move_extents.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * move_extents.c
   4 *
   5 * Copyright (C) 2011 Oracle.  All rights reserved.
   6 */
   7#include <linux/fs.h>
   8#include <linux/types.h>
   9#include <linux/mount.h>
  10#include <linux/swap.h>
  11
  12#include <cluster/masklog.h>
  13
  14#include "ocfs2.h"
  15#include "ocfs2_ioctl.h"
  16
  17#include "alloc.h"
  18#include "localalloc.h"
  19#include "aops.h"
  20#include "dlmglue.h"
  21#include "extent_map.h"
  22#include "inode.h"
  23#include "journal.h"
  24#include "suballoc.h"
  25#include "uptodate.h"
  26#include "super.h"
  27#include "dir.h"
  28#include "buffer_head_io.h"
  29#include "sysfile.h"
  30#include "refcounttree.h"
  31#include "move_extents.h"
  32
  33struct ocfs2_move_extents_context {
  34	struct inode *inode;
  35	struct file *file;
  36	int auto_defrag;
  37	int partial;
  38	int credits;
  39	u32 new_phys_cpos;
  40	u32 clusters_moved;
  41	u64 refcount_loc;
  42	struct ocfs2_move_extents *range;
  43	struct ocfs2_extent_tree et;
  44	struct ocfs2_alloc_context *meta_ac;
  45	struct ocfs2_alloc_context *data_ac;
  46	struct ocfs2_cached_dealloc_ctxt dealloc;
  47};
  48
  49static int __ocfs2_move_extent(handle_t *handle,
  50			       struct ocfs2_move_extents_context *context,
  51			       u32 cpos, u32 len, u32 p_cpos, u32 new_p_cpos,
  52			       int ext_flags)
  53{
  54	int ret = 0, index;
  55	struct inode *inode = context->inode;
  56	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  57	struct ocfs2_extent_rec *rec, replace_rec;
  58	struct ocfs2_path *path = NULL;
  59	struct ocfs2_extent_list *el;
  60	u64 ino = ocfs2_metadata_cache_owner(context->et.et_ci);
  61	u64 old_blkno = ocfs2_clusters_to_blocks(inode->i_sb, p_cpos);
  62
  63	ret = ocfs2_duplicate_clusters_by_page(handle, inode, cpos,
  64					       p_cpos, new_p_cpos, len);
  65	if (ret) {
  66		mlog_errno(ret);
  67		goto out;
  68	}
  69
  70	memset(&replace_rec, 0, sizeof(replace_rec));
  71	replace_rec.e_cpos = cpu_to_le32(cpos);
  72	replace_rec.e_leaf_clusters = cpu_to_le16(len);
  73	replace_rec.e_blkno = cpu_to_le64(ocfs2_clusters_to_blocks(inode->i_sb,
  74								   new_p_cpos));
  75
  76	path = ocfs2_new_path_from_et(&context->et);
  77	if (!path) {
  78		ret = -ENOMEM;
  79		mlog_errno(ret);
  80		goto out;
  81	}
  82
  83	ret = ocfs2_find_path(INODE_CACHE(inode), path, cpos);
  84	if (ret) {
  85		mlog_errno(ret);
  86		goto out;
  87	}
  88
  89	el = path_leaf_el(path);
  90
  91	index = ocfs2_search_extent_list(el, cpos);
  92	if (index == -1) {
  93		ret = ocfs2_error(inode->i_sb,
  94				  "Inode %llu has an extent at cpos %u which can no longer be found\n",
  95				  (unsigned long long)ino, cpos);
  96		goto out;
  97	}
  98
  99	rec = &el->l_recs[index];
 100
 101	if (ext_flags != rec->e_flags) {
 102		ret = ocfs2_error(inode->i_sb,
 103				  "Inode %llu has corrupted extent %d with flags 0x%x at cpos %u\n",
 104				  (unsigned long long)ino, index, rec->e_flags, cpos);
 105		goto out;
 106	}
 107
 108	/*
 109	 * after moving/defraging to new location, the extent is not going
 110	 * to be refcounted anymore.
 111	 */
 112	replace_rec.e_flags = ext_flags & ~OCFS2_EXT_REFCOUNTED;
 113
 114	ret = ocfs2_split_extent(handle, &context->et, path, index,
 115				 &replace_rec, context->meta_ac,
 116				 &context->dealloc);
 117	if (ret) {
 118		mlog_errno(ret);
 119		goto out;
 120	}
 121
 122	context->new_phys_cpos = new_p_cpos;
 123
 124	/*
 125	 * need I to append truncate log for old clusters?
