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_dev_t *hmp,
  62                         hammer2_key_t key, hammer2_chain_t *chain);
  63 static int hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_dev_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         if ((bref->data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) {
 116                 index = 0;
 117         } else {
 118                 off = bref->data_off & ~HAMMER2_OFF_MASK_RADIX &
 119                       (((hammer2_off_t)1 <<
 120                         HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 121                 off = off / HAMMER2_PBUFSIZE;
 122                 KKASSERT(off >= HAMMER2_ZONE_FREEMAP_00 &&
 123                          off < HAMMER2_ZONE_FREEMAP_END);
 124                 index = (int)(off - HAMMER2_ZONE_FREEMAP_00) / 4;
 125                 KKASSERT(index >= 0 && index < HAMMER2_NFREEMAPS);
 126                 if (++index == HAMMER2_NFREEMAPS)
 127                         index = 0;
 128         }
 129
 130         /*
 131          * Calculate the block offset of the reserved block.  This will
 132          * point into the 4MB reserved area at the base of the appropriate
 133          * 2GB zone, once added to the FREEMAP_x selection above.
 134          */
 135         switch(bref->keybits) {
 136         /* case HAMMER2_FREEMAP_LEVEL5_RADIX: not applicable */
 137         case HAMMER2_FREEMAP_LEVEL4_RADIX:      /* 2EB */
 138                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 139                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 140                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL4_RADIX) +
 141                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 142                        HAMMER2_ZONEFM_LEVEL4) * HAMMER2_PBUFSIZE;
 143                 break;
 144         case HAMMER2_FREEMAP_LEVEL3_RADIX:      /* 2PB */
 145                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 146                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 147                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL3_RADIX) +
 148                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 149                        HAMMER2_ZONEFM_LEVEL3) * HAMMER2_PBUFSIZE;
 150                 break;
 151         case HAMMER2_FREEMAP_LEVEL2_RADIX:      /* 2TB */
 152                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE);
 153                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 154                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL2_RADIX) +
 155                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 156                        HAMMER2_ZONEFM_LEVEL2) * HAMMER2_PBUFSIZE;
 157                 break;
 158         case HAMMER2_FREEMAP_LEVEL1_RADIX:      /* 2GB */
 159                 KKASSERT(bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 160                 KKASSERT(bytes == HAMMER2_FREEMAP_LEVELN_PSIZE);
 161                 off = H2FMBASE(bref->key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 162                       (index * 4 + HAMMER2_ZONE_FREEMAP_00 +
 163                        HAMMER2_ZONEFM_LEVEL1) * HAMMER2_PBUFSIZE;
 164                 break;
 165         default:
 166                 panic("freemap: bad radix(2) %p %d\n", bref, bref->keybits);
 167                 /* NOT REACHED */
 168                 off = (hammer2_off_t)-1;
 169                 break;
 170         }
 171         bref->data_off = off | radix;
 172 #if FREEMAP_DEBUG
 173         kprintf("FREEMAP BLOCK TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 174                 bref->type, bref->key, bref->keybits, bref->data_off);
 175 #endif
 176         return (0);
 177 }
 178
 179 /*
 180  * Normal freemap allocator
 181  *
 182  * Use available hints to allocate space using the freemap.  Create missing
 183  * freemap infrastructure on-the-fly as needed (including marking initial
 184  * allocations using the iterator as allocated, instantiating new 2GB zones,
 185  * and dealing with the end-of-media edge case).
 186  *
 187  * ip and bpref are only used as a heuristic to determine locality of
 188  * reference.  bref->key may also be used heuristically.
 189  */
 190 int
 191 hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 192                       size_t bytes)
 193 {
 194         hammer2_dev_t *hmp = chain->hmp;
 195         hammer2_blockref_t *bref = &chain->bref;
 196         hammer2_chain_t *parent;
 197         int radix;
 198         int error;
 199         unsigned int hindex;
 200         hammer2_fiterate_t iter;
 201
 202         /*
 203          * Validate the allocation size.  It must be a power of 2.
 204          *
 205          * For now require that the caller be aware of the minimum
 206          * allocation (1K).
