sys/vfs/hammer2/hammer2_bulkfree.c

   1 /*
   2  * Copyright (c) 2013-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/kernel.h>
  37 #include <sys/fcntl.h>
  38 #include <sys/buf.h>
  39 #include <sys/proc.h>
  40 #include <sys/namei.h>
  41 #include <sys/mount.h>
  42 #include <sys/vnode.h>
  43 #include <sys/mountctl.h>
  44 #include <vm/vm_kern.h>
  45 #include <vm/vm_extern.h>
  46
  47 #include "hammer2.h"
  48
  49 #define H2FMBASE(key, radix)    ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
  50 #define H2FMSHIFT(radix)        ((hammer2_off_t)1 << (radix))
  51
  52 /*
  53  * breadth-first search
  54  */
  55 typedef struct hammer2_chain_save {
  56         TAILQ_ENTRY(hammer2_chain_save) entry;
  57         hammer2_chain_t *chain;
  58         int pri;
  59 } hammer2_chain_save_t;
  60
  61 TAILQ_HEAD(hammer2_chain_save_list, hammer2_chain_save);
  62 typedef struct hammer2_chain_save_list hammer2_chain_save_list_t;
  63
  64 typedef struct hammer2_bulkfree_info {
  65         hammer2_dev_t           *hmp;
  66         kmem_anon_desc_t        kp;
  67         hammer2_off_t           sbase;          /* sub-loop iteration */
  68         hammer2_off_t           sstop;
  69         hammer2_bmap_data_t     *bmap;
  70         int                     depth;
  71         long                    count_10_00;    /* staged->free      */
  72         long                    count_11_10;    /* allocated->staged */
  73         long                    count_00_11;    /* (should not happen) */
  74         long                    count_01_11;    /* (should not happen) */
  75         long                    count_10_11;    /* staged->allocated */
  76         long                    count_l0cleans;
  77         long                    count_linadjusts;
  78         long                    count_inodes_scanned;
  79         long                    count_dedup_factor;
  80         long                    bytes_scanned;
  81         hammer2_off_t           adj_free;
  82         hammer2_tid_t           mtid;
  83         hammer2_tid_t           saved_mirror_tid;
  84         time_t                  save_time;
  85         hammer2_chain_save_list_t list;
  86         hammer2_dedup_t         *dedup;
  87         int                     pri;
  88 } hammer2_bulkfree_info_t;
  89
  90 static int h2_bulkfree_test(hammer2_bulkfree_info_t *info,
  91                         hammer2_blockref_t *bref, int pri);
  92
  93 /*
  94  * General bulk scan function with callback.  Called with a referenced
  95  * but UNLOCKED parent.  The parent is returned in the same state.
  96  */
  97 static
  98 int
  99 hammer2_bulk_scan(hammer2_chain_t *parent,
 100                   int (*func)(hammer2_bulkfree_info_t *info,
 101                               hammer2_blockref_t *bref),
 102                   hammer2_bulkfree_info_t *info)
 103 {
 104         hammer2_blockref_t bref;
 105         hammer2_chain_t *chain;
 106         int cache_index = -1;
 107         int doabort = 0;
 108         int first = 1;
 109
 110         ++info->pri;
 111
 112         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
 113                                    HAMMER2_RESOLVE_SHARED);
 114         chain = NULL;
 115
 116         /*
 117          * Generally loop on the contents if we have not been flagged
 118          * for abort.
 119          *
 120          * Remember that these chains are completely isolated from
 121          * the frontend, so we can release locks temporarily without
 122          * imploding.
 123          */
 124         while ((doabort & HAMMER2_BULK_ABORT) == 0 &&
 125                hammer2_chain_scan(parent, &chain, &bref, &first,
 126                                   &cache_index,
 127                                   HAMMER2_LOOKUP_NODATA |
 128                                   HAMMER2_LOOKUP_SHARED) != NULL) {
 129                 /*
 130                  * Process bref, chain is only non-NULL if the bref
 131                  * might be recursable (its possible that we sometimes get
 132                  * a non-NULL chain where the bref cannot be recursed).
 133                  */
 134 #if 0
 135                 kprintf("SCAN %016jx\n", bref.data_off);
 136                 int xerr = tsleep(&info->pri, PCATCH, "slp", hz / 10);
 137                 if (xerr == EINTR || xerr == ERESTART) {
 138                         doabort |= HAMMER2_BULK_ABORT;
 139                 }
 140 #endif
 141                 ++info->pri;
 142                 if (h2_bulkfree_test(info, &bref, 1))
 143                         continue;
 144
 145                 doabort |= func(info, &bref);
 146
 147                 if (doabort & HAMMER2_BULK_ABORT)
 148                         break;
 149
 150                 /*
 151                  * A non-null chain is always returned if it is
 152                  * recursive, otherwise a non-null chain might be
 153                  * returned but usually is not when not recursive.
