sys/vfs/hammer2/hammer2_bulkfree.c

   1 /*
   2  * Copyright (c) 2013-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/kernel.h>
  37 #include <sys/fcntl.h>
  38 #include <sys/buf.h>
  39 #include <sys/proc.h>
  40 #include <sys/namei.h>
  41 #include <sys/mount.h>
  42 #include <sys/vnode.h>
  43 #include <sys/mountctl.h>
  44 #include <vm/vm_kern.h>
  45 #include <vm/vm_extern.h>
  46
  47 #include "hammer2.h"
  48
  49 /*
  50  * breadth-first search
  51  */
  52 typedef struct hammer2_chain_save {
  53         TAILQ_ENTRY(hammer2_chain_save) entry;
  54         hammer2_chain_t *chain;
  55         int pri;
  56 } hammer2_chain_save_t;
  57
  58 TAILQ_HEAD(hammer2_chain_save_list, hammer2_chain_save);
  59 typedef struct hammer2_chain_save_list hammer2_chain_save_list_t;
  60
  61 typedef struct hammer2_bulkfree_info {
  62         hammer2_dev_t           *hmp;
  63         kmem_anon_desc_t        kp;
  64         hammer2_off_t           sbase;          /* sub-loop iteration */
  65         hammer2_off_t           sstop;
  66         hammer2_bmap_data_t     *bmap;
  67         int                     depth;
  68         long                    count_10_00;
  69         long                    count_11_10;
  70         long                    count_10_11;
  71         long                    count_l0cleans;
  72         long                    count_linadjusts;
  73         long                    count_inodes_scanned;
  74         long                    count_dedup_factor;
  75         long                    bytes_scanned;
  76         hammer2_off_t           adj_free;
  77         hammer2_tid_t           mtid;
  78         hammer2_tid_t           saved_mirror_tid;
  79         time_t                  save_time;
  80         hammer2_chain_save_list_t list;
  81         hammer2_dedup_t         *dedup;
  82         int                     pri;
  83 } hammer2_bulkfree_info_t;
  84
  85 static int h2_bulkfree_test(hammer2_bulkfree_info_t *info,
  86                         hammer2_blockref_t *bref, int pri);
  87
  88 /*
  89  * General bulk scan function with callback.  Called with a referenced
  90  * but UNLOCKED parent.  The parent is returned in the same state.
  91  */
  92 static
  93 int
  94 hammer2_bulk_scan(hammer2_chain_t *parent,
  95                   int (*func)(hammer2_bulkfree_info_t *info,
  96                               hammer2_blockref_t *bref),
  97                   hammer2_bulkfree_info_t *info)
  98 {
  99         hammer2_blockref_t bref;
 100         hammer2_chain_t *chain;
 101         int cache_index = -1;
 102         int doabort = 0;
 103         int first = 1;
 104
 105         ++info->pri;
 106
 107         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
 108                                    HAMMER2_RESOLVE_SHARED);
 109         chain = NULL;
 110
 111         /*
 112          * Generally loop on the contents if we have not been flagged
 113          * for abort.
 114          *
 115          * Remember that these chains are completely isolated from
 116          * the frontend, so we can release locks temporarily without
 117          * imploding.
 118          */
 119         while ((doabort & HAMMER2_BULK_ABORT) == 0 &&
 120                hammer2_chain_scan(parent, &chain, &bref, &first,
 121                                   &cache_index,
 122                                   HAMMER2_LOOKUP_NODATA |
 123                                   HAMMER2_LOOKUP_SHARED) != NULL) {
 124                 /*
 125                  * Process bref, chain is only non-NULL if the bref
 126                  * might be recursable (its possible that we sometimes get
 127                  * a non-NULL chain where the bref cannot be recursed).
 128                  */
 129 #if 0
 130                 kprintf("SCAN %016jx\n", bref.data_off);
 131                 int xerr = tsleep(&info->pri, PCATCH, "slp", hz / 10);
 132                 if (xerr == EINTR || xerr == ERESTART) {
 133                         doabort |= HAMMER2_BULK_ABORT;
 134                 }
 135 #endif
 136                 ++info->pri;
 137                 if (h2_bulkfree_test(info, &bref, 1))
 138                         continue;
 139
 140                 doabort |= func(info, &bref);
 141
 142                 if (doabort & HAMMER2_BULK_ABORT)
 143                         break;
 144
 145                 /*
 146                  * A non-null chain is always returned if it is
 147                  * recursive, otherwise a non-null chain might be
 148                  * returned but usually is not when not recursive.
