sys/vfs/hammer2/hammer2_freemap.c

   1 /*
   2  * Copyright (c) 2011-2018 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  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/fcntl.h>
  39 #include <sys/buf.h>
  40 #include <sys/proc.h>
  41 #include <sys/namei.h>
  42 #include <sys/mount.h>
  43 #include <sys/vnode.h>
  44 #include <sys/mountctl.h>
  45
  46 #include "hammer2.h"
  47
  48 #define FREEMAP_DEBUG   0
  49
  50 struct hammer2_fiterate {
  51         hammer2_off_t   bpref;
  52         hammer2_off_t   bnext;
  53         int             loops;
  54         int             relaxed;
  55 };
  56
  57 typedef struct hammer2_fiterate hammer2_fiterate_t;
  58
  59 static int hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
  60                         hammer2_blockref_t *bref, int radix,
  61                         hammer2_fiterate_t *iter, hammer2_tid_t mtid);
  62 static void hammer2_freemap_init(hammer2_dev_t *hmp,
  63                         hammer2_key_t key, hammer2_chain_t *chain);
  64 static int hammer2_bmap_alloc(hammer2_dev_t *hmp,
  65                         hammer2_bmap_data_t *bmap, uint16_t class,
  66                         int n, int sub_key, int radix, hammer2_key_t *basep);
  67 static int hammer2_freemap_iterate(hammer2_chain_t **parentp,
  68                         hammer2_chain_t **chainp,
  69                         hammer2_fiterate_t *iter);
  70
  71 /*
  72  * Calculate the device offset for the specified FREEMAP_NODE or FREEMAP_LEAF
  73  * bref.  Return a combined media offset and physical size radix.  Freemap
  74  * chains use fixed storage offsets in the 4MB reserved area at the
  75  * beginning of each 2GB zone
  76  *
  77  * Rotate between four possibilities.  Theoretically this means we have three
  78  * good freemaps in case of a crash which we can use as a base for the fixup
  79  * scan at mount-time.
  80  */
  81 static
  82 int
  83 hammer2_freemap_reserve(hammer2_chain_t *chain, int radix)
  84 {
  85         hammer2_blockref_t *bref = &chain->bref;
  86         hammer2_off_t off;
  87         int index;
  88         int index_inc;
  89         size_t bytes;
  90
  91         /*
  92          * Physical allocation size.
  93          */
  94         bytes = (size_t)1 << radix;
  95
  96         /*
  97          * Calculate block selection index 0..7 of current block.  If this
  98          * is the first allocation of the block (verses a modification of an
  99          * existing block), we use index 0, otherwise we use the next rotating
 100          * index.
 101          */
 102         if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 103                 index = 0;
 104         } else {
 105                 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 106                       HAMMER2_FREEMAP_LEVEL1_MASK;
 107                 off = off / HAMMER2_PBUFSIZE;
 108                 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
 109                          off < HAMMER2_ZONE_FREEMAP_END);
 110                 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) /
 111                         HAMMER2_ZONE_FREEMAP_INC;
 112                 KKASSERT(index >= 0 && index < HAMMER2_NFREEMAPS);
 113                 if (++index == HAMMER2_NFREEMAPS)
 114                         index = 0;
 115         }
 116
 117         /*
 118          * Calculate the block offset of the reserved block.  This will
 119          * point into the 4MB reserved area at the base of the appropriate
 120          * 2GB zone, once added to the FREEMAP_x selection above.
 121          */
 122         index_inc = index * HAMMER2_ZONE_FREEMAP_INC;
 123
 124         switch(bref->keybits) {
 125         /* case HAMMER2_FREEMAP_LEVEL6_RADIX: not applicable */
 126         case HAMMER2_FREEMAP_LEVEL5_RADIX:      /* 4EB */
 127                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 128                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 129                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL5_RADIX) +
 130                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 131                        HAMMER2_ZONEFM_LEVEL5) * HAMMER2_PBUFSIZE;
 132                 break;
 133         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 16PB */
 134                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 135                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 136                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 137                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 138                        HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
 139                 break;
 140         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 64TB */
 141                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 142                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 143                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 144                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 145                        HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
 146                 break;
 147         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 256GB */
 148                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 149                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 150                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 151                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 152                        HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
 153                 break;
 154         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 1GB */
 155                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 156                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 157                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 158                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 159                        HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
 160                 break;
 161         default:
 162                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 163                 /* NOT REACHED */
 164                 off = (hammer2_off_t)-1;
 165                 break;
 166         }
 167         bref->data_off = off | radix;
 168 #if FREEMAP_DEBUG
 169         kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 170                 bref->type, bref->key, bref->keybits, bref->data_off);
 171 #endif
 172         return (0);
 173 }
 174
 175 /*
 176  * Normal freemap allocator
 177  *
 178  * Use available hints to allocate space using the freemap.  Create missing
 179  * freemap infrastructure on-the-fly as needed (including marking initial
 180  * allocations using the iterator as allocated, instantiating new 2GB zones,
 181  * and dealing with the end-of-media edge case).
 182  *
 183  * ip and bpref are only used as a heuristic to determine locality of
 184  * reference.  bref->key may also be used heuristically.
 185  *
 186  * This function is a NOP if bytes is 0.
 187  */
 188 int
 189 hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes)
 190 {
 191         hammer2_dev_t *hmp = chain->hmp;
 192         hammer2_blockref_t *bref = &chain->bref;
 193         hammer2_chain_t *parent;
 194         hammer2_tid_t mtid;
 195         int radix;
 196         int error;
 197         unsigned int hindex;
 198         hammer2_fiterate_t iter;
 199
 200         /*
 201          * If allocating or downsizing to zero we just get rid of whatever
 202          * data_off we had.