 207          */
 208         radix = hammer2_getradix(bytes);
 209         KKASSERT((size_t)1 << radix == bytes);
 210
 211         if (bref->type == HAMMER2_BREF_TYPE_FREEMAP_NODE ||
 212             bref->type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) {
 213                 /*
 214                  * Freemap blocks themselves are assigned from the reserve
 215                  * area, not allocated from the freemap.
 216                  */
 217                 error = hammer2_freemap_reserve(trans, chain, radix);
 218                 return error;
 219         }
 220
 221         KKASSERT(bytes >= HAMMER2_ALLOC_MIN && bytes <= HAMMER2_ALLOC_MAX);
 222
 223         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 224                 ++trans->sync_xid;
 225
 226         /*
 227          * Calculate the starting point for our allocation search.
 228          *
 229          * Each freemap leaf is dedicated to a specific freemap_radix.
 230          * The freemap_radix can be more fine-grained than the device buffer
 231          * radix which results in inodes being grouped together in their
 232          * own segment, terminal-data (16K or less) and initial indirect
 233          * block being grouped together, and then full-indirect and full-data
 234          * blocks (64K) being grouped together.
 235          *
 236          * The single most important aspect of this is the inode grouping
 237          * because that is what allows 'find' and 'ls' and other filesystem
 238          * topology operations to run fast.
 239          */
 240 #if 0
 241         if (bref->data_off & ~HAMMER2_OFF_MASK_RADIX)
 242                 bpref = bref->data_off & ~HAMMER2_OFF_MASK_RADIX;
 243         else if (trans->tmp_bpref)
 244                 bpref = trans->tmp_bpref;
 245         else if (trans->tmp_ip)
 246                 bpref = trans->tmp_ip->chain->bref.data_off;
 247         else
 248 #endif
 249         /*
 250          * Heuristic tracking index.  We would like one for each distinct
 251          * bref type if possible.  heur_freemap[] has room for two classes
 252          * for each type.  At a minimum we have to break-up our heuristic
 253          * by device block sizes.
 254          */
 255         hindex = hammer2_devblkradix(radix) - HAMMER2_MINIORADIX;
 256         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR_NRADIX);
 257         hindex += bref->type * HAMMER2_FREEMAP_HEUR_NRADIX;
 258         hindex &= HAMMER2_FREEMAP_HEUR_TYPES * HAMMER2_FREEMAP_HEUR_NRADIX - 1;
 259         KKASSERT(hindex < HAMMER2_FREEMAP_HEUR);
 260
 261         iter.bpref = hmp->heur_freemap[hindex];
 262
 263         /*
 264          * Make sure bpref is in-bounds.  It's ok if bpref covers a zone's
 265          * reserved area, the try code will iterate past it.
 266          */
 267         if (iter.bpref > hmp->voldata.volu_size)
 268                 iter.bpref = hmp->voldata.volu_size - 1;
 269
 270         /*
 271          * Iterate the freemap looking for free space before and after.
 272          */
 273         parent = &hmp->fchain;
 274         hammer2_chain_ref(parent);
 275         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 276         error = EAGAIN;
 277         iter.bnext = iter.bpref;
 278         iter.loops = 0;
 279
 280         while (error == EAGAIN) {
 281                 error = hammer2_freemap_try_alloc(trans, &parent, bref,
 282                                                   radix, &iter);
 283         }
 284         hmp->heur_freemap[hindex] = iter.bnext;
 285         hammer2_chain_unlock(parent);
 286         hammer2_chain_drop(parent);
 287
 288         if (trans->flags & (HAMMER2_TRANS_ISFLUSH | HAMMER2_TRANS_PREFLUSH))
 289                 --trans->sync_xid;
 290
 291         return (error);
 292 }
 293
 294 static int
 295 hammer2_freemap_try_alloc(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 296                           hammer2_blockref_t *bref, int radix,
 297                           hammer2_fiterate_t *iter)
 298 {
 299         hammer2_dev_t *hmp = (*parentp)->hmp;
 300         hammer2_off_t l0size;
 301         hammer2_off_t l1size;
 302         hammer2_off_t l1mask;
 303         hammer2_key_t key_dummy;
 304         hammer2_chain_t *chain;
 305         hammer2_off_t key;
 306         size_t bytes;
 307         uint16_t class;
 308         int error = 0;
 309         int cache_index = -1;
 310
 311         /*
 312          * Calculate the number of bytes being allocated, the number
 313          * of contiguous bits of bitmap being allocated, and the bitmap
 314          * mask.
