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  * XXX I made a mistake and made the reserved area begin at each LEVEL1 zone,
  78  *     which is on a 1GB demark.  This will eat a little more space but for
  79  *     now we retain compatibility and make FMZONEBASE every 1GB
  80  *
  81  *     (see same thing in hammer2_bulkfree.c near the top, as well as in
  82  *     newfs_hammer2).
  83  *
  84  * Rotate between four possibilities.  Theoretically this means we have three
  85  * good freemaps in case of a crash which we can use as a base for the fixup
  86  * scan at mount-time.
  87  */
  88 #define H2FMZONEBASE(key)       ((key) & ~HAMMER2_FREEMAP_LEVEL1_MASK)
  89 #define H2FMBASE(key, radix)    ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
  90 #define H2FMSHIFT(radix)        ((hammer2_off_t)1 << (radix))
  91
  92 static
  93 int
  94 hammer2_freemap_reserve(hammer2_chain_t *chain, int radix)
  95 {
  96         hammer2_blockref_t *bref = &chain->bref;
  97         hammer2_off_t off;
  98         int index;
  99         int index_inc;
 100         size_t bytes;
 101
 102         /*
 103          * Physical allocation size.
 104          */
 105         bytes = (size_t)1 << radix;
 106
 107         /*
 108          * Calculate block selection index 0..7 of current block.  If this
 109          * is the first allocation of the block (verses a modification of an
 110          * existing block), we use index 0, otherwise we use the next rotating
 111          * index.
 112          */
 113         if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 114                 index = 0;
 115         } else {
 116                 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 117                       (HAMMER2_FREEMAP_LEVEL1_SIZE - 1);
 118                 off = off / HAMMER2_PBUFSIZE;
 119                 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
 120                          off < HAMMER2_ZONE_FREEMAP_END);
 121                 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) /
 122                         HAMMER2_ZONE_FREEMAP_INC;
 123                 KKASSERT(index >= 0 && index < HAMMER2_NFREEMAPS);
 124                 if (++index == HAMMER2_NFREEMAPS)
 125                         index = 0;
 126         }
 127
 128         /*
 129          * Calculate the block offset of the reserved block.  This will
 130          * point into the 4MB reserved area at the base of the appropriate
 131          * 2GB zone, once added to the FREEMAP_x selection above.
 132          */
 133         index_inc = index * HAMMER2_ZONE_FREEMAP_INC;
 134
 135         switch(bref->keybits) {
 136         /* case HAMMER2_FREEMAP_LEVEL6_RADIX: not applicable */
 137         case HAMMER2_FREEMAP_LEVEL5_RADIX:      /* 4EB */
 138                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 139                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 140                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL5_RADIX) +
 141                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 142                        HAMMER2_ZONEFM_LEVEL5) * HAMMER2_PBUFSIZE;
 143                 break;
 144         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 16PB */
 145                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 146                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 147                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 148                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 149                        HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
 150                 break;
 151         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 64TB */
 152                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 153                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 154                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 155                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 156                        HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
 157                 break;
 158         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 256GB */
 159                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 160                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 161                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 162                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 163                        HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
 164                 break;
 165         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 1GB */
 166                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 167                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 168                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 169                       (index_inc + HAMMER2_ZONE_FREEMAP_00 +
 170                        HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
 171                 break;
 172         default:
 173                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 174                 /* NOT REACHED */
 175                 off = (hammer2_off_t)-1;
 176                 break;
 177         }
 178         bref->data_off = off | radix;
 179 #if FREEMAP_DEBUG
 180         kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 181                 bref->type, bref->key, bref->keybits, bref->data_off);
 182 #endif
 183         return (0);
 184 }
 185
 186 /*
 187  * Normal freemap allocator
 188  *
 189  * Use available hints to allocate space using the freemap.  Create missing
 190  * freemap infrastructure on-the-fly as needed (including marking initial
 191  * allocations using the iterator as allocated, instantiating new 2GB zones,
 192  * and dealing with the end-of-media edge case).
 193  *
 194  * ip and bpref are only used as a heuristic to determine locality of
 195  * reference.  bref->key may also be used heuristically.
 196  *
 197  * This function is a NOP if bytes is 0.
 198  */
 199 int
 200 hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes)
 201 {
 202         hammer2_dev_t *hmp = chain->hmp;
 203         hammer2_blockref_t *bref = &chain->bref;
 204         hammer2_chain_t *parent;
 205         hammer2_tid_t mtid;
 206         int radix;
 207         int error;
 208         unsigned int hindex;
 209         hammer2_fiterate_t iter;
 210
 211         /*
 212          * If allocating or downsizing to zero we just get rid of whatever
 213          * data_off we had.
 214          */
 215         if (bytes == 0) {
 216                 chain->bref.data_off = 0;
 217                 return 0;
 218         }
 219
 220         KKASSERT(hmp->spmp);
 221         mtid = hammer2_trans_sub(hmp->spmp);
 222
 223         /*
 224          * Validate the allocation size.  It must be a power of 2.
 225          *
 226          * For now require that the caller be aware of the minimum
 227          * allocation (1K).
 228          */
 229         radix = hammer2_getradix(bytes);
 230         KKASSERT((size_t)1 << radix == bytes);
 231
 232         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 233             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 234                 /*
 235                  * Freemap blocks themselves are assigned from the reserve
 236                  * area, not allocated from the freemap.
 237                  */
 238                 error = hammer2_freemap_reserve(chain, radix);
 239
 240                 return error;
 241         }
 242
 243         KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
 244
 245         /*
 246          * Calculate the starting point for our allocation search.