 203          */
 204         if (bytes == 0) {
 205                 chain->bref.data_off = 0;
 206                 return 0;
 207         }
 208
 209         KKASSERT(hmp->spmp);
 210         mtid = hammer2_trans_sub(hmp->spmp);
 211
 212         /*
 213          * Validate the allocation size.  It must be a power of 2.
 214          *
 215          * For now require that the caller be aware of the minimum
 216          * allocation (1K).
 217          */
 218         radix = hammer2_getradix(bytes);
 219         KKASSERT((size_t)1 << radix == bytes);
 220
 221         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 222             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 223                 /*
 224                  * Freemap blocks themselves are assigned from the reserve
 225                  * area, not allocated from the freemap.
 226                  */
 227                 error = hammer2_freemap_reserve(chain, radix);
 228
 229                 return error;
 230         }
 231
 232         KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
 233
 234         /*
 235          * Calculate the starting point for our allocation search.
 236          *
 237          * Each freemap leaf is dedicated to a specific freemap_radix.
 238          * The freemap_radix can be more fine-grained than the device buffer
 239          * radix which results in inodes being grouped together in their
 240          * own segment, terminal-data (16K or less) and initial indirect
 241          * block being grouped together, and then full-indirect and full-data
 242          * blocks (64K) being grouped together.
 243          *
 244          * The single most important aspect of this is the inode grouping
 245          * because that is what allows 'find' and 'ls' and other filesystem
 246          * topology operations to run fast.
 247          */
 248 #if 0
 249         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 250                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 251         else if (trans->tmp_bpref)
 252                 bpref = trans->tmp_bpref;
 253         else if (trans->tmp_ip)
 254                 bpref = trans->tmp_ip->chain->bref.data_off;
 255         else
 256 #endif
 257         /*
 258          * Heuristic tracking index.  We would like one for each distinct
 259          * bref type if possible.  heur_freemap[] has room for two classes
 260          * for each type.  At a minimum we have to break-up our heuristic
 261          * by device block sizes.
 262          */
 263         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 264         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 265         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 266         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 267         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_SIZE);
 268
 269         iter.bpref = hmp->heur_freemap[hindex];
 270         iter.relaxed = hmp->freemap_relaxed;
 271
 272         /*
 273          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 274          * reserved area, the try code will iterate past it.
 275          */
 276         if (iter.bpref > hmp->voldata.volu_size)
 277                 iter.bpref = hmp->voldata.volu_size - 1;
 278
 279         /*
 280          * Iterate the freemap looking for free space before and after.
 281          */
 282         parent = &hmp->fchain;
 283         hammer2_chain_ref(parent);
 284         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 285         error = HAMMER2_ERROR_EAGAIN;
 286         iter.bnext = iter.bpref;
 287         iter.loops = 0;
 288
 289         while (error == HAMMER2_ERROR_EAGAIN) {
 290                 error = hammer2_freemap_try_alloc(&parent, bref, radix,
 291                                                   &iter, mtid);
 292         }
 293         hmp->freemap_relaxed |= iter.relaxed;   /* heuristical, SMP race ok */
 294         hmp->heur_freemap[hindex] = iter.bnext;
 295         hammer2_chain_unlock(parent);
 296         hammer2_chain_drop(parent);
 297
 298         return (error);
 299 }
 300
 301 static int
 302 hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
 303                           hammer2_blockref_t *bref, int radix,
 304                           hammer2_fiterate_t *iter, hammer2_tid_t mtid)
 305 {
 306         hammer2_dev_t *hmp = (*parentp)->hmp;
 307         hammer2_off_t l0size;
 308         hammer2_off_t l1size;
 309         hammer2_off_t l1mask;
 310         hammer2_key_t key_dummy;
 311         hammer2_chain_t *chain;
 312         hammer2_off_t key;
 313         size_t bytes;
 314         uint16_t class;
 315         int error;
 316
 317         /*
 318          * Calculate the number of bytes being allocated, the number
 319          * of contiguous bits of bitmap being allocated, and the bitmap
 320          * mask.
 321          *
 322          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 323          *          mask calculation.
 324          */
 325         bytes = (size_t)1 << radix;
 326         class = (bref->type << 8) | hammer2_devblkradix(radix);
 327
 328         /*
 329          * Lookup the level1 freemap chain, creating and initializing one
 330          * if necessary.  Intermediate levels will be created automatically
 331          * when necessary by hammer2_chain_create().
 332          */
 333         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 334         l0size = HAMMER2_FREEMAP_LEVEL0_SIZE;
 335         l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
 336         l1mask = l1size - 1;
 337
 338         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 339                                      &error,
 340                                      HAMMER2_LOOKUP_ALWAYS |
 341                                      HAMMER2_LOOKUP_MATCHIND);
 342
 343         if (chain == NULL) {
 344                 /*
 345                  * Create the missing leaf, be sure to initialize
 346                  * the auxillary freemap tracking information in
 347                  * the bref.check.freemap structure.
 348                  */
 349 #if 0
 350                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 351                         key, iter->bpref);
 352 #endif
 353                 error = hammer2_chain_create(parentp, &chain, NULL, hmp->spmp,
 354                                      HAMMER2_METH_DEFAULT,
 355                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 356                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 357                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 358                                      mtid, 0, 0);
 359                 KKASSERT(error == 0);
 360                 if (error == 0) {
 361                         hammer2_chain_modify(chain, mtid, 0, 0);
 362                         bzero(&chain->data->bmdata[0],
 363                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 364                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 365                         chain->bref.check.freemap.avail = l1size;
 366                         /* bref.methods should already be inherited */
 367
 368                         hammer2_freemap_init(hmp, key, chain);
 369                 }
 370         } else if (chain->error) {
 371                 /*
 372                  * Error during lookup.