 149                  */
 150                 if (chain == NULL)
 151                         continue;
 152
 153                 /*
 154                  * Else check type and setup depth-first scan.
 155                  *
 156                  * Account for bytes actually read.
 157                  */
 158                 info->bytes_scanned += chain->bytes;
 159
 160                 switch(chain->bref.type) {
 161                 case HAMMER2_BREF_TYPE_INODE:
 162                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 163                 case HAMMER2_BREF_TYPE_INDIRECT:
 164                 case HAMMER2_BREF_TYPE_VOLUME:
 165                 case HAMMER2_BREF_TYPE_FREEMAP:
 166                         ++info->depth;
 167                         if (info->depth > 16) {
 168                                 hammer2_chain_save_t *save;
 169                                 save = kmalloc(sizeof(*save), M_HAMMER2,
 170                                                M_WAITOK | M_ZERO);
 171                                 save->chain = chain;
 172                                 hammer2_chain_ref(chain);
 173                                 TAILQ_INSERT_TAIL(&info->list, save, entry);
 174
 175                                 /* guess */
 176                                 info->pri += 10;
 177                         } else {
 178                                 int savepri = info->pri;
 179
 180                                 hammer2_chain_unlock(chain);
 181                                 info->pri = 0;
 182                                 doabort |= hammer2_bulk_scan(chain, func, info);
 183                                 info->pri += savepri;
 184                                 hammer2_chain_lock(chain,
 185                                                    HAMMER2_RESOLVE_ALWAYS |
 186                                                    HAMMER2_RESOLVE_SHARED);
 187                         }
 188                         --info->depth;
 189                         break;
 190                 default:
 191                         /* does not recurse */
 192                         break;
 193                 }
 194         }
 195         if (chain) {
 196                 hammer2_chain_unlock(chain);
 197                 hammer2_chain_drop(chain);
 198         }
 199
 200         /*
 201          * Save with higher pri now that we know what it is.
 202          */
 203         h2_bulkfree_test(info, &parent->bref, info->pri + 1);
 204
 205         hammer2_chain_unlock(parent);
 206
 207         return doabort;
 208 }
 209
 210 /*
 211  * Bulkfree algorithm
 212  *
 213  * Repeat {
 214  *      Chain flush (partial synchronization) XXX removed
 215  *      Scan the whole topology - build in-memory freemap (mark 11)
 216  *      Reconcile the in-memory freemap against the on-disk freemap.
 217  *              ondisk xx -> ondisk 11 (if allocated)
 218  *              ondisk 11 -> ondisk 10 (if free in-memory)
 219  *              ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
 220  * }
 221  *
 222  * The topology scan may have to be performed multiple times to window
 223  * freemaps which are too large to fit in kernel memory.
 224  *
 225  * Races are handled using a double-transition (11->10, 10->00).  The bulkfree
 226  * scan snapshots the volume root's blockset and thus can run concurrent with
 227  * normal operations, as long as a full flush is made between each pass to
 228  * synchronize any modified chains (otherwise their blocks might be improperly
 229  * freed).
 230  *
 231  * Temporary memory in multiples of 64KB is required to reconstruct the leaf
 232  * hammer2_bmap_data blocks so they can later be compared against the live
 233  * freemap.  Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
 234  * A 32MB save area thus represents around ~1 TB.  The temporary memory
 235  * allocated can be specified.  If it is not sufficient multiple topology
 236  * passes will be made.
 237  */
 238
 239 /*
 240  * Bulkfree callback info
 241  */
 242 static int h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo,
 243                         hammer2_blockref_t *bref);
 244 static void h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo);
 245 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 246                         hammer2_bmap_data_t *live, hammer2_bmap_data_t *bmap,
 247                         int nofree);
 248
 249 int
 250 hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_ioc_bulkfree_t *bfi)
 251 {
 252         hammer2_bulkfree_info_t cbinfo;
 253         hammer2_chain_t *vchain;
 254         hammer2_chain_save_t *save;
 255         hammer2_off_t incr;
 256         size_t size;
 257         int doabort = 0;
 258
 259         /*
 260          * A bulkfree operations lock is required for the duration.  We
 261          * must hold it across our flushes to guarantee that we never run
 262          * two bulkfree passes in a row without a flush in the middle.
