Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * Copyright (C) 2007 Oracle.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/fs.h>
  20#include <linux/slab.h>
  21#include <linux/sched.h>
  22#include <linux/writeback.h>
  23#include <linux/pagemap.h>
  24#include <linux/blkdev.h>
  25#include "ctree.h"
  26#include "disk-io.h"
  27#include "transaction.h"
  28#include "locking.h"
  29#include "tree-log.h"
  30#include "inode-map.h"
 
  31
  32#define BTRFS_ROOT_TRANS_TAG 0
  33
  34static noinline void put_transaction(struct btrfs_transaction *transaction)
  35{
  36	WARN_ON(atomic_read(&transaction->use_count) == 0);
  37	if (atomic_dec_and_test(&transaction->use_count)) {
  38		BUG_ON(!list_empty(&transaction->list));
 
 
  39		memset(transaction, 0, sizeof(*transaction));
  40		kmem_cache_free(btrfs_transaction_cachep, transaction);
  41	}
  42}
  43
  44static noinline void switch_commit_root(struct btrfs_root *root)
  45{
  46	free_extent_buffer(root->commit_root);
  47	root->commit_root = btrfs_root_node(root);
  48}
  49
  50/*
  51 * either allocate a new transaction or hop into the existing one
  52 */
  53static noinline int join_transaction(struct btrfs_root *root, int nofail)
  54{
  55	struct btrfs_transaction *cur_trans;
 
  56
  57	spin_lock(&root->fs_info->trans_lock);
  58	if (root->fs_info->trans_no_join) {
 
 
 
 
 
 
 
  59		if (!nofail) {
  60			spin_unlock(&root->fs_info->trans_lock);
  61			return -EBUSY;
  62		}
  63	}
  64
  65	cur_trans = root->fs_info->running_transaction;
  66	if (cur_trans) {
 
 
 
 
  67		atomic_inc(&cur_trans->use_count);
  68		atomic_inc(&cur_trans->num_writers);
  69		cur_trans->num_joined++;
  70		spin_unlock(&root->fs_info->trans_lock);
  71		return 0;
  72	}
  73	spin_unlock(&root->fs_info->trans_lock);
  74
  75	cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  76	if (!cur_trans)
  77		return -ENOMEM;
  78	spin_lock(&root->fs_info->trans_lock);
  79	if (root->fs_info->running_transaction) {
 
 
 
 
 
  80		kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  81		cur_trans = root->fs_info->running_transaction;
  82		atomic_inc(&cur_trans->use_count);
  83		atomic_inc(&cur_trans->num_writers);
  84		cur_trans->num_joined++;
  85		spin_unlock(&root->fs_info->trans_lock);
  86		return 0;
 
  87	}
 
  88	atomic_set(&cur_trans->num_writers, 1);
  89	cur_trans->num_joined = 0;
  90	init_waitqueue_head(&cur_trans->writer_wait);
  91	init_waitqueue_head(&cur_trans->commit_wait);
  92	cur_trans->in_commit = 0;
  93	cur_trans->blocked = 0;
  94	/*
  95	 * One for this trans handle, one so it will live on until we
  96	 * commit the transaction.
  97	 */
  98	atomic_set(&cur_trans->use_count, 2);
  99	cur_trans->commit_done = 0;
 100	cur_trans->start_time = get_seconds();
 101
 102	cur_trans->delayed_refs.root = RB_ROOT;
 103	cur_trans->delayed_refs.num_entries = 0;
 104	cur_trans->delayed_refs.num_heads_ready = 0;
 105	cur_trans->delayed_refs.num_heads = 0;
 106	cur_trans->delayed_refs.flushing = 0;
 107	cur_trans->delayed_refs.run_delayed_start = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 108	spin_lock_init(&cur_trans->commit_lock);
 109	spin_lock_init(&cur_trans->delayed_refs.lock);
 
 110
 111	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 112	list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 113	extent_io_tree_init(&cur_trans->dirty_pages,
 114			     root->fs_info->btree_inode->i_mapping);
 115	root->fs_info->generation++;
 116	cur_trans->transid = root->fs_info->generation;
 117	root->fs_info->running_transaction = cur_trans;
 118	spin_unlock(&root->fs_info->trans_lock);
 
 119
 120	return 0;
 121}
 122
 123/*
 124 * this does all the record keeping required to make sure that a reference
 125 * counted root is properly recorded in a given transaction.  This is required
 126 * to make sure the old root from before we joined the transaction is deleted
 127 * when the transaction commits
 128 */
 129static int record_root_in_trans(struct btrfs_trans_handle *trans,
 130			       struct btrfs_root *root)
 131{
 132	if (root->ref_cows && root->last_trans < trans->transid) {
 133		WARN_ON(root == root->fs_info->extent_root);
 134		WARN_ON(root->commit_root != root->node);
 135
 136		/*
 137		 * see below for in_trans_setup usage rules
 138		 * we have the reloc mutex held now, so there
 139		 * is only one writer in this function
 140		 */
 141		root->in_trans_setup = 1;
 142
 143		/* make sure readers find in_trans_setup before
 144		 * they find our root->last_trans update
 145		 */
 146		smp_wmb();
 147
 148		spin_lock(&root->fs_info->fs_roots_radix_lock);
 149		if (root->last_trans == trans->transid) {
 150			spin_unlock(&root->fs_info->fs_roots_radix_lock);
 151			return 0;
 152		}
 153		radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 154			   (unsigned long)root->root_key.objectid,
 155			   BTRFS_ROOT_TRANS_TAG);
 156		spin_unlock(&root->fs_info->fs_roots_radix_lock);
 157		root->last_trans = trans->transid;
 158
 159		/* this is pretty tricky.  We don't want to
 160		 * take the relocation lock in btrfs_record_root_in_trans
 161		 * unless we're really doing the first setup for this root in
 162		 * this transaction.
 163		 *
 164		 * Normally we'd use root->last_trans as a flag to decide
 165		 * if we want to take the expensive mutex.
 166		 *
 167		 * But, we have to set root->last_trans before we
 168		 * init the relocation root, otherwise, we trip over warnings
 169		 * in ctree.c.  The solution used here is to flag ourselves
 170		 * with root->in_trans_setup.  When this is 1, we're still
 171		 * fixing up the reloc trees and everyone must wait.
 172		 *
 173		 * When this is zero, they can trust root->last_trans and fly
 174		 * through btrfs_record_root_in_trans without having to take the
 175		 * lock.  smp_wmb() makes sure that all the writes above are
 176		 * done before we pop in the zero below
 177		 */
 178		btrfs_init_reloc_root(trans, root);
 179		smp_wmb();
 180		root->in_trans_setup = 0;
 181	}
 182	return 0;
 183}
 184
 185
 186int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 187			       struct btrfs_root *root)
 188{
 189	if (!root->ref_cows)
 190		return 0;
 191
 192	/*
 193	 * see record_root_in_trans for comments about in_trans_setup usage
 194	 * and barriers
 195	 */
 196	smp_rmb();
 197	if (root->last_trans == trans->transid &&
 198	    !root->in_trans_setup)
 199		return 0;
 200
 201	mutex_lock(&root->fs_info->reloc_mutex);
 202	record_root_in_trans(trans, root);
 203	mutex_unlock(&root->fs_info->reloc_mutex);
 204
 205	return 0;
 206}
 207
 208/* wait for commit against the current transaction to become unblocked
 209 * when this is done, it is safe to start a new transaction, but the current
 210 * transaction might not be fully on disk.
 211 */
 212static void wait_current_trans(struct btrfs_root *root)
 213{
 214	struct btrfs_transaction *cur_trans;
 215
 216	spin_lock(&root->fs_info->trans_lock);
 217	cur_trans = root->fs_info->running_transaction;
 218	if (cur_trans && cur_trans->blocked) {
 219		atomic_inc(&cur_trans->use_count);
 220		spin_unlock(&root->fs_info->trans_lock);
 221
 222		wait_event(root->fs_info->transaction_wait,
 223			   !cur_trans->blocked);
 224		put_transaction(cur_trans);
 225	} else {
 226		spin_unlock(&root->fs_info->trans_lock);
 227	}
 228}
 229
 230enum btrfs_trans_type {
 231	TRANS_START,
 232	TRANS_JOIN,
 233	TRANS_USERSPACE,
 234	TRANS_JOIN_NOLOCK,
 235};
 236
 237static int may_wait_transaction(struct btrfs_root *root, int type)
 238{
 239	if (root->fs_info->log_root_recovering)
 240		return 0;
 241
 242	if (type == TRANS_USERSPACE)
 243		return 1;
 244
 245	if (type == TRANS_START &&
 246	    !atomic_read(&root->fs_info->open_ioctl_trans))
 247		return 1;
 248
 249	return 0;
 250}
 251
 252static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 253						    u64 num_items, int type)
 254{
 255	struct btrfs_trans_handle *h;
 256	struct btrfs_transaction *cur_trans;
 257	u64 num_bytes = 0;
 258	int ret;
 259
 260	if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 261		return ERR_PTR(-EROFS);
 262
 263	if (current->journal_info) {
 264		WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 265		h = current->journal_info;
 266		h->use_count++;
 267		h->orig_rsv = h->block_rsv;
 268		h->block_rsv = NULL;
 269		goto got_it;
 270	}
 271
 272	/*
 273	 * Do the reservation before we join the transaction so we can do all
 274	 * the appropriate flushing if need be.
 275	 */
 276	if (num_items > 0 && root != root->fs_info->chunk_root) {
 277		num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 278		ret = btrfs_block_rsv_add(NULL, root,
 279					  &root->fs_info->trans_block_rsv,
 280					  num_bytes);
 281		if (ret)
 282			return ERR_PTR(ret);
 283	}
 284again:
 285	h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 286	if (!h)
 287		return ERR_PTR(-ENOMEM);
 288
 289	if (may_wait_transaction(root, type))
 290		wait_current_trans(root);
 291
 292	do {
 293		ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 294		if (ret == -EBUSY)
 295			wait_current_trans(root);
 296	} while (ret == -EBUSY);
 297
 298	if (ret < 0) {
 299		kmem_cache_free(btrfs_trans_handle_cachep, h);
 300		return ERR_PTR(ret);
 301	}
 302
 303	cur_trans = root->fs_info->running_transaction;
 304
 305	h->transid = cur_trans->transid;
 306	h->transaction = cur_trans;
 307	h->blocks_used = 0;
 308	h->bytes_reserved = 0;
 309	h->delayed_ref_updates = 0;
 310	h->use_count = 1;
 311	h->block_rsv = NULL;
 312	h->orig_rsv = NULL;
 
 313
 314	smp_mb();
 315	if (cur_trans->blocked && may_wait_transaction(root, type)) {
 316		btrfs_commit_transaction(h, root);
 317		goto again;
 318	}
 319
 320	if (num_bytes) {
 
