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 #define H2FMBASE(key, radix)    ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
  50 #define H2FMSHIFT(radix)        ((hammer2_off_t)1 << (radix))
  51
  52 /*
  53  * breadth-first search
  54  */
  55 typedef struct hammer2_chain_save {
  56         TAILQ_ENTRY(hammer2_chain_save) entry;
  57         hammer2_chain_t *chain;
  58         int pri;
  59 } hammer2_chain_save_t;
  60
  61 TAILQ_HEAD(hammer2_chain_save_list, hammer2_chain_save);
  62 typedef struct hammer2_chain_save_list hammer2_chain_save_list_t;
  63
  64 typedef struct hammer2_bulkfree_info {
  65         hammer2_dev_t           *hmp;
  66         kmem_anon_desc_t        kp;
  67         hammer2_off_t           sbase;          /* sub-loop iteration */
  68         hammer2_off_t           sstop;
  69         hammer2_bmap_data_t     *bmap;
  70         int                     depth;
  71         long                    count_10_00;    /* staged->free      */
  72         long                    count_11_10;    /* allocated->staged */
  73         long                    count_00_11;    /* (should not happen) */
  74         long                    count_01_11;    /* (should not happen) */
  75         long                    count_10_11;    /* staged->allocated */
  76         long                    count_l0cleans;
  77         long                    count_linadjusts;
  78         long                    count_inodes_scanned;
  79         long                    count_dedup_factor;
  80         long                    bytes_scanned;
  81         hammer2_off_t           adj_free;
  82         hammer2_tid_t           mtid;
  83         hammer2_tid_t           saved_mirror_tid;
  84         time_t                  save_time;
  85         hammer2_chain_save_list_t list;
  86         hammer2_dedup_t         *dedup;
  87         int                     pri;
  88 } hammer2_bulkfree_info_t;
  89
  90 static int h2_bulkfree_test(hammer2_bulkfree_info_t *info,
  91                         hammer2_blockref_t *bref, int pri);
  92
  93 /*
  94  * General bulk scan function with callback.  Called with a referenced
  95  * but UNLOCKED parent.  The parent is returned in the same state.
  96  */
  97 static
  98 int
  99 hammer2_bulk_scan(hammer2_chain_t *parent,
 100                   int (*func)(hammer2_bulkfree_info_t *info,
 101                               hammer2_blockref_t *bref),
 102                   hammer2_bulkfree_info_t *info)
 103 {
 104         hammer2_blockref_t bref;
 105         hammer2_chain_t *chain;
 106         int cache_index = -1;
 107         int doabort = 0;
 108         int first = 1;
 109
 110         ++info->pri;
 111
 112         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
 113                                    HAMMER2_RESOLVE_SHARED);
 114         chain = NULL;
 115
 116         /*
 117          * Generally loop on the contents if we have not been flagged
 118          * for abort.
 119          *
 120          * Remember that these chains are completely isolated from
 121          * the frontend, so we can release locks temporarily without
 122          * imploding.
 123          */
 124         while ((doabort & HAMMER2_BULK_ABORT) == 0 &&
 125                hammer2_chain_scan(parent, &chain, &bref, &first,
 126                                   &cache_index,
 127                                   HAMMER2_LOOKUP_NODATA |
 128                                   HAMMER2_LOOKUP_SHARED) != NULL) {
 129                 /*
 130                  * Process bref, chain is only non-NULL if the bref
 131                  * might be recursable (its possible that we sometimes get
 132                  * a non-NULL chain where the bref cannot be recursed).
 133                  */
 134 #if 0
 135                 kprintf("SCAN %016jx\n", bref.data_off);
 136                 int xerr = tsleep(&info->pri, PCATCH, "slp", hz / 10);
 137                 if (xerr == EINTR || xerr == ERESTART) {
 138                         doabort |= HAMMER2_BULK_ABORT;
 139                 }
 140 #endif
 141                 ++info->pri;
 142                 if (h2_bulkfree_test(info, &bref, 1))
 143                         continue;
 144
 145                 doabort |= func(info, &bref);
 146
 147                 if (doabort & HAMMER2_BULK_ABORT)
 148                         break;
 149
 150                 /*
 151                  * A non-null chain is always returned if it is
 152                  * recursive, otherwise a non-null chain might be
 153                  * returned but usually is not when not recursive.
 154                  */
 155                 if (chain == NULL)
 156                         continue;
 157
 158                 /*
 159                  * Else check type and setup depth-first scan.
 160                  *
 161                  * Account for bytes actually read.