 263          */
 264         lockmgr(&hmp->bulklk, LK_EXCLUSIVE);
 265
 266         /*
 267          * We have to clear the live dedup cache as it might have entries
 268          * that are freeable as of now.  Any new entries in the dedup cache
 269          * made after this point, even if they become freeable, will have
 270          * previously been fully allocated and will be protected by the
 271          * 2-stage bulkfree.
 272          */
 273         hammer2_dedup_clear(hmp);
 274
 275 #if 0
 276         /*
 277          * XXX This has been removed.  Instead of trying to flush, which
 278          * appears to have a ton of races against life chains even with
 279          * the two-stage scan, we simply refuse to free any blocks
 280          * related to freemap chains modified after the last filesystem
 281          * sync.
 282          *
 283          * Do a quick flush so we can snapshot vchain for any blocks that
 284          * have been allocated prior to this point.  We don't need to
 285          * flush vnodes, logical buffers, or dirty inodes that have not
 286          * allocated blocks yet.  We do not want to flush the device buffers
 287          * nor do we want to flush the actual volume root to disk here,
 288          * that is not needed to perform the snapshot.
 289          */
 290         hammer2_flush_quick(hmp);
 291 #endif
 292
 293         /*
 294          * Setup for free pass
 295          */
 296         bzero(&cbinfo, sizeof(cbinfo));
 297         size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
 298                ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
 299         cbinfo.hmp = hmp;
 300         cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size);
 301         cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
 302
 303         cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
 304                                M_HAMMER2, M_WAITOK | M_ZERO);
 305
 306         /*
 307          * Normalize start point to a 2GB boundary.  We operate on a
 308          * 64KB leaf bitmap boundary which represents 2GB of storage.
 309          */
 310         cbinfo.sbase = bfi->sbase;
 311         if (cbinfo.sbase > hmp->voldata.volu_size)
 312                 cbinfo.sbase = hmp->voldata.volu_size;
 313         cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 314
 315         /*
 316          * The primary storage scan must use a snapshot of the volume
 317          * root to avoid racing renames and other frontend work.
 318          *
 319          * Note that snapshots only snap synchronized storage, so
 320          * we have to flush between each pass or we risk freeing
 321          * storage allocated by the frontend.
 322          */
 323         vchain = hammer2_chain_bulksnap(&hmp->vchain);
 324         TAILQ_INIT(&cbinfo.list);
 325
 326         /*
 327          * Loop on a full meta-data scan as many times as required to
 328          * get through all available storage.
 329          */
 330         while (cbinfo.sbase < hmp->voldata.volu_size) {
 331                 /*
 332                  * We have enough ram to represent (incr) bytes of storage.
 333                  * Each 64KB of ram represents 2GB of storage.
 334                  */
 335                 bzero(cbinfo.bmap, size);
 336                 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
 337                        HAMMER2_FREEMAP_LEVEL1_SIZE;
 338                 if (hmp->voldata.volu_size - cbinfo.sbase < incr)
 339                         cbinfo.sstop = hmp->voldata.volu_size;
 340                 else
 341                         cbinfo.sstop = cbinfo.sbase + incr;
 342                 if (hammer2_debug & 1)
 343                 kprintf("bulkfree pass %016jx/%jdGB\n",
 344                         (intmax_t)cbinfo.sbase,
 345                         (intmax_t)incr / HAMMER2_FREEMAP_LEVEL1_SIZE);
 346
 347                 hammer2_trans_init(hmp->spmp, 0);
 348                 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
 349                 cbinfo.pri = 0;
 350                 doabort |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
 351                                              &cbinfo);
 352
 353                 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
 354                        doabort == 0) {
 355                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 356                         cbinfo.pri = 0;
 357                         doabort |= hammer2_bulk_scan(save->chain,
 358                                                      h2_bulkfree_callback,
 359                                                      &cbinfo);
 360                         hammer2_chain_drop(save->chain);
 361                         kfree(save, M_HAMMER2);
 362                 }
 363                 while (save) {
 364                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 365                         hammer2_chain_drop(save->chain);
 366                         kfree(save, M_HAMMER2);
 367                         save = TAILQ_FIRST(&cbinfo.list);
 368                 }
 369
 370                 kprintf("bulkfree lastdrop %d %d\n",
 371                         vchain->refs, vchain->core.chain_count);
 372
 373                 /*
 374                  * If complete scan succeeded we can synchronize our
 375                  * in-memory freemap against live storage.  If an abort
 376                  * did occur we cannot safely synchronize our partially
 377                  * filled-out in-memory freemap.
