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 %p/%p %016jx.%02x\n",
 136                         parent, chain, bref.data_off, bref.type);
 137                 int xerr = tsleep(&info->pri, PCATCH, "slp", hz / 10);
 138                 if (xerr == EINTR || xerr == ERESTART) {
 139                         doabort |= HAMMER2_BULK_ABORT;
 140                 }
 141 #endif
 142                 ++info->pri;
 143                 if (h2_bulkfree_test(info, &bref, 1))
 144                         continue;
 145
 146                 doabort |= func(info, &bref);
 147
 148                 if (doabort & HAMMER2_BULK_ABORT)
 149                         break;
 150
 151                 /*
 152                  * A non-null chain is always returned if it is
 153                  * recursive, otherwise a non-null chain might be
 154                  * returned but usually is not when not recursive.
 155                  */
 156                 if (chain == NULL)
 157                         continue;
 158
 159                 /*
 160                  * Else check type and setup depth-first scan.
 161                  *
 162                  * Account for bytes actually read.
 163                  */
 164                 info->bytes_scanned += chain->bytes;
 165
 166                 switch(chain->bref.type) {
 167                 case HAMMER2_BREF_TYPE_INODE:
 168                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 169                 case HAMMER2_BREF_TYPE_INDIRECT:
 170                 case HAMMER2_BREF_TYPE_VOLUME:
 171                 case HAMMER2_BREF_TYPE_FREEMAP:
 172                         ++info->depth;
 173                         if (info->depth > 16) {
 174                                 hammer2_chain_save_t *save;
 175                                 save = kmalloc(sizeof(*save), M_HAMMER2,
 176                                                M_WAITOK | M_ZERO);
 177                                 save->chain = chain;
 178                                 hammer2_chain_ref(chain);
 179                                 TAILQ_INSERT_TAIL(&info->list, save, entry);
 180
 181                                 /* guess */
 182                                 info->pri += 10;
 183                         } else {
 184                                 int savepri = info->pri;
 185
 186                                 hammer2_chain_unlock(chain);
 187                                 info->pri = 0;
 188                                 doabort |= hammer2_bulk_scan(chain, func, info);
 189                                 info->pri += savepri;
 190                                 hammer2_chain_lock(chain,
 191                                                    HAMMER2_RESOLVE_ALWAYS |
 192                                                    HAMMER2_RESOLVE_SHARED);
 193                         }
 194                         --info->depth;
 195                         break;
 196                 default:
 197                         /* does not recurse */
 198                         break;
 199                 }
 200         }
 201         if (chain) {
 202                 hammer2_chain_unlock(chain);
 203                 hammer2_chain_drop(chain);
 204         }
 205
 206         /*
 207          * Save with higher pri now that we know what it is.
 208          */
 209         h2_bulkfree_test(info, &parent->bref, info->pri + 1);
 210
 211         hammer2_chain_unlock(parent);
 212
 213         return doabort;
 214 }
 215
 216 /*
 217  * Bulkfree algorithm
 218  *
 219  * Repeat {
 220  *      Chain flush (partial synchronization) XXX removed
 221  *      Scan the whole topology - build in-memory freemap (mark 11)
 222  *      Reconcile the in-memory freemap against the on-disk freemap.
 223  *              ondisk xx -> ondisk 11 (if allocated)
 224  *              ondisk 11 -> ondisk 10 (if free in-memory)
 225  *              ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
 226  * }
 227  *
 228  * The topology scan may have to be performed multiple times to window
 229  * freemaps which are too large to fit in kernel memory.
 230  *
 231  * Races are handled using a double-transition (11->10, 10->00).  The bulkfree
 232  * scan snapshots the volume root's blockset and thus can run concurrent with
 233  * normal operations, as long as a full flush is made between each pass to
 234  * synchronize any modified chains (otherwise their blocks might be improperly
 235  * freed).
 236  *
 237  * Temporary memory in multiples of 64KB is required to reconstruct the leaf
 238  * hammer2_bmap_data blocks so they can later be compared against the live
 239  * freemap.  Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
 240  * A 32MB save area thus represents around ~1 TB.  The temporary memory
 241  * allocated can be specified.  If it is not sufficient multiple topology
 242  * passes will be made.