 373                  */
 374                 kprintf("hammer2_freemap_try_alloc: %016jx: error %s\n",
 375                         (intmax_t)bref->data_off,
 376                         hammer2_error_str(chain->error));
 377                 error = HAMMER2_ERROR_EIO;
 378         } else if ((chain->bref.check.freemap.bigmask &
 379                    ((size_t)1 << radix)) == 0) {
 380                 /*
 381                  * Already flagged as not having enough space
 382                  */
 383                 error = HAMMER2_ERROR_ENOSPC;
 384         } else {
 385                 /*
 386                  * Modify existing chain to setup for adjustment.
 387                  */
 388                 hammer2_chain_modify(chain, mtid, 0, 0);
 389         }
 390
 391         /*
 392          * Scan 4MB entries.
 393          */
 394         if (error == 0) {
 395                 hammer2_bmap_data_t *bmap;
 396                 hammer2_key_t base_key;
 397                 int count;
 398                 int start;
 399                 int n;
 400
 401                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 402                 start = (int)((iter->bnext - key) >>
 403                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 404                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 405                 hammer2_chain_modify(chain, mtid, 0, 0);
 406
 407                 error = HAMMER2_ERROR_ENOSPC;
 408                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 409                         int availchk;
 410
 411                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 412                             start - count < 0) {
 413                                 break;
 414                         }
 415
 416                         /*
 417                          * Calculate bmap pointer from thart starting index
 418                          * forwards.
 419                          *
 420                          * NOTE: bmap pointer is invalid if n >= FREEMAP_COUNT.
 421                          */
 422                         n = start + count;
 423                         bmap = &chain->data->bmdata[n];
 424
 425                         if (n >= HAMMER2_FREEMAP_COUNT) {
 426                                 availchk = 0;
 427                         } else if (bmap->avail) {
 428                                 availchk = 1;
 429                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 430                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 431                                 availchk = 1;
 432                         } else {
 433                                 availchk = 0;
 434                         }
 435
 436                         /*
 437                          * Try to allocate from a matching freemap class
 438                          * superblock.  If we are in relaxed mode we allocate
 439                          * from any freemap class superblock.
 440                          */
 441                         if (availchk &&
 442                             (bmap->class == 0 || bmap->class == class ||
 443                              iter->relaxed)) {
 444                                 base_key = key + n * l0size;
 445                                 error = hammer2_bmap_alloc(hmp, bmap,
 446                                                            class, n,
 447                                                            (int)bref->key,
 448                                                            radix,
 449                                                            &base_key);
 450                                 if (error != HAMMER2_ERROR_ENOSPC) {
 451                                         key = base_key;
 452                                         break;
 453                                 }
 454                         }
 455
 456                         /*
 457                          * Calculate bmap pointer from thart starting index
 458                          * backwards (locality).
 459                          *
 460                          * Must recalculate after potentially having called
 461                          * hammer2_bmap_alloc() above in case chain was
 462                          * reallocated.
 463                          *
 464                          * NOTE: bmap pointer is invalid if n < 0.
 465                          */
 466                         n = start - count;
 467                         bmap = &chain->data->bmdata[n];
 468                         if (n < 0) {
 469                                 availchk = 0;
 470                         } else if (bmap->avail) {
 471                                 availchk = 1;
 472                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 473                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 474                                 availchk = 1;
 475                         } else {
 476                                 availchk = 0;
 477                         }
 478
 479                         /*
 480                          * Try to allocate from a matching freemap class
 481                          * superblock.  If we are in relaxed mode we allocate
 482                          * from any freemap class superblock.
 483                          */
 484                         if (availchk &&
 485                             (bmap->class == 0 || bmap->class == class ||
 486                             iter->relaxed)) {
 487                                 base_key = key + n * l0size;
 488                                 error = hammer2_bmap_alloc(hmp, bmap,
 489                                                            class, n,
 490                                                            (int)bref->key,
 491                                                            radix,
 492                                                            &base_key);
 493                                 if (error != HAMMER2_ERROR_ENOSPC) {
 494                                         key = base_key;
 495                                         break;
 496                                 }
 497                         }
 498                 }
 499
 500                 /*
 501                  * We only know for sure that we can clear the bitmap bit
 502                  * if we scanned the entire array (start == 0).
 503                  */
 504                 if (error == HAMMER2_ERROR_ENOSPC && start == 0) {
 505                         chain->bref.check.freemap.bigmask &=
 506                                 (uint32_t)~((size_t)1 << radix);
 507                 }
 508                 /* XXX also scan down from original count */
 509         }
 510
 511         if (error == 0) {
 512                 /*
 513                  * Assert validity.  Must be beyond the static allocator used
 514                  * by newfs_hammer2 (and thus also beyond the aux area),
 515                  * not go past the volume size, and must not be in the
 516                  * reserved segment area for a zone.
 517                  */
 518                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 519                          key + bytes <= hmp->voldata.volu_size);
 520                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 521                 bref->data_off = key | radix;
 522
 523                 /*
 524                  * Record dedupability.  The dedup bits are cleared
 525                  * when bulkfree transitions the freemap from 11->10,
 526                  * and asserted to be clear on the 10->00 transition.
 527                  *
 528                  * We must record the bitmask with the chain locked
 529                  * at the time we set the allocation bits to avoid
 530                  * racing a bulkfree.
 531                  */
 532                 if (bref->type == HAMMER2_BREF_TYPE_DATA)
 533                         hammer2_io_dedup_set(hmp, bref);
 534 #if 0
 535                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 536                         chain,
 537                         bref->key, bref->data_off, chain->bref.data_off);
 538 #endif
 539         } else if (error == HAMMER2_ERROR_ENOSPC) {
 540                 /*
 541                  * Return EAGAIN with next iteration in iter->bnext, or
 542                  * return ENOSPC if the allocation map has been exhausted.