 
 321		h->block_rsv = &root->fs_info->trans_block_rsv;
 322		h->bytes_reserved = num_bytes;
 323	}
 324
 325got_it:
 326	btrfs_record_root_in_trans(h, root);
 327
 328	if (!current->journal_info && type != TRANS_USERSPACE)
 329		current->journal_info = h;
 330	return h;
 331}
 332
 333struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 334						   int num_items)
 335{
 336	return start_transaction(root, num_items, TRANS_START);
 337}
 338struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 339{
 340	return start_transaction(root, 0, TRANS_JOIN);
 341}
 342
 343struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 344{
 345	return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 346}
 347
 348struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 349{
 350	return start_transaction(root, 0, TRANS_USERSPACE);
 351}
 352
 353/* wait for a transaction commit to be fully complete */
 354static noinline void wait_for_commit(struct btrfs_root *root,
 355				    struct btrfs_transaction *commit)
 356{
 357	wait_event(commit->commit_wait, commit->commit_done);
 358}
 359
 360int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 361{
 362	struct btrfs_transaction *cur_trans = NULL, *t;
 363	int ret;
 364
 365	ret = 0;
 366	if (transid) {
 367		if (transid <= root->fs_info->last_trans_committed)
 368			goto out;
 369
 370		/* find specified transaction */
 371		spin_lock(&root->fs_info->trans_lock);
 372		list_for_each_entry(t, &root->fs_info->trans_list, list) {
 373			if (t->transid == transid) {
 374				cur_trans = t;
 375				atomic_inc(&cur_trans->use_count);
 376				break;
 377			}
 378			if (t->transid > transid)
 379				break;
 380		}
 381		spin_unlock(&root->fs_info->trans_lock);
 382		ret = -EINVAL;
 383		if (!cur_trans)
 384			goto out;  /* bad transid */
 385	} else {
 386		/* find newest transaction that is committing | committed */
 387		spin_lock(&root->fs_info->trans_lock);
 388		list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 389					    list) {
 390			if (t->in_commit) {
 391				if (t->commit_done)
 392					break;
 393				cur_trans = t;
 394				atomic_inc(&cur_trans->use_count);
 395				break;
 396			}
 397		}
 398		spin_unlock(&root->fs_info->trans_lock);
 399		if (!cur_trans)
 400			goto out;  /* nothing committing|committed */
 401	}
 402
 403	wait_for_commit(root, cur_trans);
 404
 405	put_transaction(cur_trans);
 406	ret = 0;
 407out:
 408	return ret;
 409}
 410
 411void btrfs_throttle(struct btrfs_root *root)
 412{
 413	if (!atomic_read(&root->fs_info->open_ioctl_trans))
 414		wait_current_trans(root);
 415}
 416
 417static int should_end_transaction(struct btrfs_trans_handle *trans,
 418				  struct btrfs_root *root)
 419{
 420	int ret;
 421	ret = btrfs_block_rsv_check(trans, root,
 422				    &root->fs_info->global_block_rsv, 0, 5);
 423	return ret ? 1 : 0;
 424}
 425
 426int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 427				 struct btrfs_root *root)
 428{
 429	struct btrfs_transaction *cur_trans = trans->transaction;
 
 430	int updates;
 
 431
 432	smp_mb();
 433	if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 434		return 1;
 435
 
 
 
 
 
 
 436	updates = trans->delayed_ref_updates;
 437	trans->delayed_ref_updates = 0;
 438	if (updates)
 439		btrfs_run_delayed_refs(trans, root, updates);
 
 
 
 
 
 440
 441	return should_end_transaction(trans, root);
 442}
 443
 444static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 445			  struct btrfs_root *root, int throttle, int lock)
 446{
 447	struct btrfs_transaction *cur_trans = trans->transaction;
 448	struct btrfs_fs_info *info = root->fs_info;
 449	int count = 0;
 
 450
 451	if (--trans->use_count) {
 452		trans->block_rsv = trans->orig_rsv;
 453		return 0;
 454	}
 455
 456	while (count < 4) {
 
 
 457		unsigned long cur = trans->delayed_ref_updates;
 458		trans->delayed_ref_updates = 0;
 459		if (cur &&
 460		    trans->transaction->delayed_refs.num_heads_ready > 64) {
 461			trans->delayed_ref_updates = 0;
 462
 463			/*
 464			 * do a full flush if the transaction is trying
 465			 * to close
 466			 */
 467			if (trans->transaction->delayed_refs.flushing)
 468				cur = 0;
 469			btrfs_run_delayed_refs(trans, root, cur);
 470		} else {
 471			break;
 472		}
 473		count++;
 474	}
 475
 476	btrfs_trans_release_metadata(trans, root);
 477
 478	if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 479	    should_end_transaction(trans, root)) {
 480		trans->transaction->blocked = 1;
 481		smp_wmb();
 482	}
 483
 484	if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 485		if (throttle) {
 486			/*
 487			 * We may race with somebody else here so end up having
 488			 * to call end_transaction on ourselves again, so inc
 489			 * our use_count.
 490			 */
 491			trans->use_count++;
 492			return btrfs_commit_transaction(trans, root);
 493		} else {
 494			wake_up_process(info->transaction_kthread);
 495		}
 496	}
 497
 498	WARN_ON(cur_trans != info->running_transaction);
 499	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 500	atomic_dec(&cur_trans->num_writers);
 501
 502	smp_mb();
 503	if (waitqueue_active(&cur_trans->writer_wait))
 504		wake_up(&cur_trans->writer_wait);
 505	put_transaction(cur_trans);
 506
 507	if (current->journal_info == trans)
 508		current->journal_info = NULL;
 509	memset(trans, 0, sizeof(*trans));
 510	kmem_cache_free(btrfs_trans_handle_cachep, trans);
 511
 512	if (throttle)
 513		btrfs_run_delayed_iputs(root);
 514
 515	return 0;
 
 
 
 
 
 
 
 516}
 517
 518int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 519			  struct btrfs_root *root)
 520{
 521	int ret;
 522
 523	ret = __btrfs_end_transaction(trans, root, 0, 1);
 524	if (ret)
 525		return ret;
 526	return 0;
 527}
 528
 529int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 530				   struct btrfs_root *root)
 531{
 532	int ret;
 533
 534	ret = __btrfs_end_transaction(trans, root, 1, 1);
 535	if (ret)
 536		return ret;
 537	return 0;
 538}
 539
 540int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 541				 struct btrfs_root *root)
 542{
 543	int ret;
 544
 545	ret = __btrfs_end_transaction(trans, root, 0, 0);
 546	if (ret)
 547		return ret;
 548	return 0;
 549}
 550
 551int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 552				struct btrfs_root *root)
 553{
 554	return __btrfs_end_transaction(trans, root, 1, 1);
 555}
 556
 557/*
 558 * when btree blocks are allocated, they have some corresponding bits set for
 559 * them in one of two extent_io trees.  This is used to make sure all of
 560 * those extents are sent to disk but does not wait on them
 561 */
 562int btrfs_write_marked_extents(struct btrfs_root *root,
 563			       struct extent_io_tree *dirty_pages, int mark)
 564{
 565	int ret;
 566	int err = 0;
 567	int werr = 0;
 568	struct page *page;
 569	struct inode *btree_inode = root->fs_info->btree_inode;
 570	u64 start = 0;
 571	u64 end;
 572	unsigned long index;
 573
 574	while (1) {
 575		ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 576					    mark);
 577		if (ret)
 578			break;
 579		while (start <= end) {
 580			cond_resched();
 581
 582			index = start >> PAGE_CACHE_SHIFT;
 583			start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 584			page = find_get_page(btree_inode->i_mapping, index);
 585			if (!page)
 586				continue;
 587
 588			btree_lock_page_hook(page);
 589			if (!page->mapping) {
 590				unlock_page(page);
 591				page_cache_release(page);
 592				continue;
 593			}
 594
 595			if (PageWriteback(page)) {
 596				if (PageDirty(page))
 597					wait_on_page_writeback(page);
 598				else {
 599					unlock_page(page);
 600					page_cache_release(page);
 601					continue;
 602				}
 603			}
 604			err = write_one_page(page, 0);
 605			if (err)
 606				werr = err;
 607			page_cache_release(page);
 608		}
 609	}
 610	if (err)
 611		werr = err;
 612	return werr;
 613}
 614
 615/*
 616 * when btree blocks are allocated, they have some corresponding bits set for
 617 * them in one of two extent_io trees.  This is used to make sure all of
 618 * those extents are on disk for transaction or log commit.  We wait
 619 * on all the pages and clear them from the dirty pages state tree
 620 */
 621int btrfs_wait_marked_extents(struct btrfs_root *root,
 622			      struct extent_io_tree *dirty_pages, int mark)
 623{
 624	int ret;
 625	int err = 0;
 626	int werr = 0;
 627	struct page *page;
 628	struct inode *btree_inode = root->fs_info->btree_inode;
 629	u64 start = 0;
 630	u64 end;
 631	unsigned long index;
 632
 633	while (1) {
 634		ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 635					    mark);
 636		if (ret)
 637			break;
 638
 639		clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 640		while (start <= end) {
 641			index = start >> PAGE_CACHE_SHIFT;
 642			start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 643			page = find_get_page(btree_inode->i_mapping, index);
 644			if (!page)
 645				continue;
 646			if (PageDirty(page)) {
 647				btree_lock_page_hook(page);
 648				wait_on_page_writeback(page);
 649				err = write_one_page(page, 0);
 650				if (err)
 651					werr = err;
 652			}
 653			wait_on_page_writeback(page);
 654			page_cache_release(page);
 655			cond_resched();
 656		}
 657	}
 658	if (err)
 659		werr = err;
 660	return werr;
 661}
 662
 663/*
 664 * when btree blocks are allocated, they have some corresponding bits set for
 665 * them in one of two extent_io trees.  This is used to make sure all of
 666 * those extents are on disk for transaction or log commit
 667 */
 668int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 669				struct extent_io_tree *dirty_pages, int mark)
 670{
 671	int ret;
 672	int ret2;
 673
 674	ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 675	ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 676	return ret || ret2;
 
 
 
 
 
 677}
 678
 679int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 680				     struct btrfs_root *root)
 681{
 682	if (!trans || !trans->transaction) {
 683		struct inode *btree_inode;
 684		btree_inode = root->fs_info->btree_inode;
 685		return filemap_write_and_wait(btree_inode->i_mapping);
 686	}
 687	return btrfs_write_and_wait_marked_extents(root,
 688					   &trans->transaction->dirty_pages,
 689					   EXTENT_DIRTY);
 690}
 691
 692/*
 693 * this is used to update the root pointer in the tree of tree roots.
 694 *
 695 * But, in the case of the extent allocation tree, updating the root
 696 * pointer may allocate blocks which may change the root of the extent
 697 * allocation tree.
 698 *
 699 * So, this loops and repeats and makes sure the cowonly root didn't
 700 * change while the root pointer was being updated in the metadata.
 701 */
 702static int update_cowonly_root(struct btrfs_trans_handle *trans,
 703			       struct btrfs_root *root)
 704{
 705	int ret;
 706	u64 old_root_bytenr;
 707	u64 old_root_used;
 708	struct btrfs_root *tree_root = root->fs_info->tree_root;
 709
 710	old_root_used = btrfs_root_used(&root->root_item);
 711	btrfs_write_dirty_block_groups(trans, root);
 712
 713	while (1) {
 714		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 715		if (old_root_bytenr == root->node->start &&
 716		    old_root_used == btrfs_root_used(&root->root_item))
 717			break;
 718
 719		btrfs_set_root_node(&root->root_item, root->node);
 720		ret = btrfs_update_root(trans, tree_root,
 721					&root->root_key,
 722					&root->root_item);
 723		BUG_ON(ret);
 
 724
 725		old_root_used = btrfs_root_used(&root->root_item);
 726		ret = btrfs_write_dirty_block_groups(trans, root);
 727		BUG_ON(ret);
 
 728	}
 729
 730	if (root != root->fs_info->extent_root)
 731		switch_commit_root(root);
 732
 733	return 0;
 734}
 735
 736/*
 737 * update all the cowonly tree roots on disk
 
 
 
 
 738 */
 739static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 740					 struct btrfs_root *root)
 741{
 742	struct btrfs_fs_info *fs_info = root->fs_info;
 743	struct list_head *next;
 744	struct extent_buffer *eb;
 745	int ret;
 746
 747	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 748	BUG_ON(ret);
 
 749
 750	eb = btrfs_lock_root_node(fs_info->tree_root);
 751	btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 
 752	btrfs_tree_unlock(eb);
 753	free_extent_buffer(eb);
 754
 
 
 
 755	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 
 
 
 
 756	BUG_ON(ret);
 757
 758	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 759		next = fs_info->dirty_cowonly_roots.next;
 760		list_del_init(next);
 761		root = list_entry(next, struct btrfs_root, dirty_list);
 762
 763		update_cowonly_root(trans, root);
 