 162                  */
 163                 info->bytes_scanned += chain->bytes;
 164
 165                 switch(chain->bref.type) {
 166                 case HAMMER2_BREF_TYPE_INODE:
 167                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 168                 case HAMMER2_BREF_TYPE_INDIRECT:
 169                 case HAMMER2_BREF_TYPE_VOLUME:
 170                 case HAMMER2_BREF_TYPE_FREEMAP:
 171                         ++info->depth;
 172                         if (info->depth > 16) {
 173                                 hammer2_chain_save_t *save;
 174                                 save = kmalloc(sizeof(*save), M_HAMMER2,
 175                                                M_WAITOK | M_ZERO);
 176                                 save->chain = chain;
 177                                 hammer2_chain_ref(chain);
 178                                 TAILQ_INSERT_TAIL(&info->list, save, entry);
 179
 180                                 /* guess */
 181                                 info->pri += 10;
 182                         } else {
 183                                 int savepri = info->pri;
 184
 185                                 hammer2_chain_unlock(chain);
 186                                 info->pri = 0;
 187                                 doabort |= hammer2_bulk_scan(chain, func, info);
 188                                 info->pri += savepri;
 189                                 hammer2_chain_lock(chain,
 190                                                    HAMMER2_RESOLVE_ALWAYS |
 191                                                    HAMMER2_RESOLVE_SHARED);
 192                         }
 193                         --info->depth;
 194                         break;
 195                 default:
 196                         /* does not recurse */
 197                         break;
 198                 }
 199         }
 200         if (chain) {
 201                 hammer2_chain_unlock(chain);
 202                 hammer2_chain_drop(chain);
 203         }
 204
 205         /*
 206          * Save with higher pri now that we know what it is.
 207          */
 208         h2_bulkfree_test(info, &parent->bref, info->pri + 1);
 209
 210         hammer2_chain_unlock(parent);
 211
 212         return doabort;
 213 }
 214
 215 /*
 216  * Bulkfree algorithm
 217  *
 218  * Repeat {
 219  *      Chain flush (partial synchronization) XXX removed
 220  *      Scan the whole topology - build in-memory freemap (mark 11)
 221  *      Reconcile the in-memory freemap against the on-disk freemap.
 222  *              ondisk xx -> ondisk 11 (if allocated)
 223  *              ondisk 11 -> ondisk 10 (if free in-memory)
 224  *              ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
 225  * }
 226  *
 227  * The topology scan may have to be performed multiple times to window
 228  * freemaps which are too large to fit in kernel memory.
 229  *
 230  * Races are handled using a double-transition (11->10, 10->00).  The bulkfree
 231  * scan snapshots the volume root's blockset and thus can run concurrent with
 232  * normal operations, as long as a full flush is made between each pass to
 233  * synchronize any modified chains (otherwise their blocks might be improperly
 234  * freed).
 235  *
 236  * Temporary memory in multiples of 64KB is required to reconstruct the leaf
 237  * hammer2_bmap_data blocks so they can later be compared against the live
 238  * freemap.  Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
 239  * A 32MB save area thus represents around ~1 TB.  The temporary memory
 240  * allocated can be specified.  If it is not sufficient multiple topology
 241  * passes will be made.
 242  */
 243
 244 /*
 245  * Bulkfree callback info
 246  */
 247 static void hammer2_bulkfree_thread(void *arg __unused);
 248 static void cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size);
 249 static int h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo,
 250                         hammer2_blockref_t *bref);
 251 static void h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo);
 252 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 253                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 254                         hammer2_bmap_data_t *bmap, int nofree);
 255
 256 void
 257 hammer2_bulkfree_init(hammer2_dev_t *hmp)
 258 {
 259         hammer2_thr_create(&hmp->bfthr, NULL, hmp,
 260                            hmp->devrepname, -1, -1,
 261                            hammer2_bulkfree_thread);
 262 }
 263
 264 void
 265 hammer2_bulkfree_uninit(hammer2_dev_t *hmp)
 266 {
 267         hammer2_thr_delete(&hmp->bfthr);
 268 }
 269
 270 static void
 271 hammer2_bulkfree_thread(void *arg)
 272 {
 273         hammer2_thread_t *thr = arg;
 274         hammer2_ioc_bulkfree_t bfi;
 275         uint32_t flags;
 276
 277         for (;;) {
 278                 hammer2_thr_wait_any(thr,
 279                                      HAMMER2_THREAD_STOP |
 280                                      HAMMER2_THREAD_FREEZE |
 281                                      HAMMER2_THREAD_UNFREEZE |
 282                                      HAMMER2_THREAD_REMASTER,
 283                                      hz * 60);
 284
 285                 flags = thr->flags;
 286                 cpu_ccfence();
 287                 if (flags & HAMMER2_THREAD_STOP)
 288                         break;