 126	 */
 127	if (old_blkno) {
 128		if (ext_flags & OCFS2_EXT_REFCOUNTED)
 129			ret = ocfs2_decrease_refcount(inode, handle,
 130					ocfs2_blocks_to_clusters(osb->sb,
 131								 old_blkno),
 132					len, context->meta_ac,
 133					&context->dealloc, 1);
 134		else
 135			ret = ocfs2_truncate_log_append(osb, handle,
 136							old_blkno, len);
 137	}
 138
 139	ocfs2_update_inode_fsync_trans(handle, inode, 0);
 140out:
 141	ocfs2_free_path(path);
 142	return ret;
 143}
 144
 145/*
 146 * lock allocator, and reserve appropriate number of bits for
 147 * meta blocks.
 148 */
 149static int ocfs2_lock_meta_allocator_move_extents(struct inode *inode,
 150					struct ocfs2_extent_tree *et,
 151					u32 clusters_to_move,
 152					u32 extents_to_split,
 153					struct ocfs2_alloc_context **meta_ac,
 154					int extra_blocks,
 155					int *credits)
 156{
 157	int ret, num_free_extents;
 158	unsigned int max_recs_needed = 2 * extents_to_split + clusters_to_move;
 159	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 160
 161	num_free_extents = ocfs2_num_free_extents(et);
 162	if (num_free_extents < 0) {
 163		ret = num_free_extents;
 164		mlog_errno(ret);
 165		goto out;
 166	}
 167
 168	if (!num_free_extents ||
 169	    (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed))
 170		extra_blocks += ocfs2_extend_meta_needed(et->et_root_el);
 171
 172	ret = ocfs2_reserve_new_metadata_blocks(osb, extra_blocks, meta_ac);
 173	if (ret) {
 174		mlog_errno(ret);
 175		goto out;
 176	}
 177
 178
 179	*credits += ocfs2_calc_extend_credits(osb->sb, et->et_root_el);
 180
 181	mlog(0, "reserve metadata_blocks: %d, data_clusters: %u, credits: %d\n",
 182	     extra_blocks, clusters_to_move, *credits);
 183out:
 184	if (ret) {
 185		if (*meta_ac) {
 186			ocfs2_free_alloc_context(*meta_ac);
 187			*meta_ac = NULL;
 188		}
 189	}
 190
 191	return ret;
 192}
 193
 194/*
 195 * Using one journal handle to guarantee the data consistency in case
 196 * crash happens anywhere.
 197 *
 198 *  XXX: defrag can end up with finishing partial extent as requested,
 199 * due to not enough contiguous clusters can be found in allocator.
 200 */
 201static int ocfs2_defrag_extent(struct ocfs2_move_extents_context *context,
 202			       u32 cpos, u32 phys_cpos, u32 *len, int ext_flags)
 203{
 204	int ret, credits = 0, extra_blocks = 0, partial = context->partial;
 205	handle_t *handle;
 206	struct inode *inode = context->inode;
 207	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 208	struct inode *tl_inode = osb->osb_tl_inode;
 209	struct ocfs2_refcount_tree *ref_tree = NULL;
 210	u32 new_phys_cpos, new_len;
 211	u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 212	int need_free = 0;
 213
 214	if ((ext_flags & OCFS2_EXT_REFCOUNTED) && *len) {
 215		BUG_ON(!ocfs2_is_refcount_inode(inode));
 216		BUG_ON(!context->refcount_loc);
 217
 218		ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 219					       &ref_tree, NULL);
 220		if (ret) {
 221			mlog_errno(ret);
 222			return ret;
 223		}
 224
 225		ret = ocfs2_prepare_refcount_change_for_del(inode,
 226							context->refcount_loc,
 227							phys_blkno,
 228							*len,
 229							&credits,
 230							&extra_blocks);
 231		if (ret) {
 232			mlog_errno(ret);
 233			goto out;
 234		}
 235	}
 236
 237	ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 238						*len, 1,
 239						&context->meta_ac,
 240						extra_blocks, &credits);
 241	if (ret) {
 242		mlog_errno(ret);
 243		goto out;
 244	}
 245
 246	/*
 247	 * should be using allocation reservation strategy there?
 248	 *
 249	 * if (context->data_ac)
 250	 *	context->data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
 251	 */
 252
 253	inode_lock(tl_inode);
 254
 255	if (ocfs2_truncate_log_needs_flush(osb)) {
 256		ret = __ocfs2_flush_truncate_log(osb);
 257		if (ret < 0) {
 258			mlog_errno(ret);
 259			goto out_unlock_mutex;
 260		}
 261	}
 262
 263	/*
 264	 * Make sure ocfs2_reserve_cluster is called after
 265	 * __ocfs2_flush_truncate_log, otherwise, dead lock may happen.
 266	 *
 267	 * If ocfs2_reserve_cluster is called
 268	 * before __ocfs2_flush_truncate_log, dead lock on global bitmap
 269	 * may happen.