 243  */
 244
 245 /*
 246  * Bulkfree callback info
 247  */
 248 static void hammer2_bulkfree_thread(void *arg __unused);
 249 static void cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size);
 250 static int h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo,
 251                         hammer2_blockref_t *bref);
 252 static void h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo);
 253 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 254                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 255                         hammer2_bmap_data_t *bmap, hammer2_key_t alloc_base);
 256
 257 void
 258 hammer2_bulkfree_init(hammer2_dev_t *hmp)
 259 {
 260         hammer2_thr_create(&hmp->bfthr, NULL, hmp,
 261                            hmp->devrepname, -1, -1,
 262                            hammer2_bulkfree_thread);
 263 }
 264
 265 void
 266 hammer2_bulkfree_uninit(hammer2_dev_t *hmp)
 267 {
 268         hammer2_thr_delete(&hmp->bfthr);
 269 }
 270
 271 static void
 272 hammer2_bulkfree_thread(void *arg)
 273 {
 274         hammer2_thread_t *thr = arg;
 275         hammer2_ioc_bulkfree_t bfi;
 276         uint32_t flags;
 277
 278         for (;;) {
 279                 hammer2_thr_wait_any(thr,
 280                                      HAMMER2_THREAD_STOP |
 281                                      HAMMER2_THREAD_FREEZE |
 282                                      HAMMER2_THREAD_UNFREEZE |
 283                                      HAMMER2_THREAD_REMASTER,
 284                                      hz * 60);
 285
 286                 flags = thr->flags;
 287                 cpu_ccfence();
 288                 if (flags & HAMMER2_THREAD_STOP)
 289                         break;
 290                 if (flags & HAMMER2_THREAD_FREEZE) {
 291                         hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN,
 292                                                  HAMMER2_THREAD_FREEZE);
 293                         continue;
 294                 }
 295                 if (flags & HAMMER2_THREAD_UNFREEZE) {
 296                         hammer2_thr_signal2(thr, 0,
 297                                                  HAMMER2_THREAD_FROZEN |
 298                                                  HAMMER2_THREAD_UNFREEZE);
 299                         continue;
 300                 }
 301                 if (flags & HAMMER2_THREAD_FROZEN)
 302                         continue;
 303                 if (flags & HAMMER2_THREAD_REMASTER) {
 304                         hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER);
 305                         bzero(&bfi, sizeof(bfi));
 306                         bfi.size = 8192 * 1024;
 307                         /* hammer2_bulkfree_pass(thr->hmp, &bfi); */
 308                 }
 309         }
 310         thr->td = NULL;
 311         hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED);
 312         /* structure can go invalid at this point */
 313 }
 314
 315 int
 316 hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_chain_t *vchain,
 317                       hammer2_ioc_bulkfree_t *bfi)
 318 {
 319         hammer2_bulkfree_info_t cbinfo;
 320         hammer2_chain_save_t *save;
 321         hammer2_off_t incr;
 322         size_t size;
 323         int doabort = 0;
 324
 325         /*
 326          * We have to clear the live dedup cache as it might have entries
 327          * that are freeable as of now.  Any new entries in the dedup cache
 328          * made after this point, even if they become freeable, will have
 329          * previously been fully allocated and will be protected by the
 330          * 2-stage bulkfree.
 331          */
 332         hammer2_dedup_clear(hmp);
 333
 334         /*
 335          * Setup for free pass
 336          */
 337         bzero(&cbinfo, sizeof(cbinfo));
 338         size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
 339                ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
 340         cbinfo.hmp = hmp;
 341         cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size, VM_SUBSYS_HAMMER);
 342         cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
 343
 344         cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
 345                                M_HAMMER2, M_WAITOK | M_ZERO);
 346
 347         /*
 348          * Normalize start point to a 2GB boundary.  We operate on a
 349          * 64KB leaf bitmap boundary which represents 2GB of storage.
 350          */
 351         cbinfo.sbase = bfi->sbase;
 352         if (cbinfo.sbase > hmp->voldata.volu_size)
 353                 cbinfo.sbase = hmp->voldata.volu_size;
 354         cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 355         TAILQ_INIT(&cbinfo.list);
 356
 357         /*
 358          * Loop on a full meta-data scan as many times as required to
 359          * get through all available storage.
 360          */
 361         while (cbinfo.sbase < hmp->voldata.volu_size) {
 362                 /*
 363                  * We have enough ram to represent (incr) bytes of storage.
 364                  * Each 64KB of ram represents 2GB of storage.
 365                  *
 366                  * We must also clean out our de-duplication heuristic for
 367                  * each (incr) bytes of storage, otherwise we wind up not
 368                  * scanning meta-data for later areas of storage because
 369                  * they had already been scanned in earlier areas of storage.