 289                 if (flags & HAMMER2_THREAD_FREEZE) {
 290                         hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN,
 291                                                  HAMMER2_THREAD_FREEZE);
 292                         continue;
 293                 }
 294                 if (flags & HAMMER2_THREAD_UNFREEZE) {
 295                         hammer2_thr_signal2(thr, 0,
 296                                                  HAMMER2_THREAD_FROZEN |
 297                                                  HAMMER2_THREAD_UNFREEZE);
 298                         continue;
 299                 }
 300                 if (flags & HAMMER2_THREAD_FROZEN)
 301                         continue;
 302                 if (flags & HAMMER2_THREAD_REMASTER) {
 303                         hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER);
 304                         bzero(&bfi, sizeof(bfi));
 305                         bfi.size = 8192 * 1024;
 306                         hammer2_bulkfree_pass(thr->hmp, &bfi);
 307                 }
 308         }
 309         thr->td = NULL;
 310         hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED);
 311         /* structure can go invalid at this point */
 312 }
 313
 314 int
 315 hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_ioc_bulkfree_t *bfi)
 316 {
 317         hammer2_bulkfree_info_t cbinfo;
 318         hammer2_chain_t *vchain;
 319         hammer2_chain_save_t *save;
 320         hammer2_off_t incr;
 321         size_t size;
 322         int doabort = 0;
 323
 324         /*
 325          * A bulk operations lock is required for the duration.  We
 326          * must hold it across our flushes to guarantee that we never run
 327          * two bulkfree passes in a row without a flush in the middle.
 328          * The 2-stage bulkfree requires this flush to work properly.
 329          */
 330         lockmgr(&hmp->bulklk, LK_EXCLUSIVE);
 331
 332         /*
 333          * We have to clear the live dedup cache as it might have entries
 334          * that are freeable as of now.  Any new entries in the dedup cache
 335          * made after this point, even if they become freeable, will have
 336          * previously been fully allocated and will be protected by the
 337          * 2-stage bulkfree.
 338          */
 339         hammer2_dedup_clear(hmp);
 340
 341         /*
 342          * Create a stable snapshot of the block tree which we can run
 343          * the bulkfree pass on.  This allows the bulkfree pass to run
 344          * concurrent with all other operations (except another bulkfree)
 345          *
 346          * This must flush all dirty chain data, but does not have to
 347          * flush dirty buffer cache buffers which have not yet been
 348          * realized and does not have to flush any newly realized dirty
 349          * chains while the bulkfree pass is running, as long as said
 350          * newly dirtied chains get flushed the next time, before the
 351          * next bulkfree pass.
 352          */
 353         vchain = hammer2_flush_quick(hmp);
 354         hammer2_chain_bulkdrop(vchain);
 355         vchain = hammer2_flush_quick(hmp);
 356
 357         /*
 358          * Setup for free pass
 359          */
 360         bzero(&cbinfo, sizeof(cbinfo));
 361         size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
 362                ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
 363         cbinfo.hmp = hmp;
 364         cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size, VM_SUBSYS_HAMMER);
 365         cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
 366
 367         cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
 368                                M_HAMMER2, M_WAITOK | M_ZERO);
 369
 370         /*
 371          * Normalize start point to a 2GB boundary.  We operate on a
 372          * 64KB leaf bitmap boundary which represents 2GB of storage.
 373          */
 374         cbinfo.sbase = bfi->sbase;
 375         if (cbinfo.sbase > hmp->voldata.volu_size)
 376                 cbinfo.sbase = hmp->voldata.volu_size;
 377         cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 378         TAILQ_INIT(&cbinfo.list);
 379
 380         /*
 381          * Loop on a full meta-data scan as many times as required to
 382          * get through all available storage.
 383          */
 384         while (cbinfo.sbase < hmp->voldata.volu_size) {
 385                 /*
 386                  * We have enough ram to represent (incr) bytes of storage.
 387                  * Each 64KB of ram represents 2GB of storage.
 388                  */
 389                 cbinfo_bmap_init(&cbinfo, size);
 390                 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
 391                        HAMMER2_FREEMAP_LEVEL1_SIZE;
 392                 if (hmp->voldata.volu_size - cbinfo.sbase < incr)
 393                         cbinfo.sstop = hmp->voldata.volu_size;
 394                 else
 395                         cbinfo.sstop = cbinfo.sbase + incr;
 396                 if (hammer2_debug & 1) {
 397                         kprintf("bulkfree pass %016jx/%jdGB\n",
 398                                 (intmax_t)cbinfo.sbase,
 399                                 (intmax_t)incr / HAMMER2_FREEMAP_LEVEL1_SIZE);
 400                 }
 401
 402                 /*
 403                  * Scan topology for stuff inside this range.