 378                  */
 379                 if (doabort == 0) {
 380                         h2_bulkfree_sync(&cbinfo);
 381
 382                         hammer2_voldata_lock(hmp);
 383                         hammer2_voldata_modify(hmp);
 384                         hmp->voldata.allocator_free += cbinfo.adj_free;
 385                         hammer2_voldata_unlock(hmp);
 386                 }
 387
 388                 /*
 389                  * Cleanup for next loop.
 390                  */
 391                 hammer2_trans_done(hmp->spmp);
 392                 if (doabort)
 393                         break;
 394                 cbinfo.sbase = cbinfo.sstop;
 395         }
 396         hammer2_chain_bulkdrop(vchain);
 397         kmem_free_swapbacked(&cbinfo.kp);
 398         kfree(cbinfo.dedup, M_HAMMER2);
 399         cbinfo.dedup = NULL;
 400
 401         bfi->sstop = cbinfo.sbase;
 402
 403         incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
 404         if (incr > 10000)
 405                 incr = 10000;
 406
 407         kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
 408                 (int)incr / 100,
 409                 (int)incr % 100);
 410
 411         kprintf("    transition->free   %ld\n", cbinfo.count_10_00);
 412         kprintf("    transition->staged %ld\n", cbinfo.count_11_10);
 413         kprintf("    raced on           %ld\n", cbinfo.count_10_11);
 414         kprintf("    ~2MB segs cleaned  %ld\n", cbinfo.count_l0cleans);
 415         kprintf("    linear adjusts     %ld\n", cbinfo.count_linadjusts);
 416         kprintf("    dedup factor       %ld\n", cbinfo.count_dedup_factor);
 417
 418         lockmgr(&hmp->bulklk, LK_RELEASE);
 419
 420         return doabort;
 421 }
 422
 423 static int
 424 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
 425 {
 426         hammer2_bmap_data_t *bmap;
 427         hammer2_off_t data_off;
 428         uint16_t class;
 429         size_t bytes;
 430         int radix;
 431         int error;
 432
 433         /*
 434          * Check for signal and allow yield to userland during scan
 435          */
 436         if (hammer2_signal_check(&cbinfo->save_time))
 437                 return HAMMER2_BULK_ABORT;
 438         if (bref->type == HAMMER2_BREF_TYPE_INODE) {
 439                 ++cbinfo->count_inodes_scanned;
 440                 if ((cbinfo->count_inodes_scanned & 1023) == 0)
 441                         kprintf(" inodes %6ld bytes %9ld\n",
 442                                 cbinfo->count_inodes_scanned,
 443                                 cbinfo->bytes_scanned);
 444         }
 445
 446
 447         /*
 448          * Calculate the data offset and determine if it is within
 449          * the current freemap range being gathered.
 450          */
 451         error = 0;
 452         data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 453         if (data_off < cbinfo->sbase || data_off > cbinfo->sstop)
 454                 return 0;
 455         if (data_off < cbinfo->hmp->voldata.allocator_beg)
 456                 return 0;
 457         if (data_off > cbinfo->hmp->voldata.volu_size)
 458                 return 0;
 459
 460         /*
 461          * Calculate the information needed to generate the in-memory
 462          * freemap record.
 463          *
 464          * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
 465          * it's a problem if it does.  (Or L0 (2MB) for that matter).
 466          */
 467         radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
 468         bytes = (size_t)1 << radix;
 469         class = (bref->type << 8) | hammer2_devblkradix(radix);
 470
 471         if (data_off + bytes > cbinfo->sstop) {
 472                 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
 473                         "%016jx %016jx/%d\n",
 474                         (intmax_t)bref->data_off,
 475                         (intmax_t)bref->key,
 476                         bref->keybits);
 477                 bytes = cbinfo->sstop - data_off;       /* XXX */
 478         }
 479
 480         /*
 481          * Convert to a storage offset relative to the beginning of the
 482          * storage range we are collecting.  Then lookup the level0 bmap entry.
 483          */
 484         data_off -= cbinfo->sbase;
 485         bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
 486
 487         /*
 488          * Convert data_off to a bmap-relative value (~2MB storage range).
 489          * Adjust linear, class, and avail.
