1/*
   2 * Copyright (c) 2010 Red Hat, Inc. 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 License as
   6 * published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it would be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11 * GNU General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public License
  14 * along with this program; if not, write the Free Software Foundation,
  15 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  16 */
  17
  18#include "xfs.h"
  19#include "xfs_fs.h"
  20#include "xfs_format.h"
  21#include "xfs_log_format.h"
  22#include "xfs_shared.h"
  23#include "xfs_trans_resv.h"
  24#include "xfs_mount.h"
  25#include "xfs_error.h"
  26#include "xfs_alloc.h"
  27#include "xfs_extent_busy.h"
  28#include "xfs_discard.h"
  29#include "xfs_trans.h"
  30#include "xfs_trans_priv.h"
  31#include "xfs_log.h"
  32#include "xfs_log_priv.h"
  33
  34/*
  35 * Allocate a new ticket. Failing to get a new ticket makes it really hard to
  36 * recover, so we don't allow failure here. Also, we allocate in a context that
  37 * we don't want to be issuing transactions from, so we need to tell the
  38 * allocation code this as well.
  39 *
  40 * We don't reserve any space for the ticket - we are going to steal whatever
  41 * space we require from transactions as they commit. To ensure we reserve all
  42 * the space required, we need to set the current reservation of the ticket to
  43 * zero so that we know to steal the initial transaction overhead from the
  44 * first transaction commit.
  45 */
  46static struct xlog_ticket *
  47xlog_cil_ticket_alloc(
  48	struct xlog	*log)
  49{
  50	struct xlog_ticket *tic;
  51
  52	tic = xlog_ticket_alloc(log, 0, 1, XFS_TRANSACTION, 0,
  53				KM_SLEEP|KM_NOFS);
 
