libxfs/xfs_defer.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
   4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5  */
   6 #include "libxfs_priv.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_bit.h"
  13 #include "xfs_sb.h"
  14 #include "xfs_mount.h"
  15 #include "xfs_defer.h"
  16 #include "xfs_trans.h"
  17 #include "xfs_inode.h"
  18 #include "xfs_trace.h"
  19
  20 /*
  21  * Deferred Operations in XFS
  22  *
  23  * Due to the way locking rules work in XFS, certain transactions (block
  24  * mapping and unmapping, typically) have permanent reservations so that
  25  * we can roll the transaction to adhere to AG locking order rules and
  26  * to unlock buffers between metadata updates.  Prior to rmap/reflink,
  27  * the mapping code had a mechanism to perform these deferrals for
  28  * extents that were going to be freed; this code makes that facility
  29  * more generic.
  30  *
  31  * When adding the reverse mapping and reflink features, it became
  32  * necessary to perform complex remapping multi-transactions to comply
  33  * with AG locking order rules, and to be able to spread a single
  34  * refcount update operation (an operation on an n-block extent can
  35  * update as many as n records!) among multiple transactions.  XFS can
  36  * roll a transaction to facilitate this, but using this facility
  37  * requires us to log "intent" items in case log recovery needs to
  38  * redo the operation, and to log "done" items to indicate that redo
  39  * is not necessary.
  40  *
  41  * Deferred work is tracked in xfs_defer_pending items.  Each pending
  42  * item tracks one type of deferred work.  Incoming work items (which
  43  * have not yet had an intent logged) are attached to a pending item
  44  * on the dop_intake list, where they wait for the caller to finish
  45  * the deferred operations.
  46  *
  47  * Finishing a set of deferred operations is an involved process.  To
  48  * start, we define "rolling a deferred-op transaction" as follows:
  49  *
  50  * > For each xfs_defer_pending item on the dop_intake list,
  51  *   - Sort the work items in AG order.  XFS locking
  52  *     order rules require us to lock buffers in AG order.
  53  *   - Create a log intent item for that type.
  54  *   - Attach it to the pending item.
  55  *   - Move the pending item from the dop_intake list to the
  56  *     dop_pending list.
  57  * > Roll the transaction.
  58  *
  59  * NOTE: To avoid exceeding the transaction reservation, we limit the
  60  * number of items that we attach to a given xfs_defer_pending.
  61  *
  62  * The actual finishing process looks like this:
  63  *
  64  * > For each xfs_defer_pending in the dop_pending list,
  65  *   - Roll the deferred-op transaction as above.
  66  *   - Create a log done item for that type, and attach it to the
  67  *     log intent item.
  68  *   - For each work item attached to the log intent item,
  69  *     * Perform the described action.
  70  *     * Attach the work item to the log done item.
  71  *     * If the result of doing the work was -EAGAIN, ->finish work
  72  *       wants a new transaction.  See the "Requesting a Fresh
  73  *       Transaction while Finishing Deferred Work" section below for
  74  *       details.
  75  *
  76  * The key here is that we must log an intent item for all pending
  77  * work items every time we roll the transaction, and that we must log
  78  * a done item as soon as the work is completed.  With this mechanism
  79  * we can perform complex remapping operations, chaining intent items
  80  * as needed.
  81  *
  82  * Requesting a Fresh Transaction while Finishing Deferred Work
  83  *
  84  * If ->finish_item decides that it needs a fresh transaction to
  85  * finish the work, it must ask its caller (xfs_defer_finish) for a
  86  * continuation.  The most likely cause of this circumstance are the
  87  * refcount adjust functions deciding that they've logged enough items
  88  * to be at risk of exceeding the transaction reservation.
  89  *
  90  * To get a fresh transaction, we want to log the existing log done
  91  * item to prevent the log intent item from replaying, immediately log
  92  * a new log intent item with the unfinished work items, roll the
  93  * transaction, and re-call ->finish_item wherever it left off.  The
  94  * log done item and the new log intent item must be in the same
  95  * transaction or atomicity cannot be guaranteed; defer_finish ensures
  96  * that this happens.
  97  *
  98  * This requires some coordination between ->finish_item and
  99  * defer_finish.  Upon deciding to request a new transaction,
 100  * ->finish_item should update the current work item to reflect the
 101  * unfinished work.  Next, it should reset the log done item's list
 102  * count to the number of items finished, and return -EAGAIN.
