sys/vfs/hammer2/hammer2_freemap.c

   1 /*
   2  * Copyright (c) 2011-2014 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 };
  55
  56 typedef struct hammer2_fiterate hammer2_fiterate_t;
  57
  58 static int hammer2_freemap_try_alloc(hammer2_trans_t *trans,
  59                         hammer2_chain_t **parentp, hammer2_blockref_t *bref,
  60                         int radix, hammer2_fiterate_t *iter);
  61 static void hammer2_freemap_init(hammer2_trans_t *trans, hammer2_mount_t *hmp,
  62                         hammer2_key_t key, hammer2_chain_t *chain);
  63 static int hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
  64                         hammer2_bmap_data_t *bmap, uint16_t class,
  65                         int n, int radix, hammer2_key_t *basep);
  66 static int hammer2_freemap_iterate(hammer2_trans_t *trans,
  67                         hammer2_chain_t **parentp, hammer2_chain_t **chainp,
  68                         hammer2_fiterate_t *iter);
  69
  70 static __inline
  71 int
  72 hammer2_freemapradix(int radix)
  73 {
  74         return(radix);
  75 }
  76
  77 /*
  78  * Calculate the device offset for the specified FREEMAP_NODE or FREEMAP_LEAF
  79  * bref.  Return a combined media offset and physical size radix.  Freemap
  80  * chains use fixed storage offsets in the 4MB reserved area at the
  81  * beginning of each 2GB zone
  82  *
  83  * Rotate between four possibilities.  Theoretically this means we have three
  84  * good freemaps in case of a crash which we can use as a base for the fixup
  85  * scan at mount-time.
  86  */
  87 #define H2FMBASE(key, radix)    ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
  88 #define H2FMSHIFT(radix)        ((hammer2_off_t)1 << (radix))
  89
  90 static
  91 int
  92 hammer2_freemap_reserve(hammer2_trans_t *trans, hammer2_chain_t *chain,
  93                         int radix)
  94 {
  95         hammer2_blockref_t *bref = &chain->bref;
  96         hammer2_off_t off;
  97         int index;
  98         size_t bytes;
  99
 100         /*
 101          * Physical allocation size -> radix.  Typically either 256 for
 102          * a level 0 freemap leaf or 65536 for a level N freemap node.
 103          *
 104          * NOTE: A 256 byte bitmap represents 256 x 8 x 1024 = 2MB of storage.
 105          *       Do not use hammer2_allocsize() here as it has a min cap.
 106          */
 107         bytes = 1 << radix;
 108
 109         /*
 110          * Calculate block selection index 0..7 of current block.
 111          */
 112         if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 113                 index = 0;
 114         } else {
 115                 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 116                       (((hammer2_off_t)1 << HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 117                 off = off / HAMMER2_PBUFSIZE;
 118                 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
 119                          off < HAMMER2_ZONE_FREEMAP_END);
 120                 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) / 4;
 121                 KKASSERT(index >= 0 && index < HAMMER2_ZONE_FREEMAP_COPIES);
 122         }
 123
 124         /*
 125          * Calculate new index (our 'allocation').
 126          */
 127         index = (index + 1) % HAMMER2_ZONE_FREEMAP_COPIES;
 128
 129         /*
 130          * Calculate the block offset of the reserved block.  This will
 131          * point into the 4MB reserved area at the base of the appropriate
 132          * 2GB zone, once added to the FREEMAP_x selection above.
