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_chain_t *vchain,
 316                       hammer2_ioc_bulkfree_t *bfi)
 317 {
 318         hammer2_bulkfree_info_t cbinfo;
 319         hammer2_chain_save_t *save;
 320         hammer2_off_t incr;
 321         size_t size;
 322         int doabort = 0;
 323
 324         /*
 325          * We have to clear the live dedup cache as it might have entries
 326          * that are freeable as of now.  Any new entries in the dedup cache
 327          * made after this point, even if they become freeable, will have
 328          * previously been fully allocated and will be protected by the
 329          * 2-stage bulkfree.
 330          */
 331         hammer2_dedup_clear(hmp);
 332
 333         /*
 334          * Setup for free pass
 335          */
 336         bzero(&cbinfo, sizeof(cbinfo));
 337         size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
 338                ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
 339         cbinfo.hmp = hmp;
 340         cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size, VM_SUBSYS_HAMMER);
 341         cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
 342
 343         cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
 344                                M_HAMMER2, M_WAITOK | M_ZERO);
 345
 346         /*
 347          * Normalize start point to a 2GB boundary.  We operate on a
 348          * 64KB leaf bitmap boundary which represents 2GB of storage.
 349          */
 350         cbinfo.sbase = bfi->sbase;
 351         if (cbinfo.sbase > hmp->voldata.volu_size)
 352                 cbinfo.sbase = hmp->voldata.volu_size;
 353         cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 354         TAILQ_INIT(&cbinfo.list);
 355
 356         /*
 357          * Loop on a full meta-data scan as many times as required to
 358          * get through all available storage.
 359          */
 360         while (cbinfo.sbase < hmp->voldata.volu_size) {
 361                 /*
 362                  * We have enough ram to represent (incr) bytes of storage.
 363                  * Each 64KB of ram represents 2GB of storage.
 364                  */
 365                 cbinfo_bmap_init(&cbinfo, size);
 366                 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
 367                        HAMMER2_FREEMAP_LEVEL1_SIZE;
 368                 if (hmp->voldata.volu_size - cbinfo.sbase < incr)
 369                         cbinfo.sstop = hmp->voldata.volu_size;
 370                 else
 371                         cbinfo.sstop = cbinfo.sbase + incr;
 372                 if (hammer2_debug & 1) {
 373                         kprintf("bulkfree pass %016jx/%jdGB\n",
 374                                 (intmax_t)cbinfo.sbase,
 375                                 (intmax_t)incr / HAMMER2_FREEMAP_LEVEL1_SIZE);
 376                 }
 377
 378                 /*
 379                  * Scan topology for stuff inside this range.
 380                  */
 381                 hammer2_trans_init(hmp->spmp, 0);
 382                 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
 383                 cbinfo.pri = 0;
 384                 doabort |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
 385                                              &cbinfo);
 386
 387                 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
 388                        doabort == 0) {
 389                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 390                         cbinfo.pri = 0;
 391                         doabort |= hammer2_bulk_scan(save->chain,
 392                                                      h2_bulkfree_callback,
 393                                                      &cbinfo);
 394                         hammer2_chain_drop(save->chain);
 395                         kfree(save, M_HAMMER2);
 396                 }
 397                 while (save) {
 398                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 399                         hammer2_chain_drop(save->chain);
 400                         kfree(save, M_HAMMER2);
 401                         save = TAILQ_FIRST(&cbinfo.list);
 402                 }
 403
 404                 kprintf("bulkfree lastdrop %d %d doabort=%d\n",
 405                         vchain->refs, vchain->core.chain_count, doabort);
 406
 407                 /*
 408                  * If complete scan succeeded we can synchronize our
 409                  * in-memory freemap against live storage.  If an abort
 410                  * did occur we cannot safely synchronize our partially
 411                  * filled-out in-memory freemap.
 412                  */
 413                 if (doabort == 0) {
 414                         h2_bulkfree_sync(&cbinfo);
 415
 416                         hammer2_voldata_lock(hmp);
 417                         hammer2_voldata_modify(hmp);
 418                         hmp->voldata.allocator_free += cbinfo.adj_free;
 419                         hammer2_voldata_unlock(hmp);
 420                 }
 421
 422                 /*
 423                  * Cleanup for next loop.
