sys/vfs/hammer2/hammer2_freemap.c

   1 /*
   2  * Copyright (c) 2011-2013 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 struct hammer2_fiterate {
  49         hammer2_off_t   bpref;
  50         hammer2_off_t   bnext;
  51         int             loops;
  52 };
  53
  54 typedef struct hammer2_fiterate hammer2_fiterate_t;
  55
  56 static int hammer2_freemap_try_alloc(hammer2_trans_t *trans,
  57                         hammer2_chain_t **parentp, hammer2_blockref_t *bref,
  58                         int radix, hammer2_fiterate_t *iter);
  59 static void hammer2_freemap_init(hammer2_trans_t *trans, hammer2_mount_t *hmp,
  60                         hammer2_key_t key, hammer2_chain_t *chain);
  61 static int hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
  62                         hammer2_bmap_data_t *bmap, uint16_t class,
  63                         int n, int radix, hammer2_key_t *basep);
  64 static int hammer2_freemap_iterate(hammer2_trans_t *trans,
  65                         hammer2_chain_t **parentp, hammer2_chain_t **chainp,
  66                         hammer2_fiterate_t *iter);
  67
  68 static __inline
  69 int
  70 hammer2_freemapradix(int radix)
  71 {
  72         return(radix);
  73 }
  74
  75 /*
  76  * Calculate the device offset for the specified FREEMAP_NODE or FREEMAP_LEAF
  77  * bref.  Return a combined media offset and physical size radix.  Freemap
  78  * chains use fixed storage offsets in the 4MB reserved area at the
  79  * beginning of each 2GB zone
  80  *
  81  * Rotate between four possibilities.  Theoretically this means we have three
  82  * good freemaps in case of a crash which we can use as a base for the fixup
  83  * scan at mount-time.
  84  */
  85 #define H2FMBASE(key, radix)    ((key) & ~(((hammer2_off_t)1 << (radix)) - 1))
  86 #define H2FMSHIFT(radix)        ((hammer2_off_t)1 << (radix))
  87
  88 static
  89 int
  90 hammer2_freemap_reserve(hammer2_trans_t *trans, hammer2_chain_t *chain,
  91                         int radix)
  92 {
  93         hammer2_blockref_t *bref = &chain->bref;
  94         hammer2_off_t off;
  95         size_t bytes;
  96
  97         /*
  98          * Physical allocation size -> radix.  Typically either 256 for
  99          * a level 0 freemap leaf or 65536 for a level N freemap node.
 100          *
 101          * NOTE: A 256 byte bitmap represents 256 x 8 x 1024 = 2MB of storage.
 102          *       Do not use hammer2_allocsize() here as it has a min cap.
 103          */
 104         bytes = 1 << radix;
 105
 106         /*
 107          * Adjust by HAMMER2_ZONE_FREEMAP_{A,B,C,D} using the existing
 108          * offset as a basis.  Start in zone A if previously unallocated.
 109          */
 110 #if 0
 111         kprintf("trans %04jx/%08x freemap chain %p.%d [%08x] %016jx/%d %016jx",
 112                 trans->sync_tid, trans->flags,
 113                 chain, chain->bref.type, chain->flags,
 114                 chain->bref.key, chain->bref.keybits,
 115                 bref->data_off);
 116 #endif
 117         if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 118                 off = HAMMER2_ZONE_FREEMAP_A;
 119         } else {
 120                 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 121                       (((hammer2_off_t)1 << HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 122                 off = off / HAMMER2_PBUFSIZE;
 123                 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_A);
 124                 KKASSERT(off < HAMMER2_ZONE_FREEMAP_D + 4);
 125         }
 126
 127         if ((trans->flags &
 128              (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_ISALLOCATING)) ==
 129             HAMMER2_TRANS_ISFLUSH) {
 130                 /*
 131                  * Delete-Duplicates while flushing the fchain topology
 132                  * itself.
 133                  */
 134 #if 0
 135                 kprintf(" flush ");
 136 #endif
 137                 if (off >= HAMMER2_ZONE_FREEMAP_D)
 138                         off = HAMMER2_ZONE_FREEMAP_B;
 139                 else if (off >= HAMMER2_ZONE_FREEMAP_C)
 140                         off = HAMMER2_ZONE_FREEMAP_A;
 141                 else if (off >= HAMMER2_ZONE_FREEMAP_B)
 142                         off = HAMMER2_ZONE_FREEMAP_D;
 143                 else
 144                         off = HAMMER2_ZONE_FREEMAP_C;
 145         } else {
 146                 /*
 147                  * Allocations from the freemap via a normal transaction
 148                  * or a flush whos sync_tid has been bumped (so effectively
 149                  * done as a normal transaction).
