23c55STao Ma /* 31766ae23c55STao Ma * There are many different situation here. 31776ae23c55STao Ma * 1. If refcount == 1, remove the flag and don't COW. 31786ae23c55STao Ma * 2. If refcount > 1, allocate clusters. 31796ae23c55STao Ma * Here we may not allocate r_len once at a time, so continue 31806ae23c55STao Ma * until we reach num_clusters. 31816ae23c55STao Ma */ 31826ae23c55STao Ma if (le32_to_cpu(rec.r_refcount) == 1) { 31836ae23c55STao Ma delete = 0; 3184913580b4STao Ma ret = ocfs2_clear_ext_refcount(handle, 3185913580b4STao Ma &context->data_et, 31866ae23c55STao Ma cpos, p_cluster, 31876ae23c55STao Ma set_len, e_flags, 31886ae23c55STao Ma context->meta_ac, 31896ae23c55STao Ma &context->dealloc); 31906ae23c55STao Ma if (ret) { 31916ae23c55STao Ma mlog_errno(ret); 31926ae23c55STao Ma goto out_commit; 31936ae23c55STao Ma } 31946ae23c55STao Ma } else { 31956ae23c55STao Ma delete = 1; 31966ae23c55STao Ma 31976ae23c55STao Ma ret = __ocfs2_claim_clusters(osb, handle, 31986ae23c55STao Ma context->data_ac, 31996ae23c55STao Ma 1, set_len, 32006f70fa51STao Ma &new_bit, &new_len); 32016f70fa51STao Ma if (ret) { 32026f70fa51STao Ma mlog_errno(ret); 32036f70fa51STao Ma goto out_commit; 32046f70fa51STao Ma } 32056f70fa51STao Ma 32066f70fa51STao Ma ret = ocfs2_replace_clusters(handle, context, 32076f70fa51STao Ma cpos, p_cluster, new_bit, 32086f70fa51STao Ma new_len, e_flags); 32096f70fa51STao Ma if (ret) { 32106f70fa51STao Ma mlog_errno(ret); 32116f70fa51STao Ma goto out_commit; 32126f70fa51STao Ma } 32136ae23c55STao Ma set_len = new_len; 32146f70fa51STao Ma } 32156f70fa51STao Ma 3216913580b4STao Ma ret = __ocfs2_decrease_refcount(handle, ref_ci, 32176f70fa51STao Ma context->ref_root_bh, 32186ae23c55STao Ma p_cluster, set_len, 32196f70fa51STao Ma context->meta_ac, 32206ae23c55STao Ma &context->dealloc, delete); 32216f70fa51STao Ma if (ret) { 32226f70fa51STao Ma mlog_errno(ret); 32236f70fa51STao Ma goto out_commit; 32246f70fa51STao Ma } 32256f70fa51STao Ma 32266ae23c55STao Ma cpos += set_len; 32276ae23c55STao Ma p_cluster += set_len; 32286ae23c55STao Ma num_clusters -= set_len; 32296ae23c55STao Ma brelse(ref_leaf_bh); 32306ae23c55STao Ma ref_leaf_bh = NULL; 32316ae23c55STao Ma } 32326ae23c55STao Ma 3233492a8a33STao Ma /* handle any post_cow action. */ 3234492a8a33STao Ma if (context->post_refcount && context->post_refcount->func) { 3235492a8a33STao Ma ret = context->post_refcount->func(context->inode, handle, 3236492a8a33STao Ma context->post_refcount->para); 3237492a8a33STao Ma if (ret) { 3238492a8a33STao Ma mlog_errno(ret); 3239492a8a33STao Ma goto out_commit; 3240492a8a33STao Ma } 3241492a8a33STao Ma } 3242492a8a33STao Ma 32436f70fa51STao Ma /* 32446f70fa51STao Ma * Here we should write the new page out first if we are 32456f70fa51STao Ma * in write-back mode. 32466f70fa51STao Ma */ 3247492a8a33STao Ma if (context->get_clusters == ocfs2_di_get_clusters) { 32486f70fa51STao Ma ret = ocfs2_cow_sync_writeback(sb, context, cpos, num_clusters); 32496f70fa51STao Ma if (ret) 32506f70fa51STao Ma mlog_errno(ret); 3251492a8a33STao Ma } 32526f70fa51STao Ma 32536f70fa51STao Ma out_commit: 32546f70fa51STao Ma ocfs2_commit_trans(osb, handle); 32556f70fa51STao Ma 32566f70fa51STao Ma out: 32576f70fa51STao Ma if (context->data_ac) { 32586f70fa51STao Ma ocfs2_free_alloc_context(context->data_ac); 32596f70fa51STao Ma context->data_ac = NULL; 32606f70fa51STao Ma } 32616f70fa51STao Ma if (context->meta_ac) { 32626f70fa51STao Ma ocfs2_free_alloc_context(context->meta_ac); 32636f70fa51STao Ma context->meta_ac = NULL; 32646f70fa51STao Ma } 32656ae23c55STao Ma brelse(ref_leaf_bh); 32666f70fa51STao Ma 32676f70fa51STao Ma return ret; 32686f70fa51STao Ma } 32696f70fa51STao Ma 3270913580b4STao Ma static int ocfs2_replace_cow(struct ocfs2_cow_context *context) 32716f70fa51STao Ma { 32726f70fa51STao Ma int ret = 0; 3273913580b4STao Ma struct inode *inode = context->inode; 3274913580b4STao Ma u32 cow_start = context->cow_start, cow_len = context->cow_len; 3275913580b4STao Ma u32 p_cluster, num_clusters; 32766f70fa51STao Ma unsigned int ext_flags; 32776f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 32786f70fa51STao Ma 32796f70fa51STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) { 32806f70fa51STao Ma ocfs2_error(inode->i_sb, "Inode %lu want to use refcount " 32816f70fa51STao Ma "tree, but the feature bit is not set in the " 32826f70fa51STao Ma "super block.", inode->i_ino); 32836f70fa51STao Ma return -EROFS; 32846f70fa51STao Ma } 32856f70fa51STao Ma 32866f70fa51STao Ma ocfs2_init_dealloc_ctxt(&context->dealloc); 32876f70fa51STao Ma 32886f70fa51STao Ma while (cow_len) { 3289913580b4STao Ma ret = context->get_clusters(context, cow_start, &p_cluster, 32906f70fa51STao Ma &num_clusters, &ext_flags); 32916f70fa51STao Ma if (ret) { 32926f70fa51STao Ma mlog_errno(ret); 32936f70fa51STao Ma break; 32946f70fa51STao Ma } 32956f70fa51STao Ma 32966f70fa51STao Ma BUG_ON(!(ext_flags & OCFS2_EXT_REFCOUNTED)); 32976f70fa51STao Ma 32986f70fa51STao Ma if (cow_len < num_clusters) 32996f70fa51STao Ma num_clusters = cow_len; 33006f70fa51STao Ma 33016f70fa51STao Ma ret = ocfs2_make_clusters_writable(inode->i_sb, context, 33026f70fa51STao Ma cow_start, p_cluster, 33036f70fa51STao Ma num_clusters, ext_flags); 33046f70fa51STao Ma if (ret) { 33056f70fa51STao Ma mlog_errno(ret); 33066f70fa51STao Ma break; 33076f70fa51STao Ma } 33086f70fa51STao Ma 33096f70fa51STao Ma cow_len -= num_clusters; 33106f70fa51STao Ma cow_start += num_clusters; 33116f70fa51STao Ma } 33126f70fa51STao Ma 33136f70fa51STao Ma if (ocfs2_dealloc_has_cluster(&context->dealloc)) { 33146f70fa51STao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 33156f70fa51STao Ma ocfs2_run_deallocs(osb, &context->dealloc); 33166f70fa51STao Ma } 33176f70fa51STao Ma 33186f70fa51STao Ma return ret; 33196f70fa51STao Ma } 33206f70fa51STao Ma 33216f70fa51STao Ma /* 332237f8a2bfSTao Ma * Starting at cpos, try to CoW write_len clusters. Don't CoW 332337f8a2bfSTao Ma * past max_cpos. This will stop when it runs into a hole or an 332437f8a2bfSTao Ma * unrefcounted extent. 