 424                  */
 425                 hammer2_trans_done(hmp->spmp);
 426                 if (doabort)
 427                         break;
 428                 cbinfo.sbase = cbinfo.sstop;
 429                 cbinfo.adj_free = 0;
 430         }
 431         kmem_free_swapbacked(&cbinfo.kp);
 432         kfree(cbinfo.dedup, M_HAMMER2);
 433         cbinfo.dedup = NULL;
 434
 435         bfi->sstop = cbinfo.sbase;
 436
 437         incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
 438         if (incr > 10000)
 439                 incr = 10000;
 440
 441         kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
 442                 (int)incr / 100,
 443                 (int)incr % 100);
 444
 445         kprintf("    transition->free   %ld\n", cbinfo.count_10_00);
 446         kprintf("    transition->staged %ld\n", cbinfo.count_11_10);
 447         kprintf("    ERR(00)->allocated %ld\n", cbinfo.count_00_11);
 448         kprintf("    ERR(01)->allocated %ld\n", cbinfo.count_01_11);
 449         kprintf("    staged->allocated  %ld\n", cbinfo.count_10_11);
 450         kprintf("    ~2MB segs cleaned  %ld\n", cbinfo.count_l0cleans);
 451         kprintf("    linear adjusts     %ld\n", cbinfo.count_linadjusts);
 452         kprintf("    dedup factor       %ld\n", cbinfo.count_dedup_factor);
 453
 454         return doabort;
 455 }
 456
 457 static void
 458 cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size)
 459 {
 460         hammer2_bmap_data_t *bmap = cbinfo->bmap;
 461         hammer2_key_t key = cbinfo->sbase;
 462         hammer2_key_t lokey;
 463         hammer2_key_t hikey;
 464
 465         lokey = (cbinfo->hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 466                 ~HAMMER2_SEGMASK64;
 467         hikey = cbinfo->hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 468
 469         bzero(bmap, size);
 470         while (size) {
 471                 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 472                             HAMMER2_ZONE_SEG64) {
 473                         lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 474                                 HAMMER2_ZONE_SEG64;
 475                 }
 476                 if (key < lokey || key >= hikey) {
 477                         memset(bmap->bitmapq, -1,
 478                                sizeof(bmap->bitmapq));
 479                         bmap->avail = 0;
 480                         bmap->linear = HAMMER2_SEGSIZE;
 481                 } else {
 482                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 483                 }
 484                 size -= sizeof(*bmap);
 485                 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
 486                 ++bmap;
 487         }
 488 }
 489
 490 static int
 491 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
 492 {
 493         hammer2_bmap_data_t *bmap;
 494         hammer2_off_t data_off;
 495         uint16_t class;
 496         size_t bytes;
 497         int radix;
 498         int error;
 499
 500         /*
 501          * Check for signal and allow yield to userland during scan
 502          */
 503         if (hammer2_signal_check(&cbinfo->save_time))
 504                 return HAMMER2_BULK_ABORT;
 505         if (bref->type == HAMMER2_BREF_TYPE_INODE) {
 506                 ++cbinfo->count_inodes_scanned;
 507                 if ((cbinfo->count_inodes_scanned & 65535) == 0)
 508                         kprintf(" inodes %6ld bytes %9ld\n",
 509                                 cbinfo->count_inodes_scanned,
 510                                 cbinfo->bytes_scanned);
 511         }
 512
 513         /*
 514          * Calculate the data offset and determine if it is within
 515          * the current freemap range being gathered.
 516          */
 517         error = 0;
 518         data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 519         if (data_off < cbinfo->sbase || data_off >= cbinfo->sstop)
 520                 return 0;
 521         if (data_off < cbinfo->hmp->voldata.allocator_beg)
 522                 return 0;
 523         if (data_off >= cbinfo->hmp->voldata.volu_size)
 524                 return 0;
 525
 526         /*
 527          * Calculate the information needed to generate the in-memory
 528          * freemap record.
 529          *
 530          * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
 531          * it's a problem if it does.  (Or L0 (2MB) for that matter).