 133          */
 134         switch(bref->keybits) {
 135         /* case HAMMER2_FREEMAP_LEVEL5_RADIX: not applicable */
 136         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 2EB */
 137                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 138                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 139                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 140                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 141                        HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
 142                 break;
 143         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 2PB */
 144                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 145                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 146                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 147                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 148                        HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
 149                 break;
 150         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 2TB */
 151                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 152                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 153                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 154                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 155                        HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
 156                 break;
 157         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 2GB */
 158                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 159                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 160                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 161                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 162                        HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
 163                 break;
 164         default:
 165                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 166                 /* NOT REACHED */
 167                 off = (hammer2_off_t)-1;
 168                 break;
 169         }
 170         bref->data_off = off | radix;
 171 #if FREEMAP_DEBUG
 172         kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 173                 bref->type, bref->key, bref->keybits, bref->data_off);
 174 #endif
 175         return (0);
 176 }
 177
 178 /*
 179  * Normal freemap allocator
 180  *
 181  * Use available hints to allocate space using the freemap.  Create missing
 182  * freemap infrastructure on-the-fly as needed (including marking initial
 183  * allocations using the iterator as allocated, instantiating new 2GB zones,
 184  * and dealing with the end-of-media edge case).
 185  *
 186  * ip and bpref are only used as a heuristic to determine locality of
 187  * reference.  bref->key may also be used heuristically.
 188  */
 189 int
 190 hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 191                       size_t bytes)
 192 {
 193         hammer2_mount_t *hmp = chain->hmp;
 194         hammer2_blockref_t *bref = &chain->bref;
 195         hammer2_chain_t *parent;
 196         int radix;
 197         int error;
 198         unsigned int hindex;
 199         hammer2_fiterate_t iter;
 200
 201         /*
 202          * Validate the allocation size.  It must be a power of 2.
 203          *
 204          * For now require that the caller be aware of the minimum
 205          * allocation (1K).
 206          */
 207         radix = hammer2_getradix(bytes);
 208         KKASSERT((size_t)1 << radix == bytes);
 209
 210         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 211             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 212                 /*
 213                  * Freemap blocks themselves are assigned from the reserve
 214                  * area, not allocated from the freemap.
 215                  */
 216                 error = hammer2_freemap_reserve(trans, chain, radix);
 217                 return error;
 218         }
 219
 220         KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
 221
 222         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 223                 ++trans->sync_xid;
 224
 225         /*
 226          * Calculate the starting point for our allocation search.
 227          *
 228          * Each freemap leaf is dedicated to a specific freemap_radix.
 229          * The freemap_radix can be more fine-grained than the device buffer
 230          * radix which results in inodes being grouped together in their
 231          * own segment, terminal-data (16K or less) and initial indirect
 232          * block being grouped together, and then full-indirect and full-data
 233          * blocks (64K) being grouped together.
 234          *
 235          * The single most important aspect of this is the inode grouping
 236          * because that is what allows 'find' and 'ls' and other filesystem
 237          * topology operations to run fast.
 238          */
 239 #if 0
 240         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 241                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 242         else if (trans->tmp_bpref)
 243                 bpref = trans->tmp_bpref;
 244         else if (trans->tmp_ip)
 245                 bpref = trans->tmp_ip->chain->bref.data_off;
 246         else
 247 #endif
 248         /*
 249          * Heuristic tracking index.  We would like one for each distinct
 250          * bref type if possible.  heur_freemap[] has room for two classes
 251          * for each type.  At a minimum we have to break-up our heuristic
 252          * by device block sizes.
 253          */
 254         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 255         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 256         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 257         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 258         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR);
 259
 260         iter.bpref = hmp->heur_freemap[hindex];
 261
 262         /*
 263          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 264          * reserved area, the try code will iterate past it.
 265          */
 266         if (iter.bpref > hmp->voldata.volu_size)
 267                 iter.bpref = hmp->voldata.volu_size - 1;
 268
 269         /*
 270          * Iterate the freemap looking for free space before and after.