 404                  */
 405                 hammer2_trans_init(hmp->spmp, 0);
 406                 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
 407                 cbinfo.pri = 0;
 408                 doabort |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
 409                                              &cbinfo);
 410
 411                 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
 412                        doabort == 0) {
 413                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 414                         cbinfo.pri = 0;
 415                         doabort |= hammer2_bulk_scan(save->chain,
 416                                                      h2_bulkfree_callback,
 417                                                      &cbinfo);
 418                         hammer2_chain_drop(save->chain);
 419                         kfree(save, M_HAMMER2);
 420                 }
 421                 while (save) {
 422                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 423                         hammer2_chain_drop(save->chain);
 424                         kfree(save, M_HAMMER2);
 425                         save = TAILQ_FIRST(&cbinfo.list);
 426                 }
 427
 428                 kprintf("bulkfree lastdrop %d %d doabort=%d\n",
 429                         vchain->refs, vchain->core.chain_count, doabort);
 430
 431                 /*
 432                  * If complete scan succeeded we can synchronize our
 433                  * in-memory freemap against live storage.  If an abort
 434                  * did occur we cannot safely synchronize our partially
 435                  * filled-out in-memory freemap.
 436                  */
 437                 if (doabort == 0) {
 438                         h2_bulkfree_sync(&cbinfo);
 439
 440                         hammer2_voldata_lock(hmp);
 441                         hammer2_voldata_modify(hmp);
 442                         hmp->voldata.allocator_free += cbinfo.adj_free;
 443                         hammer2_voldata_unlock(hmp);
 444                 }
 445
 446                 /*
 447                  * Cleanup for next loop.
 448                  */
 449                 hammer2_trans_done(hmp->spmp);
 450                 if (doabort)
 451                         break;
 452                 cbinfo.sbase = cbinfo.sstop;
 453                 cbinfo.adj_free = 0;
 454         }
 455         hammer2_chain_bulkdrop(vchain);
 456         kmem_free_swapbacked(&cbinfo.kp);
 457         kfree(cbinfo.dedup, M_HAMMER2);
 458         cbinfo.dedup = NULL;
 459
 460         bfi->sstop = cbinfo.sbase;
 461
 462         incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
 463         if (incr > 10000)
 464                 incr = 10000;
 465
 466         kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
 467                 (int)incr / 100,
 468                 (int)incr % 100);
 469
 470         kprintf("    transition->free   %ld\n", cbinfo.count_10_00);
 471         kprintf("    transition->staged %ld\n", cbinfo.count_11_10);
 472         kprintf("    ERR(00)->allocated %ld\n", cbinfo.count_00_11);
 473         kprintf("    ERR(01)->allocated %ld\n", cbinfo.count_01_11);
 474         kprintf("    staged->allocated  %ld\n", cbinfo.count_10_11);
 475         kprintf("    ~2MB segs cleaned  %ld\n", cbinfo.count_l0cleans);
 476         kprintf("    linear adjusts     %ld\n", cbinfo.count_linadjusts);
 477         kprintf("    dedup factor       %ld\n", cbinfo.count_dedup_factor);
 478
 479         lockmgr(&hmp->bulklk, LK_RELEASE);
 480         /* hammer2_vfs_sync(mp, MNT_WAIT); sync needed */
 481
 482         return doabort;
 483 }
 484
 485 static void
 486 cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size)
 487 {
 488         hammer2_bmap_data_t *bmap = cbinfo->bmap;
 489         hammer2_key_t key = cbinfo->sbase;
 490         hammer2_key_t lokey;
 491         hammer2_key_t hikey;
 492
 493         lokey = (cbinfo->hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 494                 ~HAMMER2_SEGMASK64;
 495         hikey = cbinfo->hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 496
 497         bzero(bmap, size);
 498         while (size) {
 499                 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 500                             HAMMER2_ZONE_SEG64) {
 501                         lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 502                                 HAMMER2_ZONE_SEG64;
 503                 }
 504                 if (key < lokey || key >= hikey) {
 505                         memset(bmap->bitmapq, -1,
 506                                sizeof(bmap->bitmapq));
 507                         bmap->avail = 0;
 508                         bmap->linear = HAMMER2_SEGSIZE;
 509                 } else {
 510                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 511                 }
 512                 size -= sizeof(*bmap);
 513                 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
 514                 ++bmap;
 515         }
 516 }
 517
 518 static int
 519 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
 520 {
 521         hammer2_bmap_data_t *bmap;
 522         hammer2_off_t data_off;
 523         uint16_t class;
 524         size_t bytes;
 525         int radix;
 526         int error;
 527
 528         /*
 529          * Check for signal and allow yield to userland during scan
 530          */
 531         if (hammer2_signal_check(&cbinfo->save_time))
 532                 return HAMMER2_BULK_ABORT;
 533         if (bref->type == HAMMER2_BREF_TYPE_INODE) {
 534                 ++cbinfo->count_inodes_scanned;
 535                 if ((cbinfo->count_inodes_scanned & 65535) == 0)
 536                         kprintf(" inodes %6ld bytes %9ld\n",
 537                                 cbinfo->count_inodes_scanned,
 538                                 cbinfo->bytes_scanned);
 539         }
 540
 541         /*
 542          * Calculate the data offset and determine if it is within
 543          * the current freemap range being gathered.