 103  * defer_finish sees the -EAGAIN, logs the new log intent item
 104  * with the remaining work items, and leaves the xfs_defer_pending
 105  * item at the head of the dop_work queue.  Then it rolls the
 106  * transaction and picks up processing where it left off.  It is
 107  * required that ->finish_item must be careful to leave enough
 108  * transaction reservation to fit the new log intent item.
 109  *
 110  * This is an example of remapping the extent (E, E+B) into file X at
 111  * offset A and dealing with the extent (C, C+B) already being mapped
 112  * there:
 113  * +-------------------------------------------------+
 114  * | Unmap file X startblock C offset A length B     | t0
 115  * | Intent to reduce refcount for extent (C, B)     |
 116  * | Intent to remove rmap (X, C, A, B)              |
 117  * | Intent to free extent (D, 1) (bmbt block)       |
 118  * | Intent to map (X, A, B) at startblock E         |
 119  * +-------------------------------------------------+
 120  * | Map file X startblock E offset A length B       | t1
 121  * | Done mapping (X, E, A, B)                       |
 122  * | Intent to increase refcount for extent (E, B)   |
 123  * | Intent to add rmap (X, E, A, B)                 |
 124  * +-------------------------------------------------+
 125  * | Reduce refcount for extent (C, B)               | t2
 126  * | Done reducing refcount for extent (C, 9)        |
 127  * | Intent to reduce refcount for extent (C+9, B-9) |
 128  * | (ran out of space after 9 refcount updates)     |
 129  * +-------------------------------------------------+
 130  * | Reduce refcount for extent (C+9, B+9)           | t3
 131  * | Done reducing refcount for extent (C+9, B-9)    |
 132  * | Increase refcount for extent (E, B)             |
 133  * | Done increasing refcount for extent (E, B)      |
 134  * | Intent to free extent (C, B)                    |
 135  * | Intent to free extent (F, 1) (refcountbt block) |
 136  * | Intent to remove rmap (F, 1, REFC)              |
 137  * +-------------------------------------------------+
 138  * | Remove rmap (X, C, A, B)                        | t4
 139  * | Done removing rmap (X, C, A, B)                 |
 140  * | Add rmap (X, E, A, B)                           |
 141  * | Done adding rmap (X, E, A, B)                   |
 142  * | Remove rmap (F, 1, REFC)                        |
 143  * | Done removing rmap (F, 1, REFC)                 |
 144  * +-------------------------------------------------+
 145  * | Free extent (C, B)                              | t5
 146  * | Done freeing extent (C, B)                      |
 147  * | Free extent (D, 1)                              |
 148  * | Done freeing extent (D, 1)                      |
 149  * | Free extent (F, 1)                              |
 150  * | Done freeing extent (F, 1)                      |
 151  * +-------------------------------------------------+
 152  *
 153  * If we should crash before t2 commits, log recovery replays
 154  * the following intent items:
 155  *
 156  * - Intent to reduce refcount for extent (C, B)
 157  * - Intent to remove rmap (X, C, A, B)
 158  * - Intent to free extent (D, 1) (bmbt block)
 159  * - Intent to increase refcount for extent (E, B)
 160  * - Intent to add rmap (X, E, A, B)
 161  *
 162  * In the process of recovering, it should also generate and take care
 163  * of these intent items:
 164  *
 165  * - Intent to free extent (C, B)
 166  * - Intent to free extent (F, 1) (refcountbt block)
 167  * - Intent to remove rmap (F, 1, REFC)
 168  *
 169  * Note that the continuation requested between t2 and t3 is likely to
 170  * reoccur.
 171  */
 172
 173 static const struct xfs_defer_op_type *defer_op_types[XFS_DEFER_OPS_TYPE_MAX];
 174
 175 /*
 176  * For each pending item in the intake list, log its intent item and the
 177  * associated extents, then add the entire intake list to the end of
 178  * the pending list.