33256f70fa51STao Ma */ 33266f70fa51STao Ma static int ocfs2_refcount_cow_hunk(struct inode *inode, 33276f70fa51STao Ma struct buffer_head *di_bh, 332837f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 33296f70fa51STao Ma { 33306f70fa51STao Ma int ret; 33316f70fa51STao Ma u32 cow_start = 0, cow_len = 0; 33326f70fa51STao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 33336f70fa51STao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 33346f70fa51STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 33356f70fa51STao Ma struct buffer_head *ref_root_bh = NULL; 33366f70fa51STao Ma struct ocfs2_refcount_tree *ref_tree; 3337913580b4STao Ma struct ocfs2_cow_context *context = NULL; 33386f70fa51STao Ma 33396f70fa51STao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 33406f70fa51STao Ma 3341913580b4STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &di->id2.i_list, 334237f8a2bfSTao Ma cpos, write_len, max_cpos, 33436f70fa51STao Ma &cow_start, &cow_len); 33446f70fa51STao Ma if (ret) { 33456f70fa51STao Ma mlog_errno(ret); 33466f70fa51STao Ma goto out; 33476f70fa51STao Ma } 334837f8a2bfSTao Ma 33496f70fa51STao Ma mlog(0, "CoW inode %lu, cpos %u, write_len %u, cow_start %u, " 33506f70fa51STao Ma "cow_len %u\n", inode->i_ino, 33516f70fa51STao Ma cpos, write_len, cow_start, cow_len); 33526f70fa51STao Ma 33536f70fa51STao Ma BUG_ON(cow_len == 0); 33546f70fa51STao Ma 3355913580b4STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3356913580b4STao Ma if (!context) { 3357913580b4STao Ma ret = -ENOMEM; 3358913580b4STao Ma mlog_errno(ret); 3359913580b4STao Ma goto out; 3360913580b4STao Ma } 3361913580b4STao Ma 33626f70fa51STao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 33636f70fa51STao Ma 1, &ref_tree, &ref_root_bh); 33646f70fa51STao Ma if (ret) { 33656f70fa51STao Ma mlog_errno(ret); 33666f70fa51STao Ma goto out; 33676f70fa51STao Ma } 33686f70fa51STao Ma 3369913580b4STao Ma context->inode = inode; 3370913580b4STao Ma context->cow_start = cow_start; 3371913580b4STao Ma context->cow_len = cow_len; 3372913580b4STao Ma context->ref_tree = ref_tree; 3373913580b4STao Ma context->ref_root_bh = ref_root_bh; 3374913580b4STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page; 3375913580b4STao Ma context->get_clusters = ocfs2_di_get_clusters; 3376913580b4STao Ma 3377913580b4STao Ma ocfs2_init_dinode_extent_tree(&context->data_et, 3378913580b4STao Ma INODE_CACHE(inode), di_bh); 3379913580b4STao Ma 3380913580b4STao Ma ret = ocfs2_replace_cow(context); 33816f70fa51STao Ma if (ret) 33826f70fa51STao Ma mlog_errno(ret); 33836f70fa51STao Ma 3384913580b4STao Ma /* 3385913580b4STao Ma * truncate the extent map here since no matter whether we meet with 3386913580b4STao Ma * any error during the action, we shouldn't trust cached extent map 3387913580b4STao Ma * any more. 3388913580b4STao Ma */ 3389913580b4STao Ma ocfs2_extent_map_trunc(inode, cow_start); 3390913580b4STao Ma 33916f70fa51STao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 33926f70fa51STao Ma brelse(ref_root_bh); 33936f70fa51STao Ma out: 3394913580b4STao Ma kfree(context); 33956f70fa51STao Ma return ret; 33966f70fa51STao Ma } 33976f70fa51STao Ma 33986f70fa51STao Ma /* 33996f70fa51STao Ma * CoW any and all clusters between cpos and cpos+write_len. 340037f8a2bfSTao Ma * Don't CoW past max_cpos. If this returns successfully, all 340137f8a2bfSTao Ma * clusters between cpos and cpos+write_len are safe to modify. 34026f70fa51STao Ma */ 34036f70fa51STao Ma int ocfs2_refcount_cow(struct inode *inode, 34046f70fa51STao Ma struct buffer_head *di_bh, 340537f8a2bfSTao Ma u32 cpos, u32 write_len, u32 max_cpos) 34066f70fa51STao Ma { 34076f70fa51STao Ma int ret = 0; 34086f70fa51STao Ma u32 p_cluster, num_clusters; 34096f70fa51STao Ma unsigned int ext_flags; 34106f70fa51STao Ma 34116f70fa51STao Ma while (write_len) { 34126f70fa51STao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 34136f70fa51STao Ma &num_clusters, &ext_flags); 34146f70fa51STao Ma if (ret) { 34156f70fa51STao Ma mlog_errno(ret); 34166f70fa51STao Ma break; 34176f70fa51STao Ma } 34186f70fa51STao Ma 34196f70fa51STao Ma if (write_len < num_clusters) 34206f70fa51STao Ma num_clusters = write_len; 34216f70fa51STao Ma 34226f70fa51STao Ma if (ext_flags & OCFS2_EXT_REFCOUNTED) { 34236f70fa51STao Ma ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos, 342437f8a2bfSTao Ma num_clusters, max_cpos); 34256f70fa51STao Ma if (ret) { 34266f70fa51STao Ma mlog_errno(ret); 34276f70fa51STao Ma break; 34286f70fa51STao Ma } 34296f70fa51STao Ma } 34306f70fa51STao Ma 34316f70fa51STao Ma write_len -= num_clusters; 34326f70fa51STao Ma cpos += num_clusters; 34336f70fa51STao Ma } 34346f70fa51STao Ma 34356f70fa51STao Ma return ret; 34366f70fa51STao Ma } 3437110a045aSTao Ma 3438492a8a33STao Ma static int ocfs2_xattr_value_get_clusters(struct ocfs2_cow_context *context, 3439492a8a33STao Ma u32 v_cluster, u32 *p_cluster, 3440492a8a33STao Ma u32 *num_clusters, 3441492a8a33STao Ma unsigned int *extent_flags) 3442492a8a33STao Ma { 3443492a8a33STao Ma struct inode *inode = context->inode; 3444492a8a33STao Ma struct ocfs2_xattr_value_root *xv = context->cow_object; 3445492a8a33STao Ma 3446492a8a33STao Ma return ocfs2_xattr_get_clusters(inode, v_cluster, p_cluster, 3447492a8a33STao Ma num_clusters, &xv->xr_list, 3448492a8a33STao Ma extent_flags); 3449492a8a33STao Ma } 3450492a8a33STao Ma 3451492a8a33STao Ma /* 3452492a8a33STao Ma * Given a xattr value root, calculate the most meta/credits we need for 3453492a8a33STao Ma * refcount tree change if we truncate it to 0. 