 532          */
 533         radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
 534         KKASSERT(radix != 0);
 535         bytes = (size_t)1 << radix;
 536         class = (bref->type << 8) | hammer2_devblkradix(radix);
 537
 538         if (data_off + bytes >= cbinfo->sstop) {
 539                 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
 540                         "%016jx %016jx/%d\n",
 541                         (intmax_t)bref->data_off,
 542                         (intmax_t)bref->key,
 543                         bref->keybits);
 544                 bytes = cbinfo->sstop - data_off;       /* XXX */
 545         }
 546
 547         /*
 548          * Convert to a storage offset relative to the beginning of the
 549          * storage range we are collecting.  Then lookup the level0 bmap entry.
 550          */
 551         data_off -= cbinfo->sbase;
 552         bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
 553
 554         /*
 555          * Convert data_off to a bmap-relative value (~2MB storage range).
 556          * Adjust linear, class, and avail.
 557          *
 558          * Hammer2 does not allow allocations to cross the L0 (2MB) boundary,
 559          */
 560         data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
 561         if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
 562                 kprintf("hammer2_bulkfree_scan: illegal 2MB boundary "
 563                         "%016jx %016jx/%d\n",
 564                         (intmax_t)bref->data_off,
 565                         (intmax_t)bref->key,
 566                         bref->keybits);
 567                 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
 568         }
 569
 570         if (bmap->class == 0) {
 571                 bmap->class = class;
 572                 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 573         }
 574         if (bmap->class != class) {
 575                 kprintf("hammer2_bulkfree_scan: illegal mixed class "
 576                         "%016jx %016jx/%d (%04x vs %04x)\n",
 577                         (intmax_t)bref->data_off,
 578                         (intmax_t)bref->key,
 579                         bref->keybits,
 580                         class, bmap->class);
 581         }
 582         if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
 583                 bmap->linear = (int32_t)data_off + (int32_t)bytes;
 584
 585         /*
 586          * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
 587          * 64-bit entries, 2 bits per entry, to code 11.
 588          *
 589          * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
 590          */
 591         while (bytes > 0) {
 592                 int bindex;
 593                 hammer2_bitmap_t bmask;
 594
 595                 bindex = (int)data_off >> (HAMMER2_FREEMAP_BLOCK_RADIX +
 596                                            HAMMER2_BMAP_INDEX_RADIX);
 597                 bmask = (hammer2_bitmap_t)3 <<
 598                         ((((int)data_off & HAMMER2_BMAP_INDEX_MASK) >>
 599                          HAMMER2_FREEMAP_BLOCK_RADIX) << 1);
 600
 601                 /*
 602                  * NOTE! The (avail) calculation is bitmap-granular.  Multiple
 603                  *       sub-granular records can wind up at the same bitmap
 604                  *       position.
 605                  */
 606                 if ((bmap->bitmapq[bindex] & bmask) == 0) {
 607                         if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE) {
 608                                 bmap->avail -= HAMMER2_FREEMAP_BLOCK_SIZE;
 609                         } else {
 610                                 bmap->avail -= bytes;
 611                         }
 612                         bmap->bitmapq[bindex] |= bmask;
 613                 }
 614                 data_off += HAMMER2_FREEMAP_BLOCK_SIZE;
 615                 if (bytes < HAMMER2_FREEMAP_BLOCK_SIZE)
 616                         bytes = 0;
 617                 else
 618                         bytes -= HAMMER2_FREEMAP_BLOCK_SIZE;
 619         }
 620         return error;
 621 }
 622
 623 /*
 624  * Synchronize the in-memory bitmap with the live freemap.  This is not a
 625  * direct copy.  Instead the bitmaps must be compared:
 626  *
 627  *      In-memory       Live-freemap
 628  *         00             11 -> 10      (do nothing if live modified)
 629  *                        10 -> 00      (do nothing if live modified)
 630  *         11             10 -> 11      handles race against live