 270	 *
 271	 */
 272	ret = ocfs2_reserve_clusters(osb, *len, &context->data_ac);
 273	if (ret) {
 274		mlog_errno(ret);
 275		goto out_unlock_mutex;
 276	}
 277
 278	handle = ocfs2_start_trans(osb, credits);
 279	if (IS_ERR(handle)) {
 280		ret = PTR_ERR(handle);
 281		mlog_errno(ret);
 282		goto out_unlock_mutex;
 283	}
 284
 285	ret = __ocfs2_claim_clusters(handle, context->data_ac, 1, *len,
 286				     &new_phys_cpos, &new_len);
 287	if (ret) {
 288		mlog_errno(ret);
 289		goto out_commit;
 290	}
 291
 292	/*
 293	 * allowing partial extent moving is kind of 'pros and cons', it makes
 294	 * whole defragmentation less likely to fail, on the contrary, the bad
 295	 * thing is it may make the fs even more fragmented after moving, let
 296	 * userspace make a good decision here.
 297	 */
 298	if (new_len != *len) {
 299		mlog(0, "len_claimed: %u, len: %u\n", new_len, *len);
 300		if (!partial) {
 301			context->range->me_flags &= ~OCFS2_MOVE_EXT_FL_COMPLETE;
 302			ret = -ENOSPC;
 303			need_free = 1;
 304			goto out_commit;
 305		}
 306	}
 307
 308	mlog(0, "cpos: %u, phys_cpos: %u, new_phys_cpos: %u\n", cpos,
 309	     phys_cpos, new_phys_cpos);
 310
 311	ret = __ocfs2_move_extent(handle, context, cpos, new_len, phys_cpos,
 312				  new_phys_cpos, ext_flags);
 313	if (ret)
 314		mlog_errno(ret);
 315
 316	if (partial && (new_len != *len))
 317		*len = new_len;
 318
 319	/*
 320	 * Here we should write the new page out first if we are
 321	 * in write-back mode.
 322	 */
 323	ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, *len);
 324	if (ret)
 325		mlog_errno(ret);
 326
 327out_commit:
 328	if (need_free && context->data_ac) {
 329		struct ocfs2_alloc_context *data_ac = context->data_ac;
 330
 331		if (context->data_ac->ac_which == OCFS2_AC_USE_LOCAL)
 332			ocfs2_free_local_alloc_bits(osb, handle, data_ac,
 333					new_phys_cpos, new_len);
 334		else
 335			ocfs2_free_clusters(handle,
 336					data_ac->ac_inode,
 337					data_ac->ac_bh,
 338					ocfs2_clusters_to_blocks(osb->sb, new_phys_cpos),
 339					new_len);
 340	}
 341
 342	ocfs2_commit_trans(osb, handle);
 343
 344out_unlock_mutex:
 345	inode_unlock(tl_inode);
 346
 347	if (context->data_ac) {
 348		ocfs2_free_alloc_context(context->data_ac);
 349		context->data_ac = NULL;
 350	}
 351
 352	if (context->meta_ac) {
 353		ocfs2_free_alloc_context(context->meta_ac);
 354		context->meta_ac = NULL;
 355	}
 356
 357out:
 358	if (ref_tree)
 359		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 360
 361	return ret;
 362}
 363
 364/*
 365 * find the victim alloc group, where #blkno fits.
 366 */
 367static int ocfs2_find_victim_alloc_group(struct inode *inode,
 368					 u64 vict_blkno,
 369					 int type, int slot,
 370					 int *vict_bit,
 371					 struct buffer_head **ret_bh)
 372{
 373	int ret, i, len, bits_per_unit = 0;
 374	u64 blkno;
 375	char namebuf[40];
 376
 377	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 378	struct buffer_head *ac_bh = NULL, *gd_bh = NULL;
 379	struct ocfs2_chain_list *cl;
 380	struct ocfs2_chain_rec *rec;
 381	struct ocfs2_dinode *ac_dinode;
 382	struct ocfs2_group_desc *bg;
 383
 384	len = ocfs2_sprintf_system_inode_name(namebuf, sizeof(namebuf), type, slot);
 385	ret = ocfs2_lookup_ino_from_name(osb->sys_root_inode, namebuf, len, &blkno);
 386
 387	if (ret) {
 388		ret = -ENOENT;
 389		goto out;
 390	}
 391
 392	ret = ocfs2_read_blocks_sync(osb, blkno, 1, &ac_bh);
 393	if (ret) {
 394		mlog_errno(ret);
 395		goto out;
 396	}
 397
 398	ac_dinode = (struct ocfs2_dinode *)ac_bh->b_data;
 399	cl = &(ac_dinode->id2.i_chain);
 400	rec = &(cl->cl_recs[0]);
 401
 402	if (type == GLOBAL_BITMAP_SYSTEM_INODE)
 403		bits_per_unit = osb->s_clustersize_bits -
 404					inode->i_sb->s_blocksize_bits;
 405	/*
 406	 * 'vict_blkno' was out of the valid range.