 315          *
 316          * WARNING! cpu hardware may mask bits == 64 -> 0 and blow up the
 317          *          mask calculation.
 318          */
 319         bytes = (size_t)1 << radix;
 320         class = (bref->type << 8) | hammer2_devblkradix(radix);
 321
 322         /*
 323          * Lookup the level1 freemap chain, creating and initializing one
 324          * if necessary.  Intermediate levels will be created automatically
 325          * when necessary by hammer2_chain_create().
 326          */
 327         key = H2FMBASE(iter->bnext, HAMMER2_FREEMAP_LEVEL1_RADIX);
 328         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 329         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 330         l1mask = l1size - 1;
 331
 332         chain = hammer2_chain_lookup(parentp, &key_dummy, key, key + l1mask,
 333                                      &cache_index,
 334                                      HAMMER2_LOOKUP_ALWAYS |
 335                                      HAMMER2_LOOKUP_MATCHIND);
 336
 337         if (chain == NULL) {
 338                 /*
 339                  * Create the missing leaf, be sure to initialize
 340                  * the auxillary freemap tracking information in
 341                  * the bref.check.freemap structure.
 342                  */
 343 #if 0
 344                 kprintf("freemap create L1 @ %016jx bpref %016jx\n",
 345                         key, iter->bpref);
 346 #endif
 347                 error = hammer2_chain_create(trans, parentp, &chain, hmp->spmp,
 348                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 349                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 350                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 351                                      0);
 352                 KKASSERT(error == 0);
 353                 if (error == 0) {
 354                         hammer2_chain_modify(trans, chain, 0);
 355                         bzero(&chain->data->bmdata[0],
 356                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 357                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 358                         chain->bref.check.freemap.avail = l1size;
 359                         /* bref.methods should already be inherited */
 360
 361                         hammer2_freemap_init(trans, hmp, key, chain);
 362                 }
 363         } else if (chain->error) {
 364                 /*
 365                  * Error during lookup.
 366                  */
 367                 kprintf("hammer2_freemap_try_alloc: %016jx: error %s\n",
 368                         (intmax_t)bref->data_off,
 369                         hammer2_error_str(chain->error));
 370                 error = EIO;
 371         } else if ((chain->bref.check.freemap.bigmask & (1 << radix)) == 0) {
 372                 /*
 373                  * Already flagged as not having enough space
 374                  */
 375                 error = ENOSPC;
 376         } else {
 377                 /*
 378                  * Modify existing chain to setup for adjustment.
 379                  */
 380                 hammer2_chain_modify(trans, chain, 0);
 381         }
 382
 383         /*
 384          * Scan 2MB entries.
 385          */
 386         if (error == 0) {
 387                 hammer2_bmap_data_t *bmap;
 388                 hammer2_key_t base_key;
 389                 int count;
 390                 int start;
 391                 int n;
 392
 393                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF);
 394                 start = (int)((iter->bnext - key) >>
 395                               HAMMER2_FREEMAP_LEVEL0_RADIX);
 396                 KKASSERT(start >= 0 && start < HAMMER2_FREEMAP_COUNT);
 397                 hammer2_chain_modify(trans, chain, 0);
 398
 399                 error = ENOSPC;
 400                 for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 401                         int availchk;
 402
 403                         if (start + count >= HAMMER2_FREEMAP_COUNT &&
 404                             start - count < 0) {
 405                                 break;
 406                         }
 407
 408                         /*
 409                          * Calculate bmap pointer
 410                          *
 411                          * NOTE: bmap pointer is invalid if n >= FREEMAP_COUNT.