 543                  */
 544                 error = hammer2_freemap_iterate(parentp, &chain, iter);
 545         }
 546
 547         /*
 548          * Cleanup
 549          */
 550         if (chain) {
 551                 hammer2_chain_unlock(chain);
 552                 hammer2_chain_drop(chain);
 553         }
 554         return (error);
 555 }
 556
 557 /*
 558  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 559  *
 560  * If the linear iterator is mid-block we use it directly (the bitmap should
 561  * already be marked allocated), otherwise we search for a block in the
 562  * bitmap that fits the allocation request.
 563  *
 564  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 565  * to fully allocated and adjusts the linear allocator to allow the
 566  * remaining space to be allocated.
 567  *
 568  * sub_key is the lower 32 bits of the chain->bref.key for the chain whos
 569  * bref is being allocated.  If the radix represents an allocation >= 16KB
 570  * (aka HAMMER2_FREEMAP_BLOCK_RADIX) we try to use this key to select the
 571  * blocks directly out of the bmap.
 572  */
 573 static
 574 int
 575 hammer2_bmap_alloc(hammer2_dev_t *hmp, hammer2_bmap_data_t *bmap,
 576                    uint16_t class, int n, int sub_key,
 577                    int radix, hammer2_key_t *basep)
 578 {
 579         size_t size;
 580         size_t bgsize;
 581         int bmradix;
 582         hammer2_bitmap_t bmmask;
 583         int offset;
 584         int i;
 585         int j;
 586
 587         /*
 588          * Take into account 2-bits per block when calculating bmradix.
 589          */
 590         size = (size_t)1 << radix;
 591
 592         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 593                 bmradix = 2;
 594                 /* (16K) 2 bits per allocation block */
 595         } else {
 596                 bmradix = (hammer2_bitmap_t)2 <<
 597                           (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 598                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 599         }
 600
 601         /*
 602          * Use the linear iterator to pack small allocations, otherwise
 603          * fall-back to finding a free 16KB chunk.  The linear iterator
 604          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 605          * If it is the bitmap must be scanned.  It can become invalid
 606          * once we pack to the boundary.  We adjust it after a bitmap
 607          * allocation only for sub-16KB allocations (so the perfectly good
 608          * previous value can still be used for fragments when 16KB+
 609          * allocations are made inbetween fragmentary allocations).
 610          *
 611          * Beware of hardware artifacts when bmradix == 64 (intermediate
 612          * result can wind up being '1' instead of '0' if hardware masks
 613          * bit-count & 63).
 614          *
 615          * NOTE: j needs to be even in the j= calculation.  As an artifact
 616          *       of the /2 division, our bitmask has to clear bit 0.
 617          *
 618          * NOTE: TODO this can leave little unallocatable fragments lying
 619          *       around.
 620          */
 621         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 622             HAMMER2_FREEMAP_BLOCK_SIZE &&
 623             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 624             bmap->linear < HAMMER2_SEGSIZE) {
 625                 /*
 626                  * Use linear iterator if it is not block-aligned to avoid
 627                  * wasting space.
 628                  *
 629                  * Calculate the bitmapq[] index (i) and calculate the
 630                  * shift count within the 64-bit bitmapq[] entry.
 631                  *
 632                  * The freemap block size is 16KB, but each bitmap
 633                  * entry is two bits so use a little trick to get
 634                  * a (j) shift of 0, 2, 4, ... 62 in 16KB chunks.
 635                  */
 636                 KKASSERT(bmap->linear >= 0 &&
 637                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 638                          (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
 639                 offset = bmap->linear;
 640                 i = offset / (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS);
 641                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 62;
 642                 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 643                          HAMMER2_BMAP_ALLONES :
 644                          ((hammer2_bitmap_t)1 << bmradix) - 1;
 645                 bmmask <<= j;
 646                 bmap->linear = offset + size;
 647         } else {
 648                 /*
 649                  * Try to index a starting point based on sub_key.  This
 650                  * attempts to restore sequential block ordering on-disk
 651                  * whenever possible, even if data is committed out of
 652                  * order.
 653                  *
 654                  * i - Index bitmapq[], full data range represented is
 655                  *     HAMMER2_BMAP_SIZE.
 656                  *
 657                  * j - Index within bitmapq[i], full data range represented is
 658                  *     HAMMER2_BMAP_INDEX_SIZE.
 659                  *
 660                  * WARNING!
 661                  */
 662                 i = -1;
 663                 j = -1;
 664
 665                 switch(class >> 8) {
 666                 case HAMMER2_BREF_TYPE_DATA:
 667                         if (radix >= HAMMER2_FREEMAP_BLOCK_RADIX) {
 668                                 i = (sub_key & HAMMER2_BMAP_MASK) /
 669                                     (HAMMER2_BMAP_SIZE / HAMMER2_BMAP_ELEMENTS);
 670                                 j = (sub_key & HAMMER2_BMAP_INDEX_MASK) /
 671                                     (HAMMER2_BMAP_INDEX_SIZE /
 672                                      HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 673                                 j = j * 2;
 674                         }
 675                         break;
 676                 case HAMMER2_BREF_TYPE_INODE:
 677                         break;
 678                 default:
 679                         break;
 680                 }
 681                 if (i >= 0) {
 682                         KKASSERT(i < HAMMER2_BMAP_ELEMENTS &&
 683                                  j < 2 * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 684                         KKASSERT(j + bmradix <= HAMMER2_BMAP_BITS_PER_ELEMENT);
 685                         bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 686                                  HAMMER2_BMAP_ALLONES :
 687                                  ((hammer2_bitmap_t)1 << bmradix) - 1;
 688                         bmmask <<= j;
 689
 690                         if ((bmap->bitmapq[i] & bmmask) == 0)
 691                                 goto success;
 692                 }
 693
 694                 /*
 695                  * General element scan.
