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.  If this
 111          * is the first allocation of the block (verses a modification of an
 112          * existing block), we use index 0, otherwise we use the next rotating
 113          * index.
 114          *
 115          * NORMAL transactions use FREEMAP sections 0-5, while FREEBATCH
 116          * transactions use sections 6 and 7.  FREEBATCH transactions are
 117          * used by the batch freeing code to spool-off in-memory structures
 118          * used to track the batch free scan.
 119          */
 120         if (trans->flags & HAMMER2_TRANS_FREEBATCH) {
 121                 if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 122                         index = (HAMMER2_ZONE_FREEMAP_06 -
 123                                  HAMMER2_ZONE_FREEMAP_00) / 4;
 124                 } else {
 125                         off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 126                               (((hammer2_off_t)1 <<
 127                                 HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 128                         off = off / HAMMER2_PBUFSIZE;
 129                         KKASSERT(off >= HAMMER2_ZONE_FREEMAP_06 &&
 130                                  off < HAMMER2_ZONE_FREEMAP_08);
 131                         index = (int)(off - HAMMER2_ZONE_FREEMAP_00) / 4;
 132                         KKASSERT(index >= 6 && index < 8);
 133                         if (++index == 8)
 134                                 index = 6;
 135                 }
 136         } else {
 137                 if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 138                         index = 0;
 139                 } else {
 140                         off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 141                               (((hammer2_off_t)1 <<
 142                                 HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 143                         off = off / HAMMER2_PBUFSIZE;
 144                         KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
 145                                  off < HAMMER2_ZONE_FREEMAP_06);
 146                         index = (int)(off - HAMMER2_ZONE_FREEMAP_00) / 4;
 147                         KKASSERT(index >= 0 && index < 6);
 148                         if (++index == 6)
 149                                 index = 0;
 150                 }
 151         }
 152
 153         /*
 154          * Calculate the block offset of the reserved block.  This will
 155          * point into the 4MB reserved area at the base of the appropriate
 156          * 2GB zone, once added to the FREEMAP_x selection above.
 157          */
 158         switch(bref->keybits) {
 159         /* case HAMMER2_FREEMAP_LEVEL5_RADIX: not applicable */
 160         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 2EB */
 161                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 162                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 163                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 164                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 165                        HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
 166                 break;
 167         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 2PB */
 168                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 169                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 170                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 171                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 172                        HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
 173                 break;
 174         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 2TB */
 175                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 176                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 177                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 178                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 179                        HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
 180                 break;
 181         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 2GB */
 182                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 183                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 184                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 185                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 186                        HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
 187                 break;
 188         default:
 189                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 190                 /* NOT REACHED */
 191                 off = (hammer2_off_t)-1;
 192                 break;
 193         }
 194         bref->data_off = off | radix;
 195 #if FREEMAP_DEBUG
 196         kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 197                 bref->type, bref->key, bref->keybits, bref->data_off);
 198 #endif
 199         return (0);
 200 }
 201
 202 /*
 203  * Normal freemap allocator
 204  *
 205  * Use available hints to allocate space using the freemap.  Create missing
 206  * freemap infrastructure on-the-fly as needed (including marking initial
 207  * allocations using the iterator as allocated, instantiating new 2GB zones,
 208  * and dealing with the end-of-media edge case).
 209  *
 210  * ip and bpref are only used as a heuristic to determine locality of
 211  * reference.  bref->key may also be used heuristically.
 212  */
 213 int
 214 hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 215                       size_t bytes)
 216 {
 217         hammer2_mount_t *hmp = chain->hmp;
 218         hammer2_blockref_t *bref = &chain->bref;
 219         hammer2_chain_t *parent;
 220         int radix;
 221         int error;
 222         unsigned int hindex;
 223         hammer2_fiterate_t iter;
 224
 225         /*
 226          * Validate the allocation size.  It must be a power of 2.
 227          *
 228          * For now require that the caller be aware of the minimum
 229          * allocation (1K).