 150                  */
 151 #if 0
 152                 kprintf(" alloc ");
 153 #endif
 154                 if (off >= HAMMER2_ZONE_FREEMAP_D)
 155                         off = HAMMER2_ZONE_FREEMAP_A;
 156                 else if (off >= HAMMER2_ZONE_FREEMAP_C)
 157                         off = HAMMER2_ZONE_FREEMAP_D;
 158                 else if (off >= HAMMER2_ZONE_FREEMAP_B)
 159                         off = HAMMER2_ZONE_FREEMAP_C;
 160                 else
 161                         off = HAMMER2_ZONE_FREEMAP_B;
 162         }
 163
 164
 165         off = off * HAMMER2_PBUFSIZE;
 166
 167         /*
 168          * Calculate the block offset of the reserved block.  This will
 169          * point into the 4MB reserved area at the base of the appropriate
 170          * 2GB zone, once added to the FREEMAP_x selection above.
 171          */
 172         switch(bref->keybits) {
 173         /* case HAMMER2_FREEMAP_LEVEL5_RADIX: not applicable */
 174         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 2EB */
 175                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 176                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 177                 off += H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 178                        HAMMER2_ZONEFM_LEVEL4 * HAMMER2_PBUFSIZE;
 179                 break;
 180         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 2PB */
 181                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 182                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 183                 off += H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 184                        HAMMER2_ZONEFM_LEVEL3 * HAMMER2_PBUFSIZE;
 185                 break;
 186         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 2TB */
 187                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 188                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 189                 off += H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 190                        HAMMER2_ZONEFM_LEVEL2 * HAMMER2_PBUFSIZE;
 191                 break;
 192         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 2GB */
 193                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 194                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 195                 off += H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 196                        HAMMER2_ZONEFM_LEVEL1 * HAMMER2_PBUFSIZE;
 197                 break;
 198         default:
 199                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 200                 /* NOT REACHED */
 201                 break;
 202         }
 203         bref->data_off = off | radix;
 204 #if 0
 205         kprintf("-> %016jx\n", bref->data_off);
 206 #endif
 207         return (0);
 208 }
 209
 210 /*
 211  * Normal freemap allocator
 212  *
 213  * Use available hints to allocate space using the freemap.  Create missing
 214  * freemap infrastructure on-the-fly as needed (including marking initial
 215  * allocations using the iterator as allocated, instantiating new 2GB zones,
 216  * and dealing with the end-of-media edge case).
 217  *
 218  * ip and bpref are only used as a heuristic to determine locality of
 219  * reference.  bref->key may also be used heuristically.
 220  *
 221  * WARNING! When called from a flush we have to use the 'live' sync_tid
 222  *          and not the flush sync_tid.  The live sync_tid is the flush
 223  *          sync_tid + 1.  That is, freemap allocations which occur during
 224  *          a flush are not part of the flush.  Crash-recovery will restore
 225  *          any lost allocations.
 226  */
 227 int
 228 hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 229                       size_t bytes)
 230 {
 231         hammer2_mount_t *hmp = chain->hmp;
 232         hammer2_blockref_t *bref = &chain->bref;
 233         hammer2_chain_t *parent;
 234         int radix;
 235         int error;
 236         unsigned int hindex;
 237         hammer2_fiterate_t iter;
 238
 239         /*
 240          * Validate the allocation size.  It must be a power of 2.
 241          *
 242          * For now require that the caller be aware of the minimum
 243          * allocation (1K).
 244          */
 245         radix = hammer2_getradix(bytes);
 246         KKASSERT((size_t)1 << radix == bytes);
 247
 248         /*
 249          * Freemap blocks themselves are simply assigned from the reserve
 250          * area, not allocated from the freemap.
 251          */
 252         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 253             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 254                 return (hammer2_freemap_reserve(trans, chain, radix));
 255         }
 256
 257         /*
 258          * Mark previously allocated block as possibly freeable.  There might
 259          * be snapshots and other races so we can't just mark it fully free.
 260          * (XXX optimize this for the current-transaction create+delete case)
 261          */
 262         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX) {
 263                 hammer2_freemap_adjust(trans, hmp, bref,
 264                                        HAMMER2_FREEMAP_DOMAYFREE);
 265         }
 266
 267         /*
 268          * Setting ISALLOCATING ensures correct operation even when the
 269          * flusher itself is making allocations.