 
 764	}
 765
 766	down_write(&fs_info->extent_commit_sem);
 767	switch_commit_root(fs_info->extent_root);
 768	up_write(&fs_info->extent_commit_sem);
 769
 770	return 0;
 771}
 772
 773/*
 774 * dead roots are old snapshots that need to be deleted.  This allocates
 775 * a dirty root struct and adds it into the list of dead roots that need to
 776 * be deleted
 777 */
 778int btrfs_add_dead_root(struct btrfs_root *root)
 779{
 780	spin_lock(&root->fs_info->trans_lock);
 781	list_add(&root->root_list, &root->fs_info->dead_roots);
 782	spin_unlock(&root->fs_info->trans_lock);
 783	return 0;
 784}
 785
 786/*
 787 * update all the cowonly tree roots on disk
 788 */
 789static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 790				    struct btrfs_root *root)
 791{
 792	struct btrfs_root *gang[8];
 793	struct btrfs_fs_info *fs_info = root->fs_info;
 794	int i;
 795	int ret;
 796	int err = 0;
 797
 798	spin_lock(&fs_info->fs_roots_radix_lock);
 799	while (1) {
 800		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 801						 (void **)gang, 0,
 802						 ARRAY_SIZE(gang),
 803						 BTRFS_ROOT_TRANS_TAG);
 804		if (ret == 0)
 805			break;
 806		for (i = 0; i < ret; i++) {
 807			root = gang[i];
 808			radix_tree_tag_clear(&fs_info->fs_roots_radix,
 809					(unsigned long)root->root_key.objectid,
 810					BTRFS_ROOT_TRANS_TAG);
 811			spin_unlock(&fs_info->fs_roots_radix_lock);
 812
 813			btrfs_free_log(trans, root);
 814			btrfs_update_reloc_root(trans, root);
 815			btrfs_orphan_commit_root(trans, root);
 816
 817			btrfs_save_ino_cache(root, trans);
 818
 
 
 
 
 819			if (root->commit_root != root->node) {
 820				mutex_lock(&root->fs_commit_mutex);
 821				switch_commit_root(root);
 822				btrfs_unpin_free_ino(root);
 823				mutex_unlock(&root->fs_commit_mutex);
 824
 825				btrfs_set_root_node(&root->root_item,
 826						    root->node);
 827			}
 828
 829			err = btrfs_update_root(trans, fs_info->tree_root,
 830						&root->root_key,
 831						&root->root_item);
 832			spin_lock(&fs_info->fs_roots_radix_lock);
 833			if (err)
 834				break;
 835		}
 836	}
 837	spin_unlock(&fs_info->fs_roots_radix_lock);
 838	return err;
 839}
 840
 841/*
 842 * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 843 * otherwise every leaf in the btree is read and defragged.
 844 */
 845int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 846{
 847	struct btrfs_fs_info *info = root->fs_info;
 848	struct btrfs_trans_handle *trans;
 849	int ret;
 850	unsigned long nr;
 851
 852	if (xchg(&root->defrag_running, 1))
 853		return 0;
 854
 855	while (1) {
 856		trans = btrfs_start_transaction(root, 0);
 857		if (IS_ERR(trans))
 858			return PTR_ERR(trans);
 859
 860		ret = btrfs_defrag_leaves(trans, root, cacheonly);
 861
 862		nr = trans->blocks_used;
 863		btrfs_end_transaction(trans, root);
 864		btrfs_btree_balance_dirty(info->tree_root, nr);
 865		cond_resched();
 866
 867		if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 868			break;
 869	}
 870	root->defrag_running = 0;
 871	return ret;
 872}
 873
 874/*
 875 * new snapshots need to be created at a very specific time in the
 876 * transaction commit.  This does the actual creation
 877 */
 878static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 879				   struct btrfs_fs_info *fs_info,
 880				   struct btrfs_pending_snapshot *pending)
 881{
 882	struct btrfs_key key;
 883	struct btrfs_root_item *new_root_item;
 884	struct btrfs_root *tree_root = fs_info->tree_root;
 885	struct btrfs_root *root = pending->root;
 886	struct btrfs_root *parent_root;
 887	struct btrfs_block_rsv *rsv;
 888	struct inode *parent_inode;
 889	struct dentry *parent;
 890	struct dentry *dentry;
 891	struct extent_buffer *tmp;
 892	struct extent_buffer *old;
 893	int ret;
 894	u64 to_reserve = 0;
 895	u64 index = 0;
 896	u64 objectid;
 897	u64 root_flags;
 898
 899	rsv = trans->block_rsv;
 900
 901	new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 902	if (!new_root_item) {
 903		pending->error = -ENOMEM;
 904		goto fail;
 905	}
 906
 907	ret = btrfs_find_free_objectid(tree_root, &objectid);
 908	if (ret) {
 909		pending->error = ret;
 910		goto fail;
 911	}
 912
 913	btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 914	btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 915
 916	if (to_reserve > 0) {
 917		ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 918					  to_reserve);
 919		if (ret) {
 920			pending->error = ret;
 921			goto fail;
 922		}
 923	}
 924
 925	key.objectid = objectid;
 926	key.offset = (u64)-1;
 927	key.type = BTRFS_ROOT_ITEM_KEY;
 928
 929	trans->block_rsv = &pending->block_rsv;
 930
 931	dentry = pending->dentry;
 932	parent = dget_parent(dentry);
 933	parent_inode = parent->d_inode;
 934	parent_root = BTRFS_I(parent_inode)->root;
 935	record_root_in_trans(trans, parent_root);
 936
 937	/*
 938	 * insert the directory item
 939	 */
 940	ret = btrfs_set_inode_index(parent_inode, &index);
 941	BUG_ON(ret);
 942	ret = btrfs_insert_dir_item(trans, parent_root,
 943				dentry->d_name.name, dentry->d_name.len,
 944				parent_inode, &key,
 945				BTRFS_FT_DIR, index);
 946	BUG_ON(ret);
 
 
 
 
 
 
 947
 948	btrfs_i_size_write(parent_inode, parent_inode->i_size +
 949					 dentry->d_name.len * 2);
 950	ret = btrfs_update_inode(trans, parent_root, parent_inode);
 951	BUG_ON(ret);
 
 952
 953	/*
 954	 * pull in the delayed directory update
 955	 * and the delayed inode item
 956	 * otherwise we corrupt the FS during
 957	 * snapshot
 958	 */
 959	ret = btrfs_run_delayed_items(trans, root);
 960	BUG_ON(ret);
 
 
 
 961
 962	record_root_in_trans(trans, root);
 963	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 964	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 965	btrfs_check_and_init_root_item(new_root_item);
 966
 967	root_flags = btrfs_root_flags(new_root_item);
 968	if (pending->readonly)
 969		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 970	else
 971		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 972	btrfs_set_root_flags(new_root_item, root_flags);
 973
 974	old = btrfs_lock_root_node(root);
 975	btrfs_cow_block(trans, root, old, NULL, 0, &old);
 
 
 
 
 
 
 976	btrfs_set_lock_blocking(old);
 977
 978	btrfs_copy_root(trans, root, old, &tmp, objectid);
 
 979	btrfs_tree_unlock(old);
 980	free_extent_buffer(old);
 
 
 
 
 
 
 981
 982	btrfs_set_root_node(new_root_item, tmp);
 983	/* record when the snapshot was created in key.offset */
 984	key.offset = trans->transid;
 985	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 986	btrfs_tree_unlock(tmp);
 987	free_extent_buffer(tmp);
 988	BUG_ON(ret);
 
 989
 990	/*
 991	 * insert root back/forward references
 992	 */
 993	ret = btrfs_add_root_ref(trans, tree_root, objectid,
 994				 parent_root->root_key.objectid,
 995				 btrfs_ino(parent_inode), index,
 996				 dentry->d_name.name, dentry->d_name.len);
 997	BUG_ON(ret);
 998	dput(parent);
 
 
 999
1000	key.offset = (u64)-1;
1001	pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1002	BUG_ON(IS_ERR(pending->snap));
 
 
 
1003
1004	btrfs_reloc_post_snapshot(trans, pending);
1005	btrfs_orphan_post_snapshot(trans, pending);
 
 
1006fail:
1007	kfree(new_root_item);
1008	trans->block_rsv = rsv;
1009	btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1010	return 0;
 
 
 
 
 
 
1011}
1012
1013/*
1014 * create all the snapshots we've scheduled for creation
1015 */
1016static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1017					     struct btrfs_fs_info *fs_info)
1018{
1019	struct btrfs_pending_snapshot *pending;
1020	struct list_head *head = &trans->transaction->pending_snapshots;
1021	int ret;
1022
1023	list_for_each_entry(pending, head, list) {
1024		ret = create_pending_snapshot(trans, fs_info, pending);
1025		BUG_ON(ret);
1026	}
1027	return 0;
1028}
1029
1030static void update_super_roots(struct btrfs_root *root)
1031{
1032	struct btrfs_root_item *root_item;
1033	struct btrfs_super_block *super;
1034
1035	super = &root->fs_info->super_copy;
1036
1037	root_item = &root->fs_info->chunk_root->root_item;
1038	super->chunk_root = root_item->bytenr;
1039	super->chunk_root_generation = root_item->generation;
1040	super->chunk_root_level = root_item->level;
1041
1042	root_item = &root->fs_info->tree_root->root_item;
1043	super->root = root_item->bytenr;
1044	super->generation = root_item->generation;
1045	super->root_level = root_item->level;
1046	if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1047		super->cache_generation = root_item->generation;
1048}
1049
1050int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1051{
1052	int ret = 0;
1053	spin_lock(&info->trans_lock);
1054	if (info->running_transaction)
1055		ret = info->running_transaction->in_commit;
1056	spin_unlock(&info->trans_lock);
1057	return ret;
1058}
1059
1060int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1061{
1062	int ret = 0;
1063	spin_lock(&info->trans_lock);
1064	if (info->running_transaction)
1065		ret = info->running_transaction->blocked;
1066	spin_unlock(&info->trans_lock);
1067	return ret;
1068}
1069
1070/*
1071 * wait for the current transaction commit to start and block subsequent
1072 * transaction joins
1073 */
1074static void wait_current_trans_commit_start(struct btrfs_root *root,
1075					    struct btrfs_transaction *trans)
1076{
1077	wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1078}
1079
1080/*
1081 * wait for the current transaction to start and then become unblocked.
1082 * caller holds ref.
1083 */
1084static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1085					 struct btrfs_transaction *trans)
1086{
1087	wait_event(root->fs_info->transaction_wait,
1088		   trans->commit_done || (trans->in_commit && !trans->blocked));
1089}
1090
1091/*
1092 * commit transactions asynchronously. once btrfs_commit_transaction_async
1093 * returns, any subsequent transaction will not be allowed to join.
1094 */
1095struct btrfs_async_commit {
1096	struct btrfs_trans_handle *newtrans;
1097	struct btrfs_root *root;
1098	struct delayed_work work;
1099};
1100
1101static void do_async_commit(struct work_struct *work)
1102{
1103	struct btrfs_async_commit *ac =
1104		container_of(work, struct btrfs_async_commit, work.work);
1105
1106	btrfs_commit_transaction(ac->newtrans, ac->root);
1107	kfree(ac);
1108}
1109
1110int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1111				   struct btrfs_root *root,
1112				   int wait_for_unblock)
1113{
1114	struct btrfs_async_commit *ac;
1115	struct btrfs_transaction *cur_trans;
1116
1117	ac = kmalloc(sizeof(*ac), GFP_NOFS);
1118	if (!ac)
1119		return -ENOMEM;
1120
1121	INIT_DELAYED_WORK(&ac->work, do_async_commit);
1122	ac->root = root;
1123	ac->newtrans = btrfs_join_transaction(root);
1124	if (IS_ERR(ac->newtrans)) {
1125		int err = PTR_ERR(ac->newtrans);
1126		kfree(ac);
1127		return err;
1128	}
1129
1130	/* take transaction reference */
1131	cur_trans = trans->transaction;
1132	atomic_inc(&cur_trans->use_count);
1133
1134	btrfs_end_transaction(trans, root);
1135	schedule_delayed_work(&ac->work, 0);
1136
1137	/* wait for transaction to start and unblock */
1138	if (wait_for_unblock)
1139		wait_current_trans_commit_start_and_unblock(root, cur_trans);
1140	else
1141		wait_current_trans_commit_start(root, cur_trans);
1142
1143	if (current->journal_info == trans)
1144		current->journal_info = NULL;
1145
1146	put_transaction(cur_trans);
1147	return 0;
1148}
1149
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1150/*
1151 * btrfs_transaction state sequence:
1152 *    in_commit = 0, blocked = 0  (initial)
1153 *    in_commit = 1, blocked = 1
1154 *    blocked = 0
1155 *    commit_done = 1
1156 */
1157int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1158			     struct btrfs_root *root)
1159{
1160	unsigned long joined = 0;
1161	struct btrfs_transaction *cur_trans;
1162	struct btrfs_transaction *prev_trans = NULL;
1163	DEFINE_WAIT(wait);
1164	int ret;
1165	int should_grow = 0;
1166	unsigned long now = get_seconds();
1167	int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1168
1169	btrfs_run_ordered_operations(root, 0);
1170
 