  54
  55	/*
  56	 * set the current reservation to zero so we know to steal the basic
  57	 * transaction overhead reservation from the first transaction commit.
  58	 */
  59	tic->t_curr_res = 0;
  60	return tic;
  61}
  62
  63/*
  64 * After the first stage of log recovery is done, we know where the head and
  65 * tail of the log are. We need this log initialisation done before we can
  66 * initialise the first CIL checkpoint context.
  67 *
  68 * Here we allocate a log ticket to track space usage during a CIL push.  This
  69 * ticket is passed to xlog_write() directly so that we don't slowly leak log
  70 * space by failing to account for space used by log headers and additional
  71 * region headers for split regions.
  72 */
  73void
  74xlog_cil_init_post_recovery(
  75	struct xlog	*log)
  76{
  77	log->l_cilp->xc_ctx->ticket = xlog_cil_ticket_alloc(log);
  78	log->l_cilp->xc_ctx->sequence = 1;
  79}
  80
  81static inline int
  82xlog_cil_iovec_space(
  83	uint	niovecs)
  84{
  85	return round_up((sizeof(struct xfs_log_vec) +
  86					niovecs * sizeof(struct xfs_log_iovec)),
  87			sizeof(uint64_t));
  88}
  89
  90/*
  91 * Allocate or pin log vector buffers for CIL insertion.
  92 *
  93 * The CIL currently uses disposable buffers for copying a snapshot of the
  94 * modified items into the log during a push. The biggest problem with this is
  95 * the requirement to allocate the disposable buffer during the commit if:
  96 *	a) does not exist; or
  97 *	b) it is too small
  98 *
  99 * If we do this allocation within xlog_cil_insert_format_items(), it is done
 100 * under the xc_ctx_lock, which means that a CIL push cannot occur during
 101 * the memory allocation. This means that we have a potential deadlock situation
 102 * under low memory conditions when we have lots of dirty metadata pinned in
 103 * the CIL and we need a CIL commit to occur to free memory.
 104 *
 105 * To avoid this, we need to move the memory allocation outside the
 106 * xc_ctx_lock, but because the log vector buffers are disposable, that opens
 107 * up a TOCTOU race condition w.r.t. the CIL committing and removing the log
 108 * vector buffers between the check and the formatting of the item into the
 109 * log vector buffer within the xc_ctx_lock.
 110 *
 111 * Because the log vector buffer needs to be unchanged during the CIL push
 112 * process, we cannot share the buffer between the transaction commit (which
 113 * modifies the buffer) and the CIL push context that is writing the changes
 114 * into the log. This means skipping preallocation of buffer space is
 115 * unreliable, but we most definitely do not want to be allocating and freeing
 116 * buffers unnecessarily during commits when overwrites can be done safely.
 117 *
 118 * The simplest solution to this problem is to allocate a shadow buffer when a
 119 * log item is committed for the second time, and then to only use this buffer
 120 * if necessary. The buffer can remain attached to the log item until such time
 121 * it is needed, and this is the buffer that is reallocated to match the size of
 122 * the incoming modification. Then during the formatting of the item we can swap
 123 * the active buffer with the new one if we can't reuse the existing buffer. We
 124 * don't free the old buffer as it may be reused on the next modification if
 125 * it's size is right, otherwise we'll free and reallocate it at that point.
 126 *
 127 * This function builds a vector for the changes in each log item in the
 128 * transaction. It then works out the length of the buffer needed for each log
 129 * item, allocates them and attaches the vector to the log item in preparation
 130 * for the formatting step which occurs under the xc_ctx_lock.
 131 *
 132 * While this means the memory footprint goes up, it avoids the repeated
 133 * alloc/free pattern that repeated modifications of an item would otherwise
 134 * cause, and hence minimises the CPU overhead of such behaviour.
 135 */
 136static void
 137xlog_cil_alloc_shadow_bufs(
 138	struct xlog		*log,
 139	struct xfs_trans	*tp)
 140{
 141	struct xfs_log_item_desc *lidp;
 142
 143	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
 144		struct xfs_log_item *lip = lidp->lid_item;
 145		struct xfs_log_vec *lv;
 146		int	niovecs = 0;
 147		int	nbytes = 0;
 148		int	buf_size;
 149		bool	ordered = false;
 150
 151		/* Skip items which aren't dirty in this transaction. */
 152		if (!(lidp->lid_flags & XFS_LID_DIRTY))
 153			continue;
 154
 155		/* get number of vecs and size of data to be stored */
 156		lip->li_ops->iop_size(lip, &niovecs, &nbytes);
 157
 158		/*
 159		 * Ordered items need to be tracked but we do not wish to write
 160		 * them. We need a logvec to track the object, but we do not
 161		 * need an iovec or buffer to be allocated for copying data.
 162		 */
 163		if (niovecs == XFS_LOG_VEC_ORDERED) {
 164			ordered = true;
 165			niovecs = 0;
 166			nbytes = 0;
 167		}
 168
 169		/*
 170		 * We 64-bit align the length of each iovec so that the start
 171		 * of the next one is naturally aligned.  We'll need to
 172		 * account for that slack space here. Then round nbytes up
 173		 * to 64-bit alignment so that the initial buffer alignment is
 174		 * easy to calculate and verify.
 175		 */
 176		nbytes += niovecs * sizeof(uint64_t);
 177		nbytes = round_up(nbytes, sizeof(uint64_t));
 178
 179		/*
 180		 * The data buffer needs to start 64-bit aligned, so round up
 181		 * that space to ensure we can align it appropriately and not
 182		 * overrun the buffer.
 183		 */
 184		buf_size = nbytes + xlog_cil_iovec_space(niovecs);
 185
 186		/*
 187		 * if we have no shadow buffer, or it is too small, we need to
 188		 * reallocate it.
 189		 */
 190		if (!lip->li_lv_shadow ||
 191		    buf_size > lip->li_lv_shadow->lv_size) {
 192
 193			/*
 194			 * We free and allocate here as a realloc would copy
 195			 * unecessary data. We don't use kmem_zalloc() for the
 196			 * same reason - we don't need to zero the data area in
 197			 * the buffer, only the log vector header and the iovec
 198			 * storage.
 199			 */
 200			kmem_free(lip->li_lv_shadow);
 201
 202			lv = kmem_alloc(buf_size, KM_SLEEP|KM_NOFS);
 203			memset(lv, 0, xlog_cil_iovec_space(niovecs));
 204
 205			lv->lv_item = lip;
 206			lv->lv_size = buf_size;
 207			if (ordered)
 208				lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
 209			else
 210				lv->lv_iovecp = (struct xfs_log_iovec *)&lv[1];
 211			lip->li_lv_shadow = lv;
 212		} else {
 213			/* same or smaller, optimise common overwrite case */
 214			lv = lip->li_lv_shadow;
 215			if (ordered)
 216				lv->lv_buf_len = XFS_LOG_VEC_ORDERED;
 217			else
 218				lv->lv_buf_len = 0;
 219			lv->lv_bytes = 0;
 220			lv->lv_next = NULL;
 221		}
 222
 223		/* Ensure the lv is set up according to ->iop_size */
 224		lv->lv_niovecs = niovecs;
 225
 226		/* The allocated data region lies beyond the iovec region */
 227		lv->lv_buf = (char *)lv + xlog_cil_iovec_space(niovecs);
 228	}
 229
 230}
 231
 232/*
 233 * Prepare the log item for insertion into the CIL. Calculate the difference in
 234 * log space and vectors it will consume, and if it is a new item pin it as
 235 * well.
 236 */
 237STATIC void
 238xfs_cil_prepare_item(
 239	struct xlog		*log,
 240	struct xfs_log_vec	*lv,
 241	struct xfs_log_vec	*old_lv,
 242	int			*diff_len,
 243	int			*diff_iovecs)
 244{
 245	/* Account for the new LV being passed in */
 246	if (lv->lv_buf_len != XFS_LOG_VEC_ORDERED) {
 247		*diff_len += lv->lv_bytes;
 248		*diff_iovecs += lv->lv_niovecs;
 249	}
 250
 251	/*
 252	 * If there is no old LV, this is the first time we've seen the item in
 253	 * this CIL context and so we need to pin it. If we are replacing the
 254	 * old_lv, then remove the space it accounts for and make it the shadow
 255	 * buffer for later freeing. In both cases we are now switching to the
 256	 * shadow buffer, so update the the pointer to it appropriately.
 257	 */
 258	if (!old_lv) {
 259		lv->lv_item->li_ops->iop_pin(lv->lv_item);
 260		lv->lv_item->li_lv_shadow = NULL;
 261	} else if (old_lv != lv) {
 262		ASSERT(lv->lv_buf_len != XFS_LOG_VEC_ORDERED);
 263
 264		*diff_len -= old_lv->lv_bytes;
 265		*diff_iovecs -= old_lv->lv_niovecs;
 266		lv->lv_item->li_lv_shadow = old_lv;
 267	}
 268
 269	/* attach new log vector to log item */
 270	lv->lv_item->li_lv = lv;
 271
 272	/*
 273	 * If this is the first time the item is being committed to the
 274	 * CIL, store the sequence number on the log item so we can
 275	 * tell in future commits whether this is the first checkpoint
 276	 * the item is being committed into.
 277	 */
 278	if (!lv->lv_item->li_seq)
 279		lv->lv_item->li_seq = log->l_cilp->xc_ctx->sequence;
 280}
 281
 282/*
 283 * Format log item into a flat buffers
 284 *
 285 * For delayed logging, we need to hold a formatted buffer containing all the
 286 * changes on the log item. This enables us to relog the item in memory and
 287 * write it out asynchronously without needing to relock the object that was
 288 * modified at the time it gets written into the iclog.
 289 *
 290 * This function takes the prepared log vectors attached to each log item, and
 291 * formats the changes into the log vector buffer. The buffer it uses is
 292 * dependent on the current state of the vector in the CIL - the shadow lv is
 293 * guaranteed to be large enough for the current modification, but we will only
 294 * use that if we can't reuse the existing lv. If we can't reuse the existing
 295 * lv, then simple swap it out for the shadow lv. We don't free it - that is
 296 * done lazily either by th enext modification or the freeing of the log item.
 297 *
 298 * We don't set up region headers during this process; we simply copy the
 299 * regions into the flat buffer. We can do this because we still have to do a
 300 * formatting step to write the regions into the iclog buffer.  Writing the
 301 * ophdrs during the iclog write means that we can support splitting large
 302 * regions across iclog boundares without needing a change in the format of the
 303 * item/region encapsulation.
 304 *
 305 * Hence what we need to do now is change the rewrite the vector array to point
 306 * to the copied region inside the buffer we just allocated. This allows us to
 307 * format the regions into the iclog as though they are being formatted
 308 * directly out of the objects themselves.
 309 */
 310static void
 311xlog_cil_insert_format_items(
 312	struct xlog		*log,
 313	struct xfs_trans	*tp,
 314	int			*diff_len,
 315	int			*diff_iovecs)
 316{
 317	struct xfs_log_item_desc *lidp;
 318
 319
 320	/* Bail out if we didn't find a log item.  */
 321	if (list_empty(&tp->t_items)) {
 322		ASSERT(0);
 323		return;
 324	}
 325
 326	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
 327		struct xfs_log_item *lip = lidp->lid_item;
 328		struct xfs_log_vec *lv;
 329		struct xfs_log_vec *old_lv = NULL;
 330		struct xfs_log_vec *shadow;
 