 271          */
 272         parent = &hmp->fchain;
 273         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 274         error = EAGAIN;
 275         iter.bnext = iter.bpref;
 276         iter.loops = 0;
 277
 278         while (error == EAGAIN) {
 279                 error = hammer2_freemap_try_alloc(trans, &parent, bref,
 280                                                   radix, &iter);
 281         }
 282         hmp->heur_freemap[hindex] = iter.bnext;
 283         hammer2_chain_unlock(parent);
 284
 285         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 286                 --trans->sync_xid;
 287
 288         return (error);
 289 }
 290
 291 static int
 292 hammer2_freemap_try_alloc(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 293                           hammer2_blockref_t *bref, int radix,
 294                           hammer2_fiterate_t *iter)
 295 {
 296         hammer2_mount_t *hmp = (*parentp)->hmp;
 297         hammer2_off_t l0size;
 298         hammer2_off_t l1size;
 299         hammer2_off_t l1mask;
 300         hammer2_key_t key_dummy;
 301         hammer2_chain_t *chain;
 302         hammer2_off_t key;
 303         size_t bytes;
 304         uint16_t class;
 305         int error = 0;
 306         int cache_index = -1;
 307         int ddflag;
 308
 309
 310         /*
 311          * Calculate the number of bytes being allocated, the number
 312          * of contiguous bits of bitmap being allocated, and the bitmap
 313          * mask.
 314          *
 315          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 316          *          mask calculation.
 317          */
 318         bytes = (size_t)1 << radix;
 319         class = (bref->type << 8) | hammer2_devblkradix(radix);
 320
 321         /*
 322          * Lookup the level1 freemap chain, creating and initializing one
 323          * if necessary.  Intermediate levels will be created automatically
 324          * when necessary by hammer2_chain_create().
 325          */
 326         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 327         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 328         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 329         l1mask = l1size - 1;
 330
 331         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 332                                      &cache_index,
 333                                      HAMMER2_LOOKUP_ALWAYS |
 334                                      HAMMER2_LOOKUP_MATCHIND, &ddflag);
 335
 336         if (chain == NULL) {
 337                 /*
 338                  * Create the missing leaf, be sure to initialize
 339                  * the auxillary freemap tracking information in
 340                  * the bref.check.freemap structure.
 341                  */
 342 #if 0
 343                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 344                         key, iter->bpref);
 345 #endif
 346                 error = hammer2_chain_create(trans, parentp, &chain, hmp->spmp,
 347                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 348                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 349                                      HAMMER2_FREEMAP_LEVELN_PSIZE);
 350                 KKASSERT(error == 0);
 351                 if (error == 0) {
 352                         hammer2_chain_modify(trans, chain, 0);
 353                         bzero(&chain->data->bmdata[0],
 354                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 355                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 356                         chain->bref.check.freemap.avail = l1size;
 357                         /* bref.methods should already be inherited */
 358
 359                         hammer2_freemap_init(trans, hmp, key, chain);
 360                 }
 361         } else if ((chain->bref.check.freemap.bigmask & (1 << radix)) == 0) {
 362                 /*
 363                  * Already flagged as not having enough space
 364                  */
 365                 error = ENOSPC;
 366         } else {
 367                 /*
 368                  * Modify existing chain to setup for adjustment.
 369                  */
 370                 hammer2_chain_modify(trans, chain, 0);
 371         }
 372
 373         /*
 374          * Scan 2MB entries.
 375          */
 376         if (error == 0) {
 377                 hammer2_bmap_data_t *bmap;
 378                 hammer2_key_t base_key;
 379                 int count;
 380                 int start;
 381                 int n;
 382
 383                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 384                 start = (int)((iter->bnext - key) >>
 385                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 386                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 387                 hammer2_chain_modify(trans, chain, 0);
 388
 389                 error = ENOSPC;
 390                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 391                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 392                             start - count < 0) {
 393                                 break;
 394                         }
 395                         n = start + count;
 396                         bmap = &chain->data->bmdata[n];
 397                         if (n < HAMMER2_FREEMAP_COUNT && bmap->avail &&
 398                             (bmap->class == 0 || bmap->class == class)) {
 399                                 base_key = key + n * l0size;
 400                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 401                                                            class, n, radix,
 402                                                            &base_key);
 403                                 if (error != ENOSPC) {
 404                                         key = base_key;
 405                                         break;
 406                                 }
 407                         }
 408                         n = start - count;
 409                         bmap = &chain->data->bmdata[n];
 410                         if (n >= 0 && bmap->avail &&
 411                             (bmap->class == 0 || bmap->class == class)) {
 412                                 base_key = key + n * l0size;
 413                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 414                                                            class, n, radix,
 415                                                            &base_key);
 416                                 if (error != ENOSPC) {
 417                                         key = base_key;
 418                                         break;
 419                                 }
 420                         }
 421                 }
 422                 if (error == ENOSPC)
 423                         chain->bref.check.freemap.bigmask &= ~(1 << radix);
 424                 /* XXX also scan down from original count */
 425         }
 426
 427         if (error == 0) {
 428                 /*
 429                  * Assert validity.  Must be beyond the static allocator used
 430                  * by newfs_hammer2 (and thus also beyond the aux area),
 431                  * not go past the volume size, and must not be in the
 432                  * reserved segment area for a zone.