 544          */
 545         error = 0;
 546         data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 547         if (data_off < cbinfo->sbase || data_off >= cbinfo->sstop)
 548                 return 0;
 549         if (data_off < cbinfo->hmp->voldata.allocator_beg)
 550                 return 0;
 551         if (data_off >= cbinfo->hmp->voldata.volu_size)
 552                 return 0;
 553
 554         /*
 555          * Calculate the information needed to generate the in-memory
 556          * freemap record.
 557          *
 558          * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
 559          * it's a problem if it does.  (Or L0 (2MB) for that matter).
 560          */
 561         radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
 562         KKASSERT(radix != 0);
 563         bytes = (size_t)1 << radix;
 564         class = (bref->type << 8) | hammer2_devblkradix(radix);
 565
 566         if (data_off + bytes >= cbinfo->sstop) {
 567                 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
 568                         "%016jx %016jx/%d\n",
 569                         (intmax_t)bref->data_off,
 570                         (intmax_t)bref->key,
 571                         bref->keybits);
 572                 bytes = cbinfo->sstop - data_off;       /* XXX */
 573         }
 574
 575         /*
 576          * Convert to a storage offset relative to the beginning of the
 577          * storage range we are collecting.  Then lookup the level0 bmap entry.
 578          */
 579         data_off -= cbinfo->sbase;
 580         bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
 581
 582         /*
 583          * Convert data_off to a bmap-relative value (~2MB storage range).
 584          * Adjust linear, class, and avail.
 585          *
 586          * Hammer2 does not allow allocations to cross the L0 (2MB) boundary,
 587          */
 588         data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
 589         if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
 590                 kprintf("hammer2_bulkfree_scan: illegal 2MB boundary "
 591                         "%016jx %016jx/%d\n",
 592                         (intmax_t)bref->data_off,
 593                         (intmax_t)bref->key,
 594                         bref->keybits);
 595                 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
 596         }
 597
 598         if (bmap->class == 0) {
 599                 bmap->class = class;
 600                 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 601         }
 602         if (bmap->class != class) {
 603                 kprintf("hammer2_bulkfree_scan: illegal mixed class "
 604                         "%016jx %016jx/%d (%04x vs %04x)\n",
 605                         (intmax_t)bref->data_off,
 606                         (intmax_t)bref->key,
 607                         bref->keybits,
 608                         class, bmap->class);
 609         }
 610         if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
 611                 bmap->linear = (int32_t)data_off + (int32_t)bytes;
 612
 613         /*
 614          * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
 615          * 64-bit entries, 2 bits per entry, to code 11.
 616          *
 617          * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
 618          */
 619         while (bytes > 0) {
 620                 int bindex;
 621                 hammer2_bitmap_t bmask;
 622
 623                 bindex = (int)data_off >> (HAMMER2_FREEMAP_BLOCK_RADIX +
 624                                            HAMMER2_BMAP_INDEX_RADIX);
 625                 bmask = (hammer2_bitmap_t)3 <<
 626                         ((((int)data_off & HAMMER2_BMAP_INDEX_MASK) >>
 627                          HAMMER2_FREEMAP_BLOCK_RADIX) << 1);
 628
 629                 /*
 630                  * NOTE! The (avail) calculation is bitmap-granular.  Multiple
 631                  *       sub-granular records can wind up at the same bitmap
 632                  *       position.