 407	 */
 408	if ((vict_blkno < le64_to_cpu(rec->c_blkno)) ||
 409	    (vict_blkno >= ((u64)le32_to_cpu(ac_dinode->id1.bitmap1.i_total) <<
 410				bits_per_unit))) {
 411		ret = -EINVAL;
 412		goto out;
 413	}
 414
 415	for (i = 0; i < le16_to_cpu(cl->cl_next_free_rec); i++) {
 416
 417		rec = &(cl->cl_recs[i]);
 418		if (!rec)
 419			continue;
 420
 421		bg = NULL;
 422
 423		do {
 424			if (!bg)
 425				blkno = le64_to_cpu(rec->c_blkno);
 426			else
 427				blkno = le64_to_cpu(bg->bg_next_group);
 428
 429			if (gd_bh) {
 430				brelse(gd_bh);
 431				gd_bh = NULL;
 432			}
 433
 434			ret = ocfs2_read_blocks_sync(osb, blkno, 1, &gd_bh);
 435			if (ret) {
 436				mlog_errno(ret);
 437				goto out;
 438			}
 439
 440			bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 441
 442			if (vict_blkno < (le64_to_cpu(bg->bg_blkno) +
 443						(le16_to_cpu(bg->bg_bits) << bits_per_unit))) {
 444
 445				*ret_bh = gd_bh;
 446				*vict_bit = (vict_blkno - blkno) >>
 447							bits_per_unit;
 448				mlog(0, "find the victim group: #%llu, "
 449				     "total_bits: %u, vict_bit: %u\n",
 450				     blkno, le16_to_cpu(bg->bg_bits),
 451				     *vict_bit);
 452				goto out;
 453			}
 454
 455		} while (le64_to_cpu(bg->bg_next_group));
 456	}
 457
 458	ret = -EINVAL;
 459out:
 460	brelse(ac_bh);
 461
 462	/*
 463	 * caller has to release the gd_bh properly.
 464	 */
 465	return ret;
 466}
 467
 468/*
 469 * XXX: helper to validate and adjust moving goal.
 470 */
 471static int ocfs2_validate_and_adjust_move_goal(struct inode *inode,
 472					       struct ocfs2_move_extents *range)
 473{
 474	int ret, goal_bit = 0;
 475
 476	struct buffer_head *gd_bh = NULL;
 477	struct ocfs2_group_desc *bg;
 478	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 479	int c_to_b = 1 << (osb->s_clustersize_bits -
 480					inode->i_sb->s_blocksize_bits);
 481
 482	/*
 483	 * make goal become cluster aligned.
 484	 */
 485	range->me_goal = ocfs2_block_to_cluster_start(inode->i_sb,
 486						      range->me_goal);
 487	/*
 488	 * validate goal sits within global_bitmap, and return the victim
 489	 * group desc
 490	 */
 491	ret = ocfs2_find_victim_alloc_group(inode, range->me_goal,
 492					    GLOBAL_BITMAP_SYSTEM_INODE,
 493					    OCFS2_INVALID_SLOT,
 494					    &goal_bit, &gd_bh);
 495	if (ret)
 496		goto out;
 497
 498	bg = (struct ocfs2_group_desc *)gd_bh->b_data;
 499
 500	/*
 501	 * moving goal is not allowed to start with a group desc blok(#0 blk)
 502	 * let's compromise to the latter cluster.
 503	 */
 504	if (range->me_goal == le64_to_cpu(bg->bg_blkno))
 505		range->me_goal += c_to_b;
 506
 507	/*
 508	 * movement is not gonna cross two groups.
 509	 */
 510	if ((le16_to_cpu(bg->bg_bits) - goal_bit) * osb->s_clustersize <
 511								range->me_len) {
 512		ret = -EINVAL;
 513		goto out;
 514	}
 515	/*
 516	 * more exact validations/adjustments will be performed later during
 517	 * moving operation for each extent range.