 230          */
 231         radix = hammer2_getradix(bytes);
 232         KKASSERT((size_t)1 << radix == bytes);
 233
 234         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 235             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 236                 /*
 237                  * Freemap blocks themselves are assigned from the reserve
 238                  * area, not allocated from the freemap.
 239                  */
 240                 error = hammer2_freemap_reserve(trans, chain, radix);
 241                 return error;
 242         }
 243
 244         KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
 245
 246         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 247                 ++trans->sync_xid;
 248
 249         /*
 250          * Calculate the starting point for our allocation search.
 251          *
 252          * Each freemap leaf is dedicated to a specific freemap_radix.
 253          * The freemap_radix can be more fine-grained than the device buffer
 254          * radix which results in inodes being grouped together in their
 255          * own segment, terminal-data (16K or less) and initial indirect
 256          * block being grouped together, and then full-indirect and full-data
 257          * blocks (64K) being grouped together.
 258          *
 259          * The single most important aspect of this is the inode grouping
 260          * because that is what allows 'find' and 'ls' and other filesystem
 261          * topology operations to run fast.
 262          */
 263 #if 0
 264         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 265                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 266         else if (trans->tmp_bpref)
 267                 bpref = trans->tmp_bpref;
 268         else if (trans->tmp_ip)
 269                 bpref = trans->tmp_ip->chain->bref.data_off;
 270         else
 271 #endif
 272         /*
 273          * Heuristic tracking index.  We would like one for each distinct
 274          * bref type if possible.  heur_freemap[] has room for two classes
 275          * for each type.  At a minimum we have to break-up our heuristic
 276          * by device block sizes.
 277          */
 278         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 279         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 280         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 281         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 282         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR);
 283
 284         iter.bpref = hmp->heur_freemap[hindex];
 285
 286         /*
 287          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 288          * reserved area, the try code will iterate past it.
 289          */
 290         if (iter.bpref > hmp->voldata.volu_size)
 291                 iter.bpref = hmp->voldata.volu_size - 1;
 292
 293         /*
 294          * Iterate the freemap looking for free space before and after.
 295          */
 296         parent = &hmp->fchain;
 297         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 298         error = EAGAIN;
 299         iter.bnext = iter.bpref;
 300         iter.loops = 0;
 301
 302         while (error == EAGAIN) {
 303                 error = hammer2_freemap_try_alloc(trans, &parent, bref,
 304                                                   radix, &iter);
 305         }
 306         hmp->heur_freemap[hindex] = iter.bnext;
 307         hammer2_chain_unlock(parent);
 308
 309         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 310                 --trans->sync_xid;
 311
 312         return (error);
 313 }
 314
 315 static int
 316 hammer2_freemap_try_alloc(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 317                           hammer2_blockref_t *bref, int radix,
 318                           hammer2_fiterate_t *iter)
 319 {
 320         hammer2_mount_t *hmp = (*parentp)->hmp;
 321         hammer2_off_t l0size;
 322         hammer2_off_t l1size;
 323         hammer2_off_t l1mask;
 324         hammer2_key_t key_dummy;
 325         hammer2_chain_t *chain;
 326         hammer2_off_t key;
 327         size_t bytes;
 328         uint16_t class;
 329         int error = 0;
 330         int cache_index = -1;
 331         int ddflag;
 332
 333
 334         /*
 335          * Calculate the number of bytes being allocated, the number
 336          * of contiguous bits of bitmap being allocated, and the bitmap
 337          * mask.
 338          *
 339          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 340          *          mask calculation.
 341          */
 342         bytes = (size_t)1 << radix;
 343         class = (bref->type << 8) | hammer2_devblkradix(radix);
 344
 345         /*
 346          * Lookup the level1 freemap chain, creating and initializing one
 347          * if necessary.  Intermediate levels will be created automatically
 348          * when necessary by hammer2_chain_create().