 696                  *
 697                  * WARNING: (j) is iterating a bit index (by 2's)
 698                  */
 699                 for (i = 0; i < HAMMER2_BMAP_ELEMENTS; ++i) {
 700                         bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 701                                  HAMMER2_BMAP_ALLONES :
 702                                  ((hammer2_bitmap_t)1 << bmradix) - 1;
 703                         for (j = 0;
 704                              j < HAMMER2_BMAP_BITS_PER_ELEMENT;
 705                              j += bmradix) {
 706                                 if ((bmap->bitmapq[i] & bmmask) == 0)
 707                                         goto success;
 708                                 bmmask <<= bmradix;
 709                         }
 710                 }
 711                 /*fragments might remain*/
 712                 /*KKASSERT(bmap->avail == 0);*/
 713                 return (HAMMER2_ERROR_ENOSPC);
 714 success:
 715                 offset = i * (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS) +
 716                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 717                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 718                         bmap->linear = offset + size;
 719         }
 720
 721         /* 8 x (64/2) -> 256 x 16K -> 4MB */
 722         KKASSERT(i >= 0 && i < HAMMER2_BMAP_ELEMENTS);
 723
 724         /*
 725          * Optimize the buffer cache to avoid unnecessary read-before-write
 726          * operations.
 727          *
 728          * The device block size could be larger than the allocation size
 729          * so the actual bitmap test is somewhat more involved.  We have
 730          * to use a compatible buffer size for this operation.
 731          */
 732         if ((bmap->bitmapq[i] & bmmask) == 0 &&
 733             hammer2_devblksize(size) != size) {
 734                 size_t psize = hammer2_devblksize(size);
 735                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 736                 int pbmradix = (hammer2_bitmap_t)2 <<
 737                                         (hammer2_devblkradix(radix) -
 738                                HAMMER2_FREEMAP_BLOCK_RADIX);
 739                 hammer2_bitmap_t pbmmask;
 740                 int pradix = hammer2_getradix(psize);
 741
 742                 pbmmask = (pbmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 743                         HAMMER2_BMAP_ALLONES :
 744                         ((hammer2_bitmap_t)1 << pbmradix) - 1;
 745                 while ((pbmmask & bmmask) == 0)
 746                         pbmmask <<= pbmradix;
 747
 748 #if 0
 749                 kprintf("%016jx mask %016jx %016jx %016jx (%zd/%zd)\n",
 750                         *basep + offset, bmap->bitmapq[i],
 751                         pbmmask, bmmask, size, psize);
 752 #endif
 753
 754                 if ((bmap->bitmapq[i] & pbmmask) == 0) {
 755                         hammer2_io_t *dio;
 756
 757                         hammer2_io_newnz(hmp, class >> 8,
 758                                         (*basep + (offset & ~pmask)) |
 759                                         pradix, psize, &dio);
 760                         hammer2_io_putblk(&dio);
 761                 }
 762         }
 763
 764 #if 0
 765         /*
 766          * When initializing a new inode segment also attempt to initialize
 767          * an adjacent segment.  Be careful not to index beyond the array
 768          * bounds.
 769          *
 770          * We do this to try to localize inode accesses to improve
 771          * directory scan rates.  XXX doesn't improve scan rates.
 772          */
 773         if (size == HAMMER2_INODE_BYTES) {
 774                 if (n & 1) {
 775                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 776                                 bmap[-1].radix = radix;
 777                 } else {
 778                         if (bmap[1].radix == 0 && bmap[1].avail)
 779                                 bmap[1].radix = radix;
 780                 }
 781         }
 782 #endif
 783         /*
 784          * Calculate the bitmap-granular change in bgsize for the volume
 785          * header.  We cannot use the fine-grained change here because
 786          * the bulkfree code can't undo it.  If the bitmap element is already
 787          * marked allocated it has already been accounted for.
 788          */
 789         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 790                 if (bmap->bitmapq[i] & bmmask)
 791                         bgsize = 0;
 792                 else
 793                         bgsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 794         } else {
 795                 bgsize = size;
 796         }
 797
 798         /*
 799          * Adjust the bitmap, set the class (it might have been 0),
 800          * and available bytes, update the allocation offset (*basep)
 801          * from the L0 base to the actual offset.
 802          *
 803          * Do not override the class if doing a relaxed class allocation.
 804          *
 805          * avail must reflect the bitmap-granular availability.  The allocator
 806          * tests will also check the linear iterator.
 807          */
 808         bmap->bitmapq[i] |= bmmask;
 809         if (bmap->class == 0)
 810                 bmap->class = class;
 811         bmap->avail -= bgsize;
 812         *basep += offset;
 813
 814         /*
 815          * Adjust the volume header's allocator_free parameter.  This
 816          * parameter has to be fixed up by bulkfree which has no way to
 817          * figure out sub-16K chunking, so it must be adjusted by the
 818          * bitmap-granular size.
 819          */
 820         if (bgsize) {
 821                 hammer2_voldata_lock(hmp);
 822                 hammer2_voldata_modify(hmp);
 823                 hmp->voldata.allocator_free -= bgsize;
 824                 hammer2_voldata_unlock(hmp);
 825         }
 826
 827         return(0);
 828 }
 829
 830 /*
 831  * Initialize a freemap for the storage area (in bytes) that begins at (key).