3454492a8a33STao Ma */ 3455492a8a33STao Ma int ocfs2_refcounted_xattr_delete_need(struct inode *inode, 3456492a8a33STao Ma struct ocfs2_caching_info *ref_ci, 3457492a8a33STao Ma struct buffer_head *ref_root_bh, 3458492a8a33STao Ma struct ocfs2_xattr_value_root *xv, 3459492a8a33STao Ma int *meta_add, int *credits) 3460492a8a33STao Ma { 3461492a8a33STao Ma int ret = 0, index, ref_blocks = 0; 3462492a8a33STao Ma u32 p_cluster, num_clusters; 3463492a8a33STao Ma u32 cpos = 0, clusters = le32_to_cpu(xv->xr_clusters); 3464492a8a33STao Ma struct ocfs2_refcount_block *rb; 3465492a8a33STao Ma struct ocfs2_refcount_rec rec; 3466492a8a33STao Ma struct buffer_head *ref_leaf_bh = NULL; 3467492a8a33STao Ma 3468492a8a33STao Ma while (cpos < clusters) { 3469492a8a33STao Ma ret = ocfs2_xattr_get_clusters(inode, cpos, &p_cluster, 3470492a8a33STao Ma &num_clusters, &xv->xr_list, 3471492a8a33STao Ma NULL); 3472492a8a33STao Ma if (ret) { 3473492a8a33STao Ma mlog_errno(ret); 3474492a8a33STao Ma goto out; 3475492a8a33STao Ma } 3476492a8a33STao Ma 3477492a8a33STao Ma cpos += num_clusters; 3478492a8a33STao Ma 3479492a8a33STao Ma while (num_clusters) { 3480492a8a33STao Ma ret = ocfs2_get_refcount_rec(ref_ci, ref_root_bh, 3481492a8a33STao Ma p_cluster, num_clusters, 3482492a8a33STao Ma &rec, &index, 3483492a8a33STao Ma &ref_leaf_bh); 3484492a8a33STao Ma if (ret) { 3485492a8a33STao Ma mlog_errno(ret); 3486492a8a33STao Ma goto out; 3487492a8a33STao Ma } 3488492a8a33STao Ma 3489492a8a33STao Ma BUG_ON(!rec.r_refcount); 3490492a8a33STao Ma 3491492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_leaf_bh->b_data; 3492492a8a33STao Ma 3493492a8a33STao Ma /* 3494492a8a33STao Ma * We really don't know whether the other clusters is in 3495492a8a33STao Ma * this refcount block or not, so just take the worst 3496492a8a33STao Ma * case that all the clusters are in this block and each 3497492a8a33STao Ma * one will split a refcount rec, so totally we need 3498492a8a33STao Ma * clusters * 2 new refcount rec. 3499492a8a33STao Ma */ 3500492a8a33STao Ma if (le64_to_cpu(rb->rf_records.rl_used) + clusters * 2 > 3501492a8a33STao Ma le16_to_cpu(rb->rf_records.rl_count)) 3502492a8a33STao Ma ref_blocks++; 3503492a8a33STao Ma 3504492a8a33STao Ma *credits += 1; 3505492a8a33STao Ma brelse(ref_leaf_bh); 3506492a8a33STao Ma ref_leaf_bh = NULL; 3507492a8a33STao Ma 3508492a8a33STao Ma if (num_clusters <= le32_to_cpu(rec.r_clusters)) 3509492a8a33STao Ma break; 3510492a8a33STao Ma else 3511492a8a33STao Ma num_clusters -= le32_to_cpu(rec.r_clusters); 3512492a8a33STao Ma p_cluster += num_clusters; 3513492a8a33STao Ma } 3514492a8a33STao Ma } 3515492a8a33STao Ma 3516492a8a33STao Ma *meta_add += ref_blocks; 3517492a8a33STao Ma if (!ref_blocks) 3518492a8a33STao Ma goto out; 3519492a8a33STao Ma 3520492a8a33STao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 3521492a8a33STao Ma if (le32_to_cpu(rb->rf_flags) & OCFS2_REFCOUNT_TREE_FL) 3522492a8a33STao Ma *credits += OCFS2_EXPAND_REFCOUNT_TREE_CREDITS; 3523492a8a33STao Ma else { 3524492a8a33STao Ma struct ocfs2_extent_tree et; 3525492a8a33STao Ma 3526492a8a33STao Ma ocfs2_init_refcount_extent_tree(&et, ref_ci, ref_root_bh); 3527492a8a33STao Ma *credits += ocfs2_calc_extend_credits(inode->i_sb, 3528492a8a33STao Ma et.et_root_el, 3529492a8a33STao Ma ref_blocks); 3530492a8a33STao Ma } 3531492a8a33STao Ma 3532492a8a33STao Ma out: 3533492a8a33STao Ma brelse(ref_leaf_bh); 3534492a8a33STao Ma return ret; 3535492a8a33STao Ma } 3536492a8a33STao Ma 3537492a8a33STao Ma /* 3538492a8a33STao Ma * Do CoW for xattr. 3539492a8a33STao Ma */ 3540492a8a33STao Ma int ocfs2_refcount_cow_xattr(struct inode *inode, 3541492a8a33STao Ma struct ocfs2_dinode *di, 3542492a8a33STao Ma struct ocfs2_xattr_value_buf *vb, 3543492a8a33STao Ma struct ocfs2_refcount_tree *ref_tree, 3544492a8a33STao Ma struct buffer_head *ref_root_bh, 3545492a8a33STao Ma u32 cpos, u32 write_len, 3546492a8a33STao Ma struct ocfs2_post_refcount *post) 3547492a8a33STao Ma { 3548492a8a33STao Ma int ret; 3549492a8a33STao Ma struct ocfs2_xattr_value_root *xv = vb->vb_xv; 3550492a8a33STao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 3551492a8a33STao Ma struct ocfs2_cow_context *context = NULL; 3552492a8a33STao Ma u32 cow_start, cow_len; 3553492a8a33STao Ma 3554492a8a33STao Ma BUG_ON(!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)); 3555492a8a33STao Ma 3556492a8a33STao Ma ret = ocfs2_refcount_cal_cow_clusters(inode, &xv->xr_list, 3557492a8a33STao Ma cpos, write_len, UINT_MAX, 3558492a8a33STao Ma &cow_start, &cow_len); 3559492a8a33STao Ma if (ret) { 3560492a8a33STao Ma mlog_errno(ret); 3561492a8a33STao Ma goto out; 3562492a8a33STao Ma } 3563492a8a33STao Ma 3564492a8a33STao Ma BUG_ON(cow_len == 0); 3565492a8a33STao Ma 3566492a8a33STao Ma context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS); 3567492a8a33STao Ma if (!context) { 3568492a8a33STao Ma ret = -ENOMEM; 3569492a8a33STao Ma mlog_errno(ret); 3570492a8a33STao Ma goto out; 3571492a8a33STao Ma } 3572492a8a33STao Ma 3573492a8a33STao Ma context->inode = inode; 3574492a8a33STao Ma context->cow_start = cow_start; 3575492a8a33STao Ma context->cow_len = cow_len; 3576492a8a33STao Ma context->ref_tree = ref_tree; 3577492a8a33STao Ma context->ref_root_bh = ref_root_bh;; 3578492a8a33STao Ma context->cow_object = xv; 3579492a8a33STao Ma 3580492a8a33STao Ma context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_jbd; 3581492a8a33STao Ma /* We need the extra credits for duplicate_clusters by jbd. */ 3582492a8a33STao Ma context->extra_credits = 3583492a8a33STao Ma ocfs2_clusters_to_blocks(inode->i_sb, 1) * cow_len; 3584492a8a33STao Ma context->get_clusters = ocfs2_xattr_value_get_clusters; 3585492a8a33STao Ma context->post_refcount = post; 3586492a8a33STao Ma 3587492a8a33STao Ma ocfs2_init_xattr_value_extent_tree(&context->data_et, 3588492a8a33STao Ma INODE_CACHE(inode), vb); 3589492a8a33STao Ma 3590492a8a33STao Ma ret = ocfs2_replace_cow(context); 3591492a8a33STao Ma if (ret) 3592492a8a33STao Ma mlog_errno(ret); 3593492a8a33STao Ma 3594492a8a33STao