 349          */
 350         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 351         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 352         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 353         l1mask = l1size - 1;
 354
 355         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 356                                      &cache_index,
 357                                      HAMMER2_LOOKUP_ALWAYS |
 358                                      HAMMER2_LOOKUP_MATCHIND, &ddflag);
 359
 360         if (chain == NULL) {
 361                 /*
 362                  * Create the missing leaf, be sure to initialize
 363                  * the auxillary freemap tracking information in
 364                  * the bref.check.freemap structure.
 365                  */
 366 #if 0
 367                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 368                         key, iter->bpref);
 369 #endif
 370                 error = hammer2_chain_create(trans, parentp, &chain, hmp->spmp,
 371                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 372                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 373                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 374                                      0);
 375                 KKASSERT(error == 0);
 376                 if (error == 0) {
 377                         hammer2_chain_modify(trans, chain, 0);
 378                         bzero(&chain->data->bmdata[0],
 379                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 380                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 381                         chain->bref.check.freemap.avail = l1size;
 382                         /* bref.methods should already be inherited */
 383
 384                         hammer2_freemap_init(trans, hmp, key, chain);
 385                 }
 386         } else if ((chain->bref.check.freemap.bigmask & (1 << radix)) == 0) {
 387                 /*
 388                  * Already flagged as not having enough space
 389                  */
 390                 error = ENOSPC;
 391         } else {
 392                 /*
 393                  * Modify existing chain to setup for adjustment.
 394                  */
 395                 hammer2_chain_modify(trans, chain, 0);
 396         }
 397
 398         /*
 399          * Scan 2MB entries.
 400          */
 401         if (error == 0) {
 402                 hammer2_bmap_data_t *bmap;
 403                 hammer2_key_t base_key;
 404                 int count;
 405                 int start;
 406                 int n;
 407
 408                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 409                 start = (int)((iter->bnext - key) >>
 410                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 411                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 412                 hammer2_chain_modify(trans, chain, 0);
 413
 414                 error = ENOSPC;
 415                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 416                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 417                             start - count < 0) {
 418                                 break;
 419                         }
 420                         n = start + count;
 421                         bmap = &chain->data->bmdata[n];
 422                         if (n < HAMMER2_FREEMAP_COUNT && bmap->avail &&
 423                             (bmap->class == 0 || bmap->class == class)) {
 424                                 base_key = key + n * l0size;
 425                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 426                                                            class, n, radix,
 427                                                            &base_key);
 428                                 if (error != ENOSPC) {
 429                                         key = base_key;
 430                                         break;
 431                                 }
 432                         }
 433                         n = start - count;
 434                         bmap = &chain->data->bmdata[n];
 435                         if (n >= 0 && bmap->avail &&
 436                             (bmap->class == 0 || bmap->class == class)) {
 437                                 base_key = key + n * l0size;
 438                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 439                                                            class, n, radix,
 440                                                            &base_key);
 441                                 if (error != ENOSPC) {
 442                                         key = base_key;
 443                                         break;
 444                                 }
 445                         }
 446                 }
 447                 if (error == ENOSPC)
 448                         chain->bref.check.freemap.bigmask &= ~(1 << radix);
 449                 /* XXX also scan down from original count */
 450         }
 451
 452         if (error == 0) {
 453                 /*
 454                  * Assert validity.  Must be beyond the static allocator used
 455                  * by newfs_hammer2 (and thus also beyond the aux area),
 456                  * not go past the volume size, and must not be in the
 457                  * reserved segment area for a zone.
 458                  */
 459                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 460                          key + bytes <= hmp->voldata.volu_size);
 461                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 462                 bref->data_off = key | radix;
 463
 464 #if 0
 465                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 466                         chain,
 467                         bref->key, bref->data_off, chain->bref.data_off);
 468 #endif
 469         } else if (error == ENOSPC) {
 470                 /*
 471                  * Return EAGAIN with next iteration in iter->bnext, or
 472                  * return ENOSPC if the allocation map has been exhausted.