 179  */
 180 STATIC void
 181 xfs_defer_create_intents(
 182         struct xfs_trans                *tp)
 183 {
 184         struct list_head                *li;
 185         struct xfs_defer_pending        *dfp;
 186
 187         list_for_each_entry(dfp, &tp->t_dfops, dfp_list) {
 188                 dfp->dfp_intent = dfp->dfp_type->create_intent(tp,
 189                                 dfp->dfp_count);
 190                 trace_xfs_defer_create_intent(tp->t_mountp, dfp);
 191                 list_sort(tp->t_mountp, &dfp->dfp_work,
 192                                 dfp->dfp_type->diff_items);
 193                 list_for_each(li, &dfp->dfp_work)
 194                         dfp->dfp_type->log_item(tp, dfp->dfp_intent, li);
 195         }
 196 }
 197
 198 /* Abort all the intents that were committed. */
 199 STATIC void
 200 xfs_defer_trans_abort(
 201         struct xfs_trans                *tp,
 202         struct list_head                *dop_pending)
 203 {
 204         struct xfs_defer_pending        *dfp;
 205
 206         trace_xfs_defer_trans_abort(tp, _RET_IP_);
 207
 208         /* Abort intent items that don't have a done item. */
 209         list_for_each_entry(dfp, dop_pending, dfp_list) {
 210                 trace_xfs_defer_pending_abort(tp->t_mountp, dfp);
 211                 if (dfp->dfp_intent && !dfp->dfp_done) {
 212                         dfp->dfp_type->abort_intent(dfp->dfp_intent);
 213                         dfp->dfp_intent = NULL;
 214                 }
 215         }
 216 }
 217
 218 /* Roll a transaction so we can do some deferred op processing. */
 219 STATIC int
 220 xfs_defer_trans_roll(
 221         struct xfs_trans                **tpp)
 222 {
 223         struct xfs_trans                *tp = *tpp;
 224         struct xfs_buf_log_item         *bli;
 225         struct xfs_inode_log_item       *ili;
 226         struct xfs_log_item             *lip;
 227         struct xfs_buf                  *bplist[XFS_DEFER_OPS_NR_BUFS];
 228         struct xfs_inode                *iplist[XFS_DEFER_OPS_NR_INODES];
 229         int                             bpcount = 0, ipcount = 0;
 230         int                             i;
 231         int                             error;
 232
 233         list_for_each_entry(lip, &tp->t_items, li_trans) {
 234                 switch (lip->li_type) {
 235                 case XFS_LI_BUF:
 236                         bli = container_of(lip, struct xfs_buf_log_item,
 237                                            bli_item);
 238                         if (bli->bli_flags & XFS_BLI_HOLD) {
 239                                 if (bpcount >= XFS_DEFER_OPS_NR_BUFS) {
 240                                         ASSERT(0);
 241                                         return -EFSCORRUPTED;
 242                                 }
 243                                 xfs_trans_dirty_buf(tp, bli->bli_buf);
 244                                 bplist[bpcount++] = bli->bli_buf;
 245                         }
 246                         break;
 247                 case XFS_LI_INODE:
 248                         ili = container_of(lip, struct xfs_inode_log_item,
 249                                            ili_item);
 250                         if (ili->ili_lock_flags == 0) {
 251                                 if (ipcount >= XFS_DEFER_OPS_NR_INODES) {
 252                                         ASSERT(0);
 253                                         return -EFSCORRUPTED;
 254                                 }
 255                                 xfs_trans_log_inode(tp, ili->ili_inode,
 256                                                     XFS_ILOG_CORE);
 257                                 iplist[ipcount++] = ili->ili_inode;
 258                         }
 259                         break;
 260                 default:
 261                         break;
 262                 }
 263         }
 264
 265         trace_xfs_defer_trans_roll(tp, _RET_IP_);
 266
 267         /* Roll the transaction. */
 268         error = xfs_trans_roll(tpp);
 269         tp = *tpp;
 270         if (error) {
 271                 trace_xfs_defer_trans_roll_error(tp, error);
 272                 return error;
 273         }
 274
 275         /* Rejoin the joined inodes. */
 276         for (i = 0; i < ipcount; i++)
 277                 xfs_trans_ijoin(tp, iplist[i], 0);
 278
 279         /* Rejoin the buffers and dirty them so the log moves forward. */
 280         for (i = 0; i < bpcount; i++) {
 281                 xfs_trans_bjoin(tp, bplist[i]);
 282                 xfs_trans_bhold(tp, bplist[i]);
 283         }
 284
 285         return error;
 286 }
 287
 288 /*
 289  * Reset an already used dfops after finish.
 290  */
 291 static void
 292 xfs_defer_reset(
 293         struct xfs_trans        *tp)
 294 {
 295         ASSERT(list_empty(&tp->t_dfops));
 296
 297         /*
 298          * Low mode state transfers across transaction rolls to mirror dfops
 299          * lifetime. Clear it now that dfops is reset.
 300          */
 301         tp->t_flags &= ~XFS_TRANS_LOWMODE;
 302 }
 303
 304 /*
 305  * Free up any items left in the list.
 306  */
 307 static void
 308 xfs_defer_cancel_list(
 309         struct xfs_mount                *mp,
 310         struct list_head                *dop_list)
 311 {
 312         struct xfs_defer_pending        *dfp;
 313         struct xfs_defer_pending        *pli;
 314         struct list_head                *pwi;
 315         struct list_head                *n;
 316
 317         /*
 318          * Free the pending items.  Caller should already have arranged
 319          * for the intent items to be released.