 370                  * Since the ranging is different, we have to restart
 371                  * the dedup heuristic too.
 372                  */
 373                 cbinfo_bmap_init(&cbinfo, size);
 374                 bzero(cbinfo.dedup, sizeof(*cbinfo.dedup) *
 375                                     HAMMER2_DEDUP_HEUR_SIZE);
 376                 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
 377                        HAMMER2_FREEMAP_LEVEL1_SIZE;
 378                 if (hmp->voldata.volu_size - cbinfo.sbase < incr)
 379                         cbinfo.sstop = hmp->voldata.volu_size;
 380                 else
 381                         cbinfo.sstop = cbinfo.sbase + incr;
 382                 if (hammer2_debug & 1) {
 383                         kprintf("bulkfree pass %016jx/%jdGB\n",
 384                                 (intmax_t)cbinfo.sbase,
 385                                 (intmax_t)incr / HAMMER2_FREEMAP_LEVEL1_SIZE);
 386                 }
 387
 388                 /*
 389                  * Scan topology for stuff inside this range.
 390                  */
 391                 hammer2_trans_init(hmp->spmp, 0);
 392                 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
 393                 cbinfo.pri = 0;
 394                 doabort |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
 395                                              &cbinfo);
 396
 397                 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
 398                        doabort == 0) {
 399                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 400                         cbinfo.pri = 0;
 401                         doabort |= hammer2_bulk_scan(save->chain,
 402                                                      h2_bulkfree_callback,
 403                                                      &cbinfo);
 404                         hammer2_chain_drop(save->chain);
 405                         kfree(save, M_HAMMER2);
 406                 }
 407                 while (save) {
 408                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 409                         hammer2_chain_drop(save->chain);
 410                         kfree(save, M_HAMMER2);
 411                         save = TAILQ_FIRST(&cbinfo.list);
 412                 }
 413
 414                 kprintf("bulkfree lastdrop %d %d doabort=%d\n",
 415                         vchain->refs, vchain->core.chain_count, doabort);
 416
 417                 /*
 418                  * If complete scan succeeded we can synchronize our
 419                  * in-memory freemap against live storage.  If an abort
 420                  * did occur we cannot safely synchronize our partially
 421                  * filled-out in-memory freemap.
 422                  */
 423                 if (doabort == 0) {
 424                         h2_bulkfree_sync(&cbinfo);
 425
 426                         hammer2_voldata_lock(hmp);
 427                         hammer2_voldata_modify(hmp);
 428                         hmp->voldata.allocator_free += cbinfo.adj_free;
 429                         hammer2_voldata_unlock(hmp);
 430                 }
 431
 432                 /*
 433                  * Cleanup for next loop.
 434                  */
 435                 hammer2_trans_done(hmp->spmp);
 436                 if (doabort)
 437                         break;
 438                 cbinfo.sbase = cbinfo.sstop;
 439                 cbinfo.adj_free = 0;
 440         }
 441         kmem_free_swapbacked(&cbinfo.kp);
 442         kfree(cbinfo.dedup, M_HAMMER2);
 443         cbinfo.dedup = NULL;
 444
 445         bfi->sstop = cbinfo.sbase;
 446
 447         incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
 448         if (incr > 10000)
 449                 incr = 10000;
 450
 451         kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
 452                 (int)incr / 100,
 453                 (int)incr % 100);
 454
 455         kprintf("    transition->free   %ld\n", cbinfo.count_10_00);
 456         kprintf("    transition->staged %ld\n", cbinfo.count_11_10);
 457         kprintf("    ERR(00)->allocated %ld\n", cbinfo.count_00_11);
 458         kprintf("    ERR(01)->allocated %ld\n", cbinfo.count_01_11);
 459         kprintf("    staged->allocated  %ld\n", cbinfo.count_10_11);
 460         kprintf("    ~2MB segs cleaned  %ld\n", cbinfo.count_l0cleans);
 461         kprintf("    linear adjusts     %ld\n", cbinfo.count_linadjusts);
 462         kprintf("    dedup factor       %ld\n", cbinfo.count_dedup_factor);
 463
 464         return doabort;
 465 }
 466
 467 static void
 468 cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size)
 469 {
 470         hammer2_bmap_data_t *bmap = cbinfo->bmap;
 471         hammer2_key_t key = cbinfo->sbase;
 472         hammer2_key_t lokey;
 473         hammer2_key_t hikey;
 474
 475         lokey = (cbinfo->hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 476                 ~HAMMER2_SEGMASK64;
 477         hikey = cbinfo->hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 478
 479         bzero(bmap, size);
 480         while (size) {