 412                          */
 413                         n = start + count;
 414                         bmap = &chain->data->bmdata[n];
 415
 416                         if (n >= HAMMER2_FREEMAP_COUNT) {
 417                                 availchk = 0;
 418                         } else if (bmap->avail) {
 419                                 availchk = 1;
 420                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 421                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 422                                 availchk = 1;
 423                         } else {
 424                                 availchk = 0;
 425                         }
 426
 427                         if (availchk &&
 428                             (bmap->class == 0 || bmap->class == class)) {
 429                                 base_key = key + n * l0size;
 430                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 431                                                            class, n, radix,
 432                                                            &base_key);
 433                                 if (error != ENOSPC) {
 434                                         key = base_key;
 435                                         break;
 436                                 }
 437                         }
 438
 439                         /*
 440                          * Must recalculate after potentially having called
 441                          * hammer2_bmap_alloc() above in case chain was
 442                          * reallocated.
 443                          *
 444                          * NOTE: bmap pointer is invalid if n < 0.
 445                          */
 446                         n = start - count;
 447                         bmap = &chain->data->bmdata[n];
 448                         if (n < 0) {
 449                                 availchk = 0;
 450                         } else if (bmap->avail) {
 451                                 availchk = 1;
 452                         } else if (radix < HAMMER2_FREEMAP_BLOCK_RADIX &&
 453                                   (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK)) {
 454                                 availchk = 1;
 455                         } else {
 456                                 availchk = 0;
 457                         }
 458
 459                         if (availchk &&
 460                             (bmap->class == 0 || bmap->class == class)) {
 461                                 base_key = key + n * l0size;
 462                                 error = hammer2_bmap_alloc(trans, hmp, bmap,
 463                                                            class, n, radix,
 464                                                            &base_key);
 465                                 if (error != ENOSPC) {
 466                                         key = base_key;
 467                                         break;
 468                                 }
 469                         }
 470                 }
 471                 if (error == ENOSPC)
 472                         chain->bref.check.freemap.bigmask &= ~(1 << radix);
 473                 /* XXX also scan down from original count */
 474         }
 475
 476         if (error == 0) {
 477                 /*
 478                  * Assert validity.  Must be beyond the static allocator used
 479                  * by newfs_hammer2 (and thus also beyond the aux area),
 480                  * not go past the volume size, and must not be in the
 481                  * reserved segment area for a zone.
 482                  */
 483                 KKASSERT(key >= hmp->voldata.allocator_beg &&
 484                          key + bytes <= hmp->voldata.volu_size);
 485                 KKASSERT((key & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 486                 bref->data_off = key | radix;
 487
 488 #if 0
 489                 kprintf("alloc cp=%p %016jx %016jx using %016jx\n",
 490                         chain,
 491                         bref->key, bref->data_off, chain->bref.data_off);
 492 #endif
 493         } else if (error == ENOSPC) {
 494                 /*
 495                  * Return EAGAIN with next iteration in iter->bnext, or
 496                  * return ENOSPC if the allocation map has been exhausted.
 497                  */
 498                 error = hammer2_freemap_iterate(trans, parentp, &chain, iter);
 499         }
 500
 501         /*
 502          * Cleanup
 503          */
 504         if (chain) {
 505                 hammer2_chain_unlock(chain);
 506                 hammer2_chain_drop(chain);
 507         }
 508         return (error);
 509 }
 510
 511 /*
 512  * Allocate (1<<radix) bytes from the bmap whos base data offset is (*basep).
 513  *
 514  * If the linear iterator is mid-block we use it directly (the bitmap should
 515  * already be marked allocated), otherwise we search for a block in the bitmap
 516  * that fits the allocation request.
 517  *
 518  * A partial bitmap allocation sets the minimum bitmap granularity (16KB)
 519  * to fully allocated and adjusts the linear allocator to allow the
 520  * remaining space to be allocated.
 521  */
 522 static
 523 int
 524 hammer2_bmap_alloc(hammer2_trans_t *trans, hammer2_dev_t *hmp,
 525                    hammer2_bmap_data_t *bmap,
 526                    uint16_t class, int n, int radix, hammer2_key_t *basep)
 527 {
 528         hammer2_io_t *dio;
 529         size_t size;
 530         size_t bgsize;
 531         int bmradix;
 532         uint32_t bmmask;
 533         int offset;
 534         int error;
 535         int i;
 536         int j;
 537
 538         /*
 539          * Take into account 2-bits per block when calculating bmradix.