 247          *
 248          * Each freemap leaf is dedicated to a specific freemap_radix.
 249          * The freemap_radix can be more fine-grained than the device buffer
 250          * radix which results in inodes being grouped together in their
 251          * own segment, terminal-data (16K or less) and initial indirect
 252          * block being grouped together, and then full-indirect and full-data
 253          * blocks (64K) being grouped together.
 254          *
 255          * The single most important aspect of this is the inode grouping
 256          * because that is what allows 'find' and 'ls' and other filesystem
 257          * topology operations to run fast.
 258          */
 259 #if 0
 260         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 261                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 262         else if (trans->tmp_bpref)
 263                 bpref = trans->tmp_bpref;
 264         else if (trans->tmp_ip)
 265                 bpref = trans->tmp_ip->chain->bref.data_off;
 266         else
 267 #endif
 268         /*
 269          * Heuristic tracking index.  We would like one for each distinct
 270          * bref type if possible.  heur_freemap[] has room for two classes
 271          * for each type.  At a minimum we have to break-up our heuristic
 272          * by device block sizes.
 273          */
 274         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 275         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 276         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 277         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 278         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_SIZE);
 279
 280         iter.bpref = hmp->heur_freemap[hindex];
 281         iter.relaxed = hmp->freemap_relaxed;
 282
 283         /*
 284          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 285          * reserved area, the try code will iterate past it.
 286          */
 287         if (iter.bpref > hmp->voldata.volu_size)
 288                 iter.bpref = hmp->voldata.volu_size - 1;
 289
 290         /*
 291          * Iterate the freemap looking for free space before and after.
 292          */
 293         parent = &hmp->fchain;
 294         hammer2_chain_ref(parent);
 295         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 296         error = HAMMER2_ERROR_EAGAIN;
 297         iter.bnext = iter.bpref;
 298         iter.loops = 0;
 299
 300         while (error == HAMMER2_ERROR_EAGAIN) {
 301                 error = hammer2_freemap_try_alloc(&parent, bref, radix,
 302                                                   &iter, mtid);
 303         }
 304         hmp->freemap_relaxed |= iter.relaxed;   /* heuristical, SMP race ok */
 305         hmp->heur_freemap[hindex] = iter.bnext;
 306         hammer2_chain_unlock(parent);
 307         hammer2_chain_drop(parent);
 308
 309         return (error);
 310 }
 311
 312 static int
 313 hammer2_freemap_try_alloc(hammer2_chain_t **parentp,
 314                           hammer2_blockref_t *bref, int radix,
 315                           hammer2_fiterate_t *iter, hammer2_tid_t mtid)
 316 {
 317         hammer2_dev_t *hmp = (*parentp)->hmp;
 318         hammer2_off_t l0size;
 319         hammer2_off_t l1size;
 320         hammer2_off_t l1mask;
 321         hammer2_key_t key_dummy;
 322         hammer2_chain_t *chain;
 323         hammer2_off_t key;
 324         size_t bytes;
 325         uint16_t class;
 326         int error;
 327
 328         /*
 329          * Calculate the number of bytes being allocated, the number
 330          * of contiguous bits of bitmap being allocated, and the bitmap
 331          * mask.
 332          *
 333          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 334          *          mask calculation.
 335          */
 336         bytes = (size_t)1 << radix;
 337         class = (bref->type << 8) | hammer2_devblkradix(radix);
 338
 339         /*
 340          * Lookup the level1 freemap chain, creating and initializing one
 341          * if necessary.  Intermediate levels will be created automatically
 342          * when necessary by hammer2_chain_create().
 343          */
 344         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 345         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 346         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 347         l1mask = l1size - 1;
 348
 349         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 350                                      &error,
 351                                      HAMMER2_LOOKUP_ALWAYS |
 352                                      HAMMER2_LOOKUP_MATCHIND);
 353
 354         if (chain == NULL) {
 355                 /*
 356                  * Create the missing leaf, be sure to initialize
 357                  * the auxillary freemap tracking information in
 358                  * the bref.check.freemap structure.
 359                  */
 360 #if 0
 361                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 362                         key, iter->bpref);
 363 #endif
 364                 error = hammer2_chain_create(parentp, &chain, NULL, hmp->spmp,
 365                                      HAMMER2_METH_DEFAULT,
 366                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 367                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 368                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 369                                      mtid, 0, 0);
 370                 KKASSERT(error == 0);
 371                 if (error == 0) {
 372                         hammer2_chain_modify(chain, mtid, 0, 0);
 373                         bzero(&chain->data->bmdata[0],
 374                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 375                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 376                         chain->bref.check.freemap.avail = l1size;
 377                         /* bref.methods should already be inherited */
 378
 379                         hammer2_freemap_init(hmp, key, chain);
 380                 }
 381         } else if (chain->error) {
 382                 /*
 383                  * Error during lookup.
 384                  */
 385                 kprintf("hammer2_freemap_try_alloc: %016jx: error %s\n",
 386                         (intmax_t)bref->data_off,
 387                         hammer2_error_str(chain->error));
 388                 error = HAMMER2_ERROR_EIO;
 389         } else if ((chain->bref.check.freemap.bigmask &
 390                    ((size_t)1 << radix)) == 0) {
 391                 /*
 392                  * Already flagged as not having enough space
 393                  */
 394                 error = HAMMER2_ERROR_ENOSPC;
 395         } else {
 396                 /*
 397                  * Modify existing chain to setup for adjustment.