 481                 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 482                             HAMMER2_ZONE_SEG64) {
 483                         lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 484                                 HAMMER2_ZONE_SEG64;
 485                 }
 486                 if (key < lokey || key >= hikey) {
 487                         memset(bmap->bitmapq, -1,
 488                                sizeof(bmap->bitmapq));
 489                         bmap->avail = 0;
 490                         bmap->linear = HAMMER2_SEGSIZE;
 491                 } else {
 492                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 493                 }
 494                 size -= sizeof(*bmap);
 495                 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
 496                 ++bmap;
 497         }
 498 }
 499
 500 static int
 501 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
 502 {
 503         hammer2_bmap_data_t *bmap;
 504         hammer2_off_t data_off;
 505         uint16_t class;
 506         size_t bytes;
 507         int radix;
 508         int error;
 509
 510         /*
 511          * Check for signal and allow yield to userland during scan
 512          */
 513         if (hammer2_signal_check(&cbinfo->save_time))
 514                 return HAMMER2_BULK_ABORT;
 515         if (bref->type == HAMMER2_BREF_TYPE_INODE) {
 516                 ++cbinfo->count_inodes_scanned;
 517                 if ((cbinfo->count_inodes_scanned & 65535) == 0)
 518                         kprintf(" inodes %6ld bytes %9ld\n",
 519                                 cbinfo->count_inodes_scanned,
 520                                 cbinfo->bytes_scanned);
 521         }
 522
 523         /*
 524          * Calculate the data offset and determine if it is within
 525          * the current freemap range being gathered.
 526          */
 527         error = 0;
 528         data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 529         if (data_off < cbinfo->sbase || data_off >= cbinfo->sstop)
 530                 return 0;
 531         if (data_off < cbinfo->hmp->voldata.allocator_beg)
 532                 return 0;
 533         if (data_off >= cbinfo->hmp->voldata.volu_size)
 534                 return 0;
 535
 536         /*
 537          * Calculate the information needed to generate the in-memory
 538          * freemap record.
 539          *
 540          * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
 541          * it's a problem if it does.  (Or L0 (2MB) for that matter).
 542          */
 543         radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
 544         KKASSERT(radix != 0);
 545         bytes = (size_t)1 << radix;
 546         class = (bref->type << 8) | hammer2_devblkradix(radix);
 547
 548         if (data_off + bytes >= cbinfo->sstop) {
 549                 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
 550                         "%016jx %016jx/%d\n",
 551                         (intmax_t)bref->data_off,
 552                         (intmax_t)bref->key,
 553                         bref->keybits);
 554                 bytes = cbinfo->sstop - data_off;       /* XXX */
 555         }
 556
 557         /*
 558          * Convert to a storage offset relative to the beginning of the
 559          * storage range we are collecting.  Then lookup the level0 bmap entry.
 560          */
 561         data_off -= cbinfo->sbase;
 562         bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
 563
 564         /*
 565          * Convert data_off to a bmap-relative value (~4MB storage range).
 566          * Adjust linear, class, and avail.
 567          *
 568          * Hammer2 does not allow allocations to cross the L0 (4MB) boundary,
 569          */
 570         data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
 571         if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
 572                 kprintf("hammer2_bulkfree_scan: illegal 4MB boundary "
 573                         "%016jx %016jx/%d\n",
 574                         (intmax_t)bref->data_off,
 575                         (intmax_t)bref->key,
 576                         bref->keybits);
 577                 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
 578         }
 579
 580         if (bmap->class == 0) {
 581                 bmap->class = class;
 582                 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 583         }
 584         if (bmap->class != class) {
 585                 kprintf("hammer2_bulkfree_scan: illegal mixed class "
 586                         "%016jx %016jx/%d (%04x vs %04x)\n",
 587                         (intmax_t)bref->data_off,
 588                         (intmax_t)bref->key,
 589                         bref->keybits,
 590                         class, bmap->class);
 591         }
 592
 593         /*
 594          * Just record the highest byte-granular offset for now.  Do not
 595          * match against allocations which are in multiples of whole blocks.
 596          *
 597          * Make sure that any in-block linear offset at least covers the
 598          * data range.  This can cause bmap->linear to become block-aligned.