 433                  */
 434                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 435                          key + bytes <= hmp->voldata.volu_size);
 436                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 437                 bref->data_off = key | radix;
 438
 439 #if 0
 440                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 441                         chain,
 442                         bref->key, bref->data_off, chain->bref.data_off);
 443 #endif
 444         } else if (error == ENOSPC) {
 445                 /*
 446                  * Return EAGAIN with next iteration in iter->bnext, or
 447                  * return ENOSPC if the allocation map has been exhausted.
 448                  */
 449                 error = hammer2_freemap_iterate(trans, parentp, &chain, iter);
 450         }
 451
 452         /*
 453          * Cleanup
 454          */
 455         if (chain)
 456                 hammer2_chain_unlock(chain);
 457         return (error);
 458 }
 459
 460 /*
 461  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 462  *
 463  * If the linear iterator is mid-block we use it directly (the bitmap should
 464  * already be marked allocated), otherwise we search for a block in the bitmap
 465  * that fits the allocation request.
 466  *
 467  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 468  * to fully allocated and adjusts the linear allocator to allow the
 469  * remaining space to be allocated.
 470  */
 471 static
 472 int
 473 hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 474                    hammer2_bmap_data_t *bmap,
 475                    uint16_t class, int n, int radix, hammer2_key_t *basep)
 476 {
 477         hammer2_io_t *dio;
 478         size_t size;
 479         size_t bsize;
 480         int bmradix;
 481         uint32_t bmmask;
 482         int offset;
 483         int error;
 484         int i;
 485         int j;
 486
 487         /*
 488          * Take into account 2-bits per block when calculating bmradix.
 489          */
 490         size = (size_t)1 << radix;
 491
 492         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 493                 bmradix = 2;
 494                 bsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 495                 /* (16K) 2 bits per allocation block */
 496         } else {
 497                 bmradix = 2 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 498                 bsize = size;
 499                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 500         }
 501
 502         /*
 503          * Use the linear iterator to pack small allocations, otherwise
 504          * fall-back to finding a free 16KB chunk.  The linear iterator
 505          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 506          * If it is the bitmap must be scanned.  It can become invalid
 507          * once we pack to the boundary.  We adjust it after a bitmap
 508          * allocation only for sub-16KB allocations (so the perfectly good
 509          * previous value can still be used for fragments when 16KB+
 510          * allocations are made).
 511          *
 512          * Beware of hardware artifacts when bmradix == 32 (intermediate
 513          * result can wind up being '1' instead of '0' if hardware masks
 514          * bit-count & 31).
 515          *
 516          * NOTE: j needs to be even in the j= calculation.  As an artifact
 517          *       of the /2 division, our bitmask has to clear bit 0.
 518          *
 519          * NOTE: TODO this can leave little unallocatable fragments lying
 520          *       around.
 521          */
 522         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 523             HAMMER2_FREEMAP_BLOCK_SIZE &&
 524             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 525             bmap->linear < HAMMER2_SEGSIZE) {
 526                 KKASSERT(bmap->linear >= 0 &&
 527                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 528                          (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
 529                 offset = bmap->linear;
 530                 i = offset / (HAMMER2_SEGSIZE / 8);
 531                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 30;
 532                 bmmask = (bmradix == 32) ?