Ma out: 3595492a8a33STao Ma kfree(context); 3596492a8a33STao Ma return ret; 3597492a8a33STao Ma } 3598492a8a33STao Ma 3599110a045aSTao Ma /* 3600110a045aSTao Ma * Insert a new extent into refcount tree and mark a extent rec 3601110a045aSTao Ma * as refcounted in the dinode tree. 3602110a045aSTao Ma */ 3603110a045aSTao Ma int ocfs2_add_refcount_flag(struct inode *inode, 3604110a045aSTao Ma struct ocfs2_extent_tree *data_et, 3605110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3606110a045aSTao Ma struct buffer_head *ref_root_bh, 3607110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 36080129241eSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc, 36090129241eSTao Ma struct ocfs2_post_refcount *post) 3610110a045aSTao Ma { 3611110a045aSTao Ma int ret; 3612110a045aSTao Ma handle_t *handle; 3613110a045aSTao Ma int credits = 1, ref_blocks = 0; 3614110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3615110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3616110a045aSTao Ma 3617110a045aSTao Ma ret = ocfs2_calc_refcount_meta_credits(inode->i_sb, 3618110a045aSTao Ma ref_ci, ref_root_bh, 3619110a045aSTao Ma p_cluster, num_clusters, 3620110a045aSTao Ma &ref_blocks, &credits); 3621110a045aSTao Ma if (ret) { 3622110a045aSTao Ma mlog_errno(ret); 3623110a045aSTao Ma goto out; 3624110a045aSTao Ma } 3625110a045aSTao Ma 3626110a045aSTao Ma mlog(0, "reserve new metadata %d, credits = %d\n", 3627110a045aSTao Ma ref_blocks, credits); 3628110a045aSTao Ma 3629110a045aSTao Ma if (ref_blocks) { 3630110a045aSTao Ma ret = ocfs2_reserve_new_metadata_blocks(OCFS2_SB(inode->i_sb), 3631110a045aSTao Ma ref_blocks, &meta_ac); 3632110a045aSTao Ma if (ret) { 3633110a045aSTao Ma mlog_errno(ret); 3634110a045aSTao Ma goto out; 3635110a045aSTao Ma } 3636110a045aSTao Ma } 3637110a045aSTao Ma 36380129241eSTao Ma if (post) 36390129241eSTao Ma credits += post->credits; 36400129241eSTao Ma 3641110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3642110a045aSTao Ma if (IS_ERR(handle)) { 3643110a045aSTao Ma ret = PTR_ERR(handle); 3644110a045aSTao Ma mlog_errno(ret); 3645110a045aSTao Ma goto out; 3646110a045aSTao Ma } 3647110a045aSTao Ma 3648110a045aSTao Ma ret = ocfs2_mark_extent_refcounted(inode, data_et, handle, 3649110a045aSTao Ma cpos, num_clusters, p_cluster, 3650110a045aSTao Ma meta_ac, dealloc); 3651110a045aSTao Ma if (ret) { 3652110a045aSTao Ma mlog_errno(ret); 3653110a045aSTao Ma goto out_commit; 3654110a045aSTao Ma } 3655110a045aSTao Ma 36567540c1a7STao Ma ret = __ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 36577540c1a7STao Ma p_cluster, num_clusters, 0, 3658110a045aSTao Ma meta_ac, dealloc); 36590129241eSTao Ma if (ret) { 36600129241eSTao Ma mlog_errno(ret); 36610129241eSTao Ma goto out_commit; 36620129241eSTao Ma } 36630129241eSTao Ma 36640129241eSTao Ma if (post && post->func) { 36650129241eSTao Ma ret = post->func(inode, handle, post->para); 3666110a045aSTao Ma if (ret) 3667110a045aSTao Ma mlog_errno(ret); 36680129241eSTao Ma } 3669110a045aSTao Ma 3670110a045aSTao Ma out_commit: 3671110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3672110a045aSTao Ma out: 3673110a045aSTao Ma if (meta_ac) 3674110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3675110a045aSTao Ma return ret; 3676110a045aSTao Ma } 3677110a045aSTao Ma 3678a9063ab9STao Ma static int ocfs2_change_ctime(struct inode *inode, 3679110a045aSTao Ma struct buffer_head *di_bh) 3680110a045aSTao Ma { 3681110a045aSTao Ma int ret; 3682a9063ab9STao Ma handle_t *handle; 3683a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3684a9063ab9STao Ma 3685a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(inode->i_sb), 3686a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 3687a9063ab9STao Ma if (IS_ERR(handle)) { 3688a9063ab9STao Ma ret = PTR_ERR(handle); 3689a9063ab9STao Ma mlog_errno(ret); 3690a9063ab9STao Ma goto out; 3691a9063ab9STao Ma } 3692a9063ab9STao Ma 3693a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 3694a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 3695a9063ab9STao Ma if (ret) { 3696a9063ab9STao Ma mlog_errno(ret); 3697a9063ab9STao Ma goto out_commit; 3698a9063ab9STao Ma } 3699a9063ab9STao Ma 3700a9063ab9STao Ma inode->i_ctime = CURRENT_TIME; 3701a9063ab9STao Ma di->i_ctime = cpu_to_le64(inode->i_ctime.tv_sec); 3702a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec); 3703a9063ab9STao Ma 3704a9063ab9STao Ma ocfs2_journal_dirty(handle, di_bh); 3705a9063ab9STao Ma 3706a9063ab9STao Ma out_commit: 3707a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle); 3708a9063ab9STao Ma out: 3709a9063ab9STao Ma return ret; 3710a9063ab9STao Ma } 3711a9063ab9STao Ma 3712a9063ab9STao Ma static int ocfs2_attach_refcount_tree(struct inode *inode, 3713a9063ab9STao Ma struct buffer_head *di_bh) 3714a9063ab9STao Ma { 3715a9063ab9STao Ma int ret, data_changed = 0; 3716110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 3717110a045aSTao Ma struct ocfs2_inode_info *oi = OCFS2_I(inode); 3718110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 3719110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3720110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 3721110a045aSTao Ma unsigned int ext_flags; 3722110a045aSTao Ma loff_t size; 3723110a045aSTao Ma u32 cpos, num_clusters, clusters, p_cluster; 3724110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 3725110a045aSTao Ma struct ocfs2_extent_tree di_et; 3726110a045aSTao Ma 3727110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 3728110a045aSTao Ma 3729110a045aSTao Ma if (!