 154                  */
 155                 if (chain == NULL)
 156                         continue;
 157
 158                 /*
 159                  * Else check type and setup depth-first scan.
 160                  *
 161                  * Account for bytes actually read.
 162                  */
 163                 info->bytes_scanned += chain->bytes;
 164
 165                 switch(chain->bref.type) {
 166                 case HAMMER2_BREF_TYPE_INODE:
 167                 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
 168                 case HAMMER2_BREF_TYPE_INDIRECT:
 169                 case HAMMER2_BREF_TYPE_VOLUME:
 170                 case HAMMER2_BREF_TYPE_FREEMAP:
 171                         ++info->depth;
 172                         if (info->depth > 16) {
 173                                 hammer2_chain_save_t *save;
 174                                 save = kmalloc(sizeof(*save), M_HAMMER2,
 175                                                M_WAITOK | M_ZERO);
 176                                 save->chain = chain;
 177                                 hammer2_chain_ref(chain);
 178                                 TAILQ_INSERT_TAIL(&info->list, save, entry);
 179
 180                                 /* guess */
 181                                 info->pri += 10;
 182                         } else {
 183                                 int savepri = info->pri;
 184
 185                                 hammer2_chain_unlock(chain);
 186                                 info->pri = 0;
 187                                 doabort |= hammer2_bulk_scan(chain, func, info);
 188                                 info->pri += savepri;
 189                                 hammer2_chain_lock(chain,
 190                                                    HAMMER2_RESOLVE_ALWAYS |
 191                                                    HAMMER2_RESOLVE_SHARED);
 192                         }
 193                         --info->depth;
 194                         break;
 195                 default:
 196                         /* does not recurse */
 197                         break;
 198                 }
 199         }
 200         if (chain) {
 201                 hammer2_chain_unlock(chain);
 202                 hammer2_chain_drop(chain);
 203         }
 204
 205         /*
 206          * Save with higher pri now that we know what it is.
 207          */
 208         h2_bulkfree_test(info, &parent->bref, info->pri + 1);
 209
 210         hammer2_chain_unlock(parent);
 211
 212         return doabort;
 213 }
 214
 215 /*
 216  * Bulkfree algorithm
 217  *
 218  * Repeat {
 219  *      Chain flush (partial synchronization) XXX removed
 220  *      Scan the whole topology - build in-memory freemap (mark 11)
 221  *      Reconcile the in-memory freemap against the on-disk freemap.
 222  *              ondisk xx -> ondisk 11 (if allocated)
 223  *              ondisk 11 -> ondisk 10 (if free in-memory)
 224  *              ondisk 10 -> ondisk 00 (if free in-memory) - on next pass
 225  * }
 226  *
 227  * The topology scan may have to be performed multiple times to window
 228  * freemaps which are too large to fit in kernel memory.
 229  *
 230  * Races are handled using a double-transition (11->10, 10->00).  The bulkfree
 231  * scan snapshots the volume root's blockset and thus can run concurrent with
 232  * normal operations, as long as a full flush is made between each pass to
 233  * synchronize any modified chains (otherwise their blocks might be improperly
 234  * freed).
 235  *
 236  * Temporary memory in multiples of 64KB is required to reconstruct the leaf
 237  * hammer2_bmap_data blocks so they can later be compared against the live
 238  * freemap.  Each 64KB block represents 128 x 16KB x 1024 = ~2 GB of storage.
 239  * A 32MB save area thus represents around ~1 TB.  The temporary memory
 240  * allocated can be specified.  If it is not sufficient multiple topology
 241  * passes will be made.