 398                  */
 399                 hammer2_chain_modify(chain, mtid, 0, 0);
 400         }
 401
 402         /*
 403          * Scan 4MB entries.
 404          */
 405         if (error == 0) {
 406                 hammer2_bmap_data_t *bmap;
 407                 hammer2_key_t base_key;
 408                 int count;
 409                 int start;
 410                 int n;
 411
 412                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 413                 start = (int)((iter->bnext - key) >>
 414                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 415                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 416                 hammer2_chain_modify(chain, mtid, 0, 0);
 417
 418                 error = HAMMER2_ERROR_ENOSPC;
 419                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 420                         int availchk;
 421
 422                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 423                             start - count < 0) {
 424                                 break;
 425                         }
 426
 427                         /*
 428                          * Calculate bmap pointer from thart starting index
 429                          * forwards.
 430                          *
 431                          * NOTE: bmap pointer is invalid if n >= FREEMAP_COUNT.
 432                          */
 433                         n = start + count;
 434                         bmap = &chain->data->bmdata[n];
 435
 436                         if (n >= HAMMER2_FREEMAP_COUNT) {
 437                                 availchk = 0;
 438                         } else if (bmap->avail) {
 439                                 availchk = 1;
 440                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 441                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 442                                 availchk = 1;
 443                         } else {
 444                                 availchk = 0;
 445                         }
 446
 447                         /*
 448                          * Try to allocate from a matching freemap class
 449                          * superblock.  If we are in relaxed mode we allocate
 450                          * from any freemap class superblock.
 451                          */
 452                         if (availchk &&
 453                             (bmap->class == 0 || bmap->class == class ||
 454                              iter->relaxed)) {
 455                                 base_key = key + n * l0size;
 456                                 error = hammer2_bmap_alloc(hmp, bmap,
 457                                                            class, n,
 458                                                            (int)bref->key,
 459                                                            radix,
 460                                                            &base_key);
 461                                 if (error != HAMMER2_ERROR_ENOSPC) {
 462                                         key = base_key;
 463                                         break;
 464                                 }
 465                         }
 466
 467                         /*
 468                          * Calculate bmap pointer from thart starting index
 469                          * backwards (locality).
 470                          *
 471                          * Must recalculate after potentially having called
 472                          * hammer2_bmap_alloc() above in case chain was
 473                          * reallocated.
 474                          *
 475                          * NOTE: bmap pointer is invalid if n < 0.
 476                          */
 477                         n = start - count;
 478                         bmap = &chain->data->bmdata[n];
 479                         if (n < 0) {
 480                                 availchk = 0;
 481                         } else if (bmap->avail) {
 482                                 availchk = 1;
 483                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 484                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 485                                 availchk = 1;
 486                         } else {
 487                                 availchk = 0;
 488                         }
 489
 490                         /*
 491                          * Try to allocate from a matching freemap class
 492                          * superblock.  If we are in relaxed mode we allocate
 493                          * from any freemap class superblock.
 494                          */
 495                         if (availchk &&
 496                             (bmap->class == 0 || bmap->class == class ||
 497                             iter->relaxed)) {
 498                                 base_key = key + n * l0size;
 499                                 error = hammer2_bmap_alloc(hmp, bmap,
 500                                                            class, n,
 501                                                            (int)bref->key,
 502                                                            radix,
 503                                                            &base_key);
 504                                 if (error != HAMMER2_ERROR_ENOSPC) {
 505                                         key = base_key;
 506                                         break;
 507                                 }
 508                         }
 509                 }
 510
 511                 /*
 512                  * We only know for sure that we can clear the bitmap bit
 513                  * if we scanned the entire array (start == 0).
 514                  */
 515                 if (error == HAMMER2_ERROR_ENOSPC && start == 0) {
 516                         chain->bref.check.freemap.bigmask &=
 517                                 (uint32_t)~((size_t)1 << radix);
 518                 }
 519                 /* XXX also scan down from original count */
 520         }
 521
 522         if (error == 0) {
 523                 /*
 524                  * Assert validity.  Must be beyond the static allocator used
 525                  * by newfs_hammer2 (and thus also beyond the aux area),
 526                  * not go past the volume size, and must not be in the
 527                  * reserved segment area for a zone.
 528                  */
 529                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 530                          key + bytes <= hmp->voldata.volu_size);
 531                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 532                 bref->data_off = key | radix;
 533
 534                 /*
 535                  * Record dedupability.  The dedup bits are cleared
 536                  * when bulkfree transitions the freemap from 11->10,
 537                  * and asserted to be clear on the 10->00 transition.
 538                  *
 539                  * We must record the bitmask with the chain locked
 540                  * at the time we set the allocation bits to avoid
 541                  * racing a bulkfree.
 542                  */
 543                 if (bref->type == HAMMER2_BREF_TYPE_DATA)
 544                         hammer2_io_dedup_set(hmp, bref);
 545 #if 0
 546                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 547                         chain,
 548                         bref->key, bref->data_off, chain->bref.data_off);
 549 #endif
 550         } else if (error == HAMMER2_ERROR_ENOSPC) {
 551                 /*
 552                  * Return EAGAIN with next iteration in iter->bnext, or
 553                  * return ENOSPC if the allocation map has been exhausted.
 554                  */
 555                 error = hammer2_freemap_iterate(parentp, &chain, iter);
 556         }
 557
 558         /*
 559          * Cleanup
 560          */
 561         if (chain) {
 562                 hammer2_chain_unlock(chain);
 563                 hammer2_chain_drop(chain);
 564         }
 565         return (error);
 566 }
 567
 568 /*
 569  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 570  *
 571  * If the linear iterator is mid-block we use it directly (the bitmap should
 572  * already be marked allocated), otherwise we search for a block in the
 573  * bitmap that fits the allocation request.