 599          */
 600         if (bytes & HAMMER2_FREEMAP_BLOCK_MASK) {
 601                 if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
 602                         bmap->linear = (int32_t)data_off + (int32_t)bytes;
 603         } else if (bmap->linear >= (int32_t)data_off &&
 604                    bmap->linear < (int32_t)data_off + (int32_t)bytes) {
 605                 bmap->linear = (int32_t)data_off + (int32_t)bytes;
 606         }
 607
 608         /*
 609          * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
 610          * 64-bit entries, 2 bits per entry, to code 11.
 611          *
 612          * NOTE: data_off mask to 524288, shift right by 14 (radix for 16384),
 613          *       and multiply shift amount by 2 for sets of 2 bits.
 614          *
 615          * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
 616          *       also, data_off may not be FREEMAP_BLOCK_SIZE aligned.
 617          */
 618         while (bytes > 0) {
 619                 hammer2_bitmap_t bmask;
 620                 int bindex;
 621
 622                 bindex = (int)data_off >> (HAMMER2_FREEMAP_BLOCK_RADIX +
 623                                            HAMMER2_BMAP_INDEX_RADIX);
 624                 bmask = (hammer2_bitmap_t)3 <<
 625                         ((((int)data_off & HAMMER2_BMAP_INDEX_MASK) >>
 626                          HAMMER2_FREEMAP_BLOCK_RADIX) << 1);
 627
 628                 /*
 629                  * NOTE! The (avail) calculation is bitmap-granular.  Multiple
 630                  *       sub-granular records can wind up at the same bitmap
 631                  *       position.
 632                  */
 633                 if ((bmap->bitmapq[bindex] & bmask) == 0) {
 634                         if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE) {
 635                                 bmap->avail -= HAMMER2_FREEMAP_BLOCK_SIZE;
 636                         } else {
 637                                 bmap->avail -= bytes;
 638                         }
 639                         bmap->bitmapq[bindex] |= bmask;
 640                 }
 641                 data_off += HAMMER2_FREEMAP_BLOCK_SIZE;
 642                 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE)
 643                         bytes = 0;
 644                 else
 645                         bytes -= HAMMER2_FREEMAP_BLOCK_SIZE;
 646         }
 647         return error;
 648 }
 649
 650 /*
 651  * Synchronize the in-memory bitmap with the live freemap.  This is not a
 652  * direct copy.  Instead the bitmaps must be compared:
 653  *
 654  *      In-memory       Live-freemap
 655  *         00             11 -> 10      (do nothing if live modified)
 656  *                        10 -> 00      (do nothing if live modified)
 657  *         11             10 -> 11      handles race against live
 658  *                        ** -> 11      nominally warn of corruption
 659  *
 660  */
 661 static void
 662 h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo)
 663 {
 664         hammer2_off_t data_off;
 665         hammer2_key_t key;
 666         hammer2_key_t key_dummy;
 667         hammer2_bmap_data_t *bmap;
 668         hammer2_bmap_data_t *live;
 669         hammer2_chain_t *live_parent;
 670         hammer2_chain_t *live_chain;
 671         int cache_index = -1;
 672         int bmapindex;
 673
 674         kprintf("hammer2_bulkfree - range ");
 675
 676         if (cbinfo->sbase < cbinfo->hmp->voldata.allocator_beg)
 677                 kprintf("%016jx-",
 678                         (intmax_t)cbinfo->hmp->voldata.allocator_beg);
 679         else
 680                 kprintf("%016jx-",
 681                         (intmax_t)cbinfo->sbase);
 682
 683         if (cbinfo->sstop > cbinfo->hmp->voldata.volu_size)
 684                 kprintf("%016jx\n",
 685                         (intmax_t)cbinfo->hmp->voldata.volu_size);
 686         else
 687                 kprintf("%016jx\n",
 688                         (intmax_t)cbinfo->sstop);
 689
 690         data_off = cbinfo->sbase;
 691         bmap = cbinfo->bmap;
 692
 693         live_parent = &cbinfo->hmp->fchain;
 694         hammer2_chain_ref(live_parent);
 695         hammer2_chain_lock(live_parent, HAMMER2_RESOLVE_ALWAYS);
 696         live_chain = NULL;
 697
 698         /*
 699          * Iterate each hammer2_bmap_data_t line (128 bytes) managing
 700          * 4MB of storage.