 631  *                        ** -> 11      nominally warn of corruption
 632  *
 633  */
 634 static void
 635 h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo)
 636 {
 637         hammer2_off_t data_off;
 638         hammer2_key_t key;
 639         hammer2_key_t key_dummy;
 640         hammer2_bmap_data_t *bmap;
 641         hammer2_bmap_data_t *live;
 642         hammer2_chain_t *live_parent;
 643         hammer2_chain_t *live_chain;
 644         int cache_index = -1;
 645         int bmapindex;
 646
 647         kprintf("hammer2_bulkfree - range ");
 648
 649         if (cbinfo->sbase < cbinfo->hmp->voldata.allocator_beg)
 650                 kprintf("%016jx-",
 651                         (intmax_t)cbinfo->hmp->voldata.allocator_beg);
 652         else
 653                 kprintf("%016jx-",
 654                         (intmax_t)cbinfo->sbase);
 655
 656         if (cbinfo->sstop > cbinfo->hmp->voldata.volu_size)
 657                 kprintf("%016jx\n",
 658                         (intmax_t)cbinfo->hmp->voldata.volu_size);
 659         else
 660                 kprintf("%016jx\n",
 661                         (intmax_t)cbinfo->sstop);
 662
 663         data_off = cbinfo->sbase;
 664         bmap = cbinfo->bmap;
 665
 666         live_parent = &cbinfo->hmp->fchain;
 667         hammer2_chain_ref(live_parent);
 668         hammer2_chain_lock(live_parent, HAMMER2_RESOLVE_ALWAYS);
 669         live_chain = NULL;
 670
 671         /*
 672          * Iterate each hammer2_bmap_data_t line (128 bytes) managing
 673          * 4MB of storage.
 674          */
 675         while (data_off < cbinfo->sstop) {
 676                 /*
 677                  * The freemap is not used below allocator_beg or beyond
 678                  * volu_size.
 679                  */
 680                 int nofree;
 681
 682                 if (data_off < cbinfo->hmp->voldata.allocator_beg)
 683                         goto next;
 684                 if (data_off >= cbinfo->hmp->voldata.volu_size)
 685                         goto next;
 686
 687                 /*
 688                  * Locate the freemap leaf on the live filesystem
 689                  */
 690                 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
 691                 nofree = 0;
 692
 693                 if (live_chain == NULL || live_chain->bref.key != key) {
 694                         if (live_chain) {
 695                                 hammer2_chain_unlock(live_chain);
 696                                 hammer2_chain_drop(live_chain);
 697                         }
 698                         live_chain = hammer2_chain_lookup(
 699                                             &live_parent,
 700                                             &key_dummy,
 701                                             key,
 702                                             key + HAMMER2_FREEMAP_LEVEL1_MASK,
 703                                             &cache_index,
 704                                             HAMMER2_LOOKUP_ALWAYS);
 705
 706 #if 0
 707                         /*
 708                          * If recent allocations were made we avoid races by
 709                          * not staging or freeing any blocks.  We can still
 710                          * remark blocks as fully allocated.
 711                          */
 712                         if (live_chain) {
 713                                 if (hammer2_debug & 1) {
 714                                         kprintf("live_chain %016jx\n",
 715                                                 (intmax_t)key);
 716                                 }
 717                                 if (live_chain->bref.mirror_tid >
 718                                     cbinfo->saved_mirror_tid) {
 719                                         kprintf("hammer2_bulkfree: "
 720                                                 "avoid %016jx\n",
 721                                                 data_off);
 722                                         nofree = 1;
 723                                 } else {
 724                                         nofree = 0;
 725                                 }
 726                         }
 727 #endif
 728                 }
 729                 if (live_chain == NULL) {
 730                         /*
 731                          * XXX if we implement a full recovery mode we need
 732                          * to create/recreate missing freemap chains if our
 733                          * bmap has any allocated blocks.