 832  */
 833 static
 834 void
 835 hammer2_freemap_init(hammer2_dev_t *hmp, hammer2_key_t key,
 836                      hammer2_chain_t *chain)
 837 {
 838         hammer2_off_t l1size;
 839         hammer2_off_t lokey;
 840         hammer2_off_t hikey;
 841         hammer2_bmap_data_t *bmap;
 842         int count;
 843
 844         /*
 845          * LEVEL1 is 1GB, there are two level1 1GB freemaps per 2GB zone.
 846          */
 847         l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
 848
 849         /*
 850          * Calculate the portion of the 1GB map that should be initialized
 851          * as free.  Portions below or after will be initialized as allocated.
 852          * SEGMASK-align the areas so we don't have to worry about sub-scans
 853          * or endianess when using memset.
 854          *
 855          * WARNING! It is possible for lokey to be larger than hikey if the
 856          *          entire 2GB segment is within the static allocation.
 857          */
 858         /*
 859          * (1) Ensure that all statically allocated space from newfs_hammer2
 860          *     is marked allocated, and take it up to the level1 base for
 861          *     this key.
 862          */
 863         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 864                 ~HAMMER2_SEGMASK64;
 865         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX))
 866                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX);
 867
 868         /*
 869          * (2) Ensure that the reserved area is marked allocated (typically
 870          *     the first 4MB of each 2GB area being represented).  Since
 871          *     each LEAF represents 1GB of storage and the zone is 2GB, we
 872          *     have to adjust lowkey upward every other LEAF sequentially.
 873          */
 874         if (lokey < H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64)
 875                 lokey = H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64;
 876
 877         /*
 878          * (3) Ensure that any trailing space at the end-of-volume is marked
 879          *     allocated.
 880          */
 881         hikey = key + HAMMER2_FREEMAP_LEVEL1_SIZE;
 882         if (hikey > hmp->voldata.volu_size) {
 883                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 884         }
 885
 886         /*
 887          * Heuristic highest possible value
 888          */
 889         chain->bref.check.freemap.avail = HAMMER2_FREEMAP_LEVEL1_SIZE;
 890         bmap = &chain->data->bmdata[0];
 891
 892         /*
 893          * Initialize bitmap (bzero'd by caller)
 894          */
 895         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 896                 if (key < lokey || key >= hikey) {
 897                         memset(bmap->bitmapq, -1,
 898                                sizeof(bmap->bitmapq));
 899                         bmap->avail = 0;
 900                         bmap->linear = HAMMER2_SEGSIZE;
 901                         chain->bref.check.freemap.avail -=
 902                                 HAMMER2_FREEMAP_LEVEL0_SIZE;
 903                 } else {
 904                         bmap->avail = HAMMER2_FREEMAP_LEVEL0_SIZE;
 905                 }
 906                 key += HAMMER2_FREEMAP_LEVEL0_SIZE;
 907                 ++bmap;
 908         }
 909 }
 910
 911 /*
 912  * The current Level 1 freemap has been exhausted, iterate to the next
 913  * one, return ENOSPC if no freemaps remain.
 914  *
 915  * At least two loops are required.  If we are not in relaxed mode and
 916  * we run out of storage we enter relaxed mode and do a third loop.
 917  * The relaxed mode is recorded back in the hmp so once we enter the mode
 918  * we remain relaxed until stuff begins to get freed and only do 2 loops.
 919  *
 920  * XXX this should rotate back to the beginning to handle freed-up space
 921  * XXX or use intermediate entries to locate free space. TODO
 922  */
 923 static int
 924 hammer2_freemap_iterate(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
 925                         hammer2_fiterate_t *iter)
 926 {
 927         hammer2_dev_t *hmp = (*parentp)->hmp;
 928
 929         iter->bnext &= ~HAMMER2_FREEMAP_LEVEL1_MASK;
 930         iter->bnext += HAMMER2_FREEMAP_LEVEL1_SIZE;
 931         if (iter->bnext >= hmp->voldata.volu_size) {
 932                 iter->bnext = 0;
 933                 if (++iter->loops >= 2) {
 934                         if (iter->relaxed == 0)
 935                                 iter->relaxed = 1;
 936                         else
 937                                 return (HAMMER2_ERROR_ENOSPC);
 938                 }
 939         }
 940         return(HAMMER2_ERROR_EAGAIN);
 941 }
 942
 943 /*
 944  * Adjust the bit-pattern for data in the freemap bitmap according to
 945  * (how).  This code is called from on-mount recovery to fixup (mark
 946  * as allocated) blocks whos freemap upates might not have been committed
 947  * in the last crash and is used by the bulk freemap scan to stage frees.
 948  *
 949  * WARNING! Cannot be called with a empty-data bref (radix == 0).
 950  *
 951  * XXX currently disabled when how == 0 (the normal real-time case).  At
 952  * the moment we depend on the bulk freescan to actually free blocks.  It
 953  * will still call this routine with a non-zero how to stage possible frees
 954  * and to do the actual free.
 955  */
 956 void
 957 hammer2_freemap_adjust(hammer2_dev_t *hmp, hammer2_blockref_t *bref,
 958                        int how)
 959 {
 960         hammer2_off_t data_off = bref->data_off;
 961         hammer2_chain_t *chain;
 962         hammer2_chain_t *parent;
 963         hammer2_bmap_data_t *bmap;
 964         hammer2_key_t key;
 965         hammer2_key_t key_dummy;
 966         hammer2_off_t l0size;
 967         hammer2_off_t l1size;
 968         hammer2_off_t l1mask;
 969         hammer2_tid_t mtid;
 970         hammer2_bitmap_t *bitmap;
 971         const hammer2_bitmap_t bmmask00 = 0;
 972         hammer2_bitmap_t bmmask01;
 973         hammer2_bitmap_t bmmask10;
 974         hammer2_bitmap_t bmmask11;
 975         size_t bytes;
 976         uint16_t class;
 977         int radix;
 978         int start;
 979         int count;
 980         int modified = 0;
 981         int error;
 982         size_t bgsize = 0;
 983
 984         KKASSERT(how == HAMMER2_FREEMAP_DORECOVER);
 985
 986         KKASSERT(hmp->spmp);
 987         mtid = hammer2_trans_sub(hmp->spmp);
 988
 989         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
 990         KKASSERT(radix != 0);
 991         data_off &= ~HAMMER2_OFF_MASK_RADIX;
 992         KKASSERT(radix <= HAMMER2_RADIX_MAX);
 993
 994         if (radix)
 995                 bytes = (size_t)1 << radix;
 996         else
 997                 bytes = 0;
 998         class = (bref->type << 8) | hammer2_devblkradix(radix);
 999
1000         /*
1001          * We can't adjust the freemap for data allocations made by
1002          * newfs_hammer2.