 270          */
 271         KKASSERT(bytes >= HAMMER2_MIN_ALLOC && bytes <= HAMMER2_MAX_ALLOC);
 272         KKASSERT((trans->flags & HAMMER2_TRANS_ISALLOCATING) == 0);
 273         atomic_set_int(&trans->flags, HAMMER2_TRANS_ISALLOCATING);
 274         if (trans->flags & HAMMER2_TRANS_ISFLUSH)
 275                 ++trans->sync_tid;
 276
 277         /*
 278          * Calculate the starting point for our allocation search.
 279          *
 280          * Each freemap leaf is dedicated to a specific freemap_radix.
 281          * The freemap_radix can be more fine-grained than the device buffer
 282          * radix which results in inodes being grouped together in their
 283          * own segment, terminal-data (16K or less) and initial indirect
 284          * block being grouped together, and then full-indirect and full-data
 285          * blocks (64K) being grouped together.
 286          *
 287          * The single most important aspect of this is the inode grouping
 288          * because that is what allows 'find' and 'ls' and other filesystem
 289          * topology operations to run fast.
 290          */
 291 #if 0
 292         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 293                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 294         else if (trans->tmp_bpref)
 295                 bpref = trans->tmp_bpref;
 296         else if (trans->tmp_ip)
 297                 bpref = trans->tmp_ip->chain->bref.data_off;
 298         else
 299 #endif
 300         /*
 301          * Heuristic tracking index.  We would like one for each distinct
 302          * bref type if possible.  heur_freemap[] has room for two classes
 303          * for each type.  At a minimum we have to break-up our heuristic
 304          * by device block sizes.
 305          */
 306         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 307         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 308         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 309         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 310         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR);
 311
 312         iter.bpref = hmp->heur_freemap[hindex];
 313
 314         /*
 315          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 316          * reserved area, the try code will iterate past it.
 317          */
 318         if (iter.bpref > hmp->voldata.volu_size)
 319                 iter.bpref = hmp->voldata.volu_size - 1;
 320
 321         /*
 322          * Iterate the freemap looking for free space before and after.
 323          */
 324         parent = &hmp->fchain;
 325         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 326         error = EAGAIN;
 327         iter.bnext = iter.bpref;
 328         iter.loops = 0;
 329
 330         while (error == EAGAIN) {
 331                 error = hammer2_freemap_try_alloc(trans, &parent, bref,
 332                                                   radix, &iter);
 333         }
 334         hmp->heur_freemap[hindex] = iter.bnext;
 335         hammer2_chain_unlock(parent);
 336
 337         atomic_clear_int(&trans->flags, HAMMER2_TRANS_ISALLOCATING);
 338         if (trans->flags & HAMMER2_TRANS_ISFLUSH)
 339                 --trans->sync_tid;
 340
 341         return (error);
 342 }
 343
 344 static int
 345 hammer2_freemap_try_alloc(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 346                           hammer2_blockref_t *bref, int radix,
 347                           hammer2_fiterate_t *iter)
 348 {
 349         hammer2_mount_t *hmp = (*parentp)->hmp;
 350         hammer2_off_t l0size;
 351         hammer2_off_t l1size;
 352         hammer2_off_t l1mask;
 353         hammer2_key_t key_dummy;
 354         hammer2_chain_t *chain;
 355         hammer2_off_t key;
 356         size_t bytes;
 357         uint16_t class;
 358         int error = 0;
 359         int cache_index = -1;
 360
 361
 362         /*
 363          * Calculate the number of bytes being allocated, the number
 364          * of contiguous bits of bitmap being allocated, and the bitmap
 365          * mask.
 366          *
 367          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 368          *          mask calculation.
 369          */
 370         bytes = (size_t)1 << radix;
 371         class = (bref->type << 8) | hammer2_devblkradix(radix);
 372
 373         /*
 374          * Lookup the level1 freemap chain, creating and initializing one
 375          * if necessary.  Intermediate levels will be created automatically
 376          * when necessary by hammer2_chain_create().
 377          */
 378         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 379         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 380         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 381         l1mask = l1size - 1;
 382
 383         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 384                                      &cache_index,
 385                                      HAMMER2_LOOKUP_FREEMAP |
 386                                      HAMMER2_LOOKUP_ALWAYS |
 387                                      HAMMER2_LOOKUP_MATCHIND);
 388
 389         if (chain == NULL) {
 390                 /*
 391                  * Create the missing leaf, be sure to initialize
 392                  * the auxillary freemap tracking information in
 393                  * the bref.check.freemap structure.