 518	 */
 519	mlog(0, "extents get ready to be moved to #%llu block\n",
 520	     range->me_goal);
 521
 522out:
 523	brelse(gd_bh);
 524
 525	return ret;
 526}
 527
 528static void ocfs2_probe_alloc_group(struct inode *inode, struct buffer_head *bh,
 529				    int *goal_bit, u32 move_len, u32 max_hop,
 530				    u32 *phys_cpos)
 531{
 532	int i, used, last_free_bits = 0, base_bit = *goal_bit;
 533	struct ocfs2_group_desc *gd = (struct ocfs2_group_desc *)bh->b_data;
 534	u32 base_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 535						 le64_to_cpu(gd->bg_blkno));
 536
 537	for (i = base_bit; i < le16_to_cpu(gd->bg_bits); i++) {
 538
 539		used = ocfs2_test_bit(i, (unsigned long *)gd->bg_bitmap);
 540		if (used) {
 541			/*
 542			 * we even tried searching the free chunk by jumping
 543			 * a 'max_hop' distance, but still failed.
 544			 */
 545			if ((i - base_bit) > max_hop) {
 546				*phys_cpos = 0;
 547				break;
 548			}
 549
 550			if (last_free_bits)
 551				last_free_bits = 0;
 552
 553			continue;
 554		} else
 555			last_free_bits++;
 556
 557		if (last_free_bits == move_len) {
 558			i -= move_len;
 559			*goal_bit = i;
 560			*phys_cpos = base_cpos + i;
 561			break;
 562		}
 563	}
 564
 565	mlog(0, "found phys_cpos: %u to fit the wanted moving.\n", *phys_cpos);
 566}
 567
 568static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 569			     u32 cpos, u32 phys_cpos, u32 *new_phys_cpos,
 570			     u32 len, int ext_flags)
 571{
 572	int ret, credits = 0, extra_blocks = 0, goal_bit = 0;
 573	handle_t *handle;
 574	struct inode *inode = context->inode;
 575	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 576	struct inode *tl_inode = osb->osb_tl_inode;
 577	struct inode *gb_inode = NULL;
 578	struct buffer_head *gb_bh = NULL;
 579	struct buffer_head *gd_bh = NULL;
 580	struct ocfs2_group_desc *gd;
 581	struct ocfs2_refcount_tree *ref_tree = NULL;
 582	u32 move_max_hop = ocfs2_blocks_to_clusters(inode->i_sb,
 583						    context->range->me_threshold);
 584	u64 phys_blkno, new_phys_blkno;
 585
 586	phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
 587
 588	if ((ext_flags & OCFS2_EXT_REFCOUNTED) && len) {
 589		BUG_ON(!ocfs2_is_refcount_inode(inode));
 590		BUG_ON(!context->refcount_loc);
 591
 592		ret = ocfs2_lock_refcount_tree(osb, context->refcount_loc, 1,
 593					       &ref_tree, NULL);
 594		if (ret) {
 595			mlog_errno(ret);
 596			return ret;
 597		}
 598
 599		ret = ocfs2_prepare_refcount_change_for_del(inode,
 600							context->refcount_loc,
 601							phys_blkno,
 602							len,
 603							&credits,
 604							&extra_blocks);
 605		if (ret) {
 606			mlog_errno(ret);
 607			goto out;
 608		}
 609	}
 610
 611	ret = ocfs2_lock_meta_allocator_move_extents(inode, &context->et,
 612						len, 1,
 613						&context->meta_ac,
 614						extra_blocks, &credits);
 615	if (ret) {
 616		mlog_errno(ret);
 617		goto out;
 618	}
 619
 620	/*
 621	 * need to count 2 extra credits for global_bitmap inode and
 622	 * group descriptor.
 623	 */
 624	credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 625
 626	inode_lock(tl_inode);
 627
 628	/*
 629	 * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 630	 * logic, while we still need to lock the global_bitmap.
 631	 */
 632	gb_inode = ocfs2_get_system_file_inode(osb, GLOBAL_BITMAP_SYSTEM_INODE,
 633					       OCFS2_INVALID_SLOT);
 634	if (!gb_inode) {
 635		mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 636		ret = -EIO;
 637		goto out_unlock_tl_inode;
 638	}
 639
 640	inode_lock(gb_inode);
 641
 642	ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 643	if (ret) {
 644		mlog_errno(ret);
 645		goto out_unlock_gb_inode;
 646	}
 647
 648	handle = ocfs2_start_trans(osb, credits);
 649	if (IS_ERR(handle)) {
 650		ret = PTR_ERR(handle);
 651		mlog_errno(ret);
 652		goto out_unlock;
 653	}
 654
 655	new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
 656	ret = ocfs2_find_victim_alloc_group(inode, new_phys_blkno,
 657					    GLOBAL_BITMAP_SYSTEM_INODE,
 658					    OCFS2_INVALID_SLOT,
 659					    &goal_bit, &gd_bh);
 660	if (ret) {
 661		mlog_errno(ret);
 662		goto out_commit;
 663	}
 664
 665	/*
 666	 * probe the victim cluster group to find a proper
 667	 * region to fit wanted movement, it even will perform
 668	 * a best-effort attempt by compromising to a threshold
 669	 * around the goal.