 490          *
 491          * Hammer2 does not allow allocations to cross the L0 (2MB) boundary,
 492          */
 493         data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
 494         if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
 495                 kprintf("hammer2_bulkfree_scan: illegal 2MB boundary "
 496                         "%016jx %016jx/%d\n",
 497                         (intmax_t)bref->data_off,
 498                         (intmax_t)bref->key,
 499                         bref->keybits);
 500                 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
 501         }
 502
 503         if (bmap->class == 0) {
 504                 bmap->class = class;
 505                 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 506         }
 507         if (bmap->class != class) {
 508                 kprintf("hammer2_bulkfree_scan: illegal mixed class "
 509                         "%016jx %016jx/%d (%04x vs %04x)\n",
 510                         (intmax_t)bref->data_off,
 511                         (intmax_t)bref->key,
 512                         bref->keybits,
 513                         class, bmap->class);
 514         }
 515         if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
 516                 bmap->linear = (int32_t)data_off + (int32_t)bytes;
 517
 518         /*
 519          * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
 520          * 64-bit entries, 2 bits per entry, to code 11.
 521          *
 522          * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
 523          */
 524         while (bytes > 0) {
 525                 int bindex;
 526                 hammer2_bitmap_t bmask;
 527
 528                 bindex = (int)data_off >> (HAMMER2_FREEMAP_BLOCK_RADIX +
 529                                            HAMMER2_BMAP_INDEX_RADIX);
 530                 bmask = (hammer2_bitmap_t)3 <<
 531                         ((((int)data_off & HAMMER2_BMAP_INDEX_MASK) >>
 532                          HAMMER2_FREEMAP_BLOCK_RADIX) << 1);
 533
 534                 /*
 535                  * NOTE! The (avail) calculation is bitmap-granular.  Multiple
 536                  *       sub-granular records can wind up at the same bitmap
 537                  *       position.
 538                  */
 539                 if ((bmap->bitmapq[bindex] & bmask) == 0) {
 540                         if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE) {
 541                                 bmap->avail -= HAMMER2_FREEMAP_BLOCK_SIZE;
 542                         } else {
 543                                 bmap->avail -= bytes;
 544                         }
 545                         bmap->bitmapq[bindex] |= bmask;
 546                 }
 547                 data_off += HAMMER2_FREEMAP_BLOCK_SIZE;
 548                 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE)
 549                         bytes = 0;
 550                 else
 551                         bytes -= HAMMER2_FREEMAP_BLOCK_SIZE;
 552         }
 553         return error;
 554 }
 555
 556 /*
 557  * Synchronize the in-memory bitmap with the live freemap.  This is not a
 558  * direct copy.  Instead the bitmaps must be compared:
 559  *
 560  *      In-memory       Live-freemap
 561  *         00             11 -> 10      (do nothing if live modified)
 562  *                        10 -> 00      (do nothing if live modified)
 563  *         11             10 -> 11      handles race against live
 564  *                        ** -> 11      nominally warn of corruption
 565  *
 566  */
 567 static void
 568 h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo)
 569 {
 570         hammer2_off_t data_off;
 571         hammer2_key_t key;
 572         hammer2_key_t key_dummy;
 573         hammer2_bmap_data_t *bmap;
 574         hammer2_bmap_data_t *live;
 575         hammer2_chain_t *live_parent;
 576         hammer2_chain_t *live_chain;
 577         int cache_index = -1;
 578         int bmapindex;
 579         int nofree;
 580
 581         kprintf("hammer2_bulkfree - range %016jx-%016jx\n",
 582                 (intmax_t)cbinfo->sbase,
 583                 (intmax_t)cbinfo->sstop);
 584
 585         data_off = cbinfo->sbase;
 586         bmap = cbinfo->bmap;
 587
 588         live_parent = &cbinfo->hmp->fchain;
 589         hammer2_chain_ref(live_parent);
 590         hammer2_chain_lock(live_parent, HAMMER2_RESOLVE_ALWAYS);
 591         live_chain = NULL;
 592         nofree = 1;     /* safety */
 593
 594         while (data_off < cbinfo->sstop) {
 595                 /*
 596                  * The freemap is not used below allocator_beg or beyond
 597                  * volu_size.