 633                  */
 634                 if ((bmap->bitmapq[bindex] & bmask) == 0) {
 635                         if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE) {
 636                                 bmap->avail -= HAMMER2_FREEMAP_BLOCK_SIZE;
 637                         } else {
 638                                 bmap->avail -= bytes;
 639                         }
 640                         bmap->bitmapq[bindex] |= bmask;
 641                 }
 642                 data_off += HAMMER2_FREEMAP_BLOCK_SIZE;
 643                 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE)
 644                         bytes = 0;
 645                 else
 646                         bytes -= HAMMER2_FREEMAP_BLOCK_SIZE;
 647         }
 648         return error;
 649 }
 650
 651 /*
 652  * Synchronize the in-memory bitmap with the live freemap.  This is not a
 653  * direct copy.  Instead the bitmaps must be compared:
 654  *
 655  *      In-memory       Live-freemap
 656  *         00             11 -> 10      (do nothing if live modified)
 657  *                        10 -> 00      (do nothing if live modified)
 658  *         11             10 -> 11      handles race against live
 659  *                        ** -> 11      nominally warn of corruption
 660  *
 661  */
 662 static void
 663 h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo)
 664 {
 665         hammer2_off_t data_off;
 666         hammer2_key_t key;
 667         hammer2_key_t key_dummy;
 668         hammer2_bmap_data_t *bmap;
 669         hammer2_bmap_data_t *live;
 670         hammer2_chain_t *live_parent;
 671         hammer2_chain_t *live_chain;
 672         int cache_index = -1;
 673         int bmapindex;
 674
 675         kprintf("hammer2_bulkfree - range ");
 676
 677         if (cbinfo->sbase < cbinfo->hmp->voldata.allocator_beg)
 678                 kprintf("%016jx-",
 679                         (intmax_t)cbinfo->hmp->voldata.allocator_beg);
 680         else
 681                 kprintf("%016jx-",
 682                         (intmax_t)cbinfo->sbase);
 683
 684         if (cbinfo->sstop > cbinfo->hmp->voldata.volu_size)
 685                 kprintf("%016jx\n",
 686                         (intmax_t)cbinfo->hmp->voldata.volu_size);
 687         else
 688                 kprintf("%016jx\n",
 689                         (intmax_t)cbinfo->sstop);
 690
 691         data_off = cbinfo->sbase;
 692         bmap = cbinfo->bmap;
 693
 694         live_parent = &cbinfo->hmp->fchain;
 695         hammer2_chain_ref(live_parent);
 696         hammer2_chain_lock(live_parent, HAMMER2_RESOLVE_ALWAYS);
 697         live_chain = NULL;
 698
 699         /*
 700          * Iterate each hammer2_bmap_data_t line (128 bytes) managing
 701          * 4MB of storage.
 702          */
 703         while (data_off < cbinfo->sstop) {
 704                 /*
 705                  * The freemap is not used below allocator_beg or beyond
 706                  * volu_size.
 707                  */
 708                 int nofree;
 709
 710                 if (data_off < cbinfo->hmp->voldata.allocator_beg)
 711                         goto next;
 712                 if (data_off >= cbinfo->hmp->voldata.volu_size)
 713                         goto next;
 714
 715                 /*
 716                  * Locate the freemap leaf on the live filesystem
 717                  */
 718                 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
 719                 nofree = 0;
 720
 721                 if (live_chain == NULL || live_chain->bref.key != key) {
 722                         if (live_chain) {
 723                                 hammer2_chain_unlock(live_chain);
 724                                 hammer2_chain_drop(live_chain);
 725                         }
 726                         live_chain = hammer2_chain_lookup(
 727                                             &live_parent,
 728                                             &key_dummy,
 729                                             key,
 730                                             key + HAMMER2_FREEMAP_LEVEL1_MASK,
 731                                             &cache_index,
 732                                             HAMMER2_LOOKUP_ALWAYS);
 733
 734 #if 0
 735                         /*
 736                          * If recent allocations were made we avoid races by
 737                          * not staging or freeing any blocks.  We can still
 738                          * remark blocks as fully allocated.
 739                          */
 740                         if (live_chain) {
 741                                 if (hammer2_debug & 1) {
 742                                         kprintf("live_chain %016jx\n",
 743                                                 (intmax_t)key);
 744                                 }
 745                                 if (live_chain->bref.mirror_tid >
 746                                     cbinfo->saved_mirror_tid) {
 747                                         kprintf("hammer2_bulkfree: "
 748                                                 "avoid %016jx\n",
 749                                                 data_off);
 750                                         nofree = 1;
 751                                 } else {
 752                                         nofree = 0;
 753                                 }
 754                         }
 755 #endif
 756                 }
 757                 if (live_chain == NULL) {
 758                         /*
 759                          * XXX if we implement a full recovery mode we need
 760                          * to create/recreate missing freemap chains if our
 761                          * bmap has any allocated blocks.