 533                          0xFFFFFFFFU : (1 << bmradix) - 1;
 534                 bmmask <<= j;
 535                 bmap->linear = offset + size;
 536         } else {
 537                 for (i = 0; i < 8; ++i) {
 538                         bmmask = (bmradix == 32) ?
 539                                  0xFFFFFFFFU : (1 << bmradix) - 1;
 540                         for (j = 0; j < 32; j += bmradix) {
 541                                 if ((bmap->bitmap[i] & bmmask) == 0)
 542                                         goto success;
 543                                 bmmask <<= bmradix;
 544                         }
 545                 }
 546                 /*fragments might remain*/
 547                 /*KKASSERT(bmap->avail == 0);*/
 548                 return (ENOSPC);
 549 success:
 550                 offset = i * (HAMMER2_SEGSIZE / 8) +
 551                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 552                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 553                         bmap->linear = offset + size;
 554         }
 555
 556         KKASSERT(i >= 0 && i < 8);      /* 8 x 16 -> 128 x 16K -> 2MB */
 557
 558         /*
 559          * Optimize the buffer cache to avoid unnecessary read-before-write
 560          * operations.
 561          *
 562          * The device block size could be larger than the allocation size
 563          * so the actual bitmap test is somewhat more involved.  We have
 564          * to use a compatible buffer size for this operation.
 565          */
 566         if ((bmap->bitmap[i] & bmmask) == 0 &&
 567             hammer2_devblksize(size) != size) {
 568                 size_t psize = hammer2_devblksize(size);
 569                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 570                 int pbmradix = 2 << (hammer2_devblkradix(radix) -
 571                                      HAMMER2_FREEMAP_BLOCK_RADIX);
 572                 uint32_t pbmmask;
 573                 int pradix = hammer2_getradix(psize);
 574
 575                 pbmmask = (pbmradix == 32) ? 0xFFFFFFFFU : (1 << pbmradix) - 1;
 576                 while ((pbmmask & bmmask) == 0)
 577                         pbmmask <<= pbmradix;
 578
 579 #if 0
 580                 kprintf("%016jx mask %08x %08x %08x (%zd/%zd)\n",
 581                         *basep + offset, bmap->bitmap[i],
 582                         pbmmask, bmmask, size, psize);
 583 #endif
 584
 585                 if ((bmap->bitmap[i] & pbmmask) == 0) {
 586                         error = hammer2_io_newq(hmp,
 587                                                 (*basep + (offset & ~pmask)) |
 588                                                  pradix,
 589                                                 psize, &dio);
 590                         hammer2_io_bqrelse(&dio);
 591                 }
 592         }
 593
 594 #if 0
 595         /*
 596          * When initializing a new inode segment also attempt to initialize
 597          * an adjacent segment.  Be careful not to index beyond the array
 598          * bounds.
 599          *
 600          * We do this to try to localize inode accesses to improve
 601          * directory scan rates.  XXX doesn't improve scan rates.
 602          */
 603         if (size == HAMMER2_INODE_BYTES) {
 604                 if (n & 1) {
 605                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 606                                 bmap[-1].radix = radix;
 607                 } else {
 608                         if (bmap[1].radix == 0 && bmap[1].avail)
 609                                 bmap[1].radix = radix;
 610                 }
 611         }
 612 #endif
 613
 614         /*
 615          * Adjust the linear iterator, set the radix if necessary (might as
 616          * well just set it unconditionally), adjust *basep to return the
 617          * allocated data offset.