 598                  */
 599                 if (data_off < cbinfo->hmp->voldata.allocator_beg)
 600                         goto next;
 601                 if (data_off > cbinfo->hmp->voldata.volu_size)
 602                         goto next;
 603
 604                 /*
 605                  * Locate the freemap leaf on the live filesystem
 606                  */
 607                 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
 608                 if (live_chain == NULL || live_chain->bref.key != key) {
 609                         if (live_chain) {
 610                                 hammer2_chain_unlock(live_chain);
 611                                 hammer2_chain_drop(live_chain);
 612                         }
 613                         live_chain = hammer2_chain_lookup(
 614                                             &live_parent,
 615                                             &key_dummy,
 616                                             key,
 617                                             key + HAMMER2_FREEMAP_LEVEL1_MASK,
 618                                             &cache_index,
 619                                             HAMMER2_LOOKUP_ALWAYS);
 620                         /*
 621                          * If recent allocations were made we avoid races by
 622                          * not freeing any blocks.
 623                          */
 624                         if (live_chain) {
 625                                 kprintf("live_chain %016jx\n", (intmax_t)key);
 626                                 if (live_chain->bref.mirror_tid >
 627                                     cbinfo->saved_mirror_tid) {
 628                                         kprintf("hammer2_bulkfree: "
 629                                                 "avoid %016jx\n",
 630                                                 data_off);
 631                                         nofree = 1;
 632                                 } else {
 633                                         nofree = 0;
 634                                 }
 635                         }
 636
 637                 }
 638                 if (live_chain == NULL) {
 639                         if (bmap->class &&
 640                             bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
 641                                 kprintf("hammer2_bulkfree: cannot locate "
 642                                         "live leaf for allocated data "
 643                                         "near %016jx\n",
 644                                         (intmax_t)data_off);
 645                         }
 646                         goto next;
 647                 }
 648                 if (live_chain->error) {
 649                         kprintf("hammer2_bulkfree: error %s looking up "
 650                                 "live leaf for allocated data near %016jx\n",
 651                                 hammer2_error_str(live_chain->error),
 652                                 (intmax_t)data_off);
 653                         hammer2_chain_unlock(live_chain);
 654                         hammer2_chain_drop(live_chain);
 655                         live_chain = NULL;
 656                         goto next;
 657                 }
 658
 659                 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
 660                             HAMMER2_FREEMAP_LEVEL0_RADIX;
 661                 live = &live_chain->data->bmdata[bmapindex];
 662
 663                 /*
 664                  * For now just handle the 11->10, 10->00, and 10->11
 665                  * transitions.
 666                  */
 667                 if (live->class == 0 ||
 668                     live->avail == HAMMER2_FREEMAP_LEVEL0_SIZE) {
 669                         goto next;
 670                 }
 671                 if (bcmp(live->bitmapq, bmap->bitmapq,
 672                          sizeof(bmap->bitmapq)) == 0) {
 673                         goto next;
 674                 }
 675                 if (hammer2_debug & 1)
 676                 kprintf("live %016jx %04d.%04x (avail=%d)\n",
 677                         data_off, bmapindex, live->class, live->avail);
 678
 679                 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
 680                 h2_bulkfree_sync_adjust(cbinfo, live, bmap, nofree);
 681 next:
 682                 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
 683                 ++bmap;
 684         }
 685         if (live_chain) {
 686                 hammer2_chain_unlock(live_chain);
 687                 hammer2_chain_drop(live_chain);
 688         }
 689         if (live_parent) {
 690                 hammer2_chain_unlock(live_parent);
 691                 hammer2_chain_drop(live_parent);
 692         }
 693 }
 694
 695 /*
 696  * Merge the bulkfree bitmap against the existing bitmap.
 697  *
 698  * If nofree is non-zero the merge will only mark free blocks as allocated
 699  * and will refuse to free any blocks.