1003          */
1004         if (data_off < hmp->voldata.allocator_beg)
1005                 return;
1006
1007         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
1008
1009         /*
1010          * Lookup the level1 freemap chain.  The chain must exist.
1011          */
1012         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
1013         l0size = HAMMER2_FREEMAP_LEVEL0_SIZE;
1014         l1size = HAMMER2_FREEMAP_LEVEL1_SIZE;
1015         l1mask = l1size - 1;
1016
1017         parent = &hmp->fchain;
1018         hammer2_chain_ref(parent);
1019         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1020
1021         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
1022                                      &error,
1023                                      HAMMER2_LOOKUP_ALWAYS |
1024                                      HAMMER2_LOOKUP_MATCHIND);
1025
1026         /*
1027          * Stop early if we are trying to free something but no leaf exists.
1028          */
1029         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
1030                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
1031                         (intmax_t)bref->data_off);
1032                 goto done;
1033         }
1034         if (chain->error) {
1035                 kprintf("hammer2_freemap_adjust: %016jx: error %s\n",
1036                         (intmax_t)bref->data_off,
1037                         hammer2_error_str(chain->error));
1038                 hammer2_chain_unlock(chain);
1039                 hammer2_chain_drop(chain);
1040                 chain = NULL;
1041                 goto done;
1042         }
1043
1044         /*
1045          * Create any missing leaf(s) if we are doing a recovery (marking
1046          * the block(s) as being allocated instead of being freed).  Be sure
1047          * to initialize the auxillary freemap tracking info in the
1048          * bref.check.freemap structure.
1049          */
1050         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
1051                 error = hammer2_chain_create(&parent, &chain, NULL, hmp->spmp,
1052                                      HAMMER2_METH_DEFAULT,
1053                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
1054                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
1055                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
1056                                      mtid, 0, 0);
1057
1058                 if (hammer2_debug & 0x0040) {
1059                         kprintf("fixup create chain %p %016jx:%d\n",
1060                                 chain, chain->bref.key, chain->bref.keybits);
1061                 }
1062
1063                 if (error == 0) {
1064                         error = hammer2_chain_modify(chain, mtid, 0, 0);
1065                         KKASSERT(error == 0);
1066                         bzero(&chain->data->bmdata[0],
1067                               HAMMER2_FREEMAP_LEVELN_PSIZE);
1068                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
1069                         chain->bref.check.freemap.avail = l1size;
1070                         /* bref.methods should already be inherited */
1071
1072                         hammer2_freemap_init(hmp, key, chain);
1073                 }
1074                 /* XXX handle error */
1075         }
1076
1077 #if FREEMAP_DEBUG
1078         kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
1079                 chain->bref.type, chain->bref.key,
1080                 chain->bref.keybits, chain->bref.data_off);
1081 #endif
1082
1083         /*
1084          * Calculate the bitmask (runs in 2-bit pairs).
1085          */
1086         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
1087         bmmask01 = (hammer2_bitmap_t)1 << start;
1088         bmmask10 = (hammer2_bitmap_t)2 << start;
1089         bmmask11 = (hammer2_bitmap_t)3 << start;
1090
1091         /*
1092          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
1093          * a bulk scan is able to roll them up.
1094          */
1095         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
1096                 count = 1;
1097                 if (how == HAMMER2_FREEMAP_DOREALFREE)
1098                         how = HAMMER2_FREEMAP_DOMAYFREE;
1099         } else {
1100                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
1101         }
1102
1103         /*
1104          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
1105          * a 4MB swath.  The bmap itself (LEVEL1) covers 2GB.
1106          *
1107          * Be sure to reset the linear iterator to ensure that the adjustment
1108          * is not ignored.
1109          */
1110 again:
1111         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
1112                                     (HAMMER2_FREEMAP_COUNT - 1)];
1113         bitmap = &bmap->bitmapq[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
1114
1115         if (modified)
1116                 bmap->linear = 0;
1117
1118         while (count) {
1119                 KKASSERT(bmmask11);
1120                 if (how == HAMMER2_FREEMAP_DORECOVER) {
1121                         /*
1122                          * Recovery request, mark as allocated.
1123                          */
1124                         if ((*bitmap & bmmask11) != bmmask11) {
1125                                 if (modified == 0) {
1126                                         hammer2_chain_modify(chain, mtid, 0, 0);
1127                                         modified = 1;
1128                                         goto again;
1129                                 }
1130                                 if ((*bitmap & bmmask11) == bmmask00) {
1131                                         bmap->avail -=
1132                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
1133                                         bgsize += HAMMER2_FREEMAP_BLOCK_SIZE;
1134                                 }
1135                                 if (bmap->class == 0)
1136                                         bmap->class = class;
1137                                 *bitmap |= bmmask11;
1138                                 if (hammer2_debug & 0x0040) {
1139                                         kprintf("hammer2_freemap_recover: "
1140                                                 "fixup type=%02x "
1141                                                 "block=%016jx/%zd\n",
1142                                                 bref->type, data_off, bytes);
1143                                 }
1144                         } else {
1145                                 /*
1146                                 kprintf("hammer2_freemap_recover:  good "
1147                                         "type=%02x block=%016jx/%zd\n",
1148                                         bref->type, data_off, bytes);
1149                                 */
1150                         }
1151                 }
1152 #if 0
1153                 /*
1154                  * XXX this stuff doesn't work, avail is miscalculated and
1155                  * code 10 means something else now.