 320          */
 321         list_for_each_entry_safe(dfp, pli, dop_list, dfp_list) {
 322                 trace_xfs_defer_cancel_list(mp, dfp);
 323                 list_del(&dfp->dfp_list);
 324                 list_for_each_safe(pwi, n, &dfp->dfp_work) {
 325                         list_del(pwi);
 326                         dfp->dfp_count--;
 327                         dfp->dfp_type->cancel_item(pwi);
 328                 }
 329                 ASSERT(dfp->dfp_count == 0);
 330                 kmem_free(dfp);
 331         }
 332 }
 333
 334 /*
 335  * Finish all the pending work.  This involves logging intent items for
 336  * any work items that wandered in since the last transaction roll (if
 337  * one has even happened), rolling the transaction, and finishing the
 338  * work items in the first item on the logged-and-pending list.
 339  *
 340  * If an inode is provided, relog it to the new transaction.
 341  */
 342 int
 343 xfs_defer_finish_noroll(
 344         struct xfs_trans                **tp)
 345 {
 346         struct xfs_defer_pending        *dfp;
 347         struct list_head                *li;
 348         struct list_head                *n;
 349         void                            *state;
 350         int                             error = 0;
 351         void                            (*cleanup_fn)(struct xfs_trans *, void *, int);
 352         LIST_HEAD(dop_pending);
 353
 354         ASSERT((*tp)->t_flags & XFS_TRANS_PERM_LOG_RES);
 355
 356         trace_xfs_defer_finish(*tp, _RET_IP_);
 357
 358         /* Until we run out of pending work to finish... */
 359         while (!list_empty(&dop_pending) || !list_empty(&(*tp)->t_dfops)) {
 360                 /* log intents and pull in intake items */
 361                 xfs_defer_create_intents(*tp);
 362                 list_splice_tail_init(&(*tp)->t_dfops, &dop_pending);
 363
 364                 /*
 365                  * Roll the transaction.
 366                  */
 367                 error = xfs_defer_trans_roll(tp);
 368                 if (error)
 369                         goto out;
 370
 371                 /* Log an intent-done item for the first pending item. */
 372                 dfp = list_first_entry(&dop_pending, struct xfs_defer_pending,
 373                                        dfp_list);
 374                 trace_xfs_defer_pending_finish((*tp)->t_mountp, dfp);
 375                 dfp->dfp_done = dfp->dfp_type->create_done(*tp, dfp->dfp_intent,
 376                                 dfp->dfp_count);
 377                 cleanup_fn = dfp->dfp_type->finish_cleanup;
 378
 379                 /* Finish the work items. */
 380                 state = NULL;
 381                 list_for_each_safe(li, n, &dfp->dfp_work) {
 382                         list_del(li);
 383                         dfp->dfp_count--;
 384                         error = dfp->dfp_type->finish_item(*tp, li,
 385                                         dfp->dfp_done, &state);
 386                         if (error == -EAGAIN) {
 387                                 /*
 388                                  * Caller wants a fresh transaction;
 389                                  * put the work item back on the list
 390                                  * and jump out.
 391                                  */
 392                                 list_add(li, &dfp->dfp_work);
 393                                 dfp->dfp_count++;
 394                                 break;
 395                         } else if (error) {
 396                                 /*
 397                                  * Clean up after ourselves and jump out.
 398                                  * xfs_defer_cancel will take care of freeing
 399                                  * all these lists and stuff.
 400                                  */
 401                                 if (cleanup_fn)
 402                                         cleanup_fn(*tp, state, error);
 403                                 goto out;
 404                         }
 405                 }
 406                 if (error == -EAGAIN) {
 407                         /*
 408                          * Caller wants a fresh transaction, so log a
 409                          * new log intent item to replace the old one
 410                          * and roll the transaction.  See "Requesting
 411                          * a Fresh Transaction while Finishing
 412                          * Deferred Work" above.