 473                  */
 474                 error = hammer2_freemap_iterate(trans, parentp, &chain, iter);
 475         }
 476
 477         /*
 478          * Cleanup
 479          */
 480         if (chain)
 481                 hammer2_chain_unlock(chain);
 482         return (error);
 483 }
 484
 485 /*
 486  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 487  *
 488  * If the linear iterator is mid-block we use it directly (the bitmap should
 489  * already be marked allocated), otherwise we search for a block in the bitmap
 490  * that fits the allocation request.
 491  *
 492  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 493  * to fully allocated and adjusts the linear allocator to allow the
 494  * remaining space to be allocated.
 495  */
 496 static
 497 int
 498 hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 499                    hammer2_bmap_data_t *bmap,
 500                    uint16_t class, int n, int radix, hammer2_key_t *basep)
 501 {
 502         hammer2_io_t *dio;
 503         size_t size;
 504         size_t bsize;
 505         int bmradix;
 506         uint32_t bmmask;
 507         int offset;
 508         int error;
 509         int i;
 510         int j;
 511
 512         /*
 513          * Take into account 2-bits per block when calculating bmradix.
 514          */
 515         size = (size_t)1 << radix;
 516
 517         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 518                 bmradix = 2;
 519                 bsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 520                 /* (16K) 2 bits per allocation block */
 521         } else {
 522                 bmradix = 2 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 523                 bsize = size;
 524                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 525         }
 526
 527         /*
 528          * Use the linear iterator to pack small allocations, otherwise
 529          * fall-back to finding a free 16KB chunk.  The linear iterator
 530          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 531          * If it is the bitmap must be scanned.  It can become invalid
 532          * once we pack to the boundary.  We adjust it after a bitmap
 533          * allocation only for sub-16KB allocations (so the perfectly good
 534          * previous value can still be used for fragments when 16KB+
 535          * allocations are made).
 536          *
 537          * Beware of hardware artifacts when bmradix == 32 (intermediate
 538          * result can wind up being '1' instead of '0' if hardware masks
 539          * bit-count & 31).
 540          *
 541          * NOTE: j needs to be even in the j= calculation.  As an artifact
 542          *       of the /2 division, our bitmask has to clear bit 0.
 543          *
 544          * NOTE: TODO this can leave little unallocatable fragments lying
 545          *       around.
 546          */
 547         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 548             HAMMER2_FREEMAP_BLOCK_SIZE &&
 549             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 550             bmap->linear < HAMMER2_SEGSIZE) {
 551                 KKASSERT(bmap->linear >= 0 &&
 552                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 553                          (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
 554                 offset = bmap->linear;
 555                 i = offset / (HAMMER2_SEGSIZE / 8);
 556                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 30;
 557                 bmmask = (bmradix == 32) ?
 558                          0xFFFFFFFFU : (1 << bmradix) - 1;
 559                 bmmask <<= j;
 560                 bmap->linear = offset + size;
 561         } else {
 562                 for (i = 0; i < 8; ++i) {
 563                         bmmask = (bmradix == 32) ?
 564                                  0xFFFFFFFFU : (1 << bmradix) - 1;
 565                         for (j = 0; j < 32; j += bmradix) {
 566                                 if ((bmap->bitmap[i] & bmmask) == 0)
 567                                         goto success;
 568                                 bmmask <<= bmradix;
 569                         }
 570                 }
 571                 /*fragments might remain*/
 572                 /*KKASSERT(bmap->avail == 0);*/
 573                 return (ENOSPC);
 574 success:
 575                 offset = i * (HAMMER2_SEGSIZE / 8) +
 576                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 577                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 578                         bmap->linear = offset + size;
 579         }
 580
 581         KKASSERT(i >= 0 && i < 8);      /* 8 x 16 -> 128 x 16K -> 2MB */
 582
 583         /*
 584          * Optimize the buffer cache to avoid unnecessary read-before-write
 585          * operations.
 586          *
 587          * The device block size could be larger than the allocation size
 588          * so the actual bitmap test is somewhat more involved.  We have
 589          * to use a compatible buffer size for this operation.