 
 331		bool	ordered = false;
 332
 333		/* Skip items which aren't dirty in this transaction. */
 334		if (!(lidp->lid_flags & XFS_LID_DIRTY))
 335			continue;
 336
 
 
 
 
 
 
 
 337		/*
 338		 * The formatting size information is already attached to
 339		 * the shadow lv on the log item.
 
 340		 */
 341		shadow = lip->li_lv_shadow;
 342		if (shadow->lv_buf_len == XFS_LOG_VEC_ORDERED)
 343			ordered = true;
 
 
 
 344
 345		/* Skip items that do not have any vectors for writing */
 346		if (!shadow->lv_niovecs && !ordered)
 347			continue;
 
 
 
 
 
 
 348
 349		/* compare to existing item size */
 350		old_lv = lip->li_lv;
 351		if (lip->li_lv && shadow->lv_size <= lip->li_lv->lv_size) {
 
 
 
 
 
 
 
 
 
 
 
 
 352			/* same or smaller, optimise common overwrite case */
 353			lv = lip->li_lv;
 354			lv->lv_next = NULL;
 355
 356			if (ordered)
 357				goto insert;
 358
 359			/*
 360			 * set the item up as though it is a new insertion so
 361			 * that the space reservation accounting is correct.
 362			 */
 363			*diff_iovecs -= lv->lv_niovecs;
 364			*diff_len -= lv->lv_bytes;
 365
 366			/* Ensure the lv is set up according to ->iop_size */
 367			lv->lv_niovecs = shadow->lv_niovecs;
 368
 369			/* reset the lv buffer information for new formatting */
 370			lv->lv_buf_len = 0;
 371			lv->lv_bytes = 0;
 372			lv->lv_buf = (char *)lv +
 373					xlog_cil_iovec_space(lv->lv_niovecs);
 374		} else {
 375			/* switch to shadow buffer! */
 376			lv = shadow;
 377			lv->lv_item = lip;
 
 378			if (ordered) {
 379				/* track as an ordered logvec */
 380				ASSERT(lip->li_lv == NULL);
 
 381				goto insert;
 382			}
 
 383		}
 384
 
 
 
 
 
 
 
 385		ASSERT(IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)));
 
 386		lip->li_ops->iop_format(lip, lv);
 387insert:
 