 
 
 
 
 
1171	/* make a pass through all the delayed refs we have so far
1172	 * any runnings procs may add more while we are here
1173	 */
1174	ret = btrfs_run_delayed_refs(trans, root, 0);
1175	BUG_ON(ret);
1176
1177	btrfs_trans_release_metadata(trans, root);
1178
1179	cur_trans = trans->transaction;
 
1180	/*
1181	 * set the flushing flag so procs in this transaction have to
1182	 * start sending their work down.
1183	 */
1184	cur_trans->delayed_refs.flushing = 1;
1185
1186	ret = btrfs_run_delayed_refs(trans, root, 0);
1187	BUG_ON(ret);
 
1188
1189	spin_lock(&cur_trans->commit_lock);
1190	if (cur_trans->in_commit) {
1191		spin_unlock(&cur_trans->commit_lock);
1192		atomic_inc(&cur_trans->use_count);
1193		btrfs_end_transaction(trans, root);
1194
1195		wait_for_commit(root, cur_trans);
1196
1197		put_transaction(cur_trans);
1198
1199		return 0;
1200	}
1201
1202	trans->transaction->in_commit = 1;
1203	trans->transaction->blocked = 1;
1204	spin_unlock(&cur_trans->commit_lock);
1205	wake_up(&root->fs_info->transaction_blocked_wait);
1206
1207	spin_lock(&root->fs_info->trans_lock);
1208	if (cur_trans->list.prev != &root->fs_info->trans_list) {
1209		prev_trans = list_entry(cur_trans->list.prev,
1210					struct btrfs_transaction, list);
1211		if (!prev_trans->commit_done) {
1212			atomic_inc(&prev_trans->use_count);
1213			spin_unlock(&root->fs_info->trans_lock);
1214
1215			wait_for_commit(root, prev_trans);
1216
1217			put_transaction(prev_trans);
1218		} else {
1219			spin_unlock(&root->fs_info->trans_lock);
1220		}
1221	} else {
1222		spin_unlock(&root->fs_info->trans_lock);
1223	}
1224
1225	if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1226		should_grow = 1;
1227
1228	do {
1229		int snap_pending = 0;
1230
1231		joined = cur_trans->num_joined;
1232		if (!list_empty(&trans->transaction->pending_snapshots))
1233			snap_pending = 1;
1234
1235		WARN_ON(cur_trans != trans->transaction);
1236
1237		if (flush_on_commit || snap_pending) {
1238			btrfs_start_delalloc_inodes(root, 1);
1239			ret = btrfs_wait_ordered_extents(root, 0, 1);
1240			BUG_ON(ret);
1241		}
1242
1243		ret = btrfs_run_delayed_items(trans, root);
1244		BUG_ON(ret);
 
1245
1246		/*
1247		 * rename don't use btrfs_join_transaction, so, once we
1248		 * set the transaction to blocked above, we aren't going
1249		 * to get any new ordered operations.  We can safely run
1250		 * it here and no for sure that nothing new will be added
1251		 * to the list
1252		 */
1253		btrfs_run_ordered_operations(root, 1);
1254
1255		prepare_to_wait(&cur_trans->writer_wait, &wait,
1256				TASK_UNINTERRUPTIBLE);
1257
1258		if (atomic_read(&cur_trans->num_writers) > 1)
1259			schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1260		else if (should_grow)
1261			schedule_timeout(1);
1262
1263		finish_wait(&cur_trans->writer_wait, &wait);
1264	} while (atomic_read(&cur_trans->num_writers) > 1 ||
1265		 (should_grow && cur_trans->num_joined != joined));
1266
1267	/*
1268	 * Ok now we need to make sure to block out any other joins while we
1269	 * commit the transaction.  We could have started a join before setting
1270	 * no_join so make sure to wait for num_writers to == 1 again.
1271	 */
1272	spin_lock(&root->fs_info->trans_lock);
1273	root->fs_info->trans_no_join = 1;
1274	spin_unlock(&root->fs_info->trans_lock);
1275	wait_event(cur_trans->writer_wait,
1276		   atomic_read(&cur_trans->num_writers) == 1);
1277
1278	/*
1279	 * the reloc mutex makes sure that we stop
1280	 * the balancing code from coming in and moving
1281	 * extents around in the middle of the commit
1282	 */
1283	mutex_lock(&root->fs_info->reloc_mutex);
1284
1285	ret = btrfs_run_delayed_items(trans, root);
1286	BUG_ON(ret);
 
 
 
1287
1288	ret = create_pending_snapshots(trans, root->fs_info);
1289	BUG_ON(ret);
 
 
 
1290
1291	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1292	BUG_ON(ret);
 
 
 
1293
1294	/*
1295	 * make sure none of the code above managed to slip in a
1296	 * delayed item
1297	 */
1298	btrfs_assert_delayed_root_empty(root);
1299
1300	WARN_ON(cur_trans != trans->transaction);
1301
1302	btrfs_scrub_pause(root);
1303	/* btrfs_commit_tree_roots is responsible for getting the
1304	 * various roots consistent with each other.  Every pointer
1305	 * in the tree of tree roots has to point to the most up to date
1306	 * root for every subvolume and other tree.  So, we have to keep
1307	 * the tree logging code from jumping in and changing any
1308	 * of the trees.
1309	 *
1310	 * At this point in the commit, there can't be any tree-log
1311	 * writers, but a little lower down we drop the trans mutex
1312	 * and let new people in.  By holding the tree_log_mutex
1313	 * from now until after the super is written, we avoid races
1314	 * with the tree-log code.
1315	 */
1316	mutex_lock(&root->fs_info->tree_log_mutex);
1317
1318	ret = commit_fs_roots(trans, root);
1319	BUG_ON(ret);
 
 
 
 
1320
1321	/* commit_fs_roots gets rid of all the tree log roots, it is now
1322	 * safe to free the root of tree log roots
1323	 */
1324	btrfs_free_log_root_tree(trans, root->fs_info);
1325
1326	ret = commit_cowonly_roots(trans, root);
1327	BUG_ON(ret);
 
 
 
 
1328
1329	btrfs_prepare_extent_commit(trans, root);
1330
1331	cur_trans = root->fs_info->running_transaction;
1332
1333	btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1334			    root->fs_info->tree_root->node);
1335	switch_commit_root(root->fs_info->tree_root);
1336
1337	btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1338			    root->fs_info->chunk_root->node);
1339	switch_commit_root(root->fs_info->chunk_root);
1340
1341	update_super_roots(root);
1342
1343	if (!root->fs_info->log_root_recovering) {
1344		btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1345		btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1346	}
1347
1348	memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1349	       sizeof(root->fs_info->super_copy));
1350
1351	trans->transaction->blocked = 0;
1352	spin_lock(&root->fs_info->trans_lock);
1353	root->fs_info->running_transaction = NULL;
1354	root->fs_info->trans_no_join = 0;
1355	spin_unlock(&root->fs_info->trans_lock);
1356	mutex_unlock(&root->fs_info->reloc_mutex);
1357
1358	wake_up(&root->fs_info->transaction_wait);
1359
1360	ret = btrfs_write_and_wait_transaction(trans, root);
1361	BUG_ON(ret);
1362	write_ctree_super(trans, root, 0);
 
 
 
 
 
 
 
 
 
 
1363
1364	/*
1365	 * the super is written, we can safely allow the tree-loggers
1366	 * to go about their business
1367	 */
1368	mutex_unlock(&root->fs_info->tree_log_mutex);
1369
1370	btrfs_finish_extent_commit(trans, root);
1371
1372	cur_trans->commit_done = 1;
1373
1374	root->fs_info->last_trans_committed = cur_trans->transid;
1375
1376	wake_up(&cur_trans->commit_wait);
1377
1378	spin_lock(&root->fs_info->trans_lock);
1379	list_del_init(&cur_trans->list);
1380	spin_unlock(&root->fs_info->trans_lock);
1381
1382	put_transaction(cur_trans);
1383	put_transaction(cur_trans);
1384
1385	trace_btrfs_transaction_commit(root);
1386
1387	btrfs_scrub_continue(root);
1388
1389	if (current->journal_info == trans)
1390		current->journal_info = NULL;
1391
1392	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1393
1394	if (current != root->fs_info->transaction_kthread)
1395		btrfs_run_delayed_iputs(root);
1396
1397	return ret;
 
 
 
 
 
 
 
 
 
1398}
1399
1400/*
1401 * interface function to delete all the snapshots we have scheduled for deletion
1402 */
1403int btrfs_clean_old_snapshots(struct btrfs_root *root)
1404{
1405	LIST_HEAD(list);
1406	struct btrfs_fs_info *fs_info = root->fs_info;
1407
1408	spin_lock(&fs_info->trans_lock);
1409	list_splice_init(&fs_info->dead_roots, &list);
1410	spin_unlock(&fs_info->trans_lock);
1411
1412	while (!list_empty(&list)) {
 
 
1413		root = list_entry(list.next, struct btrfs_root, root_list);
1414		list_del(&root->root_list);
1415
1416		btrfs_kill_all_delayed_nodes(root);
1417
1418		if (btrfs_header_backref_rev(root->node) <
1419		    BTRFS_MIXED_BACKREF_REV)
1420			btrfs_drop_snapshot(root, NULL, 0);
1421		else
1422			btrfs_drop_snapshot(root, NULL, 1);
 