 242  */
 243
 244 /*
 245  * Bulkfree callback info
 246  */
 247 static void hammer2_bulkfree_thread(void *arg __unused);
 248 static void cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size);
 249 static int h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo,
 250                         hammer2_blockref_t *bref);
 251 static void h2_bulkfree_sync(hammer2_bulkfree_info_t *cbinfo);
 252 static void h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 253                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 254                         hammer2_bmap_data_t *bmap, int nofree);
 255
 256 void
 257 hammer2_bulkfree_init(hammer2_dev_t *hmp)
 258 {
 259         hammer2_thr_create(&hmp->bfthr, NULL, hmp,
 260                            hmp->devrepname, -1, -1,
 261                            hammer2_bulkfree_thread);
 262 }
 263
 264 void
 265 hammer2_bulkfree_uninit(hammer2_dev_t *hmp)
 266 {
 267         hammer2_thr_delete(&hmp->bfthr);
 268 }
 269
 270 static void
 271 hammer2_bulkfree_thread(void *arg)
 272 {
 273         hammer2_thread_t *thr = arg;
 274         hammer2_ioc_bulkfree_t bfi;
 275         uint32_t flags;
 276
 277         for (;;) {
 278                 hammer2_thr_wait_any(thr,
 279                                      HAMMER2_THREAD_STOP |
 280                                      HAMMER2_THREAD_FREEZE |
 281                                      HAMMER2_THREAD_UNFREEZE |
 282                                      HAMMER2_THREAD_REMASTER,
 283                                      hz * 60);
 284
 285                 flags = thr->flags;
 286                 cpu_ccfence();
 287                 if (flags & HAMMER2_THREAD_STOP)
 288                         break;
 289                 if (flags & HAMMER2_THREAD_FREEZE) {
 290                         hammer2_thr_signal2(thr, HAMMER2_THREAD_FROZEN,
 291                                                  HAMMER2_THREAD_FREEZE);
 292                         continue;
 293                 }
 294                 if (flags & HAMMER2_THREAD_UNFREEZE) {
 295                         hammer2_thr_signal2(thr, 0,
 296                                                  HAMMER2_THREAD_FROZEN |
 297                                                  HAMMER2_THREAD_UNFREEZE);
 298                         continue;
 299                 }
 300                 if (flags & HAMMER2_THREAD_FROZEN)
 301                         continue;
 302                 if (flags & HAMMER2_THREAD_REMASTER) {
 303                         hammer2_thr_signal2(thr, 0, HAMMER2_THREAD_REMASTER);
 304                         bzero(&bfi, sizeof(bfi));
 305                         bfi.size = 8192 * 1024;
 306                         hammer2_bulkfree_pass(thr->hmp, &bfi);
 307                 }
 308         }
 309         thr->td = NULL;
 310         hammer2_thr_signal(thr, HAMMER2_THREAD_STOPPED);
 311         /* structure can go invalid at this point */
 312 }
 313
 314 int
 315 hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_ioc_bulkfree_t *bfi)
 316 {
 317         hammer2_bulkfree_info_t cbinfo;
 318         hammer2_chain_t *vchain;
 319         hammer2_chain_save_t *save;
 320         hammer2_off_t incr;
 321         size_t size;
 322         int doabort = 0;
 323
 324         /*
 325          * A bulkfree operations lock is required for the duration.  We
 326          * must hold it across our flushes to guarantee that we never run
 327          * two bulkfree passes in a row without a flush in the middle.
 328          */
 329         lockmgr(&hmp->bulklk, LK_EXCLUSIVE);
 330
 331         /*
 332          * We have to clear the live dedup cache as it might have entries
 333          * that are freeable as of now.  Any new entries in the dedup cache
 334          * made after this point, even if they become freeable, will have
 335          * previously been fully allocated and will be protected by the
 336          * 2-stage bulkfree.
 337          */
 338         hammer2_dedup_clear(hmp);
 339
 340         /*
 341          * Create a stable snapshot of the block tree which we can run
 342          * the bulkfree pass on.  This allows the bulkfree pass to run
 343          * concurrent with all other operations (except another bulkfree)
 344          *
 345          * This must flush all dirty chain data, but does not have to
 346          * flush dirty buffer cache buffers which have not yet been
 347          * realized and does not have to flush any newly realized dirty
 348          * chains while the bulkfree pass is running, as long as said
 349          * newly dirtied chains get flushed the next time, before the
 350          * next bulkfree pass.
 351          */
 352         vchain = hammer2_flush_quick(hmp);
 353         hammer2_chain_bulkdrop(vchain);
 354         vchain = hammer2_flush_quick(hmp);
 355
 356         /*
 357          * Setup for free pass
 358          */
 359         bzero(&cbinfo, sizeof(cbinfo));
 360         size = (bfi->size + HAMMER2_FREEMAP_LEVELN_PSIZE - 1) &
 361                ~(size_t)(HAMMER2_FREEMAP_LEVELN_PSIZE - 1);
 362         cbinfo.hmp = hmp;
 363         cbinfo.bmap = kmem_alloc_swapbacked(&cbinfo.kp, size, VM_SUBSYS_HAMMER);
 364         cbinfo.saved_mirror_tid = hmp->voldata.mirror_tid;
 365
 366         cbinfo.dedup = kmalloc(sizeof(*cbinfo.dedup) * HAMMER2_DEDUP_HEUR_SIZE,
 367                                M_HAMMER2, M_WAITOK | M_ZERO);
 368
 369         /*
 370          * Normalize start point to a 2GB boundary.  We operate on a
 371          * 64KB leaf bitmap boundary which represents 2GB of storage.