 388		xfs_cil_prepare_item(log, lv, old_lv, diff_len, diff_iovecs);
 389	}
 390}
 391
 392/*
 393 * Insert the log items into the CIL and calculate the difference in space
 394 * consumed by the item. Add the space to the checkpoint ticket and calculate
 395 * if the change requires additional log metadata. If it does, take that space
 396 * as well. Remove the amount of space we added to the checkpoint ticket from
 397 * the current transaction ticket so that the accounting works out correctly.
 398 */
 399static void
 400xlog_cil_insert_items(
 401	struct xlog		*log,
 402	struct xfs_trans	*tp)
 403{
 404	struct xfs_cil		*cil = log->l_cilp;
 405	struct xfs_cil_ctx	*ctx = cil->xc_ctx;
 406	struct xfs_log_item_desc *lidp;
 407	int			len = 0;
 408	int			diff_iovecs = 0;
 409	int			iclog_space;
 410
 411	ASSERT(tp);
 412
 413	/*
 414	 * We can do this safely because the context can't checkpoint until we
 415	 * are done so it doesn't matter exactly how we update the CIL.
 416	 */
 417	xlog_cil_insert_format_items(log, tp, &len, &diff_iovecs);
 418
 419	/*
 420	 * Now (re-)position everything modified at the tail of the CIL.
 421	 * We do this here so we only need to take the CIL lock once during
 422	 * the transaction commit.
 423	 */
 424	spin_lock(&cil->xc_cil_lock);
 425	list_for_each_entry(lidp, &tp->t_items, lid_trans) {
 426		struct xfs_log_item	*lip = lidp->lid_item;
 427
 428		/* Skip items which aren't dirty in this transaction. */
 429		if (!(lidp->lid_flags & XFS_LID_DIRTY))
 430			continue;
 431
 432		/*
 433		 * Only move the item if it isn't already at the tail. This is
 434		 * to prevent a transient list_empty() state when reinserting
 435		 * an item that is already the only item in the CIL.
 436		 */
 437		if (!list_is_last(&lip->li_cil, &cil->xc_cil))
 438			list_move_tail(&lip->li_cil, &cil->xc_cil);
 439	}
 440
 441	/* account for space used by new iovec headers  */
 442	len += diff_iovecs * sizeof(xlog_op_header_t);
 443	ctx->nvecs += diff_iovecs;
 444
 445	/* attach the transaction to the CIL if it has any busy extents */
 446	if (!list_empty(&tp->t_busy))
 447		list_splice_init(&tp->t_busy, &ctx->busy_extents);
 448
 449	/*
 450	 * Now transfer enough transaction reservation to the context ticket
 451	 * for the checkpoint. The context ticket is special - the unit
 452	 * reservation has to grow as well as the current reservation as we
 453	 * steal from tickets so we can correctly determine the space used
 454	 * during the transaction commit.
 455	 */
 456	if (ctx->ticket->t_curr_res == 0) {
 457		ctx->ticket->t_curr_res = ctx->ticket->t_unit_res;
 458		tp->t_ticket->t_curr_res -= ctx->ticket->t_unit_res;
 459	}
 460
 461	/* do we need space for more log record headers? */
 462	iclog_space = log->l_iclog_size - log->l_iclog_hsize;
 463	if (len > 0 && (ctx->space_used / iclog_space !=
 464				(ctx->space_used + len) / iclog_space)) {
 465		int hdrs;
 466
 467		hdrs = (len + iclog_space - 1) / iclog_space;
 468		/* need to take into account split region headers, too */
 469		hdrs *= log->l_iclog_hsize + sizeof(struct xlog_op_header);
 470		ctx->ticket->t_unit_res += hdrs;
 471		ctx->ticket->t_curr_res += hdrs;
 472		tp->t_ticket->t_curr_res -= hdrs;
 473		ASSERT(tp->t_ticket->t_curr_res >= len);
 474	}
 475	tp->t_ticket->t_curr_res -= len;
 476	ctx->space_used += len;
 477
 478	spin_unlock(&cil->xc_cil_lock);
 479}
 480
 481static void
 482xlog_cil_free_logvec(
 483	struct xfs_log_vec	*log_vector)
 484{
 485	struct xfs_log_vec	*lv;
 486
 487	for (lv = log_vector; lv; ) {
 488		struct xfs_log_vec *next = lv->lv_next;
 489		kmem_free(lv);
 490		lv = next;
 491	}
 492}
 493
 494/*
 495 * Mark all items committed and clear busy extents. We free the log vector
 496 * chains in a separate pass so that we unpin the log items as quickly as
 497 * possible.
 498 */
 499static void
 500xlog_cil_committed(
 501	void	*args,
 502	int	abort)
 503{
 504	struct xfs_cil_ctx	*ctx = args;
 505	struct xfs_mount	*mp = ctx->cil->xc_log->l_mp;
 506
 507	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
 508					ctx->start_lsn, abort);
 509
 510	xfs_extent_busy_sort(&ctx->busy_extents);
 511	xfs_extent_busy_clear(mp, &ctx->busy_extents,
 512			     (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
 513
 514	/*
 515	 * If we are aborting the commit, wake up anyone waiting on the
 516	 * committing list.  If we don't, then a shutdown we can leave processes
 517	 * waiting in xlog_cil_force_lsn() waiting on a sequence commit that
 518	 * will never happen because we aborted it.
 519	 */
 520	spin_lock(&ctx->cil->xc_push_lock);
 521	if (abort)
 522		wake_up_all(&ctx->cil->xc_commit_wait);
 523	list_del(&ctx->committing);
 524	spin_unlock(&ctx->cil->xc_push_lock);
 525
 526	xlog_cil_free_logvec(ctx->lv_chain);
 527
 528	if (!list_empty(&ctx->busy_extents)) {
 529		ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
 530
 531		xfs_discard_extents(mp, &ctx->busy_extents);
 532		xfs_extent_busy_clear(mp, &ctx->busy_extents, false);
 533	}
 534
 535	kmem_free(ctx);
 536}
 537
 538/*
 539 * Push the Committed Item List to the log. If @push_seq flag is zero, then it