 762                          */
 763                         if (bmap->class &&
 764                             bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
 765                                 kprintf("hammer2_bulkfree: cannot locate "
 766                                         "live leaf for allocated data "
 767                                         "near %016jx\n",
 768                                         (intmax_t)data_off);
 769                         }
 770                         goto next;
 771                 }
 772                 if (live_chain->error) {
 773                         kprintf("hammer2_bulkfree: error %s looking up "
 774                                 "live leaf for allocated data near %016jx\n",
 775                                 hammer2_error_str(live_chain->error),
 776                                 (intmax_t)data_off);
 777                         hammer2_chain_unlock(live_chain);
 778                         hammer2_chain_drop(live_chain);
 779                         live_chain = NULL;
 780                         goto next;
 781                 }
 782
 783                 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
 784                             HAMMER2_FREEMAP_LEVEL0_RADIX;
 785                 live = &live_chain->data->bmdata[bmapindex];
 786
 787                 /*
 788                  * TODO - we could shortcut this by testing that both
 789                  * live->class and bmap->class are 0, and both avails are
 790                  * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
 791                  */
 792                 if (bcmp(live->bitmapq, bmap->bitmapq,
 793                          sizeof(bmap->bitmapq)) == 0) {
 794                         goto next;
 795                 }
 796                 if (hammer2_debug & 1) {
 797                         kprintf("live %016jx %04d.%04x (avail=%d)\n",
 798                                 data_off, bmapindex, live->class, live->avail);
 799                 }
 800
 801                 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
 802                 live = &live_chain->data->bmdata[bmapindex];
 803
 804                 h2_bulkfree_sync_adjust(cbinfo, data_off, live, bmap, nofree);
 805 next:
 806                 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
 807                 ++bmap;
 808         }
 809         if (live_chain) {
 810                 hammer2_chain_unlock(live_chain);
 811                 hammer2_chain_drop(live_chain);
 812         }
 813         if (live_parent) {
 814                 hammer2_chain_unlock(live_parent);
 815                 hammer2_chain_drop(live_parent);
 816         }
 817 }
 818
 819 /*
 820  * When bulkfree is finally able to free a block it must make sure that
 821  * the INVALOK bit in any cached DIO is cleared prior to the block being
 822  * reused.
 823  */
 824 static
 825 void
 826 fixup_dio(hammer2_dev_t *hmp, hammer2_off_t data_off, int bindex, int scount)
 827 {
 828         data_off += (scount >> 1) * HAMMER2_FREEMAP_BLOCK_SIZE;
 829         data_off += bindex *
 830                 (HAMMER2_FREEMAP_BLOCK_SIZE * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 831         hammer2_io_resetinval(hmp, data_off);
 832 }
 833
 834 /*
 835  * Merge the bulkfree bitmap against the existing bitmap.
 836  *
 837  * If nofree is non-zero the merge will only mark free blocks as allocated
 838  * and will refuse to free any blocks.
 839  */
 840 static
 841 void
 842 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 843                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 844                         hammer2_bmap_data_t *bmap, int nofree)
 845 {
 846         int bindex;
 847         int scount;
 848         hammer2_bitmap_t lmask;
 849         hammer2_bitmap_t mmask;
 850
 851         for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
 852                 lmask = live->bitmapq[bindex];  /* live */
 853                 mmask = bmap->bitmapq[bindex];  /* snapshotted bulkfree */
 854                 if (lmask == mmask)
 855                         continue;
 856
 857                 for (scount = 0;
 858                      scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
 859                      scount += 2) {
 860                         if ((mmask & 3) == 0) {
 861                                 /*
 862                                  * in-memory 00         live 11 -> 10
 863                                  *                      live 10 -> 00
 864                                  *
 865                                  * Storage might be marked allocated or
 866                                  * staged and must be remarked staged or
 867                                  * free.