 372          */
 373         cbinfo.sbase = bfi->sbase;
 374         if (cbinfo.sbase > hmp->voldata.volu_size)
 375                 cbinfo.sbase = hmp->voldata.volu_size;
 376         cbinfo.sbase &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 377         TAILQ_INIT(&cbinfo.list);
 378
 379         /*
 380          * Loop on a full meta-data scan as many times as required to
 381          * get through all available storage.
 382          */
 383         while (cbinfo.sbase < hmp->voldata.volu_size) {
 384                 /*
 385                  * We have enough ram to represent (incr) bytes of storage.
 386                  * Each 64KB of ram represents 2GB of storage.
 387                  */
 388                 cbinfo_bmap_init(&cbinfo, size);
 389                 incr = size / HAMMER2_FREEMAP_LEVELN_PSIZE *
 390                        HAMMER2_FREEMAP_LEVEL1_SIZE;
 391                 if (hmp->voldata.volu_size - cbinfo.sbase < incr)
 392                         cbinfo.sstop = hmp->voldata.volu_size;
 393                 else
 394                         cbinfo.sstop = cbinfo.sbase + incr;
 395                 if (hammer2_debug & 1) {
 396                         kprintf("bulkfree pass %016jx/%jdGB\n",
 397                                 (intmax_t)cbinfo.sbase,
 398                                 (intmax_t)incr / HAMMER2_FREEMAP_LEVEL1_SIZE);
 399                 }
 400
 401                 /*
 402                  * Scan topology for stuff inside this range.
 403                  */
 404                 hammer2_trans_init(hmp->spmp, 0);
 405                 cbinfo.mtid = hammer2_trans_sub(hmp->spmp);
 406                 cbinfo.pri = 0;
 407                 doabort |= hammer2_bulk_scan(vchain, h2_bulkfree_callback,
 408                                              &cbinfo);
 409
 410                 while ((save = TAILQ_FIRST(&cbinfo.list)) != NULL &&
 411                        doabort == 0) {
 412                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 413                         cbinfo.pri = 0;
 414                         doabort |= hammer2_bulk_scan(save->chain,
 415                                                      h2_bulkfree_callback,
 416                                                      &cbinfo);
 417                         hammer2_chain_drop(save->chain);
 418                         kfree(save, M_HAMMER2);
 419                 }
 420                 while (save) {
 421                         TAILQ_REMOVE(&cbinfo.list, save, entry);
 422                         hammer2_chain_drop(save->chain);
 423                         kfree(save, M_HAMMER2);
 424                         save = TAILQ_FIRST(&cbinfo.list);
 425                 }
 426
 427                 kprintf("bulkfree lastdrop %d %d doabort=%d\n",
 428                         vchain->refs, vchain->core.chain_count, doabort);
 429
 430                 /*
 431                  * If complete scan succeeded we can synchronize our
 432                  * in-memory freemap against live storage.  If an abort
 433                  * did occur we cannot safely synchronize our partially
 434                  * filled-out in-memory freemap.
 435                  */
 436                 if (doabort == 0) {
 437                         h2_bulkfree_sync(&cbinfo);
 438
 439                         hammer2_voldata_lock(hmp);
 440                         hammer2_voldata_modify(hmp);
 441                         hmp->voldata.allocator_free += cbinfo.adj_free;
 442                         hammer2_voldata_unlock(hmp);
 443                 }
 444
 445                 /*
 446                  * Cleanup for next loop.