 540 * is a background flush and so we can chose to ignore it. Otherwise, if the
 541 * current sequence is the same as @push_seq we need to do a flush. If
 542 * @push_seq is less than the current sequence, then it has already been
 543 * flushed and we don't need to do anything - the caller will wait for it to
 544 * complete if necessary.
 545 *
 546 * @push_seq is a value rather than a flag because that allows us to do an
 547 * unlocked check of the sequence number for a match. Hence we can allows log
 548 * forces to run racily and not issue pushes for the same sequence twice. If we
 549 * get a race between multiple pushes for the same sequence they will block on
 550 * the first one and then abort, hence avoiding needless pushes.
 551 */
 552STATIC int
 553xlog_cil_push(
 554	struct xlog		*log)
 555{
 556	struct xfs_cil		*cil = log->l_cilp;
 557	struct xfs_log_vec	*lv;
 558	struct xfs_cil_ctx	*ctx;
 559	struct xfs_cil_ctx	*new_ctx;
 560	struct xlog_in_core	*commit_iclog;
 561	struct xlog_ticket	*tic;
 562	int			num_iovecs;
 563	int			error = 0;
 564	struct xfs_trans_header thdr;
 565	struct xfs_log_iovec	lhdr;
 566	struct xfs_log_vec	lvhdr = { NULL };
 567	xfs_lsn_t		commit_lsn;
 568	xfs_lsn_t		push_seq;
 569
 570	if (!cil)
 571		return 0;
 572
 573	new_ctx = kmem_zalloc(sizeof(*new_ctx), KM_SLEEP|KM_NOFS);
 574	new_ctx->ticket = xlog_cil_ticket_alloc(log);
 575
 576	down_write(&cil->xc_ctx_lock);
 577	ctx = cil->xc_ctx;
 578
 579	spin_lock(&cil->xc_push_lock);
 580	push_seq = cil->xc_push_seq;
 581	ASSERT(push_seq <= ctx->sequence);
 582
 583	/*
 584	 * Check if we've anything to push. If there is nothing, then we don't
 585	 * move on to a new sequence number and so we have to be able to push
 586	 * this sequence again later.
 587	 */
 588	if (list_empty(&cil->xc_cil)) {
 589		cil->xc_push_seq = 0;
 590		spin_unlock(&cil->xc_push_lock);
 591		goto out_skip;
 592	}
 593
 594
 595	/* check for a previously pushed seqeunce */
 596	if (push_seq < cil->xc_ctx->sequence) {
 597		spin_unlock(&cil->xc_push_lock);
 598		goto out_skip;
 599	}
 600
 601	/*
 602	 * We are now going to push this context, so add it to the committing
 603	 * list before we do anything else. This ensures that anyone waiting on
 604	 * this push can easily detect the difference between a "push in
 605	 * progress" and "CIL is empty, nothing to do".
 606	 *
 607	 * IOWs, a wait loop can now check for:
 608	 *	the current sequence not being found on the committing list;
 609	 *	an empty CIL; and
 610	 *	an unchanged sequence number
 611	 * to detect a push that had nothing to do and therefore does not need
 612	 * waiting on. If the CIL is not empty, we get put on the committing
 613	 * list before emptying the CIL and bumping the sequence number. Hence
 614	 * an empty CIL and an unchanged sequence number means we jumped out
 615	 * above after doing nothing.
 616	 *
 617	 * Hence the waiter will either find the commit sequence on the
 618	 * committing list or the sequence number will be unchanged and the CIL
 619	 * still dirty. In that latter case, the push has not yet started, and
 620	 * so the waiter will have to continue trying to check the CIL
 621	 * committing list until it is found. In extreme cases of delay, the
 622	 * sequence may fully commit between the attempts the wait makes to wait
 623	 * on the commit sequence.
 624	 */
 625	list_add(&ctx->committing, &cil->xc_committing);
 626	spin_unlock(&cil->xc_push_lock);
 627
 628	/*
 629	 * pull all the log vectors off the items in the CIL, and
 630	 * remove the items from the CIL. We don't need the CIL lock
 631	 * here because it's only needed on the transaction commit
 632	 * side which is currently locked out by the flush lock.
 633	 */
 634	lv = NULL;
 635	num_iovecs = 0;
 636	while (!list_empty(&cil->xc_cil)) {
 637		struct xfs_log_item	*item;
 638
 639		item = list_first_entry(&cil->xc_cil,
 640					struct xfs_log_item, li_cil);
 641		list_del_init(&item->li_cil);
 642		if (!ctx->lv_chain)
 643			ctx->lv_chain = item->li_lv;
 644		else
 645			lv->lv_next = item->li_lv;
 646		lv = item->li_lv;
 647		item->li_lv = NULL;
 648		num_iovecs += lv->lv_niovecs;
 649	}
 650
 651	/*
 652	 * initialise the new context and attach it to the CIL. Then attach
 653	 * the current context to the CIL committing lsit so it can be found
 654	 * during log forces to extract the commit lsn of the sequence that
 655	 * needs to be forced.
 656	 */
 657	INIT_LIST_HEAD(&new_ctx->committing);
 658	INIT_LIST_HEAD(&new_ctx->busy_extents);
 659	new_ctx->sequence = ctx->sequence + 1;
 660	new_ctx->cil = cil;
 661	cil->xc_ctx = new_ctx;
 662
 663	/*
 664	 * The switch is now done, so we can drop the context lock and move out
 665	 * of a shared context. We can't just go straight to the commit record,
 666	 * though - we need to synchronise with previous and future commits so
 667	 * that the commit records are correctly ordered in the log to ensure
 668	 * that we process items during log IO completion in the correct order.
 669	 *
 670	 * For example, if we get an EFI in one checkpoint and the EFD in the