 701          */
 702         while (data_off < cbinfo->sstop) {
 703                 /*
 704                  * The freemap is not used below allocator_beg or beyond
 705                  * volu_size.
 706                  */
 707
 708                 if (data_off < cbinfo->hmp->voldata.allocator_beg)
 709                         goto next;
 710                 if (data_off >= cbinfo->hmp->voldata.volu_size)
 711                         goto next;
 712
 713                 /*
 714                  * Locate the freemap leaf on the live filesystem
 715                  */
 716                 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
 717
 718                 if (live_chain == NULL || live_chain->bref.key != key) {
 719                         if (live_chain) {
 720                                 hammer2_chain_unlock(live_chain);
 721                                 hammer2_chain_drop(live_chain);
 722                         }
 723                         live_chain = hammer2_chain_lookup(
 724                                             &live_parent,
 725                                             &key_dummy,
 726                                             key,
 727                                             key + HAMMER2_FREEMAP_LEVEL1_MASK,
 728                                             &cache_index,
 729                                             HAMMER2_LOOKUP_ALWAYS);
 730
 731 #if 0
 732                         /*
 733                          * If recent allocations were made we avoid races by
 734                          * not staging or freeing any blocks.  We can still
 735                          * remark blocks as fully allocated.
 736                          */
 737                         if (live_chain) {
 738                                 if (hammer2_debug & 1) {
 739                                         kprintf("live_chain %016jx\n",
 740                                                 (intmax_t)key);
 741                                 }
 742                                 if (live_chain->bref.mirror_tid >
 743                                     cbinfo->saved_mirror_tid) {
 744                                         kprintf("hammer2_bulkfree: "
 745                                                 "avoid %016jx\n",
 746                                                 data_off);
 747                                         nofree = 1;
 748                                 } else {
 749                                         nofree = 0;
 750                                 }
 751                         }
 752 #endif
 753                 }
 754                 if (live_chain == NULL) {
 755                         /*
 756                          * XXX if we implement a full recovery mode we need
 757                          * to create/recreate missing freemap chains if our
 758                          * bmap has any allocated blocks.
 759                          */
 760                         if (bmap->class &&
 761                             bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
 762                                 kprintf("hammer2_bulkfree: cannot locate "
 763                                         "live leaf for allocated data "
 764                                         "near %016jx\n",
 765                                         (intmax_t)data_off);
 766                         }
 767                         goto next;
 768                 }
 769                 if (live_chain->error) {
 770                         kprintf("hammer2_bulkfree: error %s looking up "
 771                                 "live leaf for allocated data near %016jx\n",
 772                                 hammer2_error_str(live_chain->error),
 773                                 (intmax_t)data_off);
 774                         hammer2_chain_unlock(live_chain);
 775                         hammer2_chain_drop(live_chain);
 776                         live_chain = NULL;
 777                         goto next;
 778                 }
 779
 780                 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
 781                             HAMMER2_FREEMAP_LEVEL0_RADIX;
 782                 live = &live_chain->data->bmdata[bmapindex];
 783
 784                 /*
 785                  * Shortcut if the bitmaps match and the live linear
 786                  * indicator is sane.  We can't do a perfect check of
 787                  * live->linear because the only real requirement is that
 788                  * if it is not block-aligned, that it not cover the space
 789                  * within its current block which overlaps one of the data
 790                  * ranges we scan.  We don't retain enough fine-grained
 791                  * data in our scan to be able to set it exactly.
 792                  *
 793                  * TODO - we could shortcut this by testing that both
 794                  * live->class and bmap->class are 0, and both avails are
 795                  * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
 796                  */
 797                 if (bcmp(live->bitmapq, bmap->bitmapq,
 798                          sizeof(bmap->bitmapq)) == 0 &&
 799                     live->linear >= bmap->linear) {
 800                         goto next;
 801                 }
 802                 if (hammer2_debug & 1) {
 803                         kprintf("live %016jx %04d.%04x (avail=%d)\n",
 804                                 data_off, bmapindex, live->class, live->avail);
 805                 }
 806
 807                 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
 808                 live = &live_chain->data->bmdata[bmapindex];
 809
 810                 h2_bulkfree_sync_adjust(cbinfo, data_off, live, bmap,
 811                                         live_chain->bref.key +
 812                                         bmapindex *
 813                                         HAMMER2_FREEMAP_LEVEL0_SIZE);
 814 next:
 815                 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
 816                 ++bmap;
 817         }
 818         if (live_chain) {
 819                 hammer2_chain_unlock(live_chain);
 820                 hammer2_chain_drop(live_chain);
 821         }
 822         if (live_parent) {
 823                 hammer2_chain_unlock(live_parent);
 824                 hammer2_chain_drop(live_parent);
 825         }
 826 }
 827
 828 /*
 829  * Merge the bulkfree bitmap against the existing bitmap.