 670	 */
 671	ocfs2_probe_alloc_group(inode, gd_bh, &goal_bit, len, move_max_hop,
 672				new_phys_cpos);
 673	if (!*new_phys_cpos) {
 674		ret = -ENOSPC;
 675		goto out_commit;
 676	}
 677
 678	ret = __ocfs2_move_extent(handle, context, cpos, len, phys_cpos,
 679				  *new_phys_cpos, ext_flags);
 680	if (ret) {
 681		mlog_errno(ret);
 682		goto out_commit;
 683	}
 684
 685	gd = (struct ocfs2_group_desc *)gd_bh->b_data;
 686	ret = ocfs2_alloc_dinode_update_counts(gb_inode, handle, gb_bh, len,
 687					       le16_to_cpu(gd->bg_chain));
 688	if (ret) {
 689		mlog_errno(ret);
 690		goto out_commit;
 691	}
 692
 693	ret = ocfs2_block_group_set_bits(handle, gb_inode, gd, gd_bh,
 694					 goal_bit, len, 0, 0);
 695	if (ret) {
 696		ocfs2_rollback_alloc_dinode_counts(gb_inode, gb_bh, len,
 697					       le16_to_cpu(gd->bg_chain));
 698		mlog_errno(ret);
 699	}
 700
 701	/*
 702	 * Here we should write the new page out first if we are
 703	 * in write-back mode.
 704	 */
 705	ret = ocfs2_cow_sync_writeback(inode->i_sb, context->inode, cpos, len);
 706	if (ret)
 707		mlog_errno(ret);
 708
 709out_commit:
 710	ocfs2_commit_trans(osb, handle);
 711	brelse(gd_bh);
 712out_unlock:
 713	ocfs2_inode_unlock(gb_inode, 1);
 714out_unlock_gb_inode:
 715	inode_unlock(gb_inode);
 716	brelse(gb_bh);
 717	iput(gb_inode);
 718out_unlock_tl_inode:
 719	inode_unlock(tl_inode);
 720
 721out:
 722	if (context->meta_ac) {
 723		ocfs2_free_alloc_context(context->meta_ac);
 724		context->meta_ac = NULL;
 725	}
 726
 727	if (ref_tree)
 728		ocfs2_unlock_refcount_tree(osb, ref_tree, 1);
 729
 730	return ret;
 731}
 732
 733/*
 734 * Helper to calculate the defraging length in one run according to threshold.
 735 */
 736static void ocfs2_calc_extent_defrag_len(u32 *alloc_size, u32 *len_defraged,
 737					 u32 threshold, int *skip)
 738{
 739	if ((*alloc_size + *len_defraged) < threshold) {
 740		/*
 741		 * proceed defragmentation until we meet the thresh
 742		 */
 743		*len_defraged += *alloc_size;
 744	} else if (*len_defraged == 0) {
 745		/*
 746		 * XXX: skip a large extent.
 747		 */
 748		*skip = 1;
 749	} else {
 750		/*
 751		 * split this extent to coalesce with former pieces as
 752		 * to reach the threshold.
 753		 *
 754		 * we're done here with one cycle of defragmentation
 755		 * in a size of 'thresh', resetting 'len_defraged'
 756		 * forces a new defragmentation.
 757		 */
 758		*alloc_size = threshold - *len_defraged;
 759		*len_defraged = 0;
 760	}
 761}
 762
 763static int __ocfs2_move_extents_range(struct buffer_head *di_bh,
 764				struct ocfs2_move_extents_context *context)
 765{
 766	int ret = 0, flags, do_defrag, skip = 0;
 767	u32 cpos, phys_cpos, move_start, len_to_move, alloc_size;
 768	u32 len_defraged = 0, defrag_thresh = 0, new_phys_cpos = 0;
 769
 770	struct inode *inode = context->inode;
 771	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 772	struct ocfs2_move_extents *range = context->range;
 773	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 774
 775	if ((i_size_read(inode) == 0) || (range->me_len == 0))
 776		return 0;
 777
 778	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 779		return 0;
 780
 781	context->refcount_loc = le64_to_cpu(di->i_refcount_loc);
 782
 783	ocfs2_init_dinode_extent_tree(&context->et, INODE_CACHE(inode), di_bh);
 784	ocfs2_init_dealloc_ctxt(&context->dealloc);
 785
 786	/*
 787	 * TO-DO XXX:
 788	 *
 789	 * - xattr extents.
 790	 */
 791
 792	do_defrag = context->auto_defrag;
 793
 794	/*
 795	 * extents moving happens in unit of clusters, for the sake
 796	 * of simplicity, we may ignore two clusters where 'byte_start'
 797	 * and 'byte_start + len' were within.