 671	 * next (e.g. due to log forces), we do not want the checkpoint with
 672	 * the EFD to be committed before the checkpoint with the EFI.  Hence
 673	 * we must strictly order the commit records of the checkpoints so
 674	 * that: a) the checkpoint callbacks are attached to the iclogs in the
 675	 * correct order; and b) the checkpoints are replayed in correct order
 676	 * in log recovery.
 677	 *
 678	 * Hence we need to add this context to the committing context list so
 679	 * that higher sequences will wait for us to write out a commit record
 680	 * before they do.
 681	 *
 682	 * xfs_log_force_lsn requires us to mirror the new sequence into the cil
 683	 * structure atomically with the addition of this sequence to the
 684	 * committing list. This also ensures that we can do unlocked checks
 685	 * against the current sequence in log forces without risking
 686	 * deferencing a freed context pointer.
 687	 */
 688	spin_lock(&cil->xc_push_lock);
 689	cil->xc_current_sequence = new_ctx->sequence;
 690	spin_unlock(&cil->xc_push_lock);
 691	up_write(&cil->xc_ctx_lock);
 692
 693	/*
 694	 * Build a checkpoint transaction header and write it to the log to
 695	 * begin the transaction. We need to account for the space used by the
 696	 * transaction header here as it is not accounted for in xlog_write().
 697	 *
 698	 * The LSN we need to pass to the log items on transaction commit is
 699	 * the LSN reported by the first log vector write. If we use the commit
 700	 * record lsn then we can move the tail beyond the grant write head.
 701	 */
 702	tic = ctx->ticket;
 703	thdr.th_magic = XFS_TRANS_HEADER_MAGIC;
 704	thdr.th_type = XFS_TRANS_CHECKPOINT;
 705	thdr.th_tid = tic->t_tid;
 706	thdr.th_num_items = num_iovecs;
 707	lhdr.i_addr = &thdr;
 708	lhdr.i_len = sizeof(xfs_trans_header_t);
 709	lhdr.i_type = XLOG_REG_TYPE_TRANSHDR;
 710	tic->t_curr_res -= lhdr.i_len + sizeof(xlog_op_header_t);
 711
 712	lvhdr.lv_niovecs = 1;
 713	lvhdr.lv_iovecp = &lhdr;
 714	lvhdr.lv_next = ctx->lv_chain;
 715
 716	error = xlog_write(log, &lvhdr, tic, &ctx->start_lsn, NULL, 0);
 717	if (error)
 718		goto out_abort_free_ticket;
 719
 720	/*
 721	 * now that we've written the checkpoint into the log, strictly
 722	 * order the commit records so replay will get them in the right order.
 723	 */
 724restart:
 725	spin_lock(&cil->xc_push_lock);
 726	list_for_each_entry(new_ctx, &cil->xc_committing, committing) {
 727		/*
 728		 * Avoid getting stuck in this loop because we were woken by the
 729		 * shutdown, but then went back to sleep once already in the
 730		 * shutdown state.
 731		 */
 732		if (XLOG_FORCED_SHUTDOWN(log)) {
 733			spin_unlock(&cil->xc_push_lock);
 734			goto out_abort_free_ticket;
 735		}
 736
 737		/*
 738		 * Higher sequences will wait for this one so skip them.
 739		 * Don't wait for our own sequence, either.
 740		 */
 741		if (new_ctx->sequence >= ctx->sequence)
 742			continue;
 743		if (!new_ctx->commit_lsn) {
 744			/*
 745			 * It is still being pushed! Wait for the push to
 746			 * complete, then start again from the beginning.
 747			 */
 748			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
 749			goto restart;
 750		}
 751	}
 752	spin_unlock(&cil->xc_push_lock);
 753
 754	/* xfs_log_done always frees the ticket on error. */
 755	commit_lsn = xfs_log_done(log->l_mp, tic, &commit_iclog, false);
 756	if (commit_lsn == -1)
 757		goto out_abort;
 758
 759	/* attach all the transactions w/ busy extents to iclog */
 760	ctx->log_cb.cb_func = xlog_cil_committed;
 761	ctx->log_cb.cb_arg = ctx;
 762	error = xfs_log_notify(log->l_mp, commit_iclog, &ctx->log_cb);
 763	if (error)
 764		goto out_abort;
 765
 766	/*
 767	 * now the checkpoint commit is complete and we've attached the
 768	 * callbacks to the iclog we can assign the commit LSN to the context
 769	 * and wake up anyone who is waiting for the commit to complete.
 770	 */
 771	spin_lock(&cil->xc_push_lock);
 772	ctx->commit_lsn = commit_lsn;
 773	wake_up_all(&cil->xc_commit_wait);
 774	spin_unlock(&cil->xc_push_lock);
 775
 776	/* release the hounds! */
 777	return xfs_log_release_iclog(log->l_mp, commit_iclog);
 778
 779out_skip:
 780	up_write(&cil->xc_ctx_lock);
 781	xfs_log_ticket_put(new_ctx->ticket);
 782	kmem_free(new_ctx);
 783	return 0;
 784
 785out_abort_free_ticket:
 786	xfs_log_ticket_put(tic);
 787out_abort:
 788	xlog_cil_committed(ctx, XFS_LI_ABORTED);
 789	return -EIO;
 790}
 791
 792static void
 793xlog_cil_push_work(
 794	struct work_struct	*work)
 795{
 796	struct xfs_cil		*cil = container_of(work, struct xfs_cil,
 797							xc_push_work);
 798	xlog_cil_push(cil->xc_log);
 799}
 800
 801/*
 802 * We need to push CIL every so often so we don't cache more than we can fit in
 803 * the log. The limit really is that a checkpoint can't be more than half the
 804 * log (the current checkpoint is not allowed to overwrite the previous
 805 * checkpoint), but commit latency and memory usage limit this to a smaller
 806 * size.
 807 */
 808static void
 809xlog_cil_push_background(
 810	struct xlog	*log)
 811{
 812	struct xfs_cil	*cil = log->l_cilp;
 813
 814	/*