 830  */
 831 static
 832 void
 833 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 834                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 835                         hammer2_bmap_data_t *bmap, hammer2_key_t alloc_base)
 836 {
 837         int bindex;
 838         int scount;
 839         hammer2_off_t tmp_off;
 840         hammer2_bitmap_t lmask;
 841         hammer2_bitmap_t mmask;
 842
 843         tmp_off = data_off;
 844
 845         for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
 846                 lmask = live->bitmapq[bindex];  /* live */
 847                 mmask = bmap->bitmapq[bindex];  /* snapshotted bulkfree */
 848                 if (lmask == mmask) {
 849                         tmp_off += HAMMER2_BMAP_INDEX_SIZE;
 850                         continue;
 851                 }
 852
 853                 for (scount = 0;
 854                      scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
 855                      scount += 2) {
 856                         if ((mmask & 3) == 0) {
 857                                 /*
 858                                  * in-memory 00         live 11 -> 10
 859                                  *                      live 10 -> 00
 860                                  *
 861                                  * Storage might be marked allocated or
 862                                  * staged and must be remarked staged or
 863                                  * free.
 864                                  */
 865                                 switch (lmask & 3) {
 866                                 case 0: /* 00 */
 867                                         break;
 868                                 case 1: /* 01 */
 869                                         kprintf("hammer2_bulkfree: cannot "
 870                                                 "transition m=00/l=01\n");
 871                                         break;
 872                                 case 2: /* 10 -> 00 */
 873                                         live->bitmapq[bindex] &=
 874                                             ~((hammer2_bitmap_t)2 << scount);
 875                                         live->avail +=
 876                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 877                                         if (live->avail >
 878                                             HAMMER2_FREEMAP_LEVEL0_SIZE) {
 879                                                 live->avail =
 880                                                     HAMMER2_FREEMAP_LEVEL0_SIZE;
 881                                         }
 882                                         cbinfo->adj_free +=
 883                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 884                                         ++cbinfo->count_10_00;
 885                                         hammer2_io_dedup_assert(
 886                                                 cbinfo->hmp,
 887                                                 tmp_off |
 888                                                 HAMMER2_FREEMAP_BLOCK_RADIX,
 889                                                 HAMMER2_FREEMAP_BLOCK_SIZE);
 890                                         break;
 891                                 case 3: /* 11 -> 10 */
 892                                         live->bitmapq[bindex] &=
 893                                             ~((hammer2_bitmap_t)1 << scount);
 894                                         ++cbinfo->count_11_10;
 895                                         hammer2_io_dedup_delete(
 896                                                 cbinfo->hmp,
 897                                                 HAMMER2_BREF_TYPE_DATA,
 898                                                 tmp_off |
 899                                                 HAMMER2_FREEMAP_BLOCK_RADIX,
 900                                                 HAMMER2_FREEMAP_BLOCK_SIZE);
 901                                         break;
 902                                 }
 903                         } else if ((mmask & 3) == 3) {
 904                                 /*
 905                                  * in-memory 11         live 10 -> 11
 906                                  *                      live ** -> 11
 907                                  *
 908                                  * Storage might be incorrectly marked free
 909                                  * or staged and must be remarked fully
 910                                  * allocated.
 911                                  */
 912                                 switch (lmask & 3) {
 913                                 case 0: /* 00 */
 914                                         ++cbinfo->count_00_11;
 915                                         cbinfo->adj_free -=
 916                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 917                                         live->avail -=
 918                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 919                                         if ((int32_t)live->avail < 0)
 920                                                 live->avail = 0;
 921                                         break;
 922                                 case 1: /* 01 */
 923                                         ++cbinfo->count_01_11;
 924                                         break;
 925                                 case 2: /* 10 -> 11 */
 926                                         ++cbinfo->count_10_11;
 927                                         break;
 928                                 case 3: /* 11 */
 929                                         break;
 930                                 }
 931                                 live->bitmapq[bindex] |=
 932                                         ((hammer2_bitmap_t)3 << scount);
 933                         }
 934                         mmask >>= 2;
 935                         lmask >>= 2;
 936                         tmp_off += HAMMER2_FREEMAP_BLOCK_SIZE;
 937                 }
 938         }
 939
 940         /*
 941          * Determine if the live bitmap is completely free and reset its
 942          * fields if so.  Otherwise check to see if we can reduce the linear
 943          * offset.