 590          */
 591         if ((bmap->bitmap[i] & bmmask) == 0 &&
 592             hammer2_devblksize(size) != size) {
 593                 size_t psize = hammer2_devblksize(size);
 594                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 595                 int pbmradix = 2 << (hammer2_devblkradix(radix) -
 596                                      HAMMER2_FREEMAP_BLOCK_RADIX);
 597                 uint32_t pbmmask;
 598                 int pradix = hammer2_getradix(psize);
 599
 600                 pbmmask = (pbmradix == 32) ? 0xFFFFFFFFU : (1 << pbmradix) - 1;
 601                 while ((pbmmask & bmmask) == 0)
 602                         pbmmask <<= pbmradix;
 603
 604 #if 0
 605                 kprintf("%016jx mask %08x %08x %08x (%zd/%zd)\n",
 606                         *basep + offset, bmap->bitmap[i],
 607                         pbmmask, bmmask, size, psize);
 608 #endif
 609
 610                 if ((bmap->bitmap[i] & pbmmask) == 0) {
 611                         error = hammer2_io_newq(hmp,
 612                                                 (*basep + (offset & ~pmask)) |
 613                                                  pradix,
 614                                                 psize, &dio);
 615                         hammer2_io_bqrelse(&dio);
 616                 }
 617         }
 618
 619 #if 0
 620         /*
 621          * When initializing a new inode segment also attempt to initialize
 622          * an adjacent segment.  Be careful not to index beyond the array
 623          * bounds.
 624          *
 625          * We do this to try to localize inode accesses to improve
 626          * directory scan rates.  XXX doesn't improve scan rates.
 627          */
 628         if (size == HAMMER2_INODE_BYTES) {
 629                 if (n & 1) {
 630                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 631                                 bmap[-1].radix = radix;
 632                 } else {
 633                         if (bmap[1].radix == 0 && bmap[1].avail)
 634                                 bmap[1].radix = radix;
 635                 }
 636         }
 637 #endif
 638
 639         /*
 640          * Adjust the linear iterator, set the radix if necessary (might as
 641          * well just set it unconditionally), adjust *basep to return the
 642          * allocated data offset.
 643          */
 644         bmap->bitmap[i] |= bmmask;
 645         bmap->class = class;
 646         bmap->avail -= size;
 647         *basep += offset;
 648
 649         hammer2_voldata_lock(hmp);
 650         hammer2_voldata_modify(hmp);
 651         hmp->voldata.allocator_free -= size;  /* XXX */
 652         hammer2_voldata_unlock(hmp);
 653
 654         return(0);
 655 }
 656
 657 static
 658 void
 659 hammer2_freemap_init(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 660                      hammer2_key_t key, hammer2_chain_t *chain)
 661 {
 662         hammer2_off_t l1size;
 663         hammer2_off_t lokey;
 664         hammer2_off_t hikey;
 665         hammer2_bmap_data_t *bmap;
 666         int count;
 667
 668         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 669
 670         /*
 671          * Calculate the portion of the 2GB map that should be initialized
 672          * as free.  Portions below or after will be initialized as allocated.
 673          * SEGMASK-align the areas so we don't have to worry about sub-scans
 674          * or endianess when using memset.
 675          *
 676          * (1) Ensure that all statically allocated space from newfs_hammer2
 677          *     is marked allocated.
 678          *
 679          * (2) Ensure that the reserved area is marked allocated (typically
 680          *     the first 4MB of the 2GB area being represented).
 681          *
 682          * (3) Ensure that any trailing space at the end-of-volume is marked
 683          *     allocated.
 684          *
 685          * WARNING! It is possible for lokey to be larger than hikey if the
 686          *          entire 2GB segment is within the static allocation.