 574  *
 575  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 576  * to fully allocated and adjusts the linear allocator to allow the
 577  * remaining space to be allocated.
 578  *
 579  * sub_key is the lower 32 bits of the chain->bref.key for the chain whos
 580  * bref is being allocated.  If the radix represents an allocation >= 16KB
 581  * (aka HAMMER2_FREEMAP_BLOCK_RADIX) we try to use this key to select the
 582  * blocks directly out of the bmap.
 583  */
 584 static
 585 int
 586 hammer2_bmap_alloc(hammer2_dev_t *hmp, hammer2_bmap_data_t *bmap,
 587                    uint16_t class, int n, int sub_key,
 588                    int radix, hammer2_key_t *basep)
 589 {
 590         size_t size;
 591         size_t bgsize;
 592         int bmradix;
 593         hammer2_bitmap_t bmmask;
 594         int offset;
 595         int i;
 596         int j;
 597
 598         /*
 599          * Take into account 2-bits per block when calculating bmradix.
 600          */
 601         size = (size_t)1 << radix;
 602
 603         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 604                 bmradix = 2;
 605                 /* (16K) 2 bits per allocation block */
 606         } else {
 607                 bmradix = (hammer2_bitmap_t)2 <<
 608                           (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 609                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 610         }
 611
 612         /*
 613          * Use the linear iterator to pack small allocations, otherwise
 614          * fall-back to finding a free 16KB chunk.  The linear iterator
 615          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 616          * If it is the bitmap must be scanned.  It can become invalid
 617          * once we pack to the boundary.  We adjust it after a bitmap
 618          * allocation only for sub-16KB allocations (so the perfectly good
 619          * previous value can still be used for fragments when 16KB+
 620          * allocations are made inbetween fragmentary allocations).
 621          *
 622          * Beware of hardware artifacts when bmradix == 64 (intermediate
 623          * result can wind up being '1' instead of '0' if hardware masks
 624          * bit-count & 63).
 625          *
 626          * NOTE: j needs to be even in the j= calculation.  As an artifact
 627          *       of the /2 division, our bitmask has to clear bit 0.
 628          *
 629          * NOTE: TODO this can leave little unallocatable fragments lying
 630          *       around.
 631          */
 632         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 633             HAMMER2_FREEMAP_BLOCK_SIZE &&
 634             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 635             bmap->linear < HAMMER2_SEGSIZE) {
 636                 /*
 637                  * Use linear iterator if it is not block-aligned to avoid
 638                  * wasting space.
 639                  *
 640                  * Calculate the bitmapq[] index (i) and calculate the
 641                  * shift count within the 64-bit bitmapq[] entry.
 642                  *
 643                  * The freemap block size is 16KB, but each bitmap
 644                  * entry is two bits so use a little trick to get
 645                  * a (j) shift of 0, 2, 4, ... 62 in 16KB chunks.
 646                  */
 647                 KKASSERT(bmap->linear >= 0 &&
 648                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 649                          (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
 650                 offset = bmap->linear;
 651                 i = offset / (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS);
 652                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 62;
 653                 bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 654                          HAMMER2_BMAP_ALLONES :
 655                          ((hammer2_bitmap_t)1 << bmradix) - 1;
 656                 bmmask <<= j;
 657                 bmap->linear = offset + size;
 658         } else {
 659                 /*
 660                  * Try to index a starting point based on sub_key.  This
 661                  * attempts to restore sequential block ordering on-disk
 662                  * whenever possible, even if data is committed out of
 663                  * order.
 664                  *
 665                  * i - Index bitmapq[], full data range represented is
 666                  *     HAMMER2_BMAP_SIZE.
 667                  *
 668                  * j - Index within bitmapq[i], full data range represented is
 669                  *     HAMMER2_BMAP_INDEX_SIZE.
 670                  *
 671                  * WARNING!
 672                  */
 673                 i = -1;
 674                 j = -1;
 675
 676                 switch(class >> 8) {
 677                 case HAMMER2_BREF_TYPE_DATA:
 678                         if (radix >= HAMMER2_FREEMAP_BLOCK_RADIX) {
 679                                 i = (sub_key & HAMMER2_BMAP_MASK) /
 680                                     (HAMMER2_BMAP_SIZE / HAMMER2_BMAP_ELEMENTS);
 681                                 j = (sub_key & HAMMER2_BMAP_INDEX_MASK) /
 682                                     (HAMMER2_BMAP_INDEX_SIZE /
 683                                      HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 684                                 j = j * 2;
 685                         }
 686                         break;
 687                 case HAMMER2_BREF_TYPE_INODE:
 688                         break;
 689                 default:
 690                         break;
 691                 }
 692                 if (i >= 0) {
 693                         KKASSERT(i < HAMMER2_BMAP_ELEMENTS &&
 694                                  j < 2 * HAMMER2_BMAP_BLOCKS_PER_ELEMENT);
 695                         KKASSERT(j + bmradix <= HAMMER2_BMAP_BITS_PER_ELEMENT);
 696                         bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 697                                  HAMMER2_BMAP_ALLONES :
 698                                  ((hammer2_bitmap_t)1 << bmradix) - 1;
 699                         bmmask <<= j;
 700
 701                         if ((bmap->bitmapq[i] & bmmask) == 0)
 702                                 goto success;
 703                 }
 704
 705                 /*
 706                  * General element scan.