 394                  */
 395 #if 0
 396                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 397                         key, iter->bpref);
 398 #endif
 399                 error = hammer2_chain_create(trans, parentp, &chain,
 400                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 401                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 402                                      HAMMER2_FREEMAP_LEVELN_PSIZE);
 403                 if (error == 0) {
 404                         hammer2_chain_modify(trans, &chain, 0);
 405                         bzero(&chain->data->bmdata[0],
 406                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 407                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 408                         chain->bref.check.freemap.avail = l1size;
 409                         /* bref.methods should already be inherited */
 410
 411                         hammer2_freemap_init(trans, hmp, key, chain);
 412                 }
 413         } else if ((chain->bref.check.freemap.bigmask & (1 << radix)) == 0) {
 414                 /*
 415                  * Already flagged as not having enough space
 416                  */
 417                 error = ENOSPC;
 418         } else {
 419                 /*
 420                  * Modify existing chain to setup for adjustment.
 421                  */
 422                 hammer2_chain_modify(trans, &chain, 0);
 423         }
 424
 425         /*
 426          * Scan 2MB entries.
 427          */
 428         if (error == 0) {
 429                 hammer2_bmap_data_t *bmap;
 430                 hammer2_key_t base_key;
 431                 int count;
 432                 int start;
 433                 int n;
 434
 435                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 436                 start = (int)((iter->bnext - key) >>
 437                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 438                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 439                 hammer2_chain_modify(trans, &chain, 0);
 440
 441                 error = ENOSPC;
 442                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 443                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 444                             start - count < 0) {
 445                                 break;
 446                         }
 447                         n = start + count;
 448                         bmap = &chain->data->bmdata[n];
 449                         if (n < HAMMER2_FREEMAP_COUNT && bmap->avail &&
 450                             (bmap->class == 0 || bmap->class == class)) {
 451                                 base_key = key + n * l0size;
 452                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 453                                                            class, n, radix,
 454                                                            &base_key);
 455                                 if (error != ENOSPC) {
 456                                         key = base_key;
 457                                         break;
 458                                 }
 459                         }
 460                         n = start - count;
 461                         bmap = &chain->data->bmdata[n];
 462                         if (n >= 0 && bmap->avail &&
 463                             (bmap->class == 0 || bmap->class == class)) {
 464                                 base_key = key + n * l0size;
 465                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 466                                                            class, n, radix,
 467                                                            &base_key);
 468                                 if (error != ENOSPC) {
 469                                         key = base_key;
 470                                         break;
 471                                 }
 472                         }
 473                 }
 474                 if (error == ENOSPC)
 475                         chain->bref.check.freemap.bigmask &= ~(1 << radix);
 476                 /* XXX also scan down from original count */
 477         }
 478
 479         if (error == 0) {
 480                 /*
 481                  * Assert validity.  Must be beyond the static allocator used
 482                  * by newfs_hammer2 (and thus also beyond the aux area),
 483                  * not go past the volume size, and must not be in the
 484                  * reserved segment area for a zone.
 485                  */
 486                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 487                          key + bytes <= hmp->voldata.volu_size);
 488                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 489                 bref->data_off = key | radix;
 490
 491 #if 0
 492                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 493                         chain,
 494                         bref->key, bref->data_off, chain->bref.data_off);
 495 #endif
 496         } else if (error == ENOSPC) {
 497                 /*
 498                  * Return EAGAIN with next iteration in iter->bnext, or
 499                  * return ENOSPC if the allocation map has been exhausted.
 500                  */
 501                 error = hammer2_freemap_iterate(trans, parentp, &chain, iter);
 502         }
 503
 504         /*
 505          * Cleanup
 506          */
 507         if (chain)
 508                 hammer2_chain_unlock(chain);
 509         return (error);
 510 }
 511
 512 /*
 513  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 514  *
 515  * If the linear iterator is mid-block we use it directly (the bitmap should
 516  * already be marked allocated), otherwise we search for a block in the bitmap
 517  * that fits the allocation request.
 518  *
 519  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 520  * to fully allocated and adjusts the linear allocator to allow the
 521  * remaining space to be allocated.
 522  */
 523 static
 524 int
 525 hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 526                    hammer2_bmap_data_t *bmap,
 527                    uint16_t class, int n, int radix, hammer2_key_t *basep)
 528 {
 529         hammer2_io_t *dio;
 530         size_t size;
 531         size_t bsize;
 532         int bmradix;
 533         uint32_t bmmask;
 534         int offset;
 535         int error;
 536         int i;
 537         int j;
 538
 539         /*
 540          * Take into account 2-bits per block when calculating bmradix.