 447                  */
 448                 hammer2_trans_done(hmp->spmp);
 449                 if (doabort)
 450                         break;
 451                 cbinfo.sbase = cbinfo.sstop;
 452                 cbinfo.adj_free = 0;
 453         }
 454         hammer2_chain_bulkdrop(vchain);
 455         kmem_free_swapbacked(&cbinfo.kp);
 456         kfree(cbinfo.dedup, M_HAMMER2);
 457         cbinfo.dedup = NULL;
 458
 459         bfi->sstop = cbinfo.sbase;
 460
 461         incr = bfi->sstop / (hmp->voldata.volu_size / 10000);
 462         if (incr > 10000)
 463                 incr = 10000;
 464
 465         kprintf("bulkfree pass statistics (%d.%02d%% storage processed):\n",
 466                 (int)incr / 100,
 467                 (int)incr % 100);
 468
 469         kprintf("    transition->free   %ld\n", cbinfo.count_10_00);
 470         kprintf("    transition->staged %ld\n", cbinfo.count_11_10);
 471         kprintf("    ERR(00)->allocated %ld\n", cbinfo.count_00_11);
 472         kprintf("    ERR(01)->allocated %ld\n", cbinfo.count_01_11);
 473         kprintf("    staged->allocated  %ld\n", cbinfo.count_10_11);
 474         kprintf("    ~2MB segs cleaned  %ld\n", cbinfo.count_l0cleans);
 475         kprintf("    linear adjusts     %ld\n", cbinfo.count_linadjusts);
 476         kprintf("    dedup factor       %ld\n", cbinfo.count_dedup_factor);
 477
 478         lockmgr(&hmp->bulklk, LK_RELEASE);
 479         /* hammer2_vfs_sync(mp, MNT_WAIT); sync needed */
 480
 481         return doabort;
 482 }
 483
 484 static void
 485 cbinfo_bmap_init(hammer2_bulkfree_info_t *cbinfo, size_t size)
 486 {
 487         hammer2_bmap_data_t *bmap = cbinfo->bmap;
 488         hammer2_key_t key = cbinfo->sbase;
 489         hammer2_key_t lokey;
 490         hammer2_key_t hikey;
 491
 492         lokey = (cbinfo->hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 493                 ~HAMMER2_SEGMASK64;
 494         hikey = cbinfo->hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 495
 496         bzero(bmap, size);
 497         while (size) {
 498                 if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 499                             HAMMER2_ZONE_SEG64) {
 500                         lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 501                                 HAMMER2_ZONE_SEG64;
 502                 }
 503                 if (key < lokey || key >= hikey) {
 504                         memset(bmap->bitmapq, -1,
 505                                sizeof(bmap->bitmapq));
 506                         bmap->avail = 0;
 507                         bmap->linear = HAMMER2_SEGSIZE;
 508                 } else {
 509                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 510                 }
 511                 size -= sizeof(*bmap);
 512                 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
 513                 ++bmap;
 514         }
 515 }
 516
 517 static int
 518 h2_bulkfree_callback(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref)
 519 {
 520         hammer2_bmap_data_t *bmap;
 521         hammer2_off_t data_off;
 522         uint16_t class;
 523         size_t bytes;
 524         int radix;
 525         int error;
 526
 527         /*
 528          * Check for signal and allow yield to userland during scan
 529          */
 530         if (hammer2_signal_check(&cbinfo->save_time))
 531                 return HAMMER2_BULK_ABORT;
 532         if (bref->type == HAMMER2_BREF_TYPE_INODE) {
 533                 ++cbinfo->count_inodes_scanned;
 534                 if ((cbinfo->count_inodes_scanned & 65535) == 0)
 535                         kprintf(" inodes %6ld bytes %9ld\n",
 536                                 cbinfo->count_inodes_scanned,
 537                                 cbinfo->bytes_scanned);
 538         }
 539
 540         /*
 541          * Calculate the data offset and determine if it is within
 542          * the current freemap range being gathered.
 543          */
 544         error = 0;
 545         data_off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 546         if (data_off < cbinfo->sbase || data_off >= cbinfo->sstop)
 547                 return 0;
 548         if (data_off < cbinfo->hmp->voldata.allocator_beg)
 549                 return 0;
 550         if (data_off >= cbinfo->hmp->voldata.volu_size)
 551                 return 0;
 552
 553         /*
 554          * Calculate the information needed to generate the in-memory
 555          * freemap record.
 556          *
 557          * Hammer2 does not allow allocations to cross the L1 (2GB) boundary,
 558          * it's a problem if it does.  (Or L0 (2MB) for that matter).
 559          */
 560         radix = (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX);
 561         KKASSERT(radix != 0);
 562         bytes = (size_t)1 << radix;
 563         class = (bref->type << 8) | hammer2_devblkradix(radix);
 564
 565         if (data_off + bytes >= cbinfo->sstop) {
 566                 kprintf("hammer2_bulkfree_scan: illegal 2GB boundary "
 567                         "%016jx %016jx/%d\n",
 568                         (intmax_t)bref->data_off,
 569                         (intmax_t)bref->key,
 570                         bref->keybits);
 571                 bytes = cbinfo->sstop - data_off;       /* XXX */
 572         }
 573
 574         /*
 575          * Convert to a storage offset relative to the beginning of the
 576          * storage range we are collecting.  Then lookup the level0 bmap entry.
 577          */
 578         data_off -= cbinfo->sbase;
 579         bmap = cbinfo->bmap + (data_off >> HAMMER2_FREEMAP_LEVEL0_RADIX);
 580
 581         /*
 582          * Convert data_off to a bmap-relative value (~2MB storage range).
 583          * Adjust linear, class, and avail.