 815	 * The cil won't be empty because we are called while holding the
 816	 * context lock so whatever we added to the CIL will still be there
 817	 */
 818	ASSERT(!list_empty(&cil->xc_cil));
 819
 820	/*
 821	 * don't do a background push if we haven't used up all the
 822	 * space available yet.
 823	 */
 824	if (cil->xc_ctx->space_used < XLOG_CIL_SPACE_LIMIT(log))
 825		return;
 826
 827	spin_lock(&cil->xc_push_lock);
 828	if (cil->xc_push_seq < cil->xc_current_sequence) {
 829		cil->xc_push_seq = cil->xc_current_sequence;
 830		queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
 831	}
 832	spin_unlock(&cil->xc_push_lock);
 833
 834}
 835
 836/*
 837 * xlog_cil_push_now() is used to trigger an immediate CIL push to the sequence
 838 * number that is passed. When it returns, the work will be queued for
 839 * @push_seq, but it won't be completed. The caller is expected to do any
 840 * waiting for push_seq to complete if it is required.
 841 */
 842static void
 843xlog_cil_push_now(
 844	struct xlog	*log,
 845	xfs_lsn_t	push_seq)
 846{
 847	struct xfs_cil	*cil = log->l_cilp;
 848
 849	if (!cil)
 850		return;
 851
 852	ASSERT(push_seq && push_seq <= cil->xc_current_sequence);
 853
 854	/* start on any pending background push to minimise wait time on it */
 855	flush_work(&cil->xc_push_work);
 856
 857	/*
 858	 * If the CIL is empty or we've already pushed the sequence then
 859	 * there's no work we need to do.
 860	 */
 861	spin_lock(&cil->xc_push_lock);
 862	if (list_empty(&cil->xc_cil) || push_seq <= cil->xc_push_seq) {
 863		spin_unlock(&cil->xc_push_lock);
 864		return;
 865	}
 866
 867	cil->xc_push_seq = push_seq;
 868	queue_work(log->l_mp->m_cil_workqueue, &cil->xc_push_work);
 869	spin_unlock(&cil->xc_push_lock);
 870}
 871
 872bool
 873xlog_cil_empty(
 874	struct xlog	*log)
 875{
 876	struct xfs_cil	*cil = log->l_cilp;
 877	bool		empty = false;
 878
 879	spin_lock(&cil->xc_push_lock);
 880	if (list_empty(&cil->xc_cil))
 881		empty = true;
 882	spin_unlock(&cil->xc_push_lock);
 883	return empty;
 884}
 885
 886/*
 887 * Commit a transaction with the given vector to the Committed Item List.
 888 *
 889 * To do this, we need to format the item, pin it in memory if required and
 890 * account for the space used by the transaction. Once we have done that we
 891 * need to release the unused reservation for the transaction, attach the
 892 * transaction to the checkpoint context so we carry the busy extents through
 893 * to checkpoint completion, and then unlock all the items in the transaction.
 894 *
 895 * Called with the context lock already held in read mode to lock out
 896 * background commit, returns without it held once background commits are
 897 * allowed again.
 898 */
 899void
 900xfs_log_commit_cil(
 901	struct xfs_mount	*mp,
 902	struct xfs_trans	*tp,
 903	xfs_lsn_t		*commit_lsn,
 904	bool			regrant)
 905{
 906	struct xlog		*log = mp->m_log;
 907	struct xfs_cil		*cil = log->l_cilp;
 908
 909	/*
 910	 * Do all necessary memory allocation before we lock the CIL.
 911	 * This ensures the allocation does not deadlock with a CIL
 912	 * push in memory reclaim (e.g. from kswapd).
 913	 */
 914	xlog_cil_alloc_shadow_bufs(log, tp);
 915
 916	/* lock out background commit */
 917	down_read(&cil->xc_ctx_lock);
 918
 919	xlog_cil_insert_items(log, tp);
 920
 921	/* check we didn't blow the reservation */
 922	if (tp->t_ticket->t_curr_res < 0)
 923		xlog_print_tic_res(mp, tp->t_ticket);
 924
 925	tp->t_commit_lsn = cil->xc_ctx->sequence;
 926	if (commit_lsn)
 927		*commit_lsn = tp->t_commit_lsn;
 928
 929	xfs_log_done(mp, tp->t_ticket, NULL, regrant);
 930	xfs_trans_unreserve_and_mod_sb(tp);
 931
 932	/*
 933	 * Once all the items of the transaction have been copied to the CIL,
 934	 * the items can be unlocked and freed.
 935	 *
 936	 * This needs to be done before we drop the CIL context lock because we
 937	 * have to update state in the log items and unlock them before they go
 938	 * to disk. If we don't, then the CIL checkpoint can race with us and
 939	 * we can run checkpoint completion before we've updated and unlocked
 940	 * the log items. This affects (at least) processing of stale buffers,
 941	 * inodes and EFIs.
 942	 */
 943	xfs_trans_free_items(tp, tp->t_commit_lsn, false);
 944
 945	xlog_cil_push_background(log);
 946
 947	up_read(&cil->xc_ctx_lock);
 948}
 949
 950/*
 951 * Conditionally push the CIL based on the sequence passed in.
 952 *
 953 * We only need to push if we haven't already pushed the sequence
 954 * number given. Hence the only time we will trigger a push here is
 955 * if the push sequence is the same as the current context.
 956 *
 957 * We return the current commit lsn to allow the callers to determine if a
 958 * iclog flush is necessary following this call.
 959 */
 960xfs_lsn_t
 961xlog_cil_force_lsn(
 962	struct xlog	*log,
 963	xfs_lsn_t	sequence)
 964{
 965	struct xfs_cil		*cil = log->l_cilp;
 966	struct xfs_cil_ctx	*ctx;
 967	xfs_lsn_t		commit_lsn = NULLCOMMITLSN;
 968
 969	ASSERT(sequence <= cil->xc_current_sequence);
 970
 971	/*
 972	 * check to see if we need to force out the current context.