1423	}
1424	return 0;
1425}
v3.5.6
   1/*
   2 * Copyright (C) 2007 Oracle.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/fs.h>
  20#include <linux/slab.h>
  21#include <linux/sched.h>
  22#include <linux/writeback.h>
  23#include <linux/pagemap.h>
  24#include <linux/blkdev.h>
  25#include "ctree.h"
  26#include "disk-io.h"
  27#include "transaction.h"
  28#include "locking.h"
  29#include "tree-log.h"
  30#include "inode-map.h"
  31#include "volumes.h"
  32
  33#define BTRFS_ROOT_TRANS_TAG 0
  34
  35void put_transaction(struct btrfs_transaction *transaction)
  36{
  37	WARN_ON(atomic_read(&transaction->use_count) == 0);
  38	if (atomic_dec_and_test(&transaction->use_count)) {
  39		BUG_ON(!list_empty(&transaction->list));
  40		WARN_ON(transaction->delayed_refs.root.rb_node);
  41		WARN_ON(!list_empty(&transaction->delayed_refs.seq_head));
  42		memset(transaction, 0, sizeof(*transaction));
  43		kmem_cache_free(btrfs_transaction_cachep, transaction);
  44	}
  45}
  46
  47static noinline void switch_commit_root(struct btrfs_root *root)
  48{
  49	free_extent_buffer(root->commit_root);
  50	root->commit_root = btrfs_root_node(root);
  51}
  52
  53/*
  54 * either allocate a new transaction or hop into the existing one
  55 */
  56static noinline int join_transaction(struct btrfs_root *root, int nofail)
  57{
  58	struct btrfs_transaction *cur_trans;
  59	struct btrfs_fs_info *fs_info = root->fs_info;
  60
  61	spin_lock(&fs_info->trans_lock);
  62loop:
  63	/* The file system has been taken offline. No new transactions. */
  64	if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  65		spin_unlock(&fs_info->trans_lock);
  66		return -EROFS;
  67	}
  68
  69	if (fs_info->trans_no_join) {
  70		if (!nofail) {
  71			spin_unlock(&fs_info->trans_lock);
  72			return -EBUSY;
  73		}
  74	}
  75
  76	cur_trans = fs_info->running_transaction;
  77	if (cur_trans) {
  78		if (cur_trans->aborted) {
  79			spin_unlock(&fs_info->trans_lock);
  80			return cur_trans->aborted;
  81		}
  82		atomic_inc(&cur_trans->use_count);
  83		atomic_inc(&cur_trans->num_writers);
  84		cur_trans->num_joined++;
  85		spin_unlock(&fs_info->trans_lock);
  86		return 0;
  87	}
  88	spin_unlock(&fs_info->trans_lock);
  89
  90	cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  91	if (!cur_trans)
  92		return -ENOMEM;
  93
  94	spin_lock(&fs_info->trans_lock);
  95	if (fs_info->running_transaction) {
  96		/*
  97		 * someone started a transaction after we unlocked.  Make sure
  98		 * to redo the trans_no_join checks above
  99		 */
 100		kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 101		cur_trans = fs_info->running_transaction;
 102		goto loop;
 103	} else if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 
 104		spin_unlock(&root->fs_info->trans_lock);
 105		kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 106		return -EROFS;
 107	}
 108
 109	atomic_set(&cur_trans->num_writers, 1);
 110	cur_trans->num_joined = 0;
 111	init_waitqueue_head(&cur_trans->writer_wait);
 112	init_waitqueue_head(&cur_trans->commit_wait);
 113	cur_trans->in_commit = 0;
 114	cur_trans->blocked = 0;
 115	/*
 116	 * One for this trans handle, one so it will live on until we
 117	 * commit the transaction.
 118	 */
 119	atomic_set(&cur_trans->use_count, 2);
 120	cur_trans->commit_done = 0;
 121	cur_trans->start_time = get_seconds();
 122
 123	cur_trans->delayed_refs.root = RB_ROOT;
 124	cur_trans->delayed_refs.num_entries = 0;
 125	cur_trans->delayed_refs.num_heads_ready = 0;
 126	cur_trans->delayed_refs.num_heads = 0;
 127	cur_trans->delayed_refs.flushing = 0;
 128	cur_trans->delayed_refs.run_delayed_start = 0;
 129	cur_trans->delayed_refs.seq = 1;
 130
 131	/*
 132	 * although the tree mod log is per file system and not per transaction,
 133	 * the log must never go across transaction boundaries.
 134	 */
 135	smp_mb();
 136	if (!list_empty(&fs_info->tree_mod_seq_list)) {
 137		printk(KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 138			"creating a fresh transaction\n");
 139		WARN_ON(1);
 140	}
 141	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log)) {
 142		printk(KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 143			"creating a fresh transaction\n");
 144		WARN_ON(1);
 145	}
 146	atomic_set(&fs_info->tree_mod_seq, 0);
 147
 148	init_waitqueue_head(&cur_trans->delayed_refs.seq_wait);
 149	spin_lock_init(&cur_trans->commit_lock);
 150	spin_lock_init(&cur_trans->delayed_refs.lock);
 151	INIT_LIST_HEAD(&cur_trans->delayed_refs.seq_head);
 152
 153	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 154	list_add_tail(&cur_trans->list, &fs_info->trans_list);
 155	extent_io_tree_init(&cur_trans->dirty_pages,
 156			     fs_info->btree_inode->i_mapping);
 157	fs_info->generation++;
 158	cur_trans->transid = fs_info->generation;
 159	fs_info->running_transaction = cur_trans;
 160	cur_trans->aborted = 0;
 161	spin_unlock(&fs_info->trans_lock);
 162
 163	return 0;
 164}
 165
 166/*
 167 * this does all the record keeping required to make sure that a reference
 168 * counted root is properly recorded in a given transaction.  This is required
 169 * to make sure the old root from before we joined the transaction is deleted
 170 * when the transaction commits
 171 */
 172static int record_root_in_trans(struct btrfs_trans_handle *trans,
 173			       struct btrfs_root *root)
 174{
 175	if (root->ref_cows && root->last_trans < trans->transid) {
 176		WARN_ON(root == root->fs_info->extent_root);
 177		WARN_ON(root->commit_root != root->node);
 178
 179		/*
 180		 * see below for in_trans_setup usage rules
 181		 * we have the reloc mutex held now, so there
 182		 * is only one writer in this function
 183		 */
 184		root->in_trans_setup = 1;
 185
 186		/* make sure readers find in_trans_setup before
 187		 * they find our root->last_trans update
 188		 */
 189		smp_wmb();
 190
 191		spin_lock(&root->fs_info->fs_roots_radix_lock);
 192		if (root->last_trans == trans->transid) {
 193			spin_unlock(&root->fs_info->fs_roots_radix_lock);
 194			return 0;
 195		}
 196		radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 197			   (unsigned long)root->root_key.objectid,
 198			   BTRFS_ROOT_TRANS_TAG);
 199		spin_unlock(&root->fs_info->fs_roots_radix_lock);
 200		root->last_trans = trans->transid;
 201
 202		/* this is pretty tricky.  We don't want to
 203		 * take the relocation lock in btrfs_record_root_in_trans
 204		 * unless we're really doing the first setup for this root in
 205		 * this transaction.
 206		 *
 207		 * Normally we'd use root->last_trans as a flag to decide
 208		 * if we want to take the expensive mutex.
 209		 *
 210		 * But, we have to set root->last_trans before we
 211		 * init the relocation root, otherwise, we trip over warnings
 212		 * in ctree.c.  The solution used here is to flag ourselves
 213		 * with root->in_trans_setup.  When this is 1, we're still
 214		 * fixing up the reloc trees and everyone must wait.
 215		 *
 216		 * When this is zero, they can trust root->last_trans and fly
 217		 * through btrfs_record_root_in_trans without having to take the
 218		 * lock.  smp_wmb() makes sure that all the writes above are
 219		 * done before we pop in the zero below
 220		 */
 221		btrfs_init_reloc_root(trans, root);
 222		smp_wmb();
 223		root->in_trans_setup = 0;
 224	}
 225	return 0;
 226}
 227
 228
 229int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 230			       struct btrfs_root *root)
 231{
 232	if (!root->ref_cows)
 233		return 0;
 234
 235	/*
 236	 * see record_root_in_trans for comments about in_trans_setup usage
 237	 * and barriers
 238	 */
 239	smp_rmb();
 240	if (root->last_trans == trans->transid &&
 241	    !root->in_trans_setup)
 242		return 0;
 243
 244	mutex_lock(&root->fs_info->reloc_mutex);
 245	record_root_in_trans(trans, root);
 246	mutex_unlock(&root->fs_info->reloc_mutex);
 247
 248	return 0;
 249}
 250
 251/* wait for commit against the current transaction to become unblocked
 252 * when this is done, it is safe to start a new transaction, but the current
 253 * transaction might not be fully on disk.
 254 */
 255static void wait_current_trans(struct btrfs_root *root)
 256{
 257	struct btrfs_transaction *cur_trans;
 258
 259	spin_lock(&root->fs_info->trans_lock);
 260	cur_trans = root->fs_info->running_transaction;
 261	if (cur_trans && cur_trans->blocked) {
 262		atomic_inc(&cur_trans->use_count);
 263		spin_unlock(&root->fs_info->trans_lock);
 264
 265		wait_event(root->fs_info->transaction_wait,
 266			   !cur_trans->blocked);
 267		put_transaction(cur_trans);
 268	} else {
 269		spin_unlock(&root->fs_info->trans_lock);
 270	}
 271}
 272
 273enum btrfs_trans_type {
 274	TRANS_START,
 275	TRANS_JOIN,
 276	TRANS_USERSPACE,
 277	TRANS_JOIN_NOLOCK,
 278};
 279
 280static int may_wait_transaction(struct btrfs_root *root, int type)
 281{
 282	if (root->fs_info->log_root_recovering)
 283		return 0;
 284
 285	if (type == TRANS_USERSPACE)
 286		return 1;
 287
 288	if (type == TRANS_START &&
 289	    !atomic_read(&root->fs_info->open_ioctl_trans))
 290		return 1;
 291
 292	return 0;
 293}
 294
 295static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 296						    u64 num_items, int type)
 297{
 298	struct btrfs_trans_handle *h;
 299	struct btrfs_transaction *cur_trans;
 300	u64 num_bytes = 0;
 301	int ret;
 302
 303	if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 304		return ERR_PTR(-EROFS);
 305
 306	if (current->journal_info) {
 307		WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 308		h = current->journal_info;
 309		h->use_count++;
 310		h->orig_rsv = h->block_rsv;
 311		h->block_rsv = NULL;
 312		goto got_it;
 313	}
 314
 315	/*
 316	 * Do the reservation before we join the transaction so we can do all
 317	 * the appropriate flushing if need be.
 318	 */
 319	if (num_items > 0 && root != root->fs_info->chunk_root) {
 320		num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 321		ret = btrfs_block_rsv_add(root,
 322					  &root->fs_info->trans_block_rsv,
 323					  num_bytes);
 324		if (ret)
 325			return ERR_PTR(ret);
 326	}
 327again:
 328	h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 329	if (!h)
 330		return ERR_PTR(-ENOMEM);
 331
 332	if (may_wait_transaction(root, type))
 333		wait_current_trans(root);
 334
 335	do {
 336		ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 337		if (ret == -EBUSY)
 338			wait_current_trans(root);
 339	} while (ret == -EBUSY);
 340
 341	if (ret < 0) {
 342		kmem_cache_free(btrfs_trans_handle_cachep, h);
 343		return ERR_PTR(ret);
 344	}
 345
 346	cur_trans = root->fs_info->running_transaction;
 347
 348	h->transid = cur_trans->transid;
 349	h->transaction = cur_trans;
 350	h->blocks_used = 0;
 351	h->bytes_reserved = 0;
 352	h->delayed_ref_updates = 0;
 353	h->use_count = 1;
 354	h->block_rsv = NULL;
 355	h->orig_rsv = NULL;
 356	h->aborted = 0;
 357
 358	smp_mb();
 359	if (cur_trans->blocked && may_wait_transaction(root, type)) {
 360		btrfs_commit_transaction(h, root);
 361		goto again;
 362	}
 363
 364	if (num_bytes) {
 365		trace_btrfs_space_reservation(root->fs_info, "transaction",
 366					      h->transid, num_bytes, 1);
 367		h->block_rsv = &root->fs_info->trans_block_rsv;
 368		h->bytes_reserved = num_bytes;
 369	}
 370
 371got_it:
 372	btrfs_record_root_in_trans(h, root);
 373
 374	if (!current->journal_info && type != TRANS_USERSPACE)
 375		current->journal_info = h;
 376	return h;
 377}
 378
 379struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 380						   int num_items)
 381{
 382	return start_transaction(root, num_items, TRANS_START);
 383}
 384struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 385{
 386	return start_transaction(root, 0, TRANS_JOIN);
 387}
 388
 389struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 390{
 391	return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 392}
 393
 394struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 395{
 396	return start_transaction(root, 0, TRANS_USERSPACE);
 397}
 398
 399/* wait for a transaction commit to be fully complete */
 400static noinline void wait_for_commit(struct btrfs_root *root,
 401				    struct btrfs_transaction *commit)
 402{
 403	wait_event(commit->commit_wait, commit->commit_done);
 404}
 405
 406int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 407{
 408	struct btrfs_transaction *cur_trans = NULL, *t;
 409	int ret;
 410
 411	ret = 0;
 412	if (transid) {
 413		if (transid <= root->fs_info->last_trans_committed)
 414			goto out;
 415
 416		/* find specified transaction */
 417		spin_lock(&root->fs_info->trans_lock);
 418		list_for_each_entry(t, &root->fs_info->trans_list, list) {
 419			if (t->transid == transid) {
 420				cur_trans = t;
 421				atomic_inc(&cur_trans->use_count);
 422				break;
 423			}
 424			if (t->transid > transid)
 425				break;
 426		}
 427		spin_unlock(&root->fs_info->trans_lock);
 428		ret = -EINVAL;
 429		if (!cur_trans)
 430			goto out;  /* bad transid */
 431	} else {
 432		/* find newest transaction that is committing | committed */
 433		spin_lock(&root->fs_info->trans_lock);
 434		list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 435					    list) {
 436			if (t->in_commit) {
 437				if (t->commit_done)
 438					break;
 439				cur_trans = t;
 440				atomic_inc(&cur_trans->use_count);
 441				break;
 442			}
 443		}
 444		spin_unlock(&root->fs_info->trans_lock);
 445		if (!cur_trans)
 446			goto out;  /* nothing committing|committed */
 447	}
 448
 449	wait_for_commit(root, cur_trans);
 450
 451	put_transaction(cur_trans);
 452	ret = 0;
 453out:
 454	return ret;
 455}
 456
 457void btrfs_throttle(struct btrfs_root *root)
 458{
 459	if (!atomic_read(&root->fs_info->open_ioctl_trans))
 460		wait_current_trans(root);
 461}
 462
 463static int should_end_transaction(struct btrfs_trans_handle *trans,
 464				  struct btrfs_root *root)
 465{
 466	int ret;
 467
 468	ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 469	return ret ? 1 : 0;
 470}
 471
 472int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 473				 struct btrfs_root *root)
 474{
 475	struct btrfs_transaction *cur_trans = trans->transaction;
 476	struct btrfs_block_rsv *rsv = trans->block_rsv;
 477	int updates;
 478	int err;
 479
 480	smp_mb();
 481	if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 482		return 1;
 483
 484	/*
 485	 * We need to do this in case we're deleting csums so the global block
 486	 * rsv get's used instead of the csum block rsv.
 487	 */
 488	trans->block_rsv = NULL;
 489
 490	updates = trans->delayed_ref_updates;
 491	trans->delayed_ref_updates = 0;
 492	if (updates) {
 493		err = btrfs_run_delayed_refs(trans, root, updates);
 494		if (err) /* Error code will also eval true */
 495			return err;
 496	}
 497
 498	trans->block_rsv = rsv;
 499
 500	return should_end_transaction(trans, root);
 501}
 502
 503static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 504			  struct btrfs_root *root, int throttle, int lock)
 505{
 506	struct btrfs_transaction *cur_trans = trans->transaction;
 507	struct btrfs_fs_info *info = root->fs_info;
 508	int count = 0;
 509	int err = 0;
 510
 511	if (--trans->use_count) {
 512		trans->block_rsv = trans->orig_rsv;
 513		return 0;
 514	}
 515
 516	btrfs_trans_release_metadata(trans, root);
 517	trans->block_rsv = NULL;
 518	while (count < 2) {
 519		unsigned long cur = trans->delayed_ref_updates;
 520		trans->delayed_ref_updates = 0;
 521		if (cur &&
 522		    trans->transaction->delayed_refs.num_heads_ready > 64) {
 523			trans->delayed_ref_updates = 0;
 