 584          *
 585          * Hammer2 does not allow allocations to cross the L0 (2MB) boundary,
 586          */
 587         data_off &= HAMMER2_FREEMAP_LEVEL0_MASK;
 588         if (data_off + bytes > HAMMER2_FREEMAP_LEVEL0_SIZE) {
 589                 kprintf("hammer2_bulkfree_scan: illegal 2MB boundary "
 590                         "%016jx %016jx/%d\n",
 591                         (intmax_t)bref->data_off,
 592                         (intmax_t)bref->key,
 593                         bref->keybits);
 594                 bytes = HAMMER2_FREEMAP_LEVEL0_SIZE - data_off;
 595         }
 596
 597         if (bmap->class == 0) {
 598                 bmap->class = class;
 599                 bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 600         }
 601         if (bmap->class != class) {
 602                 kprintf("hammer2_bulkfree_scan: illegal mixed class "
 603                         "%016jx %016jx/%d (%04x vs %04x)\n",
 604                         (intmax_t)bref->data_off,
 605                         (intmax_t)bref->key,
 606                         bref->keybits,
 607                         class, bmap->class);
 608         }
 609         if (bmap->linear < (int32_t)data_off + (int32_t)bytes)
 610                 bmap->linear = (int32_t)data_off + (int32_t)bytes;
 611
 612         /*
 613          * Adjust the hammer2_bitmap_t bitmap[HAMMER2_BMAP_ELEMENTS].
 614          * 64-bit entries, 2 bits per entry, to code 11.
 615          *
 616          * NOTE: The allocation can be smaller than HAMMER2_FREEMAP_BLOCK_SIZE.
 617          */
 618         while (bytes > 0) {
 619                 int bindex;
 620                 hammer2_bitmap_t bmask;
 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                 int nofree;
 708
 709                 if (data_off < cbinfo->hmp->voldata.allocator_beg)
 710                         goto next;
 711                 if (data_off >= cbinfo->hmp->voldata.volu_size)
 712                         goto next;
 713
 714                 /*
 715                  * Locate the freemap leaf on the live filesystem
 716                  */
 717                 key = (data_off & ~HAMMER2_FREEMAP_LEVEL1_MASK);
 718                 nofree = 0;
 719
 720                 if (live_chain == NULL || live_chain->bref.key != key) {
 721                         if (live_chain) {
 722                                 hammer2_chain_unlock(live_chain);
 723                                 hammer2_chain_drop(live_chain);
 724                         }
 725                         live_chain = hammer2_chain_lookup(
 726                                             &live_parent,
 727                                             &key_dummy,
 728                                             key,
 729                                             key + HAMMER2_FREEMAP_LEVEL1_MASK,
 730                                             &cache_index,
 731                                             HAMMER2_LOOKUP_ALWAYS);
 732
 733 #if 0
 734                         /*
 735                          * If recent allocations were made we avoid races by
 736                          * not staging or freeing any blocks.  We can still
 737                          * remark blocks as fully allocated.
 738                          */
 739                         if (live_chain) {
 740                                 if (hammer2_debug & 1) {
 741                                         kprintf("live_chain %016jx\n",
 742                                                 (intmax_t)key);
 743                                 }
 744                                 if (live_chain->bref.mirror_tid >
 745                                     cbinfo->saved_mirror_tid) {
 746                                         kprintf("hammer2_bulkfree: "
 747                                                 "avoid %016jx\n",
 748                                                 data_off);
 749                                         nofree = 1;
 750                                 } else {
 751                                         nofree = 0;
 752                                 }
 753                         }
 754 #endif
 755                 }
 756                 if (live_chain == NULL) {
 757                         /*
 758                          * XXX if we implement a full recovery mode we need
 759                          * to create/recreate missing freemap chains if our
 760                          * bmap has any allocated blocks.
 761                          */
 762                         if (bmap->class &&
 763                             bmap->avail != HAMMER2_FREEMAP_LEVEL0_SIZE) {
 764                                 kprintf("hammer2_bulkfree: cannot locate "
 765                                         "live leaf for allocated data "
 766                                         "near %016jx\n",
 767                                         (intmax_t)data_off);
 768                         }
 769                         goto next;
 770                 }
 771                 if (live_chain->error) {
 772                         kprintf("hammer2_bulkfree: error %s looking up "
 773                                 "live leaf for allocated data near %016jx\n",
 774                                 hammer2_error_str(live_chain->error),
 775                                 (intmax_t)data_off);
 776                         hammer2_chain_unlock(live_chain);
 777                         hammer2_chain_drop(live_chain);
 778                         live_chain = NULL;
 779                         goto next;
 780                 }
 781
 782                 bmapindex = (data_off & HAMMER2_FREEMAP_LEVEL1_MASK) >>
 783                             HAMMER2_FREEMAP_LEVEL0_RADIX;
 784                 live = &live_chain->data->bmdata[bmapindex];
 785
 786                 /*
 787                  * TODO - we could shortcut this by testing that both
 788                  * live->class and bmap->class are 0, and both avails are
 789                  * set to HAMMER2_FREEMAP_LEVEL0_SIZE (4MB).