(oi->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL)) { 3730110a045aSTao Ma ret = ocfs2_create_refcount_tree(inode, di_bh); 3731110a045aSTao Ma if (ret) { 3732110a045aSTao Ma mlog_errno(ret); 3733110a045aSTao Ma goto out; 3734110a045aSTao Ma } 3735110a045aSTao Ma } 3736110a045aSTao Ma 3737110a045aSTao Ma BUG_ON(!di->i_refcount_loc); 3738110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, 3739110a045aSTao Ma le64_to_cpu(di->i_refcount_loc), 1, 3740110a045aSTao Ma &ref_tree, &ref_root_bh); 3741110a045aSTao Ma if (ret) { 3742110a045aSTao Ma mlog_errno(ret); 3743110a045aSTao Ma goto out; 3744110a045aSTao Ma } 3745110a045aSTao Ma 3746110a045aSTao Ma ocfs2_init_dinode_extent_tree(&di_et, INODE_CACHE(inode), di_bh); 3747110a045aSTao Ma 3748110a045aSTao Ma size = i_size_read(inode); 3749110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(inode->i_sb, size); 3750110a045aSTao Ma 3751110a045aSTao Ma cpos = 0; 3752110a045aSTao Ma while (cpos < clusters) { 3753110a045aSTao Ma ret = ocfs2_get_clusters(inode, cpos, &p_cluster, 3754110a045aSTao Ma &num_clusters, &ext_flags); 3755110a045aSTao Ma 3756110a045aSTao Ma if (p_cluster && !(ext_flags & OCFS2_EXT_REFCOUNTED)) { 3757110a045aSTao Ma ret = ocfs2_add_refcount_flag(inode, &di_et, 3758110a045aSTao Ma &ref_tree->rf_ci, 3759110a045aSTao Ma ref_root_bh, cpos, 3760110a045aSTao Ma p_cluster, num_clusters, 37610129241eSTao Ma &dealloc, NULL); 3762110a045aSTao Ma if (ret) { 3763110a045aSTao Ma mlog_errno(ret); 3764a9063ab9STao Ma goto unlock; 3765110a045aSTao Ma } 3766a9063ab9STao Ma 3767a9063ab9STao Ma data_changed = 1; 3768110a045aSTao Ma } 3769110a045aSTao Ma cpos += num_clusters; 3770110a045aSTao Ma } 3771110a045aSTao Ma 37720129241eSTao Ma if (oi->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 37730129241eSTao Ma ret = ocfs2_xattr_attach_refcount_tree(inode, di_bh, 37740129241eSTao Ma &ref_tree->rf_ci, 37750129241eSTao Ma ref_root_bh, 37760129241eSTao Ma &dealloc); 37770129241eSTao Ma if (ret) { 37780129241eSTao Ma mlog_errno(ret); 37790129241eSTao Ma goto unlock; 37800129241eSTao Ma } 37810129241eSTao Ma } 37820129241eSTao Ma 3783a9063ab9STao Ma if (data_changed) { 3784a9063ab9STao Ma ret = ocfs2_change_ctime(inode, di_bh); 3785a9063ab9STao Ma if (ret) 3786a9063ab9STao Ma mlog_errno(ret); 3787a9063ab9STao Ma } 3788a9063ab9STao Ma 3789a9063ab9STao Ma unlock: 3790110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 3791110a045aSTao Ma brelse(ref_root_bh); 3792110a045aSTao Ma 3793110a045aSTao Ma if (!ret && ocfs2_dealloc_has_cluster(&dealloc)) { 3794110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 3795110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 3796110a045aSTao Ma } 3797110a045aSTao Ma out: 3798110a045aSTao Ma /* 3799110a045aSTao Ma * Empty the extent map so that we may get the right extent 3800110a045aSTao Ma * record from the disk. 3801110a045aSTao Ma */ 3802110a045aSTao Ma ocfs2_extent_map_trunc(inode, 0); 3803110a045aSTao Ma 3804110a045aSTao Ma return ret; 3805110a045aSTao Ma } 3806110a045aSTao Ma 3807110a045aSTao Ma static int ocfs2_add_refcounted_extent(struct inode *inode, 3808110a045aSTao Ma struct ocfs2_extent_tree *et, 3809110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3810110a045aSTao Ma struct buffer_head *ref_root_bh, 3811110a045aSTao Ma u32 cpos, u32 p_cluster, u32 num_clusters, 3812110a045aSTao Ma unsigned int ext_flags, 3813110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 3814110a045aSTao Ma { 3815110a045aSTao Ma int ret; 3816110a045aSTao Ma handle_t *handle; 3817110a045aSTao Ma int credits = 0; 3818110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 3819110a045aSTao Ma struct ocfs2_alloc_context *meta_ac = NULL; 3820110a045aSTao Ma 3821110a045aSTao Ma ret = ocfs2_lock_refcount_allocators(inode->i_sb, 3822110a045aSTao Ma p_cluster, num_clusters, 3823110a045aSTao Ma et, ref_ci, 3824110a045aSTao Ma ref_root_bh, &meta_ac, 3825110a045aSTao Ma NULL, &credits); 3826110a045aSTao Ma if (ret) { 3827110a045aSTao Ma mlog_errno(ret); 3828110a045aSTao Ma goto out; 3829110a045aSTao Ma } 3830110a045aSTao Ma 3831110a045aSTao Ma handle = ocfs2_start_trans(osb, credits); 3832110a045aSTao Ma if (IS_ERR(handle)) { 3833110a045aSTao Ma ret = PTR_ERR(handle); 3834110a045aSTao Ma mlog_errno(ret); 3835110a045aSTao Ma goto out; 3836110a045aSTao Ma } 3837110a045aSTao Ma 3838110a045aSTao Ma ret = ocfs2_insert_extent(handle, et, cpos, 3839110a045aSTao Ma cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb, 3840110a045aSTao Ma p_cluster)), 3841110a045aSTao Ma num_clusters, ext_flags, meta_ac); 3842110a045aSTao Ma if (ret) { 3843110a045aSTao Ma mlog_errno(ret); 3844110a045aSTao Ma goto out_commit; 3845110a045aSTao Ma } 3846110a045aSTao Ma 38472999d12fSTao Ma ret = ocfs2_increase_refcount(handle, ref_ci, ref_root_bh, 3848110a045aSTao Ma p_cluster, num_clusters, 3849110a045aSTao Ma meta_ac, dealloc); 3850110a045aSTao Ma if (ret) 3851110a045aSTao Ma mlog_errno(ret); 3852110a045aSTao Ma 3853110a045aSTao Ma out_commit: 3854110a045aSTao Ma ocfs2_commit_trans(osb, handle); 3855110a045aSTao Ma out: 3856110a045aSTao Ma if (meta_ac) 3857110a045aSTao Ma ocfs2_free_alloc_context(meta_ac); 3858110a045aSTao Ma return ret; 3859110a045aSTao Ma } 3860110a045aSTao Ma 3861110a045aSTao Ma static int ocfs2_duplicate_extent_list(struct inode *s_inode, 3862110a045aSTao Ma struct inode *t_inode, 3863110a045aSTao Ma struct buffer_head *t_bh, 3864110a045aSTao Ma struct ocfs2_caching_info *ref_ci, 3865110a045aSTao Ma struct buffer_head *ref_root_bh, 3866110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt *dealloc) 3867110a045aSTao Ma { 3868110a045aSTao Ma int ret = 0; 3869110a045aSTao Ma u32 p_cluster, num_clusters, clusters, cpos; 3870110a045aSTao Ma loff_t size; 3871110a045aSTao Ma unsigned int ext_flags; 3872110a045aSTao Ma struct ocfs2_extent_tree et; 3873110a045aSTao Ma 3874110a045aSTao Ma ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(t_inode), t_bh); 3875110a045aSTao Ma 3876110a045aSTao Ma size = i_size_read(s_inode); 3877110a045aSTao Ma clusters = ocfs2_clusters_for_bytes(s_inode->i_sb, size); 3878110a045aSTao Ma 3879110a045aSTao Ma cpos = 0; 3880110a045aSTao Ma while (cpos < clusters) { 3881110a045aSTao Ma ret = ocfs2_get_clusters(s_inode, cpos, &p_cluster, 