 798	 */
 799	move_start = ocfs2_clusters_for_bytes(osb->sb, range->me_start);
 800	len_to_move = (range->me_start + range->me_len) >>
 801						osb->s_clustersize_bits;
 802	if (len_to_move >= move_start)
 803		len_to_move -= move_start;
 804	else
 805		len_to_move = 0;
 806
 807	if (do_defrag) {
 808		defrag_thresh = range->me_threshold >> osb->s_clustersize_bits;
 809		if (defrag_thresh <= 1)
 810			goto done;
 811	} else
 812		new_phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
 813							 range->me_goal);
 814
 815	mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u, "
 816	     "thresh: %u\n",
 817	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
 818	     (unsigned long long)range->me_start,
 819	     (unsigned long long)range->me_len,
 820	     move_start, len_to_move, defrag_thresh);
 821
 822	cpos = move_start;
 823	while (len_to_move) {
 824		ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &alloc_size,
 825					 &flags);
 826		if (ret) {
 827			mlog_errno(ret);
 828			goto out;
 829		}
 830
 831		if (alloc_size > len_to_move)
 832			alloc_size = len_to_move;
 833
 834		/*
 835		 * XXX: how to deal with a hole:
 836		 *
 837		 * - skip the hole of course
 838		 * - force a new defragmentation
 839		 */
 840		if (!phys_cpos) {
 841			if (do_defrag)
 842				len_defraged = 0;
 843
 844			goto next;
 845		}
 846
 847		if (do_defrag) {
 848			ocfs2_calc_extent_defrag_len(&alloc_size, &len_defraged,
 849						     defrag_thresh, &skip);
 850			/*
 851			 * skip large extents
 852			 */
 853			if (skip) {
 854				skip = 0;
 855				goto next;
 856			}
 857
 858			mlog(0, "#Defrag: cpos: %u, phys_cpos: %u, "
 859			     "alloc_size: %u, len_defraged: %u\n",
 860			     cpos, phys_cpos, alloc_size, len_defraged);
 861
 862			ret = ocfs2_defrag_extent(context, cpos, phys_cpos,
 863						  &alloc_size, flags);
 864		} else {
 865			ret = ocfs2_move_extent(context, cpos, phys_cpos,
 866						&new_phys_cpos, alloc_size,
 867						flags);
 868
 869			new_phys_cpos += alloc_size;
 870		}
 871
 872		if (ret < 0) {
 873			mlog_errno(ret);
 874			goto out;
 875		}
 876		/*
 877		 * Invalidate extent cache after moving/defragging to prevent
 878		 * stale cached data with outdated extent flags.
 879		 */
 880		ocfs2_extent_map_trunc(inode, cpos);
 881
 882		context->clusters_moved += alloc_size;
 883next:
 884		cpos += alloc_size;
 885		len_to_move -= alloc_size;
 886	}
 887
 888done:
 889	range->me_flags |= OCFS2_MOVE_EXT_FL_COMPLETE;
 890
 891out:
 892	range->me_moved_len = ocfs2_clusters_to_bytes(osb->sb,
 893						      context->clusters_moved);
 894	range->me_new_offset = ocfs2_clusters_to_bytes(osb->sb,
 895						       context->new_phys_cpos);
 896
 897	ocfs2_schedule_truncate_log_flush(osb, 1);
 898	ocfs2_run_deallocs(osb, &context->dealloc);
 899
 900	return ret;
 901}
 902
 903static int ocfs2_move_extents(struct ocfs2_move_extents_context *context)
 904{
 905	int status;
 906	handle_t *handle;
 907	struct inode *inode = context->inode;
 908	struct ocfs2_dinode *di;
 909	struct buffer_head *di_bh = NULL;
 910	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 911
 912	if (unlikely(ocfs2_emergency_state(osb)))
 913		return -EROFS;
 914
 915	inode_lock(inode);
 916
 917	/*
 918	 * This prevents concurrent writes from other nodes
 919	 */
 920	status = ocfs2_rw_lock(inode, 1);
 921	if (status) {
 922		mlog_errno(status);
 923		goto out;
 924	}
 925
 926	status = ocfs2_inode_lock(inode, &di_bh, 1);
 927	if (status) {
 928		mlog_errno(status);
 929		goto out_rw_unlock;
 930	}
 931
 932	/*
 933	 * remember ip_xattr_sem also needs to be held if necessary
 934	 */
 935	down_write(&OCFS2_I(inode)->ip_alloc_sem);
 936
 937	status = __ocfs2_move_extents_range(di_bh, context);
 938
 939	up_write(&OCFS2_I(inode)->ip_alloc_sem);
 940	if (status) {
 941		mlog_errno(status);