 540          */
 541         size = (size_t)1 << radix;
 542
 543         if (radix <= HAMMER2_FREEMAP_BLOCK_RADIX) {
 544                 bmradix = 2;
 545                 /* (16K) 2 bits per allocation block */
 546         } else {
 547                 bmradix = 2 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 548                 /* (32K-256K) 4, 8, 16, 32 bits per allocation block */
 549         }
 550
 551         /*
 552          * Use the linear iterator to pack small allocations, otherwise
 553          * fall-back to finding a free 16KB chunk.  The linear iterator
 554          * is only valid when *NOT* on a freemap chunking boundary (16KB).
 555          * If it is the bitmap must be scanned.  It can become invalid
 556          * once we pack to the boundary.  We adjust it after a bitmap
 557          * allocation only for sub-16KB allocations (so the perfectly good
 558          * previous value can still be used for fragments when 16KB+
 559          * allocations are made).
 560          *
 561          * Beware of hardware artifacts when bmradix == 32 (intermediate
 562          * result can wind up being '1' instead of '0' if hardware masks
 563          * bit-count & 31).
 564          *
 565          * NOTE: j needs to be even in the j= calculation.  As an artifact
 566          *       of the /2 division, our bitmask has to clear bit 0.
 567          *
 568          * NOTE: TODO this can leave little unallocatable fragments lying
 569          *       around.
 570          */
 571         if (((uint32_t)bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) + size <=
 572             HAMMER2_FREEMAP_BLOCK_SIZE &&
 573             (bmap->linear & HAMMER2_FREEMAP_BLOCK_MASK) &&
 574             bmap->linear < HAMMER2_SEGSIZE) {
 575                 KKASSERT(bmap->linear >= 0 &&
 576                          bmap->linear + size <= HAMMER2_SEGSIZE &&
 577                          (bmap->linear & (HAMMER2_ALLOC_MIN - 1)) == 0);
 578                 offset = bmap->linear;
 579                 i = offset / (HAMMER2_SEGSIZE / 8);
 580                 j = (offset / (HAMMER2_FREEMAP_BLOCK_SIZE / 2)) & 30;
 581                 bmmask = (bmradix == 32) ?
 582                          0xFFFFFFFFU : (1 << bmradix) - 1;
 583                 bmmask <<= j;
 584                 bmap->linear = offset + size;
 585         } else {
 586                 for (i = 0; i < 8; ++i) {
 587                         bmmask = (bmradix == 32) ?
 588                                  0xFFFFFFFFU : (1 << bmradix) - 1;
 589                         for (j = 0; j < 32; j += bmradix) {
 590                                 if ((bmap->bitmap[i] & bmmask) == 0)
 591                                         goto success;
 592                                 bmmask <<= bmradix;
 593                         }
 594                 }
 595                 /*fragments might remain*/
 596                 /*KKASSERT(bmap->avail == 0);*/
 597                 return (ENOSPC);
 598 success:
 599                 offset = i * (HAMMER2_SEGSIZE / 8) +
 600                          (j * (HAMMER2_FREEMAP_BLOCK_SIZE / 2));
 601                 if (size & HAMMER2_FREEMAP_BLOCK_MASK)
 602                         bmap->linear = offset + size;
 603         }
 604
 605         KKASSERT(i >= 0 && i < 8);      /* 8 x 16 -> 128 x 16K -> 2MB */
 606
 607         /*
 608          * Optimize the buffer cache to avoid unnecessary read-before-write
 609          * operations.
 610          *
 611          * The device block size could be larger than the allocation size
 612          * so the actual bitmap test is somewhat more involved.  We have
 613          * to use a compatible buffer size for this operation.