 541          */
 542         size = (size_t)1 << radix;
 543
 544         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 545                 bmradix = 2;
 546                 bsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 547                 /* (16K) 2 bits per allocation block */
 548         } else {
 549                 bmradix = 2 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 550                 bsize = size;
 551                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 552         }
 553
 554         /*
 555          * Use the linear iterator to pack small allocations, otherwise
 556          * fall-back to finding a free 16KB chunk.  The linear iterator
 557          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 558          * If it is the bitmap must be scanned.  It can become invalid
 559          * once we pack to the boundary.  We adjust it after a bitmap
 560          * allocation only for sub-16KB allocations (so the perfectly good
 561          * previous value can still be used for fragments when 16KB+
 562          * allocations are made).
 563          *
 564          * Beware of hardware artifacts when bmradix == 32 (intermediate
 565          * result can wind up being '1' instead of '0' if hardware masks
 566          * bit-count & 31).
 567          *
 568          * NOTE: j needs to be even in the j= calculation.  As an artifact
 569          *       of the /2 division, our bitmask has to clear bit 0.
 570          *
 571          * NOTE: TODO this can leave little unallocatable fragments lying
 572          *       around.
 573          */
 574         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 575             HAMMER2_FREEMAP_BLOCK_SIZE &&
 576             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 577             bmap->linear < HAMMER2_SEGSIZE) {
 578                 KKASSERT(bmap->linear >= 0 &&
 579                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 580                          (bmap->linear & (HAMMER2_MIN_ALLOC - 1)) == 0);
 581                 offset = bmap->linear;
 582                 i = offset / (HAMMER2_SEGSIZE / 8);
 583                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 30;
 584                 bmmask = (bmradix == 32) ?
 585                          0xFFFFFFFFU : (1 << bmradix) - 1;
 586                 bmmask <<= j;
 587                 bmap->linear = offset + size;
 588         } else {
 589                 for (i = 0; i < 8; ++i) {
 590                         bmmask = (bmradix == 32) ?
 591                                  0xFFFFFFFFU : (1 << bmradix) - 1;
 592                         for (j = 0; j < 32; j += bmradix) {
 593                                 if ((bmap->bitmap[i] & bmmask) == 0)
 594                                         goto success;
 595                                 bmmask <<= bmradix;
 596                         }
 597                 }
 598                 /*fragments might remain*/
 599                 /*KKASSERT(bmap->avail == 0);*/
 600                 return (ENOSPC);
 601 success:
 602                 offset = i * (HAMMER2_SEGSIZE / 8) +
 603                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 604                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 605                         bmap->linear = offset + size;
 606         }
 607
 608         KKASSERT(i >= 0 && i < 8);      /* 8 x 16 -> 128 x 16K -> 2MB */
 609
 610         /*
 611          * Optimize the buffer cache to avoid unnecessary read-before-write
 612          * operations.
 613          *
 614          * The device block size could be larger than the allocation size
 615          * so the actual bitmap test is somewhat more involved.  We have
 616          * to use a compatible buffer size for this operation.
 617          */
 618         if ((bmap->bitmap[i] & bmmask) == 0 &&
 619             hammer2_devblksize(size) != size) {
 620                 size_t psize = hammer2_devblksize(size);
 621                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 622                 int pbmradix = 2 << (hammer2_devblkradix(radix) -
 623                                      HAMMER2_FREEMAP_BLOCK_RADIX);
 624                 uint32_t pbmmask;
 625                 int pradix = hammer2_getradix(psize);
 626
 627                 pbmmask = (pbmradix == 32) ? 0xFFFFFFFFU : (1 << pbmradix) - 1;
 628                 while ((pbmmask & bmmask) == 0)
 629                         pbmmask <<= pbmradix;
 630
 631 #if 0
 632                 kprintf("%016jx mask %08x %08x %08x (%zd/%zd)\n",
 633                         *basep + offset, bmap->bitmap[i],
 634                         pbmmask, bmmask, size, psize);
 635 #endif
 636
 637                 if ((bmap->bitmap[i] & pbmmask) == 0) {
 638                         error = hammer2_io_newq(hmp,
 639                                                 (*basep + (offset & ~pmask)) |
 640                                                  pradix,
 641                                                 psize, &dio);
 642                         hammer2_io_bqrelse(&dio);
 643                 }
 644         }
 645
 646 #if 0
 647         /*
 648          * When initializing a new inode segment also attempt to initialize
 649          * an adjacent segment.  Be careful not to index beyond the array
 650          * bounds.