 790                  */
 791                 if (bcmp(live->bitmapq, bmap->bitmapq,
 792                          sizeof(bmap->bitmapq)) == 0) {
 793                         goto next;
 794                 }
 795                 if (hammer2_debug & 1) {
 796                         kprintf("live %016jx %04d.%04x (avail=%d)\n",
 797                                 data_off, bmapindex, live->class, live->avail);
 798                 }
 799
 800                 hammer2_chain_modify(live_chain, cbinfo->mtid, 0, 0);
 801                 live = &live_chain->data->bmdata[bmapindex];
 802
 803                 h2_bulkfree_sync_adjust(cbinfo, data_off, live, bmap, nofree);
 804 next:
 805                 data_off += HAMMER2_FREEMAP_LEVEL0_SIZE;
 806                 ++bmap;
 807         }
 808         if (live_chain) {
 809                 hammer2_chain_unlock(live_chain);
 810                 hammer2_chain_drop(live_chain);
 811         }
 812         if (live_parent) {
 813                 hammer2_chain_unlock(live_parent);
 814                 hammer2_chain_drop(live_parent);
 815         }
 816 }
 817
 818 /*
 819  * When bulkfree is finally able to free a block it must make sure that
 820  * the INVALOK bit in any cached DIO is cleared prior to the block being
 821  * reused.
 822  */
 823 static
 824 void
 825 fixup_dio(hammer2_dev_t *hmp, hammer2_off_t data_off, int bindex, int scount)
 826 {
 827         data_off += (scount >> 1) * HAMMER2_FREEMAP_BLOCK_SIZE;
 828         data_off += bindex *
 829                 (HAMMER2_FREEMAP_BLOCK_SIZE * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 830         hammer2_io_resetinval(hmp, data_off);
 831 }
 832
 833 /*
 834  * Merge the bulkfree bitmap against the existing bitmap.
 835  *
 836  * If nofree is non-zero the merge will only mark free blocks as allocated
 837  * and will refuse to free any blocks.
 838  */
 839 static
 840 void
 841 h2_bulkfree_sync_adjust(hammer2_bulkfree_info_t *cbinfo,
 842                         hammer2_off_t data_off, hammer2_bmap_data_t *live,
 843                         hammer2_bmap_data_t *bmap, int nofree)
 844 {
 845         int bindex;
 846         int scount;
 847         hammer2_bitmap_t lmask;
 848         hammer2_bitmap_t mmask;
 849
 850         for (bindex = 0; bindex < HAMMER2_BMAP_ELEMENTS; ++bindex) {
 851                 lmask = live->bitmapq[bindex];  /* live */
 852                 mmask = bmap->bitmapq[bindex];  /* snapshotted bulkfree */
 853                 if (lmask == mmask)
 854                         continue;
 855
 856                 for (scount = 0;
 857                      scount < HAMMER2_BMAP_BITS_PER_ELEMENT;
 858                      scount += 2) {
 859                         if ((mmask & 3) == 0) {
 860                                 /*
 861                                  * in-memory 00         live 11 -> 10
 862                                  *                      live 10 -> 00
 863                                  *
 864                                  * Storage might be marked allocated or
 865                                  * staged and must be remarked staged or
 866                                  * free.
 867                                  */
 868                                 switch (lmask & 3) {
 869                                 case 0: /* 00 */
 870                                         break;
 871                                 case 1: /* 01 */
 872                                         kprintf("hammer2_bulkfree: cannot "
 873                                                 "transition m=00/l=01\n");
 874                                         break;
 875                                 case 2: /* 10 -> 00 */
 876                                         if (nofree)
 877                                                 break;
 878                                         live->bitmapq[bindex] &=
 879                                             ~((hammer2_bitmap_t)2 << scount);
 880                                         live->avail +=
 881                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 882                                         if (live->avail >
 883                                             HAMMER2_FREEMAP_LEVEL0_SIZE) {
 884                                                 live->avail =
 885                                                     HAMMER2_FREEMAP_LEVEL0_SIZE;
 886                                         }
 887                                         cbinfo->adj_free +=
 888                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 889                                         ++cbinfo->count_10_00;
 890                                         break;
 891                                 case 3: /* 11 -> 10 */
 892                                         live->bitmapq[bindex] &=
 893                                             ~((hammer2_bitmap_t)1 << scount);
 894                                         ++cbinfo->count_11_10;
 895                                         fixup_dio(cbinfo->hmp, data_off,
 896                                                   bindex, scount);
 897                                         break;
 898                                 }
 899                         } else if ((mmask & 3) == 3) {
 900                                 /*
 901                                  * in-memory 11         live 10 -> 11
 902                                  *                      live ** -> 11
 903                                  *
 904                                  * Storage might be incorrectly marked free
 905                                  * or staged and must be remarked fully
 906                                  * allocated.