 614          */
 615         if ((bmap->bitmap[i] & bmmask) == 0 &&
 616             hammer2_devblksize(size) != size) {
 617                 size_t psize = hammer2_devblksize(size);
 618                 hammer2_off_t pmask = (hammer2_off_t)psize - 1;
 619                 int pbmradix = 2 << (hammer2_devblkradix(radix) -
 620                                      HAMMER2_FREEMAP_BLOCK_RADIX);
 621                 uint32_t pbmmask;
 622                 int pradix = hammer2_getradix(psize);
 623
 624                 pbmmask = (pbmradix == 32) ? 0xFFFFFFFFU : (1 << pbmradix) - 1;
 625                 while ((pbmmask & bmmask) == 0)
 626                         pbmmask <<= pbmradix;
 627
 628 #if 0
 629                 kprintf("%016jx mask %08x %08x %08x (%zd/%zd)\n",
 630                         *basep + offset, bmap->bitmap[i],
 631                         pbmmask, bmmask, size, psize);
 632 #endif
 633
 634                 if ((bmap->bitmap[i] & pbmmask) == 0) {
 635                         error = hammer2_io_newq(hmp,
 636                                                 (*basep + (offset & ~pmask)) |
 637                                                  pradix,
 638                                                 psize, &dio);
 639                         hammer2_io_bqrelse(&dio);
 640                 }
 641         }
 642
 643 #if 0
 644         /*
 645          * When initializing a new inode segment also attempt to initialize
 646          * an adjacent segment.  Be careful not to index beyond the array
 647          * bounds.
 648          *
 649          * We do this to try to localize inode accesses to improve
 650          * directory scan rates.  XXX doesn't improve scan rates.
 651          */
 652         if (size == HAMMER2_INODE_BYTES) {
 653                 if (n & 1) {
 654                         if (bmap[-1].radix == 0 && bmap[-1].avail)
 655                                 bmap[-1].radix = radix;
 656                 } else {
 657                         if (bmap[1].radix == 0 && bmap[1].avail)
 658                                 bmap[1].radix = radix;
 659                 }
 660         }
 661 #endif
 662         /*
 663          * Calculate the bitmap-granular change in bgsize for the volume
 664          * header.  We cannot use the fine-grained change here because
 665          * the bulkfree code can't undo it.  If the bitmap element is already
 666          * marked allocated it has already been accounted for.
 667          */
 668         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 669                 if (bmap->bitmap[i] & bmmask)
 670                         bgsize = 0;
 671                 else
 672                         bgsize = HAMMER2_FREEMAP_BLOCK_SIZE;
 673         } else {
 674                 bgsize = size;
 675         }
 676
 677         /*
 678          * Adjust the bitmap, set the class (it might have been 0),
 679          * and available bytes, update the allocation offset (*basep)
 680          * from the L0 base to the actual offset.
 681          *
 682          * avail must reflect the bitmap-granular availability.  The allocator
 683          * tests will also check the linear iterator.
 684          */
 685         bmap->bitmap[i] |= bmmask;
 686         bmap->class = class;
 687         bmap->avail -= bgsize;
 688         *basep += offset;
 689
 690         /*
 691          * Adjust the volume header's allocator_free parameter.  This
 692          * parameter has to be fixed up by bulkfree which has no way to
 693          * figure out sub-16K chunking, so it must be adjusted by the
 694          * bitmap-granular size.
 695          */
 696         if (bgsize) {
 697                 hammer2_voldata_lock(hmp);
 698                 hammer2_voldata_modify(hmp);
 699                 hmp->voldata.allocator_free -= bgsize;
 700                 hammer2_voldata_unlock(hmp);
 701         }
 702
 703         return(0);
 704 }
 705
 706 static
 707 void
 708 hammer2_freemap_init(hammer2_trans_t *trans, hammer2_dev_t *hmp,
 709                      hammer2_key_t key, hammer2_chain_t *chain)
 710 {
 711         hammer2_off_t l1size;
 712         hammer2_off_t lokey;
 713         hammer2_off_t hikey;
 714         hammer2_bmap_data_t *bmap;
 715         int count;
 716
 717         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 718
 719         /*
 720          * Calculate the portion of the 2GB map that should be initialized
 721          * as free.  Portions below or after will be initialized as allocated.
 722          * SEGMASK-align the areas so we don't have to worry about sub-scans
 723          * or endianess when using memset.
 724          *
 725          * (1) Ensure that all statically allocated space from newfs_hammer2
 726          *     is marked allocated.
 727          *
 728          * (2) Ensure that the reserved area is marked allocated (typically
 729          *     the first 4MB of the 2GB area being represented).
 730          *
 731          * (3) Ensure that any trailing space at the end-of-volume is marked
 732          *     allocated.