3882110a045aSTao Ma &num_clusters, &ext_flags); 3883110a045aSTao Ma 3884110a045aSTao Ma if (p_cluster) { 3885110a045aSTao Ma ret = ocfs2_add_refcounted_extent(t_inode, &et, 3886110a045aSTao Ma ref_ci, ref_root_bh, 3887110a045aSTao Ma cpos, p_cluster, 3888110a045aSTao Ma num_clusters, 3889110a045aSTao Ma ext_flags, 3890110a045aSTao Ma dealloc); 3891110a045aSTao Ma if (ret) { 3892110a045aSTao Ma mlog_errno(ret); 3893110a045aSTao Ma goto out; 3894110a045aSTao Ma } 3895110a045aSTao Ma } 3896110a045aSTao Ma 3897110a045aSTao Ma cpos += num_clusters; 3898110a045aSTao Ma } 3899110a045aSTao Ma 3900110a045aSTao Ma out: 3901110a045aSTao Ma return ret; 3902110a045aSTao Ma } 3903110a045aSTao Ma 3904a9063ab9STao Ma /* 3905a9063ab9STao Ma * change the new file's attributes to the src. 3906a9063ab9STao Ma * 3907a9063ab9STao Ma * reflink creates a snapshot of a file, that means the attributes 3908a9063ab9STao Ma * must be identical except for three exceptions - nlink, ino, and ctime. 3909a9063ab9STao Ma */ 3910a9063ab9STao Ma static int ocfs2_complete_reflink(struct inode *s_inode, 3911a9063ab9STao Ma struct buffer_head *s_bh, 3912a9063ab9STao Ma struct inode *t_inode, 39130fe9b66cSTao Ma struct buffer_head *t_bh, 39140fe9b66cSTao Ma bool preserve) 3915a9063ab9STao Ma { 3916a9063ab9STao Ma int ret; 3917a9063ab9STao Ma handle_t *handle; 3918a9063ab9STao Ma struct ocfs2_dinode *s_di = (struct ocfs2_dinode *)s_bh->b_data; 3919a9063ab9STao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)t_bh->b_data; 3920a9063ab9STao Ma loff_t size = i_size_read(s_inode); 3921a9063ab9STao Ma 3922a9063ab9STao Ma handle = ocfs2_start_trans(OCFS2_SB(t_inode->i_sb), 3923a9063ab9STao Ma OCFS2_INODE_UPDATE_CREDITS); 3924a9063ab9STao Ma if (IS_ERR(handle)) { 3925a9063ab9STao Ma ret = PTR_ERR(handle); 3926a9063ab9STao Ma mlog_errno(ret); 3927a9063ab9STao Ma return ret; 3928a9063ab9STao Ma } 3929a9063ab9STao Ma 3930a9063ab9STao Ma ret = ocfs2_journal_access_di(handle, INODE_CACHE(t_inode), t_bh, 3931a9063ab9STao Ma OCFS2_JOURNAL_ACCESS_WRITE); 3932a9063ab9STao Ma if (ret) { 3933a9063ab9STao Ma mlog_errno(ret); 3934a9063ab9STao Ma goto out_commit; 3935a9063ab9STao Ma } 3936a9063ab9STao Ma 3937a9063ab9STao Ma spin_lock(&OCFS2_I(t_inode)->ip_lock); 3938a9063ab9STao Ma OCFS2_I(t_inode)->ip_clusters = OCFS2_I(s_inode)->ip_clusters; 3939a9063ab9STao Ma OCFS2_I(t_inode)->ip_attr = OCFS2_I(s_inode)->ip_attr; 3940a9063ab9STao Ma OCFS2_I(t_inode)->ip_dyn_features = OCFS2_I(s_inode)->ip_dyn_features; 3941a9063ab9STao Ma spin_unlock(&OCFS2_I(t_inode)->ip_lock); 3942a9063ab9STao Ma i_size_write(t_inode, size); 3943a9063ab9STao Ma 3944a9063ab9STao Ma di->i_xattr_inline_size = s_di->i_xattr_inline_size; 3945a9063ab9STao Ma di->i_clusters = s_di->i_clusters; 3946a9063ab9STao Ma di->i_size = s_di->i_size; 3947a9063ab9STao Ma di->i_dyn_features = s_di->i_dyn_features; 3948a9063ab9STao Ma di->i_attr = s_di->i_attr; 39490fe9b66cSTao Ma 39500fe9b66cSTao Ma if (preserve) { 3951a9063ab9STao Ma di->i_uid = s_di->i_uid; 3952a9063ab9STao Ma di->i_gid = s_di->i_gid; 3953a9063ab9STao Ma di->i_mode = s_di->i_mode; 3954a9063ab9STao Ma 3955a9063ab9STao Ma /* 3956a9063ab9STao Ma * update time. 39570fe9b66cSTao Ma * we want mtime to appear identical to the source and 39580fe9b66cSTao Ma * update ctime. 3959a9063ab9STao Ma */ 3960a9063ab9STao Ma t_inode->i_ctime = CURRENT_TIME; 3961a9063ab9STao Ma 3962a9063ab9STao Ma di->i_ctime = cpu_to_le64(t_inode->i_ctime.tv_sec); 3963a9063ab9STao Ma di->i_ctime_nsec = cpu_to_le32(t_inode->i_ctime.tv_nsec); 3964a9063ab9STao Ma 3965a9063ab9STao Ma t_inode->i_mtime = s_inode->i_mtime; 3966a9063ab9STao Ma di->i_mtime = s_di->i_mtime; 3967a9063ab9STao Ma di->i_mtime_nsec = s_di->i_mtime_nsec; 39680fe9b66cSTao Ma } 3969a9063ab9STao Ma 3970a9063ab9STao Ma ocfs2_journal_dirty(handle, t_bh); 3971a9063ab9STao Ma 3972a9063ab9STao Ma out_commit: 3973a9063ab9STao Ma ocfs2_commit_trans(OCFS2_SB(t_inode->i_sb), handle); 3974a9063ab9STao Ma return ret; 3975a9063ab9STao Ma } 3976a9063ab9STao Ma 3977110a045aSTao Ma static int ocfs2_create_reflink_node(struct inode *s_inode, 3978110a045aSTao Ma struct buffer_head *s_bh, 3979110a045aSTao Ma struct inode *t_inode, 39800fe9b66cSTao Ma struct buffer_head *t_bh, 39810fe9b66cSTao Ma bool preserve) 3982110a045aSTao Ma { 3983110a045aSTao Ma int ret; 3984110a045aSTao Ma struct buffer_head *ref_root_bh = NULL; 3985110a045aSTao Ma struct ocfs2_cached_dealloc_ctxt dealloc; 3986110a045aSTao Ma struct ocfs2_super *osb = OCFS2_SB(s_inode->i_sb); 3987110a045aSTao Ma struct ocfs2_refcount_block *rb; 3988110a045aSTao Ma struct ocfs2_dinode *di = (struct ocfs2_dinode *)s_bh->b_data; 3989110a045aSTao Ma struct ocfs2_refcount_tree *ref_tree; 3990110a045aSTao Ma 3991110a045aSTao Ma ocfs2_init_dealloc_ctxt(&dealloc); 3992110a045aSTao Ma 3993110a045aSTao Ma ret = ocfs2_set_refcount_tree(t_inode, t_bh, 3994110a045aSTao Ma le64_to_cpu(di->i_refcount_loc)); 3995110a045aSTao Ma if (ret) { 3996110a045aSTao Ma mlog_errno(ret); 3997110a045aSTao Ma goto out; 3998110a045aSTao Ma } 3999110a045aSTao Ma 4000110a045aSTao Ma ret = ocfs2_lock_refcount_tree(osb, le64_to_cpu(di->i_refcount_loc), 4001110a045aSTao Ma 1, &ref_tree, &ref_root_bh); 4002110a045aSTao Ma if (ret) { 4003110a045aSTao Ma mlog_errno(ret); 4004110a045aSTao Ma goto out; 4005110a045aSTao Ma } 4006110a045aSTao Ma rb = (struct ocfs2_refcount_block *)ref_root_bh->b_data; 4007110a045aSTao Ma 4008110a045aSTao Ma ret = ocfs2_duplicate_extent_list(s_inode, t_inode, t_bh, 4009110a045aSTao Ma &ref_tree->rf_ci, ref_root_bh, 4010110a045aSTao Ma &dealloc); 4011a9063ab9STao Ma if (ret) { 4012a9063ab9STao Ma mlog_errno(ret); 4013a9063ab9STao Ma goto out_unlock_refcount; 4014a9063ab9STao Ma } 4015a9063ab9STao Ma 40160fe9b66cSTao Ma ret = ocfs2_complete_reflink(s_inode, s_bh, t_inode, t_bh, preserve); 4017110a045aSTao Ma if (ret) 4018110a045aSTao Ma mlog_errno(ret); 4019110a045aSTao