 651          *
 652          * We do this to try to localize inode accesses to improve
 653          * directory scan rates.  XXX doesn't improve scan rates.
 654          */
 655         if (size == HAMMER2_INODE_BYTES) {
 656                 if (n & 1) {
 657                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 658                                 bmap[-1].radix = radix;
 659                 } else {
 660                         if (bmap[1].radix == 0 && bmap[1].avail)
 661                                 bmap[1].radix = radix;
 662                 }
 663         }
 664 #endif
 665
 666         /*
 667          * Adjust the linear iterator, set the radix if necessary (might as
 668          * well just set it unconditionally), adjust *basep to return the
 669          * allocated data offset.
 670          */
 671         bmap->bitmap[i] |= bmmask;
 672         bmap->class = class;
 673         bmap->avail -= size;
 674         *basep += offset;
 675
 676         hammer2_voldata_lock(hmp);
 677         hmp->voldata.allocator_free -= size;  /* XXX */
 678         hammer2_voldata_unlock(hmp, 1);
 679
 680         return(0);
 681 }
 682
 683 static
 684 void
 685 hammer2_freemap_init(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 686                      hammer2_key_t key, hammer2_chain_t *chain)
 687 {
 688         hammer2_off_t l1size;
 689         hammer2_off_t lokey;
 690         hammer2_off_t hikey;
 691         hammer2_bmap_data_t *bmap;
 692         int count;
 693
 694         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 695
 696         /*
 697          * Calculate the portion of the 2GB map that should be initialized
 698          * as free.  Portions below or after will be initialized as allocated.
 699          * SEGMASK-align the areas so we don't have to worry about sub-scans
 700          * or endianess when using memset.
 701          *
 702          * (1) Ensure that all statically allocated space from newfs_hammer2
 703          *     is marked allocated.
 704          *
 705          * (2) Ensure that the reserved area is marked allocated (typically
 706          *     the first 4MB of the 2GB area being represented).
 707          *
 708          * (3) Ensure that any trailing space at the end-of-volume is marked
 709          *     allocated.
 710          *
 711          * WARNING! It is possible for lokey to be larger than hikey if the
 712          *          entire 2GB segment is within the static allocation.
 713          */
 714         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 715                 ~HAMMER2_SEGMASK64;
 716
 717         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 718                   HAMMER2_ZONE_SEG64) {
 719                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 720                         HAMMER2_ZONE_SEG64;
 721         }
 722
 723         hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 724         if (hikey > hmp->voldata.volu_size) {
 725                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 726         }
 727
 728         chain->bref.check.freemap.avail =
 729                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 730         bmap = &chain->data->bmdata[0];
 731
 732         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 733                 if (key < lokey || key >= hikey) {
 734                         memset(bmap->bitmap, -1,
 735                                sizeof(bmap->bitmap));
 736                         bmap->avail = 0;
 737                         bmap->linear = HAMMER2_SEGSIZE;
 738                         chain->bref.check.freemap.avail -=
 739                                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 740                 } else {
 741                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 742                 }
 743                 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 744                 ++bmap;
 745         }
 746 }
 747
 748 /*
 749  * The current Level 1 freemap has been exhausted, iterate to the next
 750  * one, return ENOSPC if no freemaps remain.
 751  *
 752  * XXX this should rotate back to the beginning to handle freed-up space
 753  * XXX or use intermediate entries to locate free space. TODO
 754  */
 755 static int
 756 hammer2_freemap_iterate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 757                         hammer2_chain_t **chainp, hammer2_fiterate_t *iter)
 758 {
 759         hammer2_mount_t *hmp = (*parentp)->hmp;
 760
 761         iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 762         iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 763         if (iter->bnext >= hmp->voldata.volu_size) {
 764                 iter->bnext = 0;
 765                 if (++iter->loops == 2)
 766                         return (ENOSPC);
 767         }
 768         return(EAGAIN);
 769 }
 770
 771 /*
 772  * Free the specified blockref.  This code is only able to fully free
 773  * blocks when (how) is non-zero, otherwise the block is marked for
 774  * the bulk freeing pass to check.
 775  *
 776  * Normal use is to only mark inodes as possibly being free.  The underlying
 777  * file blocks are not necessarily marked.  The bulk freescan can
 778  * theoretically handle the case.