 618          */
 619         bmap->bitmap[i] |= bmmask;
 620         bmap->class = class;
 621         bmap->avail -= size;
 622         *basep += offset;
 623
 624         hammer2_voldata_lock(hmp);
 625         hammer2_voldata_modify(hmp);
 626         hmp->voldata.allocator_free -= size;  /* XXX */
 627         hammer2_voldata_unlock(hmp);
 628
 629         return(0);
 630 }
 631
 632 static
 633 void
 634 hammer2_freemap_init(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 635                      hammer2_key_t key, hammer2_chain_t *chain)
 636 {
 637         hammer2_off_t l1size;
 638         hammer2_off_t lokey;
 639         hammer2_off_t hikey;
 640         hammer2_bmap_data_t *bmap;
 641         int count;
 642
 643         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 644
 645         /*
 646          * Calculate the portion of the 2GB map that should be initialized
 647          * as free.  Portions below or after will be initialized as allocated.
 648          * SEGMASK-align the areas so we don't have to worry about sub-scans
 649          * or endianess when using memset.
 650          *
 651          * (1) Ensure that all statically allocated space from newfs_hammer2
 652          *     is marked allocated.
 653          *
 654          * (2) Ensure that the reserved area is marked allocated (typically
 655          *     the first 4MB of the 2GB area being represented).
 656          *
 657          * (3) Ensure that any trailing space at the end-of-volume is marked
 658          *     allocated.
 659          *
 660          * WARNING! It is possible for lokey to be larger than hikey if the
 661          *          entire 2GB segment is within the static allocation.
 662          */
 663         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 664                 ~HAMMER2_SEGMASK64;
 665
 666         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 667                   HAMMER2_ZONE_SEG64) {
 668                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 669                         HAMMER2_ZONE_SEG64;
 670         }
 671
 672         hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 673         if (hikey > hmp->voldata.volu_size) {
 674                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 675         }
 676
 677         chain->bref.check.freemap.avail =
 678                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 679         bmap = &chain->data->bmdata[0];
 680
 681         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 682                 if (key < lokey || key >= hikey) {
 683                         memset(bmap->bitmap, -1,
 684                                sizeof(bmap->bitmap));
 685                         bmap->avail = 0;
 686                         bmap->linear = HAMMER2_SEGSIZE;
 687                         chain->bref.check.freemap.avail -=
 688                                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 689                 } else {
 690                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 691                 }
 692                 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 693                 ++bmap;
 694         }
 695 }
 696
 697 /*
 698  * The current Level 1 freemap has been exhausted, iterate to the next
 699  * one, return ENOSPC if no freemaps remain.
 700  *
 701  * XXX this should rotate back to the beginning to handle freed-up space
 702  * XXX or use intermediate entries to locate free space. TODO
 703  */
 704 static int
 705 hammer2_freemap_iterate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 706                         hammer2_chain_t **chainp, hammer2_fiterate_t *iter)
 707 {
 708         hammer2_mount_t *hmp = (*parentp)->hmp;
 709
 710         iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 711         iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 712         if (iter->bnext >= hmp->voldata.volu_size) {
 713                 iter->bnext = 0;
 714                 if (++iter->loops == 2)
 715                         return (ENOSPC);
 716         }
 717         return(EAGAIN);
 718 }
 719
 720 /*
 721  * Adjust the bit-pattern for data in the freemap bitmap according to
 722  * (how).  This code is called from on-mount recovery to fixup (mark
 723  * as allocated) blocks whos freemap upates might not have been committed
 724  * in the last crash and is used by the bulk freemap scan to stage frees.
 725  *
 726  * XXX currently disabled when how == 0 (the normal real-time case).  At
 727  * the moment we depend on the bulk freescan to actually free blocks.  It
 728  * will still call this routine with a non-zero how to stage possible frees
 729  * and to do the actual free.