1156                  */
1157                 else if ((*bitmap & bmmask11) == bmmask11) {
1158                         /*
1159                          * Mayfree/Realfree request and bitmap is currently
1160                          * marked as being fully allocated.
1161                          */
1162                         if (!modified) {
1163                                 hammer2_chain_modify(chain, 0);
1164                                 modified = 1;
1165                                 goto again;
1166                         }
1167                         if (how == HAMMER2_FREEMAP_DOREALFREE)
1168                                 *bitmap &= ~bmmask11;
1169                         else
1170                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
1171                 } else if ((*bitmap & bmmask11) == bmmask10) {
1172                         /*
1173                          * Mayfree/Realfree request and bitmap is currently
1174                          * marked as being possibly freeable.
1175                          */
1176                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
1177                                 if (!modified) {
1178                                         hammer2_chain_modify(chain, 0);
1179                                         modified = 1;
1180                                         goto again;
1181                                 }
1182                                 *bitmap &= ~bmmask11;
1183                         }
1184                 } else {
1185                         /*
1186                          * 01 - Not implemented, currently illegal state
1187                          * 00 - Not allocated at all, illegal free.
1188                          */
1189                         panic("hammer2_freemap_adjust: "
1190                               "Illegal state %08x(%08x)",
1191                               *bitmap, *bitmap & bmmask11);
1192                 }
1193 #endif
1194                 --count;
1195                 bmmask01 <<= 2;
1196                 bmmask10 <<= 2;
1197                 bmmask11 <<= 2;
1198         }
1199 #if HAMMER2_BMAP_ELEMENTS != 8
1200 #error "hammer2_freemap.c: HAMMER2_BMAP_ELEMENTS expected to be 8"
1201 #endif
1202         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
1203                 bmap->avail += 1 << radix;
1204                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
1205                 if (bmap->avail == HAMMER2_SEGSIZE &&
1206                     bmap->bitmapq[0] == 0 &&
1207                     bmap->bitmapq[1] == 0 &&
1208                     bmap->bitmapq[2] == 0 &&
1209                     bmap->bitmapq[3] == 0 &&
1210                     bmap->bitmapq[4] == 0 &&
1211                     bmap->bitmapq[5] == 0 &&
1212                     bmap->bitmapq[6] == 0 &&
1213                     bmap->bitmapq[7] == 0) {
1214                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1215                         kprintf("Freeseg %016jx\n", (intmax_t)key);
1216                         bmap->class = 0;
1217                 }
1218         }
1219
1220         /*
1221          * chain->bref.check.freemap.bigmask (XXX)
1222          *
1223          * Setting bigmask is a hint to the allocation code that there might
1224          * be something allocatable.  We also set this in recovery... it
1225          * doesn't hurt and we might want to use the hint for other validation
1226          * operations later on.
1227          *
1228          * We could calculate the largest possible allocation and set the
1229          * radii that could fit, but its easier just to set bigmask to -1.
1230          */
1231         if (modified) {
1232                 chain->bref.check.freemap.bigmask = -1;
1233                 hmp->freemap_relaxed = 0;       /* reset heuristic */
1234         }
1235
1236         hammer2_chain_unlock(chain);
1237         hammer2_chain_drop(chain);
1238 done:
1239         hammer2_chain_unlock(parent);
1240         hammer2_chain_drop(parent);
1241
1242         if (bgsize) {
1243                 hammer2_voldata_lock(hmp);
1244                 hammer2_voldata_modify(hmp);
1245                 hmp->voldata.allocator_free -= bgsize;
1246                 hammer2_voldata_unlock(hmp);
1247         }
1248 }
1249
1250 /*
1251  * Validate the freemap, in three stages.
1252  *
1253  * stage-1      ALLOCATED     -> POSSIBLY FREE
1254  *              POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1255  *
1256  *      This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1257  *      The POSSIBLY FREE state does not mean that a block is actually free
1258  *      and may be transitioned back to ALLOCATED in stage-2.
1259  *
1260  *      This is typically done during normal filesystem operations when
1261  *      something is deleted or a block is replaced.
1262  *
1263  *      This is done by bulkfree in-bulk after a memory-bounded meta-data
1264  *      scan to try to determine what might be freeable.
1265  *
1266  *      This can be done unconditionally through a freemap scan when the
1267  *      intention is to brute-force recover the proper state of the freemap.
1268  *
1269  * stage-2      POSSIBLY FREE -> ALLOCATED      (scan metadata topology)
1270  *
1271  *      This is done by bulkfree during a meta-data scan to ensure that
1272  *      all blocks still actually allocated by the filesystem are marked
1273  *      as such.
1274  *
1275  *      NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1276  *            the bulkfree stage-2 and stage-3 is running.  The live filesystem
1277  *            will use the alternative POSSIBLY FREE type (2) to prevent
1278  *            stage-3 from improperly transitioning unvetted possibly-free
1279  *            blocks to FREE.
1280  *
1281  * stage-3      POSSIBLY FREE (type 1) -> FREE  (scan freemap)
1282  *
1283  *      This is done by bulkfree to finalize POSSIBLY FREE states.
1284  *
1285  */