 413                          */
 414                         dfp->dfp_intent = dfp->dfp_type->create_intent(*tp,
 415                                         dfp->dfp_count);
 416                         dfp->dfp_done = NULL;
 417                         list_for_each(li, &dfp->dfp_work)
 418                                 dfp->dfp_type->log_item(*tp, dfp->dfp_intent,
 419                                                 li);
 420                 } else {
 421                         /* Done with the dfp, free it. */
 422                         list_del(&dfp->dfp_list);
 423                         kmem_free(dfp);
 424                 }
 425
 426                 if (cleanup_fn)
 427                         cleanup_fn(*tp, state, error);
 428         }
 429
 430 out:
 431         if (error) {
 432                 xfs_defer_trans_abort(*tp, &dop_pending);
 433                 xfs_force_shutdown((*tp)->t_mountp, SHUTDOWN_CORRUPT_INCORE);
 434                 trace_xfs_defer_finish_error(*tp, error);
 435                 xfs_defer_cancel_list((*tp)->t_mountp, &dop_pending);
 436                 xfs_defer_cancel(*tp);
 437                 return error;
 438         }
 439
 440         trace_xfs_defer_finish_done(*tp, _RET_IP_);
 441         return 0;
 442 }
 443
 444 int
 445 xfs_defer_finish(
 446         struct xfs_trans        **tp)
 447 {
 448         int                     error;
 449
 450         /*
 451          * Finish and roll the transaction once more to avoid returning to the
 452          * caller with a dirty transaction.
 453          */
 454         error = xfs_defer_finish_noroll(tp);
 455         if (error)
 456                 return error;
 457         if ((*tp)->t_flags & XFS_TRANS_DIRTY) {
 458                 error = xfs_defer_trans_roll(tp);
 459                 if (error) {
 460                         xfs_force_shutdown((*tp)->t_mountp,
 461                                            SHUTDOWN_CORRUPT_INCORE);
 462                         return error;
 463                 }
 464         }
 465         xfs_defer_reset(*tp);
 466         return 0;
 467 }
 468
 469 void
 470 xfs_defer_cancel(
 471         struct xfs_trans        *tp)
 472 {
 473         struct xfs_mount        *mp = tp->t_mountp;
 474
 475         trace_xfs_defer_cancel(tp, _RET_IP_);
 476         xfs_defer_cancel_list(mp, &tp->t_dfops);
 477 }
 478
 479 /* Add an item for later deferred processing. */
 480 void
 481 xfs_defer_add(
 482         struct xfs_trans                *tp,
 483         enum xfs_defer_ops_type         type,
 484         struct list_head                *li)
 485 {
 486         struct xfs_defer_pending        *dfp = NULL;
 487
 488         ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
 489
 490         /*
 491          * Add the item to a pending item at the end of the intake list.
 492          * If the last pending item has the same type, reuse it.  Else,
 493          * create a new pending item at the end of the intake list.
 494          */
 495         if (!list_empty(&tp->t_dfops)) {
 496                 dfp = list_last_entry(&tp->t_dfops,
 497                                 struct xfs_defer_pending, dfp_list);
 498                 if (dfp->dfp_type->type != type ||
 499                     (dfp->dfp_type->max_items &&
 500                      dfp->dfp_count >= dfp->dfp_type->max_items))
 501                         dfp = NULL;
 502         }
 503         if (!dfp) {
 504                 dfp = kmem_alloc(sizeof(struct xfs_defer_pending),
 505                                 KM_SLEEP | KM_NOFS);
 506                 dfp->dfp_type = defer_op_types[type];
 507                 dfp->dfp_intent = NULL;
 508                 dfp->dfp_done = NULL;
 509                 dfp->dfp_count = 0;
 510                 INIT_LIST_HEAD(&dfp->dfp_work);
 511                 list_add_tail(&dfp->dfp_list, &tp->t_dfops);
 512         }
 513
 514         list_add_tail(li, &dfp->dfp_work);
 515         dfp->dfp_count++;
 516 }
 517
 518 /* Initialize a deferred operation list. */
 519 void
 520 xfs_defer_init_op_type(
 521         const struct xfs_defer_op_type  *type)
 522 {
 523         defer_op_types[type->type] = type;
 524 }
 525
 526 /*
 527  * Move deferred ops from one transaction to another and reset the source to
 528  * initial state. This is primarily used to carry state forward across
 529  * transaction rolls with pending dfops.
 530  */
 531 void
 532 xfs_defer_move(
 533         struct xfs_trans        *dtp,
 534         struct xfs_trans        *stp)
 535 {
 536         list_splice_init(&stp->t_dfops, &dtp->t_dfops);
 537
 538         /*
 539          * Low free space mode was historically controlled by a dfops field.
 540          * This meant that low mode state potentially carried across multiple
 541          * transaction rolls. Transfer low mode on a dfops move to preserve
 542          * that behavior.
 543          */
 544         dtp->t_flags |= (stp->t_flags & XFS_TRANS_LOWMODE);
 545
 546         xfs_defer_reset(stp);
 547 }