 707                  *
 708                  * WARNING: (j) is iterating a bit index (by 2's)
 709                  */
 710                 for (i = 0; i < HAMMER2_BMAP_ELEMENTS; ++i) {
 711                         bmmask = (bmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 712                                  HAMMER2_BMAP_ALLONES :
 713                                  ((hammer2_bitmap_t)1 << bmradix) - 1;
 714                         for (j = 0;
 715                              j < HAMMER2_BMAP_BITS_PER_ELEMENT;
 716                              j += bmradix) {
 717                                 if ((bmap->bitmapq[i] & bmmask) == 0)
 718                                         goto success;
 719                                 bmmask <<= bmradix;
 720                         }
 721                 }
 722                 /*fragments might remain*/
 723                 /*KKASSERT(bmap->avail == 0);*/
 724                 return (HAMMER2_ERROR_ENOSPC);
 725 success:
 726                 offset = i * (HAMMER2_SEGSIZE / HAMMER2_BMAP_ELEMENTS) +
 727                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 728                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 729                         bmap->linear = offset + size;
 730         }
 731
 732         /* 8 x (64/2) -> 256 x 16K -> 4MB */
 733         KKASSERT(i >= 0 && i < HAMMER2_BMAP_ELEMENTS);
 734
 735         /*
 736          * Optimize the buffer cache to avoid unnecessary read-before-write
 737          * operations.
 738          *
 739          * The device block size could be larger than the allocation size
 740          * so the actual bitmap test is somewhat more involved.  We have
 741          * to use a compatible buffer size for this operation.
 742          */
 743         if ((bmap->bitmapq[i] & bmmask) == 0 &&
 744             hammer2_devblksize(size) != size) {
 745                 size_t psize = hammer2_devblksize(size);
 746                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 747                 int pbmradix = (hammer2_bitmap_t)2 <<
 748                                         (hammer2_devblkradix(radix) -
 749                                HAMMER2_FREEMAP_BLOCK_RADIX);
 750                 hammer2_bitmap_t pbmmask;
 751                 int pradix = hammer2_getradix(psize);
 752
 753                 pbmmask = (pbmradix == HAMMER2_BMAP_BITS_PER_ELEMENT) ?
 754                         HAMMER2_BMAP_ALLONES :
 755                         ((hammer2_bitmap_t)1 << pbmradix) - 1;
 756                 while ((pbmmask & bmmask) == 0)
 757                         pbmmask <<= pbmradix;
 758
 759 #if 0
 760                 kprintf("%016jx mask %016jx %016jx %016jx (%zd/%zd)\n",
 761                         *basep + offset, bmap->bitmapq[i],
 762                         pbmmask, bmmask, size, psize);
 763 #endif
 764
 765                 if ((bmap->bitmapq[i] & pbmmask) == 0) {
 766                         hammer2_io_t *dio;
 767
 768                         hammer2_io_newnz(hmp, class >> 8,
 769                                         (*basep + (offset & ~pmask)) |
 770                                         pradix, psize, &dio);
 771                         hammer2_io_putblk(&dio);
 772                 }
 773         }
 774
 775 #if 0
 776         /*
 777          * When initializing a new inode segment also attempt to initialize
 778          * an adjacent segment.  Be careful not to index beyond the array
 779          * bounds.
 780          *
 781          * We do this to try to localize inode accesses to improve
 782          * directory scan rates.  XXX doesn't improve scan rates.
 783          */
 784         if (size == HAMMER2_INODE_BYTES) {
 785                 if (n & 1) {
 786                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 787                                 bmap[-1].radix = radix;
 788                 } else {
 789                         if (bmap[1].radix == 0 && bmap[1].avail)
 790                                 bmap[1].radix = radix;
 791                 }
 792         }
 793 #endif
 794         /*
 795          * Calculate the bitmap-granular change in bgsize for the volume
 796          * header.  We cannot use the fine-grained change here because
 797          * the bulkfree code can't undo it.  If the bitmap element is already
 798          * marked allocated it has already been accounted for.
 799          */
 800         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 801                 if (bmap->bitmapq[i] & bmmask)
 802                         bgsize = 0;
 803                 else
 804                         bgsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 805         } else {
 806                 bgsize = size;
 807         }
 808
 809         /*
 810          * Adjust the bitmap, set the class (it might have been 0),
 811          * and available bytes, update the allocation offset (*basep)
 812          * from the L0 base to the actual offset.
 813          *
 814          * Do not override the class if doing a relaxed class allocation.
 815          *
 816          * avail must reflect the bitmap-granular availability.  The allocator
 817          * tests will also check the linear iterator.
 818          */
 819         bmap->bitmapq[i] |= bmmask;
 820         if (bmap->class == 0)
 821                 bmap->class = class;
 822         bmap->avail -= bgsize;
 823         *basep += offset;
 824
 825         /*
 826          * Adjust the volume header's allocator_free parameter.  This
 827          * parameter has to be fixed up by bulkfree which has no way to
 828          * figure out sub-16K chunking, so it must be adjusted by the
 829          * bitmap-granular size.
 830          */
 831         if (bgsize) {
 832                 hammer2_voldata_lock(hmp);
 833                 hammer2_voldata_modify(hmp);
 834                 hmp->voldata.allocator_free -= bgsize;
 835                 hammer2_voldata_unlock(hmp);
 836         }
 837
 838         return(0);
 839 }
 840
 841 /*
 842  * Initialize a freemap for the storage area (in bytes) that begins at (key).