 942		goto out_inode_unlock;
 943	}
 944
 945	/*
 946	 * We update ctime for these changes
 947	 */
 948	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 949	if (IS_ERR(handle)) {
 950		status = PTR_ERR(handle);
 951		mlog_errno(status);
 952		goto out_inode_unlock;
 953	}
 954
 955	status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 956					 OCFS2_JOURNAL_ACCESS_WRITE);
 957	if (status) {
 958		mlog_errno(status);
 959		goto out_commit;
 960	}
 961
 962	di = (struct ocfs2_dinode *)di_bh->b_data;
 963	inode_set_ctime_current(inode);
 964	di->i_ctime = cpu_to_le64(inode_get_ctime_sec(inode));
 965	di->i_ctime_nsec = cpu_to_le32(inode_get_ctime_nsec(inode));
 966	ocfs2_update_inode_fsync_trans(handle, inode, 0);
 967
 968	ocfs2_journal_dirty(handle, di_bh);
 969
 970out_commit:
 971	ocfs2_commit_trans(osb, handle);
 972
 973out_inode_unlock:
 974	brelse(di_bh);
 975	ocfs2_inode_unlock(inode, 1);
 976out_rw_unlock:
 977	ocfs2_rw_unlock(inode, 1);
 978out:
 979	inode_unlock(inode);
 980
 981	return status;
 982}
 983
 984int ocfs2_ioctl_move_extents(struct file *filp, void __user *argp)
 985{
 986	int status;
 987
 988	struct inode *inode = file_inode(filp);
 989	struct ocfs2_move_extents range;
 990	struct ocfs2_move_extents_context *context;
 991
 992	if (!argp)
 993		return -EINVAL;
 994
 995	status = mnt_want_write_file(filp);
 996	if (status)
 997		return status;
 998
 999	if ((!S_ISREG(inode->i_mode)) || !(filp->f_mode & FMODE_WRITE)) {
1000		status = -EPERM;
1001		goto out_drop;
1002	}
1003
1004	if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1005		status = -EPERM;
1006		goto out_drop;
1007	}
1008
1009	context = kzalloc(sizeof(struct ocfs2_move_extents_context), GFP_NOFS);
1010	if (!context) {
1011		status = -ENOMEM;
1012		mlog_errno(status);
1013		goto out_drop;
1014	}
1015
1016	context->inode = inode;
1017	context->file = filp;
1018
1019	if (copy_from_user(&range, argp, sizeof(range))) {
1020		status = -EFAULT;
1021		goto out_free;
1022	}
1023
1024	if (range.me_start > i_size_read(inode)) {
1025		status = -EINVAL;
1026		goto out_free;
1027	}
1028
1029	if (range.me_start + range.me_len > i_size_read(inode))
1030			range.me_len = i_size_read(inode) - range.me_start;
1031
1032	context->range = &range;
1033
1034	/*
1035	 * ok, the default threshold for the defragmentation
1036	 * is 1M, since our maximum clustersize was 1M also.
1037	 * any thought?
1038	 */
1039	if (!range.me_threshold)
1040		range.me_threshold = 1024 * 1024;
1041
1042	if (range.me_threshold > i_size_read(inode))
1043		range.me_threshold = i_size_read(inode);
1044
1045	if (range.me_flags & ~(OCFS2_MOVE_EXT_FL_AUTO_DEFRAG |
1046			       OCFS2_MOVE_EXT_FL_PART_DEFRAG)) {
1047		status = -EINVAL;
1048		goto out_free;
1049	}
1050
1051	if (range.me_flags & OCFS2_MOVE_EXT_FL_AUTO_DEFRAG) {
1052		context->auto_defrag = 1;
1053
1054		if (range.me_flags & OCFS2_MOVE_EXT_FL_PART_DEFRAG)
1055			context->partial = 1;
1056	} else {
1057		/*
1058		 * first best-effort attempt to validate and adjust the goal
1059		 * (physical address in block), while it can't guarantee later
1060		 * operation can succeed all the time since global_bitmap may
1061		 * change a bit over time.
1062		 */
1063
1064		status = ocfs2_validate_and_adjust_move_goal(inode, &range);
1065		if (status)
1066			goto out_copy;
1067	}
1068
1069	status = ocfs2_move_extents(context);
1070	if (status)
1071		mlog_errno(status);
1072out_copy:
1073	/*
1074	 * movement/defragmentation may end up being partially completed,
1075	 * that's the reason why we need to return userspace the finished
1076	 * length and new_offset even if failure happens somewhere.
1077	 */
1078	if (copy_to_user(argp, &range, sizeof(range)))
1079		status = -EFAULT;
1080
1081out_free:
1082	kfree(context);
1083out_drop:
1084	mnt_drop_write_file(filp);
1085
1086	return status;
1087}