 
 
 
 
 
 
 524			btrfs_run_delayed_refs(trans, root, cur);
 525		} else {
 526			break;
 527		}
 528		count++;
 529	}
 530
 
 
 531	if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 532	    should_end_transaction(trans, root)) {
 533		trans->transaction->blocked = 1;
 534		smp_wmb();
 535	}
 536
 537	if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 538		if (throttle) {
 539			/*
 540			 * We may race with somebody else here so end up having
 541			 * to call end_transaction on ourselves again, so inc
 542			 * our use_count.
 543			 */
 544			trans->use_count++;
 545			return btrfs_commit_transaction(trans, root);
 546		} else {
 547			wake_up_process(info->transaction_kthread);
 548		}
 549	}
 550
 551	WARN_ON(cur_trans != info->running_transaction);
 552	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 553	atomic_dec(&cur_trans->num_writers);
 554
 555	smp_mb();
 556	if (waitqueue_active(&cur_trans->writer_wait))
 557		wake_up(&cur_trans->writer_wait);
 558	put_transaction(cur_trans);
 559
 560	if (current->journal_info == trans)
 561		current->journal_info = NULL;
 
 
 562
 563	if (throttle)
 564		btrfs_run_delayed_iputs(root);
 565
 566	if (trans->aborted ||
 567	    root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 568		err = -EIO;
 569	}
 570
 571	memset(trans, 0, sizeof(*trans));
 572	kmem_cache_free(btrfs_trans_handle_cachep, trans);
 573	return err;
 574}
 575
 576int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 577			  struct btrfs_root *root)
 578{
 579	int ret;
 580
 581	ret = __btrfs_end_transaction(trans, root, 0, 1);
 582	if (ret)
 583		return ret;
 584	return 0;
 585}
 586
 587int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 588				   struct btrfs_root *root)
 589{
 590	int ret;
 591
 592	ret = __btrfs_end_transaction(trans, root, 1, 1);
 593	if (ret)
 594		return ret;
 595	return 0;
 596}
 597
 598int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 599				 struct btrfs_root *root)
 600{
 601	int ret;
 602
 603	ret = __btrfs_end_transaction(trans, root, 0, 0);
 604	if (ret)
 605		return ret;
 606	return 0;
 607}
 608
 609int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 610				struct btrfs_root *root)
 611{
 612	return __btrfs_end_transaction(trans, root, 1, 1);
 613}
 614
 615/*
 616 * when btree blocks are allocated, they have some corresponding bits set for
 617 * them in one of two extent_io trees.  This is used to make sure all of
 618 * those extents are sent to disk but does not wait on them
 619 */
 620int btrfs_write_marked_extents(struct btrfs_root *root,
 621			       struct extent_io_tree *dirty_pages, int mark)
 622{
 
 623	int err = 0;
 624	int werr = 0;
 625	struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 
 626	u64 start = 0;
 627	u64 end;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 628
 629	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 630				      mark)) {
 631		convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT, mark,
 632				   GFP_NOFS);
 633		err = filemap_fdatawrite_range(mapping, start, end);
 634		if (err)
 635			werr = err;
 636		cond_resched();
 637		start = end + 1;
 
 
 
 
 
 638	}
 639	if (err)
 640		werr = err;
 641	return werr;
 642}
 643
 644/*
 645 * when btree blocks are allocated, they have some corresponding bits set for
 646 * them in one of two extent_io trees.  This is used to make sure all of
 647 * those extents are on disk for transaction or log commit.  We wait
 648 * on all the pages and clear them from the dirty pages state tree
 649 */
 650int btrfs_wait_marked_extents(struct btrfs_root *root,
 651			      struct extent_io_tree *dirty_pages, int mark)
 652{
 
 653	int err = 0;
 654	int werr = 0;
 655	struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 
 656	u64 start = 0;
 657	u64 end;
 