 687          */
 688         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 689                 ~HAMMER2_SEGMASK64;
 690
 691         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 692                   HAMMER2_ZONE_SEG64) {
 693                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 694                         HAMMER2_ZONE_SEG64;
 695         }
 696
 697         hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 698         if (hikey > hmp->voldata.volu_size) {
 699                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 700         }
 701
 702         chain->bref.check.freemap.avail =
 703                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 704         bmap = &chain->data->bmdata[0];
 705
 706         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 707                 if (key < lokey || key >= hikey) {
 708                         memset(bmap->bitmap, -1,
 709                                sizeof(bmap->bitmap));
 710                         bmap->avail = 0;
 711                         bmap->linear = HAMMER2_SEGSIZE;
 712                         chain->bref.check.freemap.avail -=
 713                                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 714                 } else {
 715                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 716                 }
 717                 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 718                 ++bmap;
 719         }
 720 }
 721
 722 /*
 723  * The current Level 1 freemap has been exhausted, iterate to the next
 724  * one, return ENOSPC if no freemaps remain.
 725  *
 726  * XXX this should rotate back to the beginning to handle freed-up space
 727  * XXX or use intermediate entries to locate free space. TODO
 728  */
 729 static int
 730 hammer2_freemap_iterate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 731                         hammer2_chain_t **chainp, hammer2_fiterate_t *iter)
 732 {
 733         hammer2_mount_t *hmp = (*parentp)->hmp;
 734
 735         iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 736         iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 737         if (iter->bnext >= hmp->voldata.volu_size) {
 738                 iter->bnext = 0;
 739                 if (++iter->loops == 2)
 740                         return (ENOSPC);
 741         }
 742         return(EAGAIN);
 743 }
 744
 745 /*
 746  * Adjust the bit-pattern for data in the freemap bitmap according to
 747  * (how).  This code is called from on-mount recovery to fixup (mark
 748  * as allocated) blocks whos freemap upates might not have been committed
 749  * in the last crash and is used by the bulk freemap scan to stage frees.
 750  *
 751  * XXX currently disabled when how == 0 (the normal real-time case).  At
 752  * the moment we depend on the bulk freescan to actually free blocks.  It
 753  * will still call this routine with a non-zero how to stage possible frees
 754  * and to do the actual free.
 755  */
 756 void
 757 hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 758                        hammer2_blockref_t *bref, int how)
 759 {
 760         hammer2_off_t data_off = bref->data_off;
 761         hammer2_chain_t *chain;
 762         hammer2_chain_t *parent;
 763         hammer2_bmap_data_t *bmap;
 764         hammer2_key_t key;
 765         hammer2_key_t key_dummy;
 766         hammer2_off_t l0size;
 767         hammer2_off_t l1size;
 768         hammer2_off_t l1mask;
 769         uint32_t *bitmap;
 770         const uint32_t bmmask00 = 0;
 771         uint32_t bmmask01;
 772         uint32_t bmmask10;
 773         uint32_t bmmask11;
 774         size_t bytes;
 775         uint16_t class;
 776         int radix;
 777         int start;
 778         int count;
 779         int modified = 0;
 780         int cache_index = -1;
 781         int error;
 782         int ddflag;
 783
 784         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
 785         data_off &= ~HAMMER2_OFF_MASK_RADIX;
 786         KKASSERT(radix <= HAMMER2_RADIX_MAX);
 787
 788         bytes = (size_t)1 << radix;
 789         class = (bref->type << 8) | hammer2_devblkradix(radix);
 790
 791         /*
 792          * We can't adjust thre freemap for data allocations made by
 793          * newfs_hammer2.
 794          */
 795         if (data_off < hmp->voldata.allocator_beg)
 796                 return;
 797
 798         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 799
 800         /*
 801          * Lookup the level1 freemap chain.  The chain must exist.
 802          */
 803         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
 804         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 805         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 806         l1mask = l1size - 1;
 807
 808         parent = &hmp->fchain;
 809         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 810
 811         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
 812                                      &cache_index,
 813                                      HAMMER2_LOOKUP_ALWAYS |
 814                                      HAMMER2_LOOKUP_MATCHIND, &ddflag);
 815
 816         /*
 817          * Stop early if we are trying to free something but no leaf exists.