 843  */
 844 static
 845 void
 846 hammer2_freemap_init(hammer2_dev_t *hmp, hammer2_key_t key,
 847                      hammer2_chain_t *chain)
 848 {
 849         hammer2_off_t l1size;
 850         hammer2_off_t lokey;
 851         hammer2_off_t hikey;
 852         hammer2_bmap_data_t *bmap;
 853         int count;
 854
 855         /*
 856          * LEVEL1 is 1GB, there are two level1 1GB freemaps per 2GB zone.
 857          */
 858         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 859
 860         /*
 861          * Calculate the portion of the 1GB map that should be initialized
 862          * as free.  Portions below or after will be initialized as allocated.
 863          * SEGMASK-align the areas so we don't have to worry about sub-scans
 864          * or endianess when using memset.
 865          *
 866          * WARNING! It is possible for lokey to be larger than hikey if the
 867          *          entire 2GB segment is within the static allocation.
 868          */
 869         /*
 870          * (1) Ensure that all statically allocated space from newfs_hammer2
 871          *     is marked allocated, and take it up to the level1 base for
 872          *     this key.
 873          */
 874         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 875                 ~HAMMER2_SEGMASK64;
 876         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX))
 877                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX);
 878
 879         /*
 880          * (2) Ensure that the reserved area is marked allocated (typically
 881          *     the first 4MB of each 2GB area being represented).  Since
 882          *     each LEAF represents 1GB of storage and the zone is 2GB, we
 883          *     have to adjust lowkey upward every other LEAF sequentially.
 884          */
 885         if (lokey < H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64)
 886                 lokey = H2FMZONEBASE(key) + HAMMER2_ZONE_SEG64;
 887
 888         /*
 889          * (3) Ensure that any trailing space at the end-of-volume is marked
 890          *     allocated.
 891          */
 892         hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 893         if (hikey > hmp->voldata.volu_size) {
 894                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 895         }
 896
 897         /*
 898          * Heuristic highest possible value
 899          */
 900         chain->bref.check.freemap.avail =
 901                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 902         bmap = &chain->data->bmdata[0];
 903
 904         /*
 905          * Initialize bitmap (bzero'd by caller)
 906          */
 907         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 908                 if (key < lokey || key >= hikey) {
 909                         memset(bmap->bitmapq, -1,
 910                                sizeof(bmap->bitmapq));
 911                         bmap->avail = 0;
 912                         bmap->linear = HAMMER2_SEGSIZE;
 913                         chain->bref.check.freemap.avail -=
 914                                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 915                 } else {
 916                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 917                 }
 918                 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 919                 ++bmap;
 920         }
 921 }
 922
 923 /*
 924  * The current Level 1 freemap has been exhausted, iterate to the next
 925  * one, return ENOSPC if no freemaps remain.
 926  *
 927  * At least two loops are required.  If we are not in relaxed mode and
 928  * we run out of storage we enter relaxed mode and do a third loop.
 929  * The relaxed mode is recorded back in the hmp so once we enter the mode
 930  * we remain relaxed until stuff begins to get freed and only do 2 loops.
 931  *
 932  * XXX this should rotate back to the beginning to handle freed-up space
 933  * XXX or use intermediate entries to locate free space. TODO
 934  */
 935 static int
 936 hammer2_freemap_iterate(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
 937                         hammer2_fiterate_t *iter)
 938 {
 939         hammer2_dev_t *hmp = (*parentp)->hmp;
 940
 941         iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 942         iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 943         if (iter->bnext >= hmp->voldata.volu_size) {
 944                 iter->bnext = 0;
 945                 if (++iter->loops >= 2) {
 946                         if (iter->relaxed == 0)
 947                                 iter->relaxed = 1;
 948                         else
 949                                 return (HAMMER2_ERROR_ENOSPC);
 950                 }
 951         }
 952         return(HAMMER2_ERROR_EAGAIN);
 953 }
 954
 955 /*
 956  * Adjust the bit-pattern for data in the freemap bitmap according to
 957  * (how).  This code is called from on-mount recovery to fixup (mark
 958  * as allocated) blocks whos freemap upates might not have been committed
 959  * in the last crash and is used by the bulk freemap scan to stage frees.
 960  *
 961  * WARNING! Cannot be called with a empty-data bref (radix == 0).
 962  *
 963  * XXX currently disabled when how == 0 (the normal real-time case).  At
 964  * the moment we depend on the bulk freescan to actually free blocks.  It
 965  * will still call this routine with a non-zero how to stage possible frees
 966  * and to do the actual free.
 967  */
 968 void
 969 hammer2_freemap_adjust(hammer2_dev_t *hmp, hammer2_blockref_t *bref,
 970                        int how)
 971 {
 972         hammer2_off_t data_off = bref->data_off;
 973         hammer2_chain_t *chain;
 974         hammer2_chain_t *parent;
 975         hammer2_bmap_data_t *bmap;
 976         hammer2_key_t key;
 977         hammer2_key_t key_dummy;
 978         hammer2_off_t l0size;
 979         hammer2_off_t l1size;
 980         hammer2_off_t l1mask;
 981         hammer2_tid_t mtid;
 982         hammer2_bitmap_t *bitmap;
 983         const hammer2_bitmap_t bmmask00 = 0;
 984         hammer2_bitmap_t bmmask01;
 985         hammer2_bitmap_t bmmask10;
 986         hammer2_bitmap_t bmmask11;
 987         size_t bytes;
 988         uint16_t class;
 989         int radix;
 990         int start;
 991         int count;
 992         int modified = 0;
 993         int error;
 994         size_t bgsize = 0;
 995
 996         KKASSERT(how == HAMMER2_FREEMAP_DORECOVER);
 997
 998         KKASSERT(hmp->spmp);
 999         mtid = hammer2_trans_sub(hmp->spmp);
1000
1001         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
1002         KKASSERT(radix != 0);
1003         data_off &= ~HAMMER2_OFF_MASK_RADIX;
1004         KKASSERT(radix <= HAMMER2_RADIX_MAX);
1005
1006         if (radix)
1007                 bytes = (size_t)1 << radix;
1008         else
1009                 bytes = 0;
1010         class = (bref->type << 8) | hammer2_devblkradix(radix);
1011
1012         /*
1013          * We can't adjust the freemap for data allocations made by
1014          * newfs_hammer2.