 658
 659	while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 660				      EXTENT_NEED_WAIT)) {
 661		clear_extent_bits(dirty_pages, start, end, EXTENT_NEED_WAIT, GFP_NOFS);
 662		err = filemap_fdatawait_range(mapping, start, end);
 663		if (err)
 664			werr = err;
 665		cond_resched();
 666		start = end + 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 667	}
 668	if (err)
 669		werr = err;
 670	return werr;
 671}
 672
 673/*
 674 * when btree blocks are allocated, they have some corresponding bits set for
 675 * them in one of two extent_io trees.  This is used to make sure all of
 676 * those extents are on disk for transaction or log commit
 677 */
 678int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 679				struct extent_io_tree *dirty_pages, int mark)
 680{
 681	int ret;
 682	int ret2;
 683
 684	ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 685	ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 686
 687	if (ret)
 688		return ret;
 689	if (ret2)
 690		return ret2;
 691	return 0;
 692}
 693
 694int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 695				     struct btrfs_root *root)
 696{
 697	if (!trans || !trans->transaction) {
 698		struct inode *btree_inode;
 699		btree_inode = root->fs_info->btree_inode;
 700		return filemap_write_and_wait(btree_inode->i_mapping);
 701	}
 702	return btrfs_write_and_wait_marked_extents(root,
 703					   &trans->transaction->dirty_pages,
 704					   EXTENT_DIRTY);
 705}
 706
 707/*
 708 * this is used to update the root pointer in the tree of tree roots.
 709 *
 710 * But, in the case of the extent allocation tree, updating the root
 711 * pointer may allocate blocks which may change the root of the extent
 712 * allocation tree.
 713 *
 714 * So, this loops and repeats and makes sure the cowonly root didn't
 715 * change while the root pointer was being updated in the metadata.
 716 */
 717static int update_cowonly_root(struct btrfs_trans_handle *trans,
 718			       struct btrfs_root *root)
 719{
 720	int ret;
 721	u64 old_root_bytenr;
 722	u64 old_root_used;
 723	struct btrfs_root *tree_root = root->fs_info->tree_root;
 724
 725	old_root_used = btrfs_root_used(&root->root_item);
 726	btrfs_write_dirty_block_groups(trans, root);
 727
 728	while (1) {
 729		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 730		if (old_root_bytenr == root->node->start &&
 731		    old_root_used == btrfs_root_used(&root->root_item))
 732			break;
 733
 734		btrfs_set_root_node(&root->root_item, root->node);
 735		ret = btrfs_update_root(trans, tree_root,
 736					&root->root_key,
 737					&root->root_item);
 738		if (ret)
 739			return ret;
 740
 741		old_root_used = btrfs_root_used(&root->root_item);
 742		ret = btrfs_write_dirty_block_groups(trans, root);
 743		if (ret)
 744			return ret;
 745	}
 746
 747	if (root != root->fs_info->extent_root)
 748		switch_commit_root(root);
 749
 750	return 0;
 751}
 752
 753/*
 754 * update all the cowonly tree roots on disk
 755 *
 756 * The error handling in this function may not be obvious. Any of the
 757 * failures will cause the file system to go offline. We still need
 758 * to clean up the delayed refs.
 759 */
 760static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 761					 struct btrfs_root *root)
 762{
 763	struct btrfs_fs_info *fs_info = root->fs_info;
 764	struct list_head *next;
 765	struct extent_buffer *eb;
 766	int ret;
 767
 768	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 769	if (ret)
 770		return ret;
 771
 772	eb = btrfs_lock_root_node(fs_info->tree_root);
 773	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 774			      0, &eb);
 775	btrfs_tree_unlock(eb);
 776	free_extent_buffer(eb);
 777
 778	if (ret)
 779		return ret;
 780
 781	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 782	if (ret)
 783		return ret;
 784
 785	ret = btrfs_run_dev_stats(trans, root->fs_info);
 786	BUG_ON(ret);
 787
 788	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 789		next = fs_info->dirty_cowonly_roots.next;
 790		list_del_init(next);
 791		root = list_entry(next, struct btrfs_root, dirty_list);
 792
 793		ret = update_cowonly_root(trans, root);
 794		if (ret)
 795			return ret;
 796	}
 797
 798	down_write(&fs_info->extent_commit_sem);
 799	switch_commit_root(fs_info->extent_root);
 800	up_write(&fs_info->extent_commit_sem);
 801
 802	return 0;
 803}
 804
 805/*
 806 * dead roots are old snapshots that need to be deleted.  This allocates
 807 * a dirty root struct and adds it into the list of dead roots that need to
 808 * be deleted
 809 */
 810int btrfs_add_dead_root(struct btrfs_root *root)
 811{
 812	spin_lock(&root->fs_info->trans_lock);
 813	list_add(&root->root_list, &root->fs_info->dead_roots);
 814	spin_unlock(&root->fs_info->trans_lock);
 815	return 0;
 816}
 817
 818/*
 819 * update all the cowonly tree roots on disk
 820 */
 821static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 822				    struct btrfs_root *root)
 823{
 824	struct btrfs_root *gang[8];
 825	struct btrfs_fs_info *fs_info = root->fs_info;
 826	int i;
 827	int ret;
 828	int err = 0;
 829
 830	spin_lock(&fs_info->fs_roots_radix_lock);
 831	while (1) {
 832		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 833						 (void **)gang, 0,
 834						 ARRAY_SIZE(gang),
 835						 BTRFS_ROOT_TRANS_TAG);
 836		if (ret == 0)
 837			break;
 838		for (i = 0; i < ret; i++) {
 839			root = gang[i];
 840			radix_tree_tag_clear(&fs_info->fs_roots_radix,
 841					(unsigned long)root->root_key.objectid,
 842					BTRFS_ROOT_TRANS_TAG);
 843			spin_unlock(&fs_info->fs_roots_radix_lock);
 844
 845			btrfs_free_log(trans, root);
 846			btrfs_update_reloc_root(trans, root);
 847			btrfs_orphan_commit_root(trans, root);
 848
 849			btrfs_save_ino_cache(root, trans);
 850
 851			/* see comments in should_cow_block() */
 852			root->force_cow = 0;
 853			smp_wmb();
 854
 855			if (root->commit_root != root->node) {
 856				mutex_lock(&root->fs_commit_mutex);
 857				switch_commit_root(root);
 858				btrfs_unpin_free_ino(root);
 859				mutex_unlock(&root->fs_commit_mutex);
 860
 861				btrfs_set_root_node(&root->root_item,
 862						    root->node);
 863			}
 864
 865			err = btrfs_update_root(trans, fs_info->tree_root,
 866						&root->root_key,
 867						&root->root_item);
 868			spin_lock(&fs_info->fs_roots_radix_lock);
 869			if (err)
 870				break;
 871		}
 872	}
 873	spin_unlock(&fs_info->fs_roots_radix_lock);
 874	return err;
 875}
 876
 877/*
 878 * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 879 * otherwise every leaf in the btree is read and defragged.
 880 */
 881int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 882{
 883	struct btrfs_fs_info *info = root->fs_info;
 884	struct btrfs_trans_handle *trans;
 885	int ret;
 886	unsigned long nr;
 887
 888	if (xchg(&root->defrag_running, 1))
 889		return 0;
 890
 891	while (1) {
 892		trans = btrfs_start_transaction(root, 0);
 893		if (IS_ERR(trans))
 894			return PTR_ERR(trans);
 895
 896		ret = btrfs_defrag_leaves(trans, root, cacheonly);
 897
 898		nr = trans->blocks_used;
 899		btrfs_end_transaction(trans, root);
 900		btrfs_btree_balance_dirty(info->tree_root, nr);
 901		cond_resched();
 902
 903		if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 904			break;
 905	}
 906	root->defrag_running = 0;
 907	return ret;
 908}
 909
 910/*
 911 * new snapshots need to be created at a very specific time in the
 912 * transaction commit.  This does the actual creation
 913 */
 914static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 915				   struct btrfs_fs_info *fs_info,
 916				   struct btrfs_pending_snapshot *pending)
 917{
 918	struct btrfs_key key;
 919	struct btrfs_root_item *new_root_item;
 920	struct btrfs_root *tree_root = fs_info->tree_root;
 921	struct btrfs_root *root = pending->root;
 922	struct btrfs_root *parent_root;
 923	struct btrfs_block_rsv *rsv;
 924	struct inode *parent_inode;
 925	struct dentry *parent;
 926	struct dentry *dentry;
 927	struct extent_buffer *tmp;
 928	struct extent_buffer *old;
 929	int ret;
 930	u64 to_reserve = 0;
 931	u64 index = 0;
 932	u64 objectid;
 933	u64 root_flags;
 934
 935	rsv = trans->block_rsv;
 936
 937	new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 938	if (!new_root_item) {
 939		ret = pending->error = -ENOMEM;
 940		goto fail;
 941	}
 942
 943	ret = btrfs_find_free_objectid(tree_root, &objectid);
 944	if (ret) {
 945		pending->error = ret;
 946		goto fail;
 947	}
 948
 949	btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 
 950
 951	if (to_reserve > 0) {
 952		ret = btrfs_block_rsv_add_noflush(root, &pending->block_rsv,
 953						  to_reserve);
 954		if (ret) {
 955			pending->error = ret;
 956			goto fail;
 957		}
 958	}
 959
 960	key.objectid = objectid;
 961	key.offset = (u64)-1;
 962	key.type = BTRFS_ROOT_ITEM_KEY;
 963
 964	trans->block_rsv = &pending->block_rsv;
 965
 966	dentry = pending->dentry;
 967	parent = dget_parent(dentry);
 968	parent_inode = parent->d_inode;
 969	parent_root = BTRFS_I(parent_inode)->root;
 970	record_root_in_trans(trans, parent_root);
 971
 972	/*
 973	 * insert the directory item
 974	 */
 975	ret = btrfs_set_inode_index(parent_inode, &index);
 976	BUG_ON(ret); /* -ENOMEM */
 977	ret = btrfs_insert_dir_item(trans, parent_root,
 978				dentry->d_name.name, dentry->d_name.len,
 979				parent_inode, &key,
 980				BTRFS_FT_DIR, index);
 981	if (ret == -EEXIST) {
 982		pending->error = -EEXIST;
 983		dput(parent);
 984		goto fail;
 985	} else if (ret) {
 986		goto abort_trans_dput;
 987	}
 988
 989	btrfs_i_size_write(parent_inode, parent_inode->i_size +
 990					 dentry->d_name.len * 2);
 991	ret = btrfs_update_inode(trans, parent_root, parent_inode);
 992	if (ret)
 993		goto abort_trans_dput;
 994
 995	/*
 996	 * pull in the delayed directory update
 997	 * and the delayed inode item
 998	 * otherwise we corrupt the FS during
 999	 * snapshot
1000	 */
1001	ret = btrfs_run_delayed_items(trans, root);
1002	if (ret) { /* Transaction aborted */
1003		dput(parent);
1004		goto fail;
1005	}
1006
1007	record_root_in_trans(trans, root);
1008	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1009	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1010	btrfs_check_and_init_root_item(new_root_item);
1011
1012	root_flags = btrfs_root_flags(new_root_item);
1013	if (pending->readonly)
1014		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1015	else
1016		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1017	btrfs_set_root_flags(new_root_item, root_flags);
1018
1019	old = btrfs_lock_root_node(root);
1020	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1021	if (ret) {
1022		btrfs_tree_unlock(old);
1023		free_extent_buffer(old);
1024		goto abort_trans_dput;
1025	}
1026
1027	btrfs_set_lock_blocking(old);
1028
1029	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1030	/* clean up in any case */
1031	btrfs_tree_unlock(old);
1032	free_extent_buffer(old);
1033	if (ret)
1034		goto abort_trans_dput;
1035
1036	/* see comments in should_cow_block() */
1037	root->force_cow = 1;
1038	smp_wmb();
1039
1040	btrfs_set_root_node(new_root_item, tmp);
1041	/* record when the snapshot was created in key.offset */
1042	key.offset = trans->transid;
1043	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1044	btrfs_tree_unlock(tmp);
1045	free_extent_buffer(tmp);
1046	if (ret)
1047		goto abort_trans_dput;
1048
1049	/*
1050	 * insert root back/forward references
1051	 */
1052	ret = btrfs_add_root_ref(trans, tree_root, objectid,
1053				 parent_root->root_key.objectid,
1054				 btrfs_ino(parent_inode), index,
1055				 dentry->d_name.name, dentry->d_name.len);
 
1056	dput(parent);
1057	if (ret)
1058		goto fail;
1059
1060	key.offset = (u64)-1;
1061	pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1062	if (IS_ERR(pending->snap)) {
1063		ret = PTR_ERR(pending->snap);
1064		goto abort_trans;
1065	}
1066
1067	ret = btrfs_reloc_post_snapshot(trans, pending);
1068	if (ret)
1069		goto abort_trans;
1070	ret = 0;
1071fail:
1072	kfree(new_root_item);
1073	trans->block_rsv = rsv;
1074	btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1075	return ret;
1076
1077abort_trans_dput:
1078	dput(parent);
1079abort_trans:
1080	btrfs_abort_transaction(trans, root, ret);
1081	goto fail;
1082}
1083
1084/*
1085 * create all the snapshots we've scheduled for creation
1086 */
1087static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1088					     struct btrfs_fs_info *fs_info)
1089{
1090	struct btrfs_pending_snapshot *pending;
1091	struct list_head *head = &trans->transaction->pending_snapshots;
 
1092
1093	list_for_each_entry(pending, head, list)
1094		create_pending_snapshot(trans, fs_info, pending);
 
 
1095	return 0;
1096}
1097
1098static void update_super_roots(struct btrfs_root *root)
1099{
1100	struct btrfs_root_item *root_item;
1101	struct btrfs_super_block *super;
1102
1103	super = root->fs_info->super_copy;
1104
1105	root_item = &root->fs_info->chunk_root->root_item;
1106	super->chunk_root = root_item->bytenr;
1107	super->chunk_root_generation = root_item->generation;
1108	super->chunk_root_level = root_item->level;
1109
1110	root_item = &root->fs_info->tree_root->root_item;
1111	super->root = root_item->bytenr;
1112	super->generation = root_item->generation;
1113	super->root_level = root_item->level;
1114	if (btrfs_test_opt(root, SPACE_CACHE))
1115		super->cache_generation = root_item->generation;
1116}
1117
1118int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1119{
1120	int ret = 0;
1121	spin_lock(&info->trans_lock);
1122	if (info->running_transaction)
1123		ret = info->running_transaction->in_commit;
1124	spin_unlock(&info->trans_lock);
1125	return ret;
1126}
1127
1128int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1129{
1130	int ret = 0;
1131	spin_lock(&info->trans_lock);
1132	if (info->running_transaction)
1133		ret = info->running_transaction->blocked;
1134	spin_unlock(&info->trans_lock);
1135	return ret;
1136}
1137
1138/*
1139 * wait for the current transaction commit to start and block subsequent
1140 * transaction joins
1141 */
1142static void wait_current_trans_commit_start(struct btrfs_root *root,
1143					    struct btrfs_transaction *trans)
1144{
1145	wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1146}
1147
1148/*
1149 * wait for the current transaction to start and then become unblocked.
1150 * caller holds ref.
1151 */
1152static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1153					 struct btrfs_transaction *trans)
1154{
1155	wait_event(root->fs_info->transaction_wait,
1156		   trans->commit_done || (trans->in_commit && !trans->blocked));
1157}
1158
1159/*
1160 * commit transactions asynchronously. once btrfs_commit_transaction_async
1161 * returns, any subsequent transaction will not be allowed to join.
1162 */
1163struct btrfs_async_commit {
1164	struct btrfs_trans_handle *newtrans;
1165	struct btrfs_root *root;
1166	struct delayed_work work;
1167};
1168
1169static void do_async_commit(struct work_struct *work)
1170{
1171	struct btrfs_async_commit *ac =
1172		container_of(work, struct btrfs_async_commit, work.work);
1173
1174	btrfs_commit_transaction(ac->newtrans, ac->root);
1175	kfree(ac);
1176}
1177
1178int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1179				   struct btrfs_root *root,
1180				   int wait_for_unblock)
1181{
1182	struct btrfs_async_commit *ac;
1183	struct btrfs_transaction *cur_trans;
1184
1185	ac = kmalloc(sizeof(*ac), GFP_NOFS);
1186	if (!ac)
1187		return -ENOMEM;
1188
1189	INIT_DELAYED_WORK(&ac->work, do_async_commit);
1190	ac->root = root;
1191	ac->newtrans = btrfs_join_transaction(root);
1192	if (IS_ERR(ac->newtrans)) {
1193		int err = PTR_ERR(ac->newtrans);
1194		kfree(ac);
1195		return err;
1196	}
1197
1198	/* take transaction reference */
1199	cur_trans = trans->transaction;
1200	atomic_inc(&cur_trans->use_count);
1201
1202	btrfs_end_transaction(trans, root);
1203	schedule_delayed_work(&ac->work, 0);
1204
1205	/* wait for transaction to start and unblock */
1206	if (wait_for_unblock)
1207		wait_current_trans_commit_start_and_unblock(root, cur_trans);
1208	else
1209		wait_current_trans_commit_start(root, cur_trans);
1210
1211	if (current->journal_info == trans)
1212		current->journal_info = NULL;
1213
1214	put_transaction(cur_trans);
1215	return 0;
1216}
1217
1218
1219static void cleanup_transaction(struct btrfs_trans_handle *trans,
1220				struct btrfs_root *root, int err)
1221{
1222	struct btrfs_transaction *cur_trans = trans->transaction;
1223
1224	WARN_ON(trans->use_count > 1);
1225
1226	btrfs_abort_transaction(trans, root, err);
1227
1228	spin_lock(&root->fs_info->trans_lock);
1229	list_del_init(&cur_trans->list);
1230	if (cur_trans == root->fs_info->running_transaction) {
1231		root->fs_info->running_transaction = NULL;
1232		root->fs_info->trans_no_join = 0;
1233	}
1234	spin_unlock(&root->fs_info->trans_lock);
1235
1236	btrfs_cleanup_one_transaction(trans->transaction, root);
1237
1238	put_transaction(cur_trans);
1239	put_transaction(cur_trans);
1240
1241	trace_btrfs_transaction_commit(root);
1242
1243	btrfs_scrub_continue(root);
1244
1245	if (current->journal_info == trans)
1246		current->journal_info = NULL;
1247
1248	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1249}
1250
1251/*
1252 * btrfs_transaction state sequence:
1253 *    in_commit = 0, blocked = 0  (initial)
1254 *    in_commit = 1, blocked = 1
1255 *    blocked = 0
1256 *    commit_done = 1
1257 */
1258int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1259			     struct btrfs_root *root)
1260{
1261	unsigned long joined = 0;
1262	struct btrfs_transaction *cur_trans = trans->transaction;
1263	struct btrfs_transaction *prev_trans = NULL;
1264	DEFINE_WAIT(wait);
1265	int ret = -EIO;
1266	int should_grow = 0;
1267	unsigned long now = get_seconds();
1268	int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1269
1270	btrfs_run_ordered_operations(root, 0);
1271
1272	btrfs_trans_release_metadata(trans, root);
1273	trans->block_rsv = NULL;
1274
1275	if (cur_trans->aborted)
1276		goto cleanup_transaction;
1277
1278	/* make a pass through all the delayed refs we have so far
1279	 * any runnings procs may add more while we are here
1280	 */
1281	ret = btrfs_run_delayed_refs(trans, root, 0);
1282	if (ret)
1283		goto cleanup_transaction;
 