 818          */
 819         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
 820                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
 821                         (intmax_t)bref->data_off);
 822                 goto done;
 823         }
 824
 825         /*
 826          * Create any missing leaf(s) if we are doing a recovery (marking
 827          * the block(s) as being allocated instead of being freed).  Be sure
 828          * to initialize the auxillary freemap tracking info in the
 829          * bref.check.freemap structure.
 830          */
 831         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
 832                 error = hammer2_chain_create(trans, &parent, &chain, hmp->spmp,
 833                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 834                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 835                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 836                                      0);
 837
 838                 if (hammer2_debug & 0x0040) {
 839                         kprintf("fixup create chain %p %016jx:%d\n",
 840                                 chain, chain->bref.key, chain->bref.keybits);
 841                 }
 842
 843                 if (error == 0) {
 844                         hammer2_chain_modify(trans, chain, 0);
 845                         bzero(&chain->data->bmdata[0],
 846                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 847                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 848                         chain->bref.check.freemap.avail = l1size;
 849                         /* bref.methods should already be inherited */
 850
 851                         hammer2_freemap_init(trans, hmp, key, chain);
 852                 }
 853                 /* XXX handle error */
 854         }
 855
 856 #if FREEMAP_DEBUG
 857         kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 858                 chain->bref.type, chain->bref.key,
 859                 chain->bref.keybits, chain->bref.data_off);
 860 #endif
 861
 862         /*
 863          * Calculate the bitmask (runs in 2-bit pairs).
 864          */
 865         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
 866         bmmask01 = 1 << start;
 867         bmmask10 = 2 << start;
 868         bmmask11 = 3 << start;
 869
 870         /*
 871          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
 872          * a bulk scan is able to roll them up.
 873          */
 874         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 875                 count = 1;
 876                 if (how == HAMMER2_FREEMAP_DOREALFREE)
 877                         how = HAMMER2_FREEMAP_DOMAYFREE;
 878         } else {
 879                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 880         }
 881
 882         /*
 883          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
 884          * a 2MB swath.  The bmap itself (LEVEL1) covers 2GB.
 885          *
 886          * Be sure to reset the linear iterator to ensure that the adjustment
 887          * is not ignored.
 888          */
 889 again:
 890         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
 891                                     (HAMMER2_FREEMAP_COUNT - 1)];
 892         bitmap = &bmap->bitmap[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
 893
 894         if (modified)
 895                 bmap->linear = 0;
 896
 897         while (count) {
 898                 KKASSERT(bmmask11);
 899                 if (how == HAMMER2_FREEMAP_DORECOVER) {
 900                         /*
 901                          * Recovery request, mark as allocated.
 902                          */
 903                         if ((*bitmap & bmmask11) != bmmask11) {
 904                                 if (modified == 0) {
 905                                         hammer2_chain_modify(trans, chain, 0);
 906                                         modified = 1;
 907                                         goto again;
 908                                 }
 909                                 if ((*bitmap & bmmask11) == bmmask00)
 910                                         bmap->avail -= 1 << radix;
 911                                 if (bmap->class == 0)
 912                                         bmap->class = class;
 913                                 *bitmap |= bmmask11;
 914                                 if (hammer2_debug & 0x0040) {
 915                                         kprintf("hammer2_freemap_recover: "
 916                                                 "fixup type=%02x "
 917                                                 "block=%016jx/%zd\n",
 918                                                 bref->type, data_off, bytes);
 919                                 }
 920                         } else {
 921                                 /*
 922                                 kprintf("hammer2_freemap_recover:  good "
 923                                         "type=%02x block=%016jx/%zd\n",
 924                                         bref->type, data_off, bytes);
 925                                 */
 926                         }
 927                 } else if ((*bitmap & bmmask11) == bmmask11) {
 928                         /*
 929                          * Mayfree/Realfree request and bitmap is currently
 930                          * marked as being fully allocated.