1015          */
1016         if (data_off < hmp->voldata.allocator_beg)
1017                 return;
1018
1019         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
1020
1021         /*
1022          * Lookup the level1 freemap chain.  The chain must exist.
1023          */
1024         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
1025         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
1026         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
1027         l1mask = l1size - 1;
1028
1029         parent = &hmp->fchain;
1030         hammer2_chain_ref(parent);
1031         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1032
1033         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
1034                                      &error,
1035                                      HAMMER2_LOOKUP_ALWAYS |
1036                                      HAMMER2_LOOKUP_MATCHIND);
1037
1038         /*
1039          * Stop early if we are trying to free something but no leaf exists.
1040          */
1041         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
1042                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
1043                         (intmax_t)bref->data_off);
1044                 goto done;
1045         }
1046         if (chain->error) {
1047                 kprintf("hammer2_freemap_adjust: %016jx: error %s\n",
1048                         (intmax_t)bref->data_off,
1049                         hammer2_error_str(chain->error));
1050                 hammer2_chain_unlock(chain);
1051                 hammer2_chain_drop(chain);
1052                 chain = NULL;
1053                 goto done;
1054         }
1055
1056         /*
1057          * Create any missing leaf(s) if we are doing a recovery (marking
1058          * the block(s) as being allocated instead of being freed).  Be sure
1059          * to initialize the auxillary freemap tracking info in the
1060          * bref.check.freemap structure.
1061          */
1062         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
1063                 error = hammer2_chain_create(&parent, &chain, NULL, hmp->spmp,
1064                                      HAMMER2_METH_DEFAULT,
1065                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
1066                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
1067                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
1068                                      mtid, 0, 0);
1069
1070                 if (hammer2_debug & 0x0040) {
1071                         kprintf("fixup create chain %p %016jx:%d\n",
1072                                 chain, chain->bref.key, chain->bref.keybits);
1073                 }
1074
1075                 if (error == 0) {
1076                         error = hammer2_chain_modify(chain, mtid, 0, 0);
1077                         KKASSERT(error == 0);
1078                         bzero(&chain->data->bmdata[0],
1079                               HAMMER2_FREEMAP_LEVELN_PSIZE);
1080                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
1081                         chain->bref.check.freemap.avail = l1size;
1082                         /* bref.methods should already be inherited */
1083
1084                         hammer2_freemap_init(hmp, key, chain);
1085                 }
1086                 /* XXX handle error */
1087         }
1088
1089 #if FREEMAP_DEBUG
1090         kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
1091                 chain->bref.type, chain->bref.key,
1092                 chain->bref.keybits, chain->bref.data_off);
1093 #endif
1094
1095         /*
1096          * Calculate the bitmask (runs in 2-bit pairs).
1097          */
1098         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
1099         bmmask01 = (hammer2_bitmap_t)1 << start;
1100         bmmask10 = (hammer2_bitmap_t)2 << start;
1101         bmmask11 = (hammer2_bitmap_t)3 << start;
1102
1103         /*
1104          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
1105          * a bulk scan is able to roll them up.
1106          */
1107         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
1108                 count = 1;
1109                 if (how == HAMMER2_FREEMAP_DOREALFREE)
1110                         how = HAMMER2_FREEMAP_DOMAYFREE;
1111         } else {
1112                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
1113         }
1114
1115         /*
1116          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
1117          * a 4MB swath.  The bmap itself (LEVEL1) covers 2GB.
1118          *
1119          * Be sure to reset the linear iterator to ensure that the adjustment
1120          * is not ignored.
1121          */
1122 again:
1123         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
1124                                     (HAMMER2_FREEMAP_COUNT - 1)];
1125         bitmap = &bmap->bitmapq[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
1126
1127         if (modified)
1128                 bmap->linear = 0;
1129
1130         while (count) {
1131                 KKASSERT(bmmask11);
1132                 if (how == HAMMER2_FREEMAP_DORECOVER) {
1133                         /*
1134                          * Recovery request, mark as allocated.
1135                          */
1136                         if ((*bitmap & bmmask11) != bmmask11) {
1137                                 if (modified == 0) {
1138                                         hammer2_chain_modify(chain, mtid, 0, 0);
1139                                         modified = 1;
1140                                         goto again;
1141                                 }
1142                                 if ((*bitmap & bmmask11) == bmmask00) {
1143                                         bmap->avail -=
1144                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
1145                                         bgsize += HAMMER2_FREEMAP_BLOCK_SIZE;
1146                                 }
1147                                 if (bmap->class == 0)
1148                                         bmap->class = class;
1149                                 *bitmap |= bmmask11;
1150                                 if (hammer2_debug & 0x0040) {
1151                                         kprintf("hammer2_freemap_recover: "
1152                                                 "fixup type=%02x "
1153                                                 "block=%016jx/%zd\n",
1154                                                 bref->type, data_off, bytes);
1155                                 }
1156                         } else {
1157                                 /*
1158                                 kprintf("hammer2_freemap_recover:  good "
1159                                         "type=%02x block=%016jx/%zd\n",
1160                                         bref->type, data_off, bytes);
1161                                 */
1162                         }
1163                 }
1164 #if 0
1165                 /*
1166                  * XXX this stuff doesn't work, avail is miscalculated and
1167                  * code 10 means something else now.