1284
1285	cur_trans = trans->transaction;
1286
1287	/*
1288	 * set the flushing flag so procs in this transaction have to
1289	 * start sending their work down.
1290	 */
1291	cur_trans->delayed_refs.flushing = 1;
1292
1293	ret = btrfs_run_delayed_refs(trans, root, 0);
1294	if (ret)
1295		goto cleanup_transaction;
1296
1297	spin_lock(&cur_trans->commit_lock);
1298	if (cur_trans->in_commit) {
1299		spin_unlock(&cur_trans->commit_lock);
1300		atomic_inc(&cur_trans->use_count);
1301		ret = btrfs_end_transaction(trans, root);
1302
1303		wait_for_commit(root, cur_trans);
1304
1305		put_transaction(cur_trans);
1306
1307		return ret;
1308	}
1309
1310	trans->transaction->in_commit = 1;
1311	trans->transaction->blocked = 1;
1312	spin_unlock(&cur_trans->commit_lock);
1313	wake_up(&root->fs_info->transaction_blocked_wait);
1314
1315	spin_lock(&root->fs_info->trans_lock);
1316	if (cur_trans->list.prev != &root->fs_info->trans_list) {
1317		prev_trans = list_entry(cur_trans->list.prev,
1318					struct btrfs_transaction, list);
1319		if (!prev_trans->commit_done) {
1320			atomic_inc(&prev_trans->use_count);
1321			spin_unlock(&root->fs_info->trans_lock);
1322
1323			wait_for_commit(root, prev_trans);
1324
1325			put_transaction(prev_trans);
1326		} else {
1327			spin_unlock(&root->fs_info->trans_lock);
1328		}
1329	} else {
1330		spin_unlock(&root->fs_info->trans_lock);
1331	}
1332
1333	if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1334		should_grow = 1;
1335
1336	do {
1337		int snap_pending = 0;
1338
1339		joined = cur_trans->num_joined;
1340		if (!list_empty(&trans->transaction->pending_snapshots))
1341			snap_pending = 1;
1342
1343		WARN_ON(cur_trans != trans->transaction);
1344
1345		if (flush_on_commit || snap_pending) {
1346			btrfs_start_delalloc_inodes(root, 1);
1347			btrfs_wait_ordered_extents(root, 0, 1);
 
1348		}
1349
1350		ret = btrfs_run_delayed_items(trans, root);
1351		if (ret)
1352			goto cleanup_transaction;
1353
1354		/*
1355		 * rename don't use btrfs_join_transaction, so, once we
1356		 * set the transaction to blocked above, we aren't going
1357		 * to get any new ordered operations.  We can safely run
1358		 * it here and no for sure that nothing new will be added
1359		 * to the list
1360		 */
1361		btrfs_run_ordered_operations(root, 1);
1362
1363		prepare_to_wait(&cur_trans->writer_wait, &wait,
1364				TASK_UNINTERRUPTIBLE);
1365
1366		if (atomic_read(&cur_trans->num_writers) > 1)
1367			schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1368		else if (should_grow)
1369			schedule_timeout(1);
1370
1371		finish_wait(&cur_trans->writer_wait, &wait);
1372	} while (atomic_read(&cur_trans->num_writers) > 1 ||
1373		 (should_grow && cur_trans->num_joined != joined));
1374
1375	/*
1376	 * Ok now we need to make sure to block out any other joins while we
1377	 * commit the transaction.  We could have started a join before setting
1378	 * no_join so make sure to wait for num_writers to == 1 again.
1379	 */
1380	spin_lock(&root->fs_info->trans_lock);
1381	root->fs_info->trans_no_join = 1;
1382	spin_unlock(&root->fs_info->trans_lock);
1383	wait_event(cur_trans->writer_wait,
1384		   atomic_read(&cur_trans->num_writers) == 1);
1385
1386	/*
1387	 * the reloc mutex makes sure that we stop
1388	 * the balancing code from coming in and moving
1389	 * extents around in the middle of the commit
1390	 */
1391	mutex_lock(&root->fs_info->reloc_mutex);
1392
1393	ret = btrfs_run_delayed_items(trans, root);
1394	if (ret) {
1395		mutex_unlock(&root->fs_info->reloc_mutex);
1396		goto cleanup_transaction;
1397	}
1398
1399	ret = create_pending_snapshots(trans, root->fs_info);
1400	if (ret) {
1401		mutex_unlock(&root->fs_info->reloc_mutex);
1402		goto cleanup_transaction;
1403	}
1404
1405	ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1406	if (ret) {
1407		mutex_unlock(&root->fs_info->reloc_mutex);
1408		goto cleanup_transaction;
1409	}
1410
1411	/*
1412	 * make sure none of the code above managed to slip in a
1413	 * delayed item
1414	 */
1415	btrfs_assert_delayed_root_empty(root);
1416
1417	WARN_ON(cur_trans != trans->transaction);
1418
1419	btrfs_scrub_pause(root);
1420	/* btrfs_commit_tree_roots is responsible for getting the
1421	 * various roots consistent with each other.  Every pointer
1422	 * in the tree of tree roots has to point to the most up to date
1423	 * root for every subvolume and other tree.  So, we have to keep
1424	 * the tree logging code from jumping in and changing any
1425	 * of the trees.
1426	 *
1427	 * At this point in the commit, there can't be any tree-log
1428	 * writers, but a little lower down we drop the trans mutex
1429	 * and let new people in.  By holding the tree_log_mutex
1430	 * from now until after the super is written, we avoid races
1431	 * with the tree-log code.
1432	 */
1433	mutex_lock(&root->fs_info->tree_log_mutex);
1434
1435	ret = commit_fs_roots(trans, root);
1436	if (ret) {
1437		mutex_unlock(&root->fs_info->tree_log_mutex);
1438		mutex_unlock(&root->fs_info->reloc_mutex);
1439		goto cleanup_transaction;
1440	}
1441
1442	/* commit_fs_roots gets rid of all the tree log roots, it is now
1443	 * safe to free the root of tree log roots
1444	 */
1445	btrfs_free_log_root_tree(trans, root->fs_info);
1446
1447	ret = commit_cowonly_roots(trans, root);
1448	if (ret) {
1449		mutex_unlock(&root->fs_info->tree_log_mutex);
1450		mutex_unlock(&root->fs_info->reloc_mutex);
1451		goto cleanup_transaction;
1452	}
1453
1454	btrfs_prepare_extent_commit(trans, root);
1455
1456	cur_trans = root->fs_info->running_transaction;
1457
1458	btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1459			    root->fs_info->tree_root->node);
1460	switch_commit_root(root->fs_info->tree_root);
1461
1462	btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1463			    root->fs_info->chunk_root->node);
1464	switch_commit_root(root->fs_info->chunk_root);
1465
1466	update_super_roots(root);
1467
1468	if (!root->fs_info->log_root_recovering) {
1469		btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1470		btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1471	}
1472
1473	memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1474	       sizeof(*root->fs_info->super_copy));
1475
1476	trans->transaction->blocked = 0;
1477	spin_lock(&root->fs_info->trans_lock);
1478	root->fs_info->running_transaction = NULL;
1479	root->fs_info->trans_no_join = 0;
1480	spin_unlock(&root->fs_info->trans_lock);
1481	mutex_unlock(&root->fs_info->reloc_mutex);
1482
1483	wake_up(&root->fs_info->transaction_wait);
1484
1485	ret = btrfs_write_and_wait_transaction(trans, root);
1486	if (ret) {
1487		btrfs_error(root->fs_info, ret,
1488			    "Error while writing out transaction.");
1489		mutex_unlock(&root->fs_info->tree_log_mutex);
1490		goto cleanup_transaction;
1491	}
1492
1493	ret = write_ctree_super(trans, root, 0);
1494	if (ret) {
1495		mutex_unlock(&root->fs_info->tree_log_mutex);
1496		goto cleanup_transaction;
1497	}
1498
1499	/*
1500	 * the super is written, we can safely allow the tree-loggers
1501	 * to go about their business
1502	 */
1503	mutex_unlock(&root->fs_info->tree_log_mutex);
1504
1505	btrfs_finish_extent_commit(trans, root);
1506
1507	cur_trans->commit_done = 1;
1508
1509	root->fs_info->last_trans_committed = cur_trans->transid;
1510
1511	wake_up(&cur_trans->commit_wait);
1512
1513	spin_lock(&root->fs_info->trans_lock);
1514	list_del_init(&cur_trans->list);
1515	spin_unlock(&root->fs_info->trans_lock);
1516
1517	put_transaction(cur_trans);
1518	put_transaction(cur_trans);
1519
1520	trace_btrfs_transaction_commit(root);
1521
1522	btrfs_scrub_continue(root);
1523
1524	if (current->journal_info == trans)
1525		current->journal_info = NULL;
1526
1527	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1528
1529	if (current != root->fs_info->transaction_kthread)
1530		btrfs_run_delayed_iputs(root);
1531
1532	return ret;
1533
1534cleanup_transaction:
1535	btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1536//	WARN_ON(1);
1537	if (current->journal_info == trans)
1538		current->journal_info = NULL;
1539	cleanup_transaction(trans, root, ret);
1540
1541	return ret;
1542}
1543
1544/*
1545 * interface function to delete all the snapshots we have scheduled for deletion
1546 */
1547int btrfs_clean_old_snapshots(struct btrfs_root *root)
1548{
1549	LIST_HEAD(list);
1550	struct btrfs_fs_info *fs_info = root->fs_info;
1551
1552	spin_lock(&fs_info->trans_lock);
1553	list_splice_init(&fs_info->dead_roots, &list);
1554	spin_unlock(&fs_info->trans_lock);
1555
1556	while (!list_empty(&list)) {
1557		int ret;
1558
1559		root = list_entry(list.next, struct btrfs_root, root_list);
1560		list_del(&root->root_list);
1561
1562		btrfs_kill_all_delayed_nodes(root);
1563
1564		if (btrfs_header_backref_rev(root->node) <
1565		    BTRFS_MIXED_BACKREF_REV)
1566			ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1567		else
1568			ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1569		BUG_ON(ret < 0);
1570	}
1571	return 0;
1572}