 700  */
 701 static
 702 void
 703 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 704                         hammer2_bmap_data_t *live, hammer2_bmap_data_t *bmap,
 705                         int nofree)
 706 {
 707         int bindex;
 708         int scount;
 709         hammer2_bitmap_t lmask;
 710         hammer2_bitmap_t mmask;
 711
 712         for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
 713                 lmask = live->bitmapq[bindex];
 714                 mmask = bmap->bitmapq[bindex];
 715                 if (lmask == mmask)
 716                         continue;
 717
 718                 for (scount = 0;
 719                      scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
 720                      scount += 2) {
 721                         if ((mmask & 3) == 0) {
 722                                 /*
 723                                  * in-memory 00         live 11 -> 10
 724                                  *                      live 10 -> 00
 725                                  */
 726                                 switch (lmask & 3) {
 727                                 case 0: /* 00 */
 728                                         break;
 729                                 case 1: /* 01 */
 730                                         kprintf("hammer2_bulkfree: cannot "
 731                                                 "transition m=00/l=01\n");
 732                                         break;
 733                                 case 2: /* 10 -> 00 */
 734                                         if (nofree)
 735                                                 break;
 736                                         live->bitmapq[bindex] &=
 737                                             ~((hammer2_bitmap_t)2 << scount);
 738                                         live->avail +=
 739                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 740                                         cbinfo->adj_free +=
 741                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 742                                         ++cbinfo->count_10_00;
 743                                         break;
 744                                 case 3: /* 11 -> 10 */
 745                                         if (nofree)
 746                                                 break;
 747                                         live->bitmapq[bindex] &=
 748                                             ~((hammer2_bitmap_t)1 << scount);
 749                                         ++cbinfo->count_11_10;
 750                                         break;
 751                                 }
 752                         } else if ((lmask & 3) == 3) {
 753                                 /*
 754                                  * in-memory 11         live 10 -> 11
 755                                  *                      live ** -> 11
 756                                  */
 757                                 switch (lmask & 3) {
 758                                 case 0: /* 00 */
 759                                         kprintf("hammer2_bulkfree: cannot "
 760                                                 "transition m=11/l=00\n");
 761                                         break;
 762                                 case 1: /* 01 */
 763                                         kprintf("hammer2_bulkfree: cannot "
 764                                                 "transition m=11/l=01\n");
 765                                         break;
 766                                 case 2: /* 10 -> 11 */
 767                                         live->bitmapq[bindex] |=
 768                                                 ((hammer2_bitmap_t)1 << scount);
 769                                         ++cbinfo->count_10_11;
 770                                         break;
 771                                 case 3: /* 11 */
 772                                         break;
 773                                 }
 774                         }
 775                         mmask >>= 2;
 776                         lmask >>= 2;
 777                 }
 778         }
 779
 780         /*
 781          * Determine if the live bitmap is completely free and reset its
 782          * fields if so.  Otherwise check to see if we can reduce the linear
 783          * offset.
 784          */
 785         for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
 786                 if (live->bitmapq[bindex] != 0)
 787                         break;
 788         }
 789         if (bindex < 0) {
 790                 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 791                 live->class = 0;
 792                 live->linear = 0;
 793                 ++cbinfo->count_l0cleans;
 794         } else if (bindex < 7) {
 795                 ++bindex;
 796                 if (live->linear > bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 797                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 798                         ++cbinfo->count_linadjusts;
 799                 }
 800         }
 801
 802 #if 0
 803         if (bmap->class) {
 804                 kprintf("%016jx %04d.%04x (avail=%7d) "
 805                         "%08x %08x %08x %08x %08x %08x %08x %08x\n",
 806                         (intmax_t)data_off,
 807                         (int)((data_off &
 808                                HAMMER2_FREEMAP_LEVEL1_MASK) >>
 809                               HAMMER2_FREEMAP_LEVEL0_RADIX),
 810                         bmap->class,
 811                         bmap->avail,
 812                         bmap->bitmap[0], bmap->bitmap[1],
 813                         bmap->bitmap[2], bmap->bitmap[3],
 814                         bmap->bitmap[4], bmap->bitmap[5],
 815                         bmap->bitmap[6], bmap->bitmap[7]);
 816         }
 817 #endif
 818 }
 819
 820 /*
 821  * BULKFREE DEDUP HEURISTIC
 822  *
 823  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
 824  *          All fields must be loaded into locals and validated.
 825  */
 826 static
 827 int
 828 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
 829                  int pri)
 830 {
 831         hammer2_dedup_t *dedup;
 832         int best;
 833         int n;
 834         int i;
 835
 836         n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
 837         dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
 838
 839         for (i = best = 0; i < 8; ++i) {
 840                 if (dedup[i].data_off == bref->data_off) {
 841                         if (dedup[i].ticks < pri)
 842                                 dedup[i].ticks = pri;
 843                         if (pri == 1)
 844                                 cbinfo->count_dedup_factor += dedup[i].ticks;
 845                         return 1;
 846                 }
 847                 if (dedup[i].ticks < dedup[best].ticks)
 848                         best = i;
 849         }
 850         dedup[best].data_off = bref->data_off;
 851         dedup[best].ticks = pri;
 852
 853         return 0;
 854 }