 944          */
 945         for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
 946                 if (live->bitmapq[bindex] != 0)
 947                         break;
 948         }
 949         if (bindex < 0) {
 950                 /*
 951                  * Completely empty, reset entire segment
 952                  */
 953 #if 0
 954                 kprintf("hammer2: cleanseg %016jx.%04x (%d)\n",
 955                         alloc_base, live->class, live->avail);
 956 #endif
 957                 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 958                 live->class = 0;
 959                 live->linear = 0;
 960                 ++cbinfo->count_l0cleans;
 961         } else if (bindex < 7) {
 962                 /*
 963                  * Partially full, bitmapq[bindex] != 0.  The live->linear
 964                  * offset can legitimately be just about anything, but
 965                  * our bulkfree pass doesn't record enough information to
 966                  * set it exactly.  Just make sure that it is set to a
 967                  * safe value that also works in our match code above (the
 968                  * bcmp and linear test).
 969                  *
 970                  * We cannot safely leave live->linear at a sub-block offset
 971                  * unless it is already in the same block as bmap->linear.
 972                  *
 973                  * If it is not in the same block, we cannot assume that
 974                  * we can set it to bmap->linear on a sub-block boundary,
 975                  * because the live system could have bounced it around.
 976                  * In that situation we satisfy our bcmp/skip requirement
 977                  * above by setting it to the nearest higher block boundary.
 978                  * This alignment effectively kills any partial allocation it
 979                  * might have been tracking before.
 980                  */
 981                 if (live->linear < bmap->linear &&
 982                     ((live->linear ^ bmap->linear) &
 983                      ~HAMMER2_FREEMAP_BLOCK_MASK) == 0) {
 984                         live->linear = bmap->linear;
 985                         ++cbinfo->count_linadjusts;
 986                 } else {
 987                         live->linear =
 988                                 (bmap->linear + HAMMER2_FREEMAP_BLOCK_MASK) &
 989                                 ~HAMMER2_FREEMAP_BLOCK_MASK;
 990                         ++cbinfo->count_linadjusts;
 991                 }
 992         } else {
 993                 /*
 994                  * Completely full, effectively disable the linear iterator
 995                  */
 996                 live->linear = HAMMER2_SEGSIZE;
 997         }
 998
 999 #if 0
1000         if (bmap->class) {
1001                 kprintf("%016jx %04d.%04x (avail=%7d) "
1002                         "%08x %08x %08x %08x %08x %08x %08x %08x\n",
1003                         (intmax_t)data_off,
1004                         (int)((data_off &
1005                                HAMMER2_FREEMAP_LEVEL1_MASK) >>
1006                               HAMMER2_FREEMAP_LEVEL0_RADIX),
1007                         bmap->class,
1008                         bmap->avail,
1009                         bmap->bitmap[0], bmap->bitmap[1],
1010                         bmap->bitmap[2], bmap->bitmap[3],
1011                         bmap->bitmap[4], bmap->bitmap[5],
1012                         bmap->bitmap[6], bmap->bitmap[7]);
1013         }
1014 #endif
1015 }
1016
1017 /*
1018  * BULKFREE DEDUP HEURISTIC
1019  *
1020  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
1021  *          All fields must be loaded into locals and validated.
1022  */
1023 static
1024 int
1025 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
1026                  int pri)
1027 {
1028         hammer2_dedup_t *dedup;
1029         int best;
1030         int n;
1031         int i;
1032
1033         n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
1034         dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
1035
1036         for (i = best = 0; i < 8; ++i) {
1037                 if (dedup[i].data_off == bref->data_off) {
1038                         if (dedup[i].ticks < pri)
1039                                 dedup[i].ticks = pri;
1040                         if (pri == 1)
1041                                 cbinfo->count_dedup_factor += dedup[i].ticks;
1042                         return 1;
1043                 }
1044                 if (dedup[i].ticks < dedup[best].ticks)
1045                         best = i;
1046         }
1047         dedup[best].data_off = bref->data_off;
1048         dedup[best].ticks = pri;
1049
1050         return 0;
1051 }