 733          *
 734          * WARNING! It is possible for lokey to be larger than hikey if the
 735          *          entire 2GB segment is within the static allocation.
 736          */
 737         lokey = (hmp->voldata.allocator_beg + HAMMER2_SEGMASK64) &
 738                 ~HAMMER2_SEGMASK64;
 739
 740         if (lokey < H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 741                   HAMMER2_ZONE_SEG64) {
 742                 lokey = H2FMBASE(key, HAMMER2_FREEMAP_LEVEL1_RADIX) +
 743                         HAMMER2_ZONE_SEG64;
 744         }
 745
 746         hikey = key + H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 747         if (hikey > hmp->voldata.volu_size) {
 748                 hikey = hmp->voldata.volu_size & ~HAMMER2_SEGMASK64;
 749         }
 750
 751         chain->bref.check.freemap.avail =
 752                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 753         bmap = &chain->data->bmdata[0];
 754
 755         for (count = 0; count < HAMMER2_FREEMAP_COUNT; ++count) {
 756                 if (key < lokey || key >= hikey) {
 757                         memset(bmap->bitmap, -1,
 758                                sizeof(bmap->bitmap));
 759                         bmap->avail = 0;
 760                         bmap->linear = HAMMER2_SEGSIZE;
 761                         chain->bref.check.freemap.avail -=
 762                                 H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 763                 } else {
 764                         bmap->avail = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 765                 }
 766                 key += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 767                 ++bmap;
 768         }
 769 }
 770
 771 /*
 772  * The current Level 1 freemap has been exhausted, iterate to the next
 773  * one, return ENOSPC if no freemaps remain.
 774  *
 775  * XXX this should rotate back to the beginning to handle freed-up space
 776  * XXX or use intermediate entries to locate free space. TODO
 777  */
 778 static int
 779 hammer2_freemap_iterate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 780                         hammer2_chain_t **chainp, hammer2_fiterate_t *iter)
 781 {
 782         hammer2_dev_t *hmp = (*parentp)->hmp;
 783
 784         iter->bnext &= ~(H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX) - 1);
 785         iter->bnext += H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 786         if (iter->bnext >= hmp->voldata.volu_size) {
 787                 iter->bnext = 0;
 788                 if (++iter->loops == 2)
 789                         return (ENOSPC);
 790         }
 791         return(EAGAIN);
 792 }
 793
 794 /*
 795  * Adjust the bit-pattern for data in the freemap bitmap according to
 796  * (how).  This code is called from on-mount recovery to fixup (mark
 797  * as allocated) blocks whos freemap upates might not have been committed
 798  * in the last crash and is used by the bulk freemap scan to stage frees.
 799  *
 800  * XXX currently disabled when how == 0 (the normal real-time case).  At
 801  * the moment we depend on the bulk freescan to actually free blocks.  It
 802  * will still call this routine with a non-zero how to stage possible frees
 803  * and to do the actual free.
 804  */
 805 void
 806 hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_dev_t *hmp,
 807                        hammer2_blockref_t *bref, int how)
 808 {
 809         hammer2_off_t data_off = bref->data_off;
 810         hammer2_chain_t *chain;
 811         hammer2_chain_t *parent;
 812         hammer2_bmap_data_t *bmap;
 813         hammer2_key_t key;
 814         hammer2_key_t key_dummy;
 815         hammer2_off_t l0size;
 816         hammer2_off_t l1size;
 817         hammer2_off_t l1mask;
 818         uint32_t *bitmap;
 819         const uint32_t bmmask00 = 0;
 820         uint32_t bmmask01;
 821         uint32_t bmmask10;
 822         uint32_t bmmask11;
 823         size_t bytes;
 824         uint16_t class;
 825         int radix;
 826         int start;
 827         int count;
 828         int modified = 0;
 829         int cache_index = -1;
 830         int error;
 831
 832         KKASSERT(how == HAMMER2_FREEMAP_DORECOVER);
 833
 834         radix = (int)data_off & HAMMER2_OFF_MASK_RADIX;
 835         data_off &= ~HAMMER2_OFF_MASK_RADIX;
 836         KKASSERT(radix <= HAMMER2_RADIX_MAX);
 837
 838         bytes = (size_t)1 << radix;
 839         class = (bref->type << 8) | hammer2_devblkradix(radix);
 840
 841         /*
 842          * We can't adjust thre freemap for data allocations made by
 843          * newfs_hammer2.