 734                          */
 735                         if (bmap->class &&
 736                             bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
 737                                 kprintf("hammer2_bulkfree: cannot locate "
 738                                         "live leaf for allocated data "
 739                                         "near %016jx\n",
 740                                         (intmax_t)data_off);
 741                         }
 742                         goto next;
 743                 }
 744                 if (live_chain->error) {
 745                         kprintf("hammer2_bulkfree: error %s looking up "
 746                                 "live leaf for allocated data near %016jx\n",
 747                                 hammer2_error_str(live_chain->error),
 748                                 (intmax_t)data_off);
 749                         hammer2_chain_unlock(live_chain);
 750                         hammer2_chain_drop(live_chain);
 751                         live_chain = NULL;
 752                         goto next;
 753                 }
 754
 755                 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
 756                             HAMMER2_FREEMAP_LEVEL0_RADIX;
 757                 live = &live_chain->data->bmdata[bmapindex];
 758
 759                 /*
 760                  * TODO - we could shortcut this by testing that both
 761                  * live->class and bmap->class are 0, and both avails are
 762                  * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
 763                  */
 764                 if (bcmp(live->bitmapq, bmap->bitmapq,
 765                          sizeof(bmap->bitmapq)) == 0) {
 766                         goto next;
 767                 }
 768                 if (hammer2_debug & 1) {
 769                         kprintf("live %016jx %04d.%04x (avail=%d)\n",
 770                                 data_off, bmapindex, live->class, live->avail);
 771                 }
 772
 773                 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
 774                 live = &live_chain->data->bmdata[bmapindex];
 775
 776                 h2_bulkfree_sync_adjust(cbinfo, data_off, live, bmap, nofree);
 777 next:
 778                 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
 779                 ++bmap;
 780         }
 781         if (live_chain) {
 782                 hammer2_chain_unlock(live_chain);
 783                 hammer2_chain_drop(live_chain);
 784         }
 785         if (live_parent) {
 786                 hammer2_chain_unlock(live_parent);
 787                 hammer2_chain_drop(live_parent);
 788         }
 789 }
 790
 791 /*
 792  * When bulkfree is finally able to free a block it must make sure that
 793  * the INVALOK bit in any cached DIO is cleared prior to the block being
 794  * reused.
 795  */
 796 static
 797 void
 798 fixup_dio(hammer2_dev_t *hmp, hammer2_off_t data_off, int bindex, int scount)
 799 {
 800         data_off += (scount >> 1) * HAMMER2_FREEMAP_BLOCK_SIZE;
 801         data_off += bindex *
 802                 (HAMMER2_FREEMAP_BLOCK_SIZE * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 803         hammer2_io_resetinval(hmp, data_off);
 804 }
 805
 806 /*
 807  * Merge the bulkfree bitmap against the existing bitmap.
 808  *
 809  * If nofree is non-zero the merge will only mark free blocks as allocated
 810  * and will refuse to free any blocks.
 811  */
 812 static
 813 void
 814 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 815                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 816                         hammer2_bmap_data_t *bmap, int nofree)
 817 {
 818         int bindex;
 819         int scount;
 820         hammer2_bitmap_t lmask;
 821         hammer2_bitmap_t mmask;
 822
 823         for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
 824                 lmask = live->bitmapq[bindex];  /* live */
 825                 mmask = bmap->bitmapq[bindex];  /* snapshotted bulkfree */
 826                 if (lmask == mmask)
 827                         continue;
 828
 829                 for (scount = 0;
 830                      scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
 831                      scount += 2) {
 832                         if ((mmask & 3) == 0) {
 833                                 /*
 834                                  * in-memory 00         live 11 -> 10
 835                                  *                      live 10 -> 00
 836                                  *
 837                                  * Storage might be marked allocated or
 838                                  * staged and must be remarked staged or
 839                                  * free.
 840                                  */
 841                                 switch (lmask & 3) {
 842                                 case 0: /* 00 */
 843                                         break;
 844                                 case 1: /* 01 */
 845                                         kprintf("hammer2_bulkfree: cannot "
 846                                                 "transition m=00/l=01\n");
 847                                         break;
 848                                 case 2: /* 10 -> 00 */
 849                                         if (nofree)
 850                                                 break;
 851                                         live->bitmapq[bindex] &=
 852                                             ~((hammer2_bitmap_t)2 << scount);
 853                                         live->avail +=
 854                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 855                                         if (live->avail >
 856                                             HAMMER2_FREEMAP_LEVEL0_SIZE) {
 857                                                 live->avail =
 858                                                     HAMMER2_FREEMAP_LEVEL0_SIZE;
 859                                         }
 860                                         cbinfo->adj_free +=
 861                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 862                                         ++cbinfo->count_10_00;
 863                                         break;
 864                                 case 3: /* 11 -> 10 */
 865                                         live->bitmapq[bindex] &=
 866                                             ~((hammer2_bitmap_t)1 << scount);
 867                                         ++cbinfo->count_11_10;
 868                                         fixup_dio(cbinfo->hmp, data_off,
 869                                                   bindex, scount);
 870                                         break;
 871                                 }
 872                         } else if ((mmask & 3) == 3) {
 873                                 /*
 874                                  * in-memory 11         live 10 -> 11
 875                                  *                      live ** -> 11
 876                                  *
 877                                  * Storage might be incorrectly marked free
 878                                  * or staged and must be remarked fully
 879                                  * allocated.