 973	 * xlog_cil_push() handles racing pushes for the same sequence,
 974	 * so no need to deal with it here.
 975	 */
 976restart:
 977	xlog_cil_push_now(log, sequence);
 978
 979	/*
 980	 * See if we can find a previous sequence still committing.
 981	 * We need to wait for all previous sequence commits to complete
 982	 * before allowing the force of push_seq to go ahead. Hence block
 983	 * on commits for those as well.
 984	 */
 985	spin_lock(&cil->xc_push_lock);
 986	list_for_each_entry(ctx, &cil->xc_committing, committing) {
 987		/*
 988		 * Avoid getting stuck in this loop because we were woken by the
 989		 * shutdown, but then went back to sleep once already in the
 990		 * shutdown state.
 991		 */
 992		if (XLOG_FORCED_SHUTDOWN(log))
 993			goto out_shutdown;
 994		if (ctx->sequence > sequence)
 995			continue;
 996		if (!ctx->commit_lsn) {
 997			/*
 998			 * It is still being pushed! Wait for the push to
 999			 * complete, then start again from the beginning.
1000			 */
1001			xlog_wait(&cil->xc_commit_wait, &cil->xc_push_lock);
1002			goto restart;
1003		}
1004		if (ctx->sequence != sequence)
1005			continue;
1006		/* found it! */
1007		commit_lsn = ctx->commit_lsn;
1008	}
1009
1010	/*
1011	 * The call to xlog_cil_push_now() executes the push in the background.
1012	 * Hence by the time we have got here it our sequence may not have been
1013	 * pushed yet. This is true if the current sequence still matches the
1014	 * push sequence after the above wait loop and the CIL still contains
1015	 * dirty objects. This is guaranteed by the push code first adding the
1016	 * context to the committing list before emptying the CIL.
1017	 *
1018	 * Hence if we don't find the context in the committing list and the
1019	 * current sequence number is unchanged then the CIL contents are
1020	 * significant.  If the CIL is empty, if means there was nothing to push
1021	 * and that means there is nothing to wait for. If the CIL is not empty,
1022	 * it means we haven't yet started the push, because if it had started
1023	 * we would have found the context on the committing list.
1024	 */
1025	if (sequence == cil->xc_current_sequence &&
1026	    !list_empty(&cil->xc_cil)) {
1027		spin_unlock(&cil->xc_push_lock);
1028		goto restart;
1029	}
1030
1031	spin_unlock(&cil->xc_push_lock);
1032	return commit_lsn;
1033
1034	/*
1035	 * We detected a shutdown in progress. We need to trigger the log force
1036	 * to pass through it's iclog state machine error handling, even though
1037	 * we are already in a shutdown state. Hence we can't return
1038	 * NULLCOMMITLSN here as that has special meaning to log forces (i.e.
1039	 * LSN is already stable), so we return a zero LSN instead.
1040	 */
1041out_shutdown:
1042	spin_unlock(&cil->xc_push_lock);
1043	return 0;
1044}
1045
1046/*
1047 * Check if the current log item was first committed in this sequence.
1048 * We can't rely on just the log item being in the CIL, we have to check
1049 * the recorded commit sequence number.
1050 *
1051 * Note: for this to be used in a non-racy manner, it has to be called with
1052 * CIL flushing locked out. As a result, it should only be used during the
1053 * transaction commit process when deciding what to format into the item.
1054 */
1055bool
1056xfs_log_item_in_current_chkpt(
1057	struct xfs_log_item *lip)
1058{
1059	struct xfs_cil_ctx *ctx;
1060
1061	if (list_empty(&lip->li_cil))
1062		return false;
1063
1064	ctx = lip->li_mountp->m_log->l_cilp->xc_ctx;
1065
1066	/*
1067	 * li_seq is written on the first commit of a log item to record the
1068	 * first checkpoint it is written to. Hence if it is different to the
1069	 * current sequence, we're in a new checkpoint.
1070	 */
1071	if (XFS_LSN_CMP(lip->li_seq, ctx->sequence) != 0)
1072		return false;
1073	return true;
1074}
1075
1076/*
1077 * Perform initial CIL structure initialisation.
1078 */
1079int
1080xlog_cil_init(
1081	struct xlog	*log)
1082{
1083	struct xfs_cil	*cil;
1084	struct xfs_cil_ctx *ctx;
1085
1086	cil = kmem_zalloc(sizeof(*cil), KM_SLEEP|KM_MAYFAIL);
1087	if (!cil)
1088		return -ENOMEM;
1089
1090	ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP|KM_MAYFAIL);
1091	if (!ctx) {
1092		kmem_free(cil);
1093		return -ENOMEM;
1094	}
1095
1096	INIT_WORK(&cil->xc_push_work, xlog_cil_push_work);
1097	INIT_LIST_HEAD(&cil->xc_cil);
1098	INIT_LIST_HEAD(&cil->xc_committing);
1099	spin_lock_init(&cil->xc_cil_lock);
1100	spin_lock_init(&cil->xc_push_lock);
1101	init_rwsem(&cil->xc_ctx_lock);
1102	init_waitqueue_head(&cil->xc_commit_wait);
1103
1104	INIT_LIST_HEAD(&ctx->committing);
1105	INIT_LIST_HEAD(&ctx->busy_extents);
1106	ctx->sequence = 1;
1107	ctx->cil = cil;
1108	cil->xc_ctx = ctx;
1109	cil->xc_current_sequence = ctx->sequence;
1110
1111	cil->xc_log = log;
1112	log->l_cilp = cil;
1113	return 0;
1114}
1115
1116void
1117xlog_cil_destroy(
1118	struct xlog	*log)
1119{
1120	if (log->l_cilp->xc_ctx) {
1121		if (log->l_cilp->xc_ctx->ticket)
1122			xfs_log_ticket_put(log->l_cilp->xc_ctx->ticket);
1123		kmem_free(log->l_cilp->xc_ctx);
1124	}
1125
1126	ASSERT(list_empty(&log->l_cilp->xc_cil));
1127	kmem_free(log->l_cilp);
1128}
1129