 844          */
 845         if (data_off < hmp->voldata.allocator_beg)
 846                 return;
 847
 848         KKASSERT((data_off & HAMMER2_ZONE_MASK64) >= HAMMER2_ZONE_SEG);
 849
 850         /*
 851          * Lookup the level1 freemap chain.  The chain must exist.
 852          */
 853         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL1_RADIX);
 854         l0size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL0_RADIX);
 855         l1size = H2FMSHIFT(HAMMER2_FREEMAP_LEVEL1_RADIX);
 856         l1mask = l1size - 1;
 857
 858         parent = &hmp->fchain;
 859         hammer2_chain_ref(parent);
 860         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
 861
 862         chain = hammer2_chain_lookup(&parent, &key_dummy, key, key + l1mask,
 863                                      &cache_index,
 864                                      HAMMER2_LOOKUP_ALWAYS |
 865                                      HAMMER2_LOOKUP_MATCHIND);
 866
 867         /*
 868          * Stop early if we are trying to free something but no leaf exists.
 869          */
 870         if (chain == NULL && how != HAMMER2_FREEMAP_DORECOVER) {
 871                 kprintf("hammer2_freemap_adjust: %016jx: no chain\n",
 872                         (intmax_t)bref->data_off);
 873                 goto done;
 874         }
 875         if (chain->error) {
 876                 kprintf("hammer2_freemap_adjust: %016jx: error %s\n",
 877                         (intmax_t)bref->data_off,
 878                         hammer2_error_str(chain->error));
 879                 hammer2_chain_unlock(chain);
 880                 hammer2_chain_drop(chain);
 881                 chain = NULL;
 882                 goto done;
 883         }
 884
 885         /*
 886          * Create any missing leaf(s) if we are doing a recovery (marking
 887          * the block(s) as being allocated instead of being freed).  Be sure
 888          * to initialize the auxillary freemap tracking info in the
 889          * bref.check.freemap structure.
 890          */
 891         if (chain == NULL && how == HAMMER2_FREEMAP_DORECOVER) {
 892                 error = hammer2_chain_create(trans, &parent, &chain, hmp->spmp,
 893                                      key, HAMMER2_FREEMAP_LEVEL1_RADIX,
 894                                      HAMMER2_BREF_TYPE_FREEMAP_LEAF,
 895                                      HAMMER2_FREEMAP_LEVELN_PSIZE,
 896                                      0);
 897
 898                 if (hammer2_debug & 0x0040) {
 899                         kprintf("fixup create chain %p %016jx:%d\n",
 900                                 chain, chain->bref.key, chain->bref.keybits);
 901                 }
 902
 903                 if (error == 0) {
 904                         hammer2_chain_modify(trans, chain, 0);
 905                         bzero(&chain->data->bmdata[0],
 906                               HAMMER2_FREEMAP_LEVELN_PSIZE);
 907                         chain->bref.check.freemap.bigmask = (uint32_t)-1;
 908                         chain->bref.check.freemap.avail = l1size;
 909                         /* bref.methods should already be inherited */
 910
 911                         hammer2_freemap_init(trans, hmp, key, chain);
 912                 }
 913                 /* XXX handle error */
 914         }
 915
 916 #if FREEMAP_DEBUG
 917         kprintf("FREEMAP ADJUST TYPE %d %016jx/%d DATA_OFF=%016jx\n",
 918                 chain->bref.type, chain->bref.key,
 919                 chain->bref.keybits, chain->bref.data_off);
 920 #endif
 921
 922         /*
 923          * Calculate the bitmask (runs in 2-bit pairs).
 924          */
 925         start = ((int)(data_off >> HAMMER2_FREEMAP_BLOCK_RADIX) & 15) * 2;
 926         bmmask01 = 1 << start;
 927         bmmask10 = 2 << start;
 928         bmmask11 = 3 << start;
 929
 930         /*
 931          * Fixup the bitmap.  Partial blocks cannot be fully freed