 779  *
 780  * XXX currently disabled when how == 0 (the normal real-time case).  At
 781  * the moment we depend on the bulk freescan to actually free blocks.  It
 782  * will still call this routine with a non-zero how to stage possible frees
 783  * and to do the actual free.
 784  *
 785  * WARNING! When called from a flush we have to use the 'live' sync_tid
 786  *          and not the flush sync_tid.  The live sync_tid is the flush
 787  *          sync_tid + 1.  That is, freemap allocations which occur during
 788  *          a flush are not part of the flush.  Crash-recovery will restore
 789  *          any lost allocations.
 790  */
 791 void
 792 hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 793                        hammer2_blockref_t *bref, int how)
 794 {
 795         hammer2_off_t data_off = bref->data_off;
 796         hammer2_chain_t *chain;
 797         hammer2_chain_t *parent;
 798         hammer2_bmap_data_t *bmap;
 799         hammer2_key_t key;
 800         hammer2_key_t key_dummy;
 801         hammer2_off_t l0size;
 802         hammer2_off_t l1size;
 803         hammer2_off_t l1mask;
 804         uint32_t *bitmap;
 805         const uint32_t bmmask00 = 0;
 806         uint32_t bmmask01;
 807         uint32_t bmmask10;
 808         uint32_t bmmask11;
 809         size_t bytes;
 810         uint16_t class;
 811         int radix;
 812         int start;
 813         int count;
 814         int modified = 0;
 815         int cache_index = -1;
 816         int error;
 817
 818         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
 819         data_off &= ~HAMMER2_OFF_MASK_RADIX;
 820         KKASSERT(radix <= HAMMER2_MAX_RADIX);
 821
 822         bytes = (size_t)1 << radix;
 823         class = (bref->type << 8) | hammer2_devblkradix(radix);
 824
 825         /*
 826          * We can't adjust thre freemap for data allocations made by
 827          * newfs_hammer2.
 828          */
 829         if (data_off < hmp->voldata.allocator_beg)
 830                 return;
 831
 832         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 833         KKASSERT((trans->flags & HAMMER2_TRANS_ISALLOCATING) == 0);
 834         atomic_set_int(&trans->flags, HAMMER2_TRANS_ISALLOCATING);
 835         if (trans->flags & HAMMER2_TRANS_ISFLUSH)
 836                 ++trans->sync_tid;
 837
 838         /*
 839          * Lookup the level1 freemap chain.  The chain must exist.
 840          */
 841         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
 842         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 843         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 844         l1mask = l1size - 1;
 845
 846         parent = &hmp->fchain;
 847         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 848
 849         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
 850                                      &cache_index,
 851                                      HAMMER2_LOOKUP_FREEMAP |
 852                                      HAMMER2_LOOKUP_ALWAYS |
 853                                      HAMMER2_LOOKUP_MATCHIND);
 854
 855         /*
 856          * Stop early if we are trying to free something but no leaf exists.
 857          */
 858         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
 859                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
 860                         (intmax_t)bref->data_off);
 861                 goto done;
 862         }
 863
 864         /*
 865          * Create any missing leaf(s) if we are doing a recovery (marking
 866          * the block(s) as being allocated instead of being freed).  Be sure
 867          * to initialize the auxillary freemap tracking info in the
 868          * bref.check.freemap structure.
 869          */
 870         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
 871                 error = hammer2_chain_create(trans, &parent, &chain,
 872                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 873                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 874                                      HAMMER2_FREEMAP_LEVELN_PSIZE);
 875                 kprintf("fixup create chain %p %016jx:%d\n", chain, chain->bref.key, chain->bref.keybits);
 876
 877                 if (error == 0) {
 878                         hammer2_chain_modify(trans, &chain, 0);
 879                         bzero(&chain->data->bmdata[0],
 880                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 881                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 882                         chain->bref.check.freemap.avail = l1size;
 883                         /* bref.methods should already be inherited */
 884
 885                         hammer2_freemap_init(trans, hmp, key, chain);
 886                 }
 887                 /* XXX handle error */
 888         }
 889
 890         /*
 891          * Calculate the bitmask (runs in 2-bit pairs).
 892          */
 893         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
 894         bmmask01 = 1 << start;
 895         bmmask10 = 2 << start;
 896         bmmask11 = 3 << start;
 897
 898         /*
 899          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
 900          * a bulk scan is able to roll them up.
 901          */
 902         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 903                 count = 1;
 904                 if (how == HAMMER2_FREEMAP_DOREALFREE)
 905                         how = HAMMER2_FREEMAP_DOMAYFREE;
 906         } else {
 907                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 908         }
 909
 910         /*
 911          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
 912          * a 2MB swath.  The bmap itself (LEVEL1) covers 2GB.