 880                                  */
 881                                 switch (lmask & 3) {
 882                                 case 0: /* 00 */
 883                                         ++cbinfo->count_00_11;
 884                                         cbinfo->adj_free -=
 885                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 886                                         live->avail -=
 887                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 888                                         if ((int32_t)live->avail < 0)
 889                                                 live->avail = 0;
 890                                         break;
 891                                 case 1: /* 01 */
 892                                         ++cbinfo->count_01_11;
 893                                         break;
 894                                 case 2: /* 10 -> 11 */
 895                                         ++cbinfo->count_10_11;
 896                                         break;
 897                                 case 3: /* 11 */
 898                                         break;
 899                                 }
 900                                 live->bitmapq[bindex] |=
 901                                         ((hammer2_bitmap_t)3 << scount);
 902                         }
 903                         mmask >>= 2;
 904                         lmask >>= 2;
 905                 }
 906         }
 907
 908         /*
 909          * Determine if the live bitmap is completely free and reset its
 910          * fields if so.  Otherwise check to see if we can reduce the linear
 911          * offset.
 912          */
 913         for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
 914                 if (live->bitmapq[bindex] != 0)
 915                         break;
 916         }
 917         if (nofree) {
 918                 /* do nothing */
 919         } else if (bindex < 0) {
 920                 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 921                 live->class = 0;
 922                 live->linear = 0;
 923                 ++cbinfo->count_l0cleans;
 924         } else if (bindex < 7) {
 925                 ++bindex;
 926                 if (live->linear > bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 927                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 928                         ++cbinfo->count_linadjusts;
 929                 }
 930
 931                 /*
 932                  * XXX this fine-grained measure still has some issues.
 933                  */
 934                 if (live->linear < bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 935                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 936                         ++cbinfo->count_linadjusts;
 937                 }
 938         } else {
 939                 live->linear = HAMMER2_SEGSIZE;
 940         }
 941
 942 #if 0
 943         if (bmap->class) {
 944                 kprintf("%016jx %04d.%04x (avail=%7d) "
 945                         "%08x %08x %08x %08x %08x %08x %08x %08x\n",
 946                         (intmax_t)data_off,
 947                         (int)((data_off &
 948                                HAMMER2_FREEMAP_LEVEL1_MASK) >>
 949                               HAMMER2_FREEMAP_LEVEL0_RADIX),
 950                         bmap->class,
 951                         bmap->avail,
 952                         bmap->bitmap[0], bmap->bitmap[1],
 953                         bmap->bitmap[2], bmap->bitmap[3],
 954                         bmap->bitmap[4], bmap->bitmap[5],
 955                         bmap->bitmap[6], bmap->bitmap[7]);
 956         }
 957 #endif
 958 }
 959
 960 /*
 961  * BULKFREE DEDUP HEURISTIC
 962  *
 963  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
 964  *          All fields must be loaded into locals and validated.
 965  */
 966 static
 967 int
 968 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
 969                  int pri)
 970 {
 971         hammer2_dedup_t *dedup;
 972         int best;
 973         int n;
 974         int i;
 975
 976         n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
 977         dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
 978
 979         for (i = best = 0; i < 8; ++i) {
 980                 if (dedup[i].data_off == bref->data_off) {
 981                         if (dedup[i].ticks < pri)
 982                                 dedup[i].ticks = pri;
 983                         if (pri == 1)
 984                                 cbinfo->count_dedup_factor += dedup[i].ticks;
 985                         return 1;
 986                 }
 987                 if (dedup[i].ticks < dedup[best].ticks)
 988                         best = i;
 989         }
 990         dedup[best].data_off = bref->data_off;
 991         dedup[best].ticks = pri;
 992
 993         return 0;
 994 }