Ma 4020a9063ab9STao Ma out_unlock_refcount: 4021110a045aSTao Ma ocfs2_unlock_refcount_tree(osb, ref_tree, 1); 4022110a045aSTao Ma brelse(ref_root_bh); 4023110a045aSTao Ma out: 4024110a045aSTao Ma if (ocfs2_dealloc_has_cluster(&dealloc)) { 4025110a045aSTao Ma ocfs2_schedule_truncate_log_flush(osb, 1); 4026110a045aSTao Ma ocfs2_run_deallocs(osb, &dealloc); 4027110a045aSTao Ma } 4028110a045aSTao Ma 4029110a045aSTao Ma return ret; 4030110a045aSTao Ma } 403109bf27a0STao Ma 403209bf27a0STao Ma static int __ocfs2_reflink(struct dentry *old_dentry, 403309bf27a0STao Ma struct buffer_head *old_bh, 403409bf27a0STao Ma struct inode *new_inode, 403509bf27a0STao Ma bool preserve) 403609bf27a0STao Ma { 403709bf27a0STao Ma int ret; 403809bf27a0STao Ma struct inode *inode = old_dentry->d_inode; 403909bf27a0STao Ma struct buffer_head *new_bh = NULL; 404009bf27a0STao Ma 404109bf27a0STao Ma ret = filemap_fdatawrite(inode->i_mapping); 404209bf27a0STao Ma if (ret) { 404309bf27a0STao Ma mlog_errno(ret); 404409bf27a0STao Ma goto out; 404509bf27a0STao Ma } 404609bf27a0STao Ma 404709bf27a0STao Ma ret = ocfs2_attach_refcount_tree(inode, old_bh); 404809bf27a0STao Ma if (ret) { 404909bf27a0STao Ma mlog_errno(ret); 405009bf27a0STao Ma goto out; 405109bf27a0STao Ma } 405209bf27a0STao Ma 405309bf27a0STao Ma mutex_lock(&new_inode->i_mutex); 405409bf27a0STao Ma ret = ocfs2_inode_lock(new_inode, &new_bh, 1); 405509bf27a0STao Ma if (ret) { 405609bf27a0STao Ma mlog_errno(ret); 405709bf27a0STao Ma goto out_unlock; 405809bf27a0STao Ma } 405909bf27a0STao Ma 406009bf27a0STao Ma ret = ocfs2_create_reflink_node(inode, old_bh, 406109bf27a0STao Ma new_inode, new_bh, preserve); 406209bf27a0STao Ma if (ret) { 406309bf27a0STao Ma mlog_errno(ret); 406409bf27a0STao Ma goto inode_unlock; 406509bf27a0STao Ma } 406609bf27a0STao Ma 406709bf27a0STao Ma if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_XATTR_FL) { 406809bf27a0STao Ma ret = ocfs2_reflink_xattrs(inode, old_bh, 406909bf27a0STao Ma new_inode, new_bh, 407009bf27a0STao Ma preserve); 407109bf27a0STao Ma if (ret) 407209bf27a0STao Ma mlog_errno(ret); 407309bf27a0STao Ma } 407409bf27a0STao Ma inode_unlock: 407509bf27a0STao Ma ocfs2_inode_unlock(new_inode, 1); 407609bf27a0STao Ma brelse(new_bh); 407709bf27a0STao Ma out_unlock: 407809bf27a0STao Ma mutex_unlock(&new_inode->i_mutex); 407909bf27a0STao Ma out: 408009bf27a0STao Ma if (!ret) { 408109bf27a0STao Ma ret = filemap_fdatawait(inode->i_mapping); 408209bf27a0STao Ma if (ret) 408309bf27a0STao Ma mlog_errno(ret); 408409bf27a0STao Ma } 408509bf27a0STao Ma return ret; 408609bf27a0STao Ma } 408709bf27a0STao Ma 408809bf27a0STao Ma static int ocfs2_reflink(struct dentry *old_dentry, struct inode *dir, 408909bf27a0STao Ma struct dentry *new_dentry, bool preserve) 409009bf27a0STao Ma { 409109bf27a0STao Ma int error; 409209bf27a0STao Ma struct inode *inode = old_dentry->d_inode; 409309bf27a0STao Ma struct buffer_head *old_bh = NULL; 409409bf27a0STao Ma struct inode *new_orphan_inode = NULL; 409509bf27a0STao Ma 409609bf27a0STao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 409709bf27a0STao Ma return -EOPNOTSUPP; 409809bf27a0STao Ma 409909bf27a0STao Ma error = ocfs2_create_inode_in_orphan(dir, inode->i_mode, 410009bf27a0STao Ma &new_orphan_inode); 410109bf27a0STao Ma if (error) { 410209bf27a0STao Ma mlog_errno(error); 410309bf27a0STao Ma goto out; 410409bf27a0STao Ma } 410509bf27a0STao Ma 410609bf27a0STao Ma error = ocfs2_inode_lock(inode, &old_bh, 1); 410709bf27a0STao Ma if (error) { 410809bf27a0STao Ma mlog_errno(error); 410909bf27a0STao Ma goto out; 411009bf27a0STao Ma } 411109bf27a0STao Ma 411209bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_xattr_sem); 411309bf27a0STao Ma down_write(&OCFS2_I(inode)->ip_alloc_sem); 411409bf27a0STao Ma error = __ocfs2_reflink(old_dentry, old_bh, 411509bf27a0STao Ma new_orphan_inode, preserve); 411609bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_alloc_sem); 411709bf27a0STao Ma up_write(&OCFS2_I(inode)->ip_xattr_sem); 411809bf27a0STao Ma 411909bf27a0STao Ma ocfs2_inode_unlock(inode, 1); 412009bf27a0STao Ma brelse(old_bh); 412109bf27a0STao Ma 412209bf27a0STao Ma if (error) { 412309bf27a0STao Ma mlog_errno(error); 412409bf27a0STao Ma goto out; 412509bf27a0STao Ma } 412609bf27a0STao Ma 412709bf27a0STao Ma /* If the security isn't preserved, we need to re-initialize them. */ 412809bf27a0STao Ma if (!preserve) { 412909bf27a0STao Ma error = ocfs2_init_security_and_acl(dir, new_orphan_inode); 413009bf27a0STao Ma if (error) 413109bf27a0STao Ma mlog_errno(error); 413209bf27a0STao Ma } 413309bf27a0STao Ma out: 413409bf27a0STao Ma if (!error) { 413509bf27a0STao Ma error = ocfs2_mv_orphaned_inode_to_new(dir, new_orphan_inode, 413609bf27a0STao Ma new_dentry); 413709bf27a0STao Ma if (error) 413809bf27a0STao Ma mlog_errno(error); 413909bf27a0STao Ma } 414009bf27a0STao Ma 414109bf27a0STao Ma if (new_orphan_inode) { 414209bf27a0STao Ma /* 414309bf27a0STao Ma * We need to open_unlock the inode no matter whether we 414409bf27a0STao Ma * succeed or not, so that other nodes can delete it later. 