 913          */
 914 again:
 915         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
 916                                     (HAMMER2_FREEMAP_COUNT - 1)];
 917         bitmap = &bmap->bitmap[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
 918
 919
 920         while (count) {
 921                 KKASSERT(bmmask11);
 922                 if (how == HAMMER2_FREEMAP_DORECOVER) {
 923                         /*
 924                          * Recovery request, mark as allocated.
 925                          */
 926                         if ((*bitmap & bmmask11) != bmmask11) {
 927                                 if (modified == 0) {
 928                                         hammer2_chain_modify(trans, &chain, 0);
 929                                         modified = 1;
 930                                         goto again;
 931                                 }
 932                                 if ((*bitmap & bmmask11) == bmmask00)
 933                                         bmap->avail -= 1 << radix;
 934                                 if (bmap->class == 0)
 935                                         bmap->class = class;
 936                                 *bitmap |= bmmask11;
 937                                 kprintf("hammer2_freemap_recover: fixup "
 938                                         "type=%02x block=%016jx/%zd\n",
 939                                         bref->type, data_off, bytes);
 940                         } else {
 941                                 /*
 942                                 kprintf("hammer2_freemap_recover:  good "
 943                                         "type=%02x block=%016jx/%zd\n",
 944                                         bref->type, data_off, bytes);
 945                                 */
 946                         }
 947                 } else if ((*bitmap & bmmask11) == bmmask11) {
 948                         /*
 949                          * Mayfree/Realfree request and bitmap is currently
 950                          * marked as being fully allocated.
 951                          */
 952                         if (!modified) {
 953                                 hammer2_chain_modify(trans, &chain, 0);
 954                                 modified = 1;
 955                                 goto again;
 956                         }
 957                         if (how == HAMMER2_FREEMAP_DOREALFREE)
 958                                 *bitmap &= ~bmmask11;
 959                         else
 960                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
 961                 } else if ((*bitmap & bmmask11) == bmmask10) {
 962                         /*
 963                          * Mayfree/Realfree request and bitmap is currently
 964                          * marked as being possibly freeable.
 965                          */
 966                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
 967                                 if (!modified) {
 968                                         hammer2_chain_modify(trans, &chain, 0);
 969                                         modified = 1;
 970                                         goto again;
 971                                 }
 972                                 *bitmap &= ~bmmask11;
 973                         }
 974                 } else {
 975                         /*
 976                          * 01 - Not implemented, currently illegal state
 977                          * 00 - Not allocated at all, illegal free.
 978                          */
 979                         panic("hammer2_freemap_adjust: "
 980                               "Illegal state %08x(%08x)",
 981                               *bitmap, *bitmap & bmmask11);
 982                 }
 983                 --count;
 984                 bmmask01 <<= 2;
 985                 bmmask10 <<= 2;
 986                 bmmask11 <<= 2;
 987         }
 988         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
 989                 bmap->avail += 1 << radix;
 990                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
 991                 if (bmap->avail == HAMMER2_SEGSIZE &&
 992                     bmap->bitmap[0] == 0 &&
 993                     bmap->bitmap[1] == 0 &&
 994                     bmap->bitmap[2] == 0 &&
 995                     bmap->bitmap[3] == 0 &&
 996                     bmap->bitmap[4] == 0 &&
 997                     bmap->bitmap[5] == 0 &&
 998                     bmap->bitmap[6] == 0 &&
 999                     bmap->bitmap[7] == 0) {
1000                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1001                         kprintf("Freeseg %016jx\n", (intmax_t)key);
1002                         bmap->class = 0;
1003                 }
1004         }
1005
1006         /*
1007          * chain->bref.check.freemap.bigmask (XXX)
1008          *
1009          * Setting bigmask is a hint to the allocation code that there might
1010          * be something allocatable.  We also set this in recovery... it
1011          * doesn't hurt and we might want to use the hint for other validation
1012          * operations later on.
1013          */
1014         if (modified)
1015                 chain->bref.check.freemap.bigmask |= 1 << radix;
1016
1017         hammer2_chain_unlock(chain);
1018 done:
1019         hammer2_chain_unlock(parent);
1020         atomic_clear_int(&trans->flags, HAMMER2_TRANS_ISALLOCATING);
1021         if (trans->flags & HAMMER2_TRANS_ISFLUSH)
1022                 --trans->sync_tid;
1023 }