 907                                  */
 908                                 switch (lmask & 3) {
 909                                 case 0: /* 00 */
 910                                         ++cbinfo->count_00_11;
 911                                         cbinfo->adj_free -=
 912                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 913                                         live->avail -=
 914                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 915                                         if ((int32_t)live->avail < 0)
 916                                                 live->avail = 0;
 917                                         break;
 918                                 case 1: /* 01 */
 919                                         ++cbinfo->count_01_11;
 920                                         break;
 921                                 case 2: /* 10 -> 11 */
 922                                         ++cbinfo->count_10_11;
 923                                         break;
 924                                 case 3: /* 11 */
 925                                         break;
 926                                 }
 927                                 live->bitmapq[bindex] |=
 928                                         ((hammer2_bitmap_t)3 << scount);
 929                         }
 930                         mmask >>= 2;
 931                         lmask >>= 2;
 932                 }
 933         }
 934
 935         /*
 936          * Determine if the live bitmap is completely free and reset its
 937          * fields if so.  Otherwise check to see if we can reduce the linear
 938          * offset.
 939          */
 940         for (bindex = HAMMER2_BMAP_ELEMENTS - 1; bindex >= 0; --bindex) {
 941                 if (live->bitmapq[bindex] != 0)
 942                         break;
 943         }
 944         if (nofree) {
 945                 /* do nothing */
 946         } else if (bindex < 0) {
 947                 live->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 948                 live->class = 0;
 949                 live->linear = 0;
 950                 ++cbinfo->count_l0cleans;
 951         } else if (bindex < 7) {
 952                 ++bindex;
 953                 if (live->linear > bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 954                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 955                         ++cbinfo->count_linadjusts;
 956                 }
 957
 958                 /*
 959                  * XXX this fine-grained measure still has some issues.
 960                  */
 961                 if (live->linear < bindex * HAMMER2_FREEMAP_BLOCK_SIZE) {
 962                         live->linear = bindex * HAMMER2_FREEMAP_BLOCK_SIZE;
 963                         ++cbinfo->count_linadjusts;
 964                 }
 965         } else {
 966                 live->linear = HAMMER2_SEGSIZE;
 967         }
 968
 969 #if 0
 970         if (bmap->class) {
 971                 kprintf("%016jx %04d.%04x (avail=%7d) "
 972                         "%08x %08x %08x %08x %08x %08x %08x %08x\n",
 973                         (intmax_t)data_off,
 974                         (int)((data_off &
 975                                HAMMER2_FREEMAP_LEVEL1_MASK) >>
 976                               HAMMER2_FREEMAP_LEVEL0_RADIX),
 977                         bmap->class,
 978                         bmap->avail,
 979                         bmap->bitmap[0], bmap->bitmap[1],
 980                         bmap->bitmap[2], bmap->bitmap[3],
 981                         bmap->bitmap[4], bmap->bitmap[5],
 982                         bmap->bitmap[6], bmap->bitmap[7]);
 983         }
 984 #endif
 985 }
 986
 987 /*
 988  * BULKFREE DEDUP HEURISTIC
 989  *
 990  * WARNING! This code is SMP safe but the heuristic allows SMP collisions.
 991  *          All fields must be loaded into locals and validated.
 992  */
 993 static
 994 int
 995 h2_bulkfree_test(hammer2_bulkfree_info_t *cbinfo, hammer2_blockref_t *bref,
 996                  int pri)
 997 {
 998         hammer2_dedup_t *dedup;
 999         int best;
1000         int n;
1001         int i;
1002
1003         n = hammer2_icrc32(&bref->data_off, sizeof(bref->data_off));
1004         dedup = cbinfo->dedup + (n & (HAMMER2_DEDUP_HEUR_MASK & ~7));
1005
1006         for (i = best = 0; i < 8; ++i) {
1007                 if (dedup[i].data_off == bref->data_off) {
1008                         if (dedup[i].ticks < pri)
1009                                 dedup[i].ticks = pri;
1010                         if (pri == 1)
1011                                 cbinfo->count_dedup_factor += dedup[i].ticks;
1012                         return 1;
1013                 }
1014                 if (dedup[i].ticks < dedup[best].ticks)
1015                         best = i;
1016         }
1017         dedup[best].data_off = bref->data_off;
1018         dedup[best].ticks = pri;
1019
1020         return 0;
1021 }