 868                                  */
 869                                 switch (lmask & 3) {
 870                                 case 0: /* 00 */
 871                                         break;
 872                                 case 1: /* 01 */
 873                                         kprintf("hammer2_bulkfree: cannot "
 874                                                 "transition m=00/l=01\n");
 875                                         break;
 876                                 case 2: /* 10 -> 00 */
 877                                         if (nofree)
 878                                                 break;
 879                                         live->bitmapq[bindex] &=
 880                                             ~((hammer2_bitmap_t)2 << scount);
 881                                         live->avail +=
 882                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 883                                         if (live->avail >
 884                                             HAMMER2_FREEMAP_LEVEL0_SIZE) {
 885                                                 live->avail =
 886                                                     HAMMER2_FREEMAP_LEVEL0_SIZE;
 887                                         }
 888                                         cbinfo->adj_free +=
 889                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 890                                         ++cbinfo->count_10_00;
 891                                         break;
 892                                 case 3: /* 11 -> 10 */
 893                                         live->bitmapq[bindex] &=
 894                                             ~((hammer2_bitmap_t)1 << scount);
 895                                         ++cbinfo->count_11_10;
 896                                         fixup_dio(cbinfo->hmp, data_off,
 897                                                   bindex, scount);
 898                                         break;
 899                                 }
 900                         } else if ((mmask & 3) == 3) {
 901                                 /*
 902                                  * in-memory 11         live 10 -> 11
 903                                  *                      live ** -> 11
 904                                  *
 905                                  * Storage might be incorrectly marked free
 906                                  * or staged and must be remarked fully
 907                                  * allocated.
 908                                  */
 909                                 switch (lmask & 3) {
 910                                 case 0: /* 00 */
 911                                         ++cbinfo->count_00_11;
 912                                         cbinfo->adj_free -=
 913                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 914                                         live->avail -=
 915                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 916                                         if ((int32_t)live->avail < 0)
 917                                                 live->avail = 0;
 918                                         break;
 919                                 case 1: /* 01 */
 920                                         ++cbinfo->count_01_11;
 921                                         break;
 922                                 case 2: /* 10 -> 11 */
 923                                         ++cbinfo->count_10_11;
 924                                         break;
 925                                 case 3: /* 11 */
 926                                         break;
 927                                 }
 928                                 live->bitmapq[bindex] |=
 929                                         ((hammer2_bitmap_t)3 << scount);
 930                         }
 931                         mmask >>= 2;
 932                         lmask >>= 2;
 933                 }
 934         }
 935
 936         /*
 937          * Determine if the live bitmap is completely free and reset its
 938          * fields if so.  Otherwise check to see if we can reduce the linear
 939          * offset.
 940          */
 941         for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
 942                 if (live->bitmapq[bindex] != 0)
 943                         break;
 944         }
 945         if (nofree) {
 946                 /* do nothing */
 947         } else if (bindex < 0) {
 948                 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 949                 live->class = 0;
 950                 live->linear = 0;
 951                 ++cbinfo->count_l0cleans;
 952         } else if (bindex < 7) {
 953                 ++bindex;
 954                 if (live->linear > bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 955                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 956                         ++cbinfo->count_linadjusts;
 957                 }
 958
 959                 /*
 960                  * XXX this fine-grained measure still has some issues.
 961                  */
 962                 if (live->linear < bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 963                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 964                         ++cbinfo->count_linadjusts;
 965                 }
 966         } else {
 967                 live->linear = HAMMER2_SEGSIZE;
 968         }
 969
 970 #if 0
 971         if (bmap->class) {
 972                 kprintf("%016jx %04d.%04x (avail=%7d) "
 973                         "%08x %08x %08x %08x %08x %08x %08x %08x\n",
 974                         (intmax_t)data_off,
 975                         (int)((data_off &
 976                                HAMMER2_FREEMAP_LEVEL1_MASK) >>
 977                               HAMMER2_FREEMAP_LEVEL0_RADIX),
 978                         bmap->class,
 979                         bmap->avail,
 980                         bmap->bitmap[0], bmap->bitmap[1],
 981                         bmap->bitmap[2], bmap->bitmap[3],
 982                         bmap->bitmap[4], bmap->bitmap[5],
 983                         bmap->bitmap[6], bmap->bitmap[7]);
 984         }
 985 #endif
 986 }
 987
 988 /*
 989  * BULKFREE DEDUP HEURISTIC
 990  *
 991  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
 992  *          All fields must be loaded into locals and validated.
 993  */
 994 static
 995 int
 996 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
 997                  int pri)
 998 {
 999         hammer2_dedup_t *dedup;
1000         int best;
1001         int n;
1002         int i;
1003
1004         n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
1005         dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
1006
1007         for (i = best = 0; i < 8; ++i) {
1008                 if (dedup[i].data_off == bref->data_off) {
1009                         if (dedup[i].ticks < pri)
1010                                 dedup[i].ticks = pri;
1011                         if (pri == 1)
1012                                 cbinfo->count_dedup_factor += dedup[i].ticks;
1013                         return 1;
1014                 }
1015                 if (dedup[i].ticks < dedup[best].ticks)
1016                         best = i;
1017         }
1018         dedup[best].data_off = bref->data_off;
1019         dedup[best].ticks = pri;
1020
1021         return 0;
1022 }