1168                  */
1169                 else if ((*bitmap & bmmask11) == bmmask11) {
1170                         /*
1171                          * Mayfree/Realfree request and bitmap is currently
1172                          * marked as being fully allocated.
1173                          */
1174                         if (!modified) {
1175                                 hammer2_chain_modify(chain, 0);
1176                                 modified = 1;
1177                                 goto again;
1178                         }
1179                         if (how == HAMMER2_FREEMAP_DOREALFREE)
1180                                 *bitmap &= ~bmmask11;
1181                         else
1182                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
1183                 } else if ((*bitmap & bmmask11) == bmmask10) {
1184                         /*
1185                          * Mayfree/Realfree request and bitmap is currently
1186                          * marked as being possibly freeable.
1187                          */
1188                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
1189                                 if (!modified) {
1190                                         hammer2_chain_modify(chain, 0);
1191                                         modified = 1;
1192                                         goto again;
1193                                 }
1194                                 *bitmap &= ~bmmask11;
1195                         }
1196                 } else {
1197                         /*
1198                          * 01 - Not implemented, currently illegal state
1199                          * 00 - Not allocated at all, illegal free.
1200                          */
1201                         panic("hammer2_freemap_adjust: "
1202                               "Illegal state %08x(%08x)",
1203                               *bitmap, *bitmap & bmmask11);
1204                 }
1205 #endif
1206                 --count;
1207                 bmmask01 <<= 2;
1208                 bmmask10 <<= 2;
1209                 bmmask11 <<= 2;
1210         }
1211 #if HAMMER2_BMAP_ELEMENTS != 8
1212 #error "hammer2_freemap.c: HAMMER2_BMAP_ELEMENTS expected to be 8"
1213 #endif
1214         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
1215                 bmap->avail += 1 << radix;
1216                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
1217                 if (bmap->avail == HAMMER2_SEGSIZE &&
1218                     bmap->bitmapq[0] == 0 &&
1219                     bmap->bitmapq[1] == 0 &&
1220                     bmap->bitmapq[2] == 0 &&
1221                     bmap->bitmapq[3] == 0 &&
1222                     bmap->bitmapq[4] == 0 &&
1223                     bmap->bitmapq[5] == 0 &&
1224                     bmap->bitmapq[6] == 0 &&
1225                     bmap->bitmapq[7] == 0) {
1226                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1227                         kprintf("Freeseg %016jx\n", (intmax_t)key);
1228                         bmap->class = 0;
1229                 }
1230         }
1231
1232         /*
1233          * chain->bref.check.freemap.bigmask (XXX)
1234          *
1235          * Setting bigmask is a hint to the allocation code that there might
1236          * be something allocatable.  We also set this in recovery... it
1237          * doesn't hurt and we might want to use the hint for other validation
1238          * operations later on.
1239          *
1240          * We could calculate the largest possible allocation and set the
1241          * radii that could fit, but its easier just to set bigmask to -1.
1242          */
1243         if (modified) {
1244                 chain->bref.check.freemap.bigmask = -1;
1245                 hmp->freemap_relaxed = 0;       /* reset heuristic */
1246         }
1247
1248         hammer2_chain_unlock(chain);
1249         hammer2_chain_drop(chain);
1250 done:
1251         hammer2_chain_unlock(parent);
1252         hammer2_chain_drop(parent);
1253
1254         if (bgsize) {
1255                 hammer2_voldata_lock(hmp);
1256                 hammer2_voldata_modify(hmp);
1257                 hmp->voldata.allocator_free -= bgsize;
1258                 hammer2_voldata_unlock(hmp);
1259         }
1260 }
1261
1262 /*
1263  * Validate the freemap, in three stages.
1264  *
1265  * stage-1      ALLOCATED     -> POSSIBLY FREE
1266  *              POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1267  *
1268  *      This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1269  *      The POSSIBLY FREE state does not mean that a block is actually free
1270  *      and may be transitioned back to ALLOCATED in stage-2.
1271  *
1272  *      This is typically done during normal filesystem operations when
1273  *      something is deleted or a block is replaced.
1274  *
1275  *      This is done by bulkfree in-bulk after a memory-bounded meta-data
1276  *      scan to try to determine what might be freeable.
1277  *
1278  *      This can be done unconditionally through a freemap scan when the
1279  *      intention is to brute-force recover the proper state of the freemap.
1280  *
1281  * stage-2      POSSIBLY FREE -> ALLOCATED      (scan metadata topology)
1282  *
1283  *      This is done by bulkfree during a meta-data scan to ensure that
1284  *      all blocks still actually allocated by the filesystem are marked
1285  *      as such.
1286  *
1287  *      NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1288  *            the bulkfree stage-2 and stage-3 is running.  The live filesystem
1289  *            will use the alternative POSSIBLY FREE type (2) to prevent
1290  *            stage-3 from improperly transitioning unvetted possibly-free
1291  *            blocks to FREE.
1292  *
1293  * stage-3      POSSIBLY FREE (type 1) -> FREE  (scan freemap)
1294  *
1295  *      This is done by bulkfree to finalize POSSIBLY FREE states.
1296  *
1297  */