414509bf27a0STao Ma */ 414609bf27a0STao Ma ocfs2_open_unlock(new_orphan_inode); 414709bf27a0STao Ma if (error) 414809bf27a0STao Ma iput(new_orphan_inode); 414909bf27a0STao Ma } 415009bf27a0STao Ma 415109bf27a0STao Ma return error; 415209bf27a0STao Ma } 4153bd50873dSTao Ma 4154bd50873dSTao Ma /* 4155bd50873dSTao Ma * Below here are the bits used by OCFS2_IOC_REFLINK() to fake 4156bd50873dSTao Ma * sys_reflink(). This will go away when vfs_reflink() exists in 4157bd50873dSTao Ma * fs/namei.c. 4158bd50873dSTao Ma */ 4159bd50873dSTao Ma 4160bd50873dSTao Ma /* copied from may_create in VFS. */ 4161bd50873dSTao Ma static inline int ocfs2_may_create(struct inode *dir, struct dentry *child) 4162bd50873dSTao Ma { 4163bd50873dSTao Ma if (child->d_inode) 4164bd50873dSTao Ma return -EEXIST; 4165bd50873dSTao Ma if (IS_DEADDIR(dir)) 4166bd50873dSTao Ma return -ENOENT; 4167bd50873dSTao Ma return inode_permission(dir, MAY_WRITE | MAY_EXEC); 4168bd50873dSTao Ma } 4169bd50873dSTao Ma 4170bd50873dSTao Ma /* copied from user_path_parent. */ 4171bd50873dSTao Ma static int ocfs2_user_path_parent(const char __user *path, 4172bd50873dSTao Ma struct nameidata *nd, char **name) 4173bd50873dSTao Ma { 4174bd50873dSTao Ma char *s = getname(path); 4175bd50873dSTao Ma int error; 4176bd50873dSTao Ma 4177bd50873dSTao Ma if (IS_ERR(s)) 4178bd50873dSTao Ma return PTR_ERR(s); 4179bd50873dSTao Ma 4180bd50873dSTao Ma error = path_lookup(s, LOOKUP_PARENT, nd); 4181bd50873dSTao Ma if (error) 4182bd50873dSTao Ma putname(s); 4183bd50873dSTao Ma else 4184bd50873dSTao Ma *name = s; 4185bd50873dSTao Ma 4186bd50873dSTao Ma return error; 4187bd50873dSTao Ma } 4188bd50873dSTao Ma 4189bd50873dSTao Ma /** 4190bd50873dSTao Ma * ocfs2_vfs_reflink - Create a reference-counted link 4191bd50873dSTao Ma * 4192bd50873dSTao Ma * @old_dentry: source dentry + inode 4193bd50873dSTao Ma * @dir: directory to create the target 4194bd50873dSTao Ma * @new_dentry: target dentry 4195bd50873dSTao Ma * @preserve: if true, preserve all file attributes 4196bd50873dSTao Ma */ 4197bd50873dSTao Ma int ocfs2_vfs_reflink(struct dentry *old_dentry, struct inode *dir, 4198bd50873dSTao Ma struct dentry *new_dentry, bool preserve) 4199bd50873dSTao Ma { 4200bd50873dSTao Ma struct inode *inode = old_dentry->d_inode; 4201bd50873dSTao Ma int error; 4202bd50873dSTao Ma 4203bd50873dSTao Ma if (!inode) 4204bd50873dSTao Ma return -ENOENT; 4205bd50873dSTao Ma 4206bd50873dSTao Ma error = ocfs2_may_create(dir, new_dentry); 4207bd50873dSTao Ma if (error) 4208bd50873dSTao Ma return error; 4209bd50873dSTao Ma 4210bd50873dSTao Ma if (dir->i_sb != inode->i_sb) 4211bd50873dSTao Ma return -EXDEV; 4212bd50873dSTao Ma 4213bd50873dSTao Ma /* 4214bd50873dSTao Ma * A reflink to an append-only or immutable file cannot be created. 4215bd50873dSTao Ma */ 4216bd50873dSTao Ma if (IS_APPEND(inode) || IS_IMMUTABLE(inode)) 4217bd50873dSTao Ma return -EPERM; 4218bd50873dSTao Ma 4219bd50873dSTao Ma /* Only regular files can be reflinked. */ 4220bd50873dSTao Ma if (!S_ISREG(inode->i_mode)) 4221bd50873dSTao Ma return -EPERM; 4222bd50873dSTao Ma 4223bd50873dSTao Ma /* 4224bd50873dSTao Ma * If the caller wants to preserve ownership, they require the 4225bd50873dSTao Ma * rights to do so. 4226bd50873dSTao Ma */ 4227bd50873dSTao Ma if (preserve) { 4228bd50873dSTao Ma if ((current_fsuid() != inode->i_uid) && !capable(CAP_CHOWN)) 4229bd50873dSTao Ma return -EPERM; 4230bd50873dSTao Ma if (!in_group_p(inode->i_gid) && !capable(CAP_CHOWN)) 4231bd50873dSTao Ma return -EPERM; 4232bd50873dSTao Ma } 4233bd50873dSTao Ma 4234bd50873dSTao Ma /* 4235bd50873dSTao Ma * If the caller is modifying any aspect of the attributes, they 4236bd50873dSTao Ma * are not creating a snapshot. They need read permission on the 4237bd50873dSTao Ma * file. 4238bd50873dSTao Ma */ 4239bd50873dSTao Ma if (!preserve) { 4240bd50873dSTao Ma error = inode_permission(inode, MAY_READ); 4241bd50873dSTao Ma if (error) 4242bd50873dSTao Ma return error; 4243bd50873dSTao Ma } 4244bd50873dSTao Ma 4245bd50873dSTao Ma mutex_lock(&inode->i_mutex); 4246bd50873dSTao Ma vfs_dq_init(dir); 4247bd50873dSTao Ma error = ocfs2_reflink(old_dentry, dir, new_dentry, preserve); 4248bd50873dSTao Ma mutex_unlock(&inode->i_mutex); 4249bd50873dSTao Ma if (!error) 4250bd50873dSTao Ma fsnotify_create(dir, new_dentry); 4251bd50873dSTao Ma return error; 4252bd50873dSTao Ma } 4253bd50873dSTao Ma /* 4254bd50873dSTao Ma * Most codes are copied from sys_linkat. 4255bd50873dSTao Ma */ 4256bd50873dSTao Ma int ocfs2_reflink_ioctl(struct inode *inode, 4257bd50873dSTao Ma const char __user *oldname, 4258bd50873dSTao Ma const char __user *newname, 4259bd50873dSTao Ma bool preserve) 4260bd50873dSTao Ma { 4261bd50873dSTao Ma struct dentry *new_dentry; 4262bd50873dSTao Ma struct nameidata nd; 4263bd50873dSTao Ma struct path old_path; 4264bd50873dSTao Ma int error; 4265bd50873dSTao Ma char *to = NULL; 4266bd50873dSTao Ma 4267bd50873dSTao Ma if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb))) 4268bd50873dSTao Ma return -EOPNOTSUPP; 4269bd50873dSTao Ma 4270bd50873dSTao Ma error = user_path_at(AT_FDCWD, oldname, 0, &old_path); 4271bd50873dSTao Ma if (error) { 4272bd50873dSTao Ma mlog_errno(error); 4273bd50873dSTao Ma return error; 4274bd50873dSTao Ma } 4275bd50873dSTao Ma 4276bd50873dSTao Ma error = ocfs2_user_path_parent(newname, &nd, &to); 4277bd50873dSTao Ma if (error) { 4278bd50873dSTao Ma mlog_errno(error); 4279bd50873dSTao Ma goto out; 4280bd50873dSTao Ma } 4281bd50873dSTao Ma 4282bd50873dSTao Ma error = -EXDEV; 4283bd50873dSTao Ma if (old_path.mnt != nd.path.mnt) 4284bd50873dSTao Ma goto out_release; 4285bd50873dSTao Ma new_dentry = lookup_create(&nd, 0); 4286bd50873dSTao Ma error = PTR_ERR(new_dentry); 4287bd50873dSTao Ma if (IS_ERR(new_dentry)) { 4288bd50873dSTao Ma mlog_errno(error); 4289bd50873dSTao Ma goto out_unlock; 4290bd50873dSTao Ma } 4291bd50873dSTao Ma 4292bd50873dSTao Ma error = mnt_want_write(nd.path.mnt); 4293bd50873dSTao Ma if (error) { 4294bd50873dSTao Ma mlog_errno(error); 4295bd50873dSTao Ma goto out_dput; 4296bd50873dSTao Ma } 4297bd50873dSTao Ma 4298bd50873dSTao Ma error = ocfs2_vfs_reflink(old_path.dentry, 4299bd50873dSTao Ma nd.path.dentry->d_inode, 4300bd50873dSTao Ma new_dentry, preserve); 4301bd50873dSTao Ma mnt_drop_write(nd.path.mnt); 4302bd50873dSTao Ma out_dput: 4303bd50873dSTao Ma dput(new_dentry); 4304bd50873dSTao Ma out_unlock: 4305bd50873dSTao Ma mutex_unlock(&nd.path.dentry->d_inode->i_mutex); 4306bd50873dSTao Ma out_release: 4307bd50873dSTao Ma path_put(&nd.path); 4308bd50873dSTao Ma putname(to); 4309bd50873dSTao Ma out: 4310bd50873dSTao Ma path_put(&old_path); 4311bd50873dSTao Ma 4312bd50873dSTao Ma return error; 4313bd50873dSTao Ma } 4314