 730  */
 731 void
 732 hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 733                        hammer2_blockref_t *bref, int how)
 734 {
 735         hammer2_off_t data_off = bref->data_off;
 736         hammer2_chain_t *chain;
 737         hammer2_chain_t *parent;
 738         hammer2_bmap_data_t *bmap;
 739         hammer2_key_t key;
 740         hammer2_key_t key_dummy;
 741         hammer2_off_t l0size;
 742         hammer2_off_t l1size;
 743         hammer2_off_t l1mask;
 744         uint32_t *bitmap;
 745         const uint32_t bmmask00 = 0;
 746         uint32_t bmmask01;
 747         uint32_t bmmask10;
 748         uint32_t bmmask11;
 749         size_t bytes;
 750         uint16_t class;
 751         int radix;
 752         int start;
 753         int count;
 754         int modified = 0;
 755         int cache_index = -1;
 756         int error;
 757         int ddflag;
 758
 759         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
 760         data_off &= ~HAMMER2_OFF_MASK_RADIX;
 761         KKASSERT(radix <= HAMMER2_RADIX_MAX);
 762
 763         bytes = (size_t)1 << radix;
 764         class = (bref->type << 8) | hammer2_devblkradix(radix);
 765
 766         /*
 767          * We can't adjust thre freemap for data allocations made by
 768          * newfs_hammer2.
 769          */
 770         if (data_off < hmp->voldata.allocator_beg)
 771                 return;
 772
 773         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 774
 775         /*
 776          * Lookup the level1 freemap chain.  The chain must exist.
 777          */
 778         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
 779         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 780         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 781         l1mask = l1size - 1;
 782
 783         parent = &hmp->fchain;
 784         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 785
 786         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
 787                                      &cache_index,
 788                                      HAMMER2_LOOKUP_ALWAYS |
 789                                      HAMMER2_LOOKUP_MATCHIND, &ddflag);
 790
 791         /*
 792          * Stop early if we are trying to free something but no leaf exists.
 793          */
 794         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
 795                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
 796                         (intmax_t)bref->data_off);
 797                 goto done;
 798         }
 799
 800         /*
 801          * Create any missing leaf(s) if we are doing a recovery (marking
 802          * the block(s) as being allocated instead of being freed).  Be sure
 803          * to initialize the auxillary freemap tracking info in the
 804          * bref.check.freemap structure.
 805          */
 806         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
 807                 error = hammer2_chain_create(trans, &parent, &chain, hmp->spmp,
 808                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 809                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 810                                      HAMMER2_FREEMAP_LEVELN_PSIZE);
 811
 812                 if (hammer2_debug & 0x0040) {
 813                         kprintf("fixup create chain %p %016jx:%d\n",
 814                                 chain, chain->bref.key, chain->bref.keybits);
 815                 }
 816
 817                 if (error == 0) {
 818                         hammer2_chain_modify(trans, chain, 0);
 819                         bzero(&chain->data->bmdata[0],
 820                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 821                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 822                         chain->bref.check.freemap.avail = l1size;
 823                         /* bref.methods should already be inherited */
 824
 825                         hammer2_freemap_init(trans, hmp, key, chain);
 826                 }
 827                 /* XXX handle error */
 828         }
 829
 830 #if FREEMAP_DEBUG
 831         kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 832                 chain->bref.type, chain->bref.key,
 833                 chain->bref.keybits, chain->bref.data_off);
 834 #endif
 835
 836         /*
 837          * Calculate the bitmask (runs in 2-bit pairs).
 838          */
 839         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
 840         bmmask01 = 1 << start;
 841         bmmask10 = 2 << start;
 842         bmmask11 = 3 << start;
 843
 844         /*
 845          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
 846          * a bulk scan is able to roll them up.
 847          */
 848         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 849                 count = 1;
 850                 if (how == HAMMER2_FREEMAP_DOREALFREE)
 851                         how = HAMMER2_FREEMAP_DOMAYFREE;
 852         } else {
 853                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 854         }
 855
 856         /*
 857          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
 858          * a 2MB swath.  The bmap itself (LEVEL1) covers 2GB.
 859          *
 860          * Be sure to reset the linear iterator to ensure that the adjustment
 861          * is not ignored.
 862          */
 863 again:
 864         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
 865                                     (HAMMER2_FREEMAP_COUNT - 1)];
 866         bitmap = &bmap->bitmap[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
 867         bmap->linear = 0;
 868
 869         while (count) {
 870                 KKASSERT(bmmask11);
 871                 if (how == HAMMER2_FREEMAP_DORECOVER) {
 872                         /*
 873                          * Recovery request, mark as allocated.