 931                          */
 932                         if (!modified) {
 933                                 hammer2_chain_modify(trans, chain, 0);
 934                                 modified = 1;
 935                                 goto again;
 936                         }
 937                         if (how == HAMMER2_FREEMAP_DOREALFREE)
 938                                 *bitmap &= ~bmmask11;
 939                         else
 940                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
 941                 } else if ((*bitmap & bmmask11) == bmmask10) {
 942                         /*
 943                          * Mayfree/Realfree request and bitmap is currently
 944                          * marked as being possibly freeable.
 945                          */
 946                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
 947                                 if (!modified) {
 948                                         hammer2_chain_modify(trans, chain, 0);
 949                                         modified = 1;
 950                                         goto again;
 951                                 }
 952                                 *bitmap &= ~bmmask11;
 953                         }
 954                 } else {
 955                         /*
 956                          * 01 - Not implemented, currently illegal state
 957                          * 00 - Not allocated at all, illegal free.
 958                          */
 959                         panic("hammer2_freemap_adjust: "
 960                               "Illegal state %08x(%08x)",
 961                               *bitmap, *bitmap & bmmask11);
 962                 }
 963                 --count;
 964                 bmmask01 <<= 2;
 965                 bmmask10 <<= 2;
 966                 bmmask11 <<= 2;
 967         }
 968         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
 969                 bmap->avail += 1 << radix;
 970                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
 971                 if (bmap->avail == HAMMER2_SEGSIZE &&
 972                     bmap->bitmap[0] == 0 &&
 973                     bmap->bitmap[1] == 0 &&
 974                     bmap->bitmap[2] == 0 &&
 975                     bmap->bitmap[3] == 0 &&
 976                     bmap->bitmap[4] == 0 &&
 977                     bmap->bitmap[5] == 0 &&
 978                     bmap->bitmap[6] == 0 &&
 979                     bmap->bitmap[7] == 0) {
 980                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
 981                         kprintf("Freeseg %016jx\n", (intmax_t)key);
 982                         bmap->class = 0;
 983                 }
 984         }
 985
 986         /*
 987          * chain->bref.check.freemap.bigmask (XXX)
 988          *
 989          * Setting bigmask is a hint to the allocation code that there might
 990          * be something allocatable.  We also set this in recovery... it
 991          * doesn't hurt and we might want to use the hint for other validation
 992          * operations later on.
 993          */
 994         if (modified)
 995                 chain->bref.check.freemap.bigmask |= 1 << radix;
 996
 997         hammer2_chain_unlock(chain);
 998 done:
 999         hammer2_chain_unlock(parent);
1000 }
1001
1002 /*
1003  * Validate the freemap, in three stages.
1004  *
1005  * stage-1      ALLOCATED     -> POSSIBLY FREE
1006  *              POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1007  *
1008  *      This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1009  *      The POSSIBLY FREE state does not mean that a block is actually free
1010  *      and may be transitioned back to ALLOCATED in stage-2.
1011  *
1012  *      This is typically done during normal filesystem operations when
1013  *      something is deleted or a block is replaced.
1014  *
1015  *      This is done by bulkfree in-bulk after a memory-bounded meta-data
1016  *      scan to try to determine what might be freeable.
1017  *
1018  *      This can be done unconditionally through a freemap scan when the
1019  *      intention is to brute-force recover the proper state of the freemap.
1020  *
1021  * stage-2      POSSIBLY FREE -> ALLOCATED      (scan metadata topology)
1022  *
1023  *      This is done by bulkfree during a meta-data scan to ensure that
1024  *      all blocks still actually allocated by the filesystem are marked
1025  *      as such.
1026  *
1027  *      NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1028  *            the bulkfree stage-2 and stage-3 is running.  The live filesystem
1029  *            will use the alternative POSSIBLY FREE type (2) to prevent
1030  *            stage-3 from improperly transitioning unvetted possibly-free
1031  *            blocks to FREE.
1032  *
1033  * stage-3      POSSIBLY FREE (type 1) -> FREE  (scan freemap)
1034  *
1035  *      This is done by bulkfree to finalize POSSIBLY FREE states.
1036  *
1037  */