 874                          */
 875                         if ((*bitmap & bmmask11) != bmmask11) {
 876                                 if (modified == 0) {
 877                                         hammer2_chain_modify(trans, chain, 0);
 878                                         modified = 1;
 879                                         goto again;
 880                                 }
 881                                 if ((*bitmap & bmmask11) == bmmask00)
 882                                         bmap->avail -= 1 << radix;
 883                                 if (bmap->class == 0)
 884                                         bmap->class = class;
 885                                 *bitmap |= bmmask11;
 886                                 if (hammer2_debug & 0x0040) {
 887                                         kprintf("hammer2_freemap_recover: "
 888                                                 "fixup type=%02x "
 889                                                 "block=%016jx/%zd\n",
 890                                                 bref->type, data_off, bytes);
 891                                 }
 892                         } else {
 893                                 /*
 894                                 kprintf("hammer2_freemap_recover:  good "
 895                                         "type=%02x block=%016jx/%zd\n",
 896                                         bref->type, data_off, bytes);
 897                                 */
 898                         }
 899                 } else if ((*bitmap & bmmask11) == bmmask11) {
 900                         /*
 901                          * Mayfree/Realfree request and bitmap is currently
 902                          * marked as being fully allocated.
 903                          */
 904                         if (!modified) {
 905                                 hammer2_chain_modify(trans, chain, 0);
 906                                 modified = 1;
 907                                 goto again;
 908                         }
 909                         if (how == HAMMER2_FREEMAP_DOREALFREE)
 910                                 *bitmap &= ~bmmask11;
 911                         else
 912                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
 913                 } else if ((*bitmap & bmmask11) == bmmask10) {
 914                         /*
 915                          * Mayfree/Realfree request and bitmap is currently
 916                          * marked as being possibly freeable.
 917                          */
 918                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
 919                                 if (!modified) {
 920                                         hammer2_chain_modify(trans, chain, 0);
 921                                         modified = 1;
 922                                         goto again;
 923                                 }
 924                                 *bitmap &= ~bmmask11;
 925                         }
 926                 } else {
 927                         /*
 928                          * 01 - Not implemented, currently illegal state
 929                          * 00 - Not allocated at all, illegal free.
 930                          */
 931                         panic("hammer2_freemap_adjust: "
 932                               "Illegal state %08x(%08x)",
 933                               *bitmap, *bitmap & bmmask11);
 934                 }
 935                 --count;
 936                 bmmask01 <<= 2;
 937                 bmmask10 <<= 2;
 938                 bmmask11 <<= 2;
 939         }
 940         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
 941                 bmap->avail += 1 << radix;
 942                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
 943                 if (bmap->avail == HAMMER2_SEGSIZE &&
 944                     bmap->bitmap[0] == 0 &&
 945                     bmap->bitmap[1] == 0 &&
 946                     bmap->bitmap[2] == 0 &&
 947                     bmap->bitmap[3] == 0 &&
 948                     bmap->bitmap[4] == 0 &&
 949                     bmap->bitmap[5] == 0 &&
 950                     bmap->bitmap[6] == 0 &&
 951                     bmap->bitmap[7] == 0) {
 952                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
 953                         kprintf("Freeseg %016jx\n", (intmax_t)key);
 954                         bmap->class = 0;
 955                 }
 956         }
 957
 958         /*
 959          * chain->bref.check.freemap.bigmask (XXX)
 960          *
 961          * Setting bigmask is a hint to the allocation code that there might
 962          * be something allocatable.  We also set this in recovery... it
 963          * doesn't hurt and we might want to use the hint for other validation
 964          * operations later on.
 965          */
 966         if (modified)
 967                 chain->bref.check.freemap.bigmask |= 1 << radix;
 968
 969         hammer2_chain_unlock(chain);
 970 done:
 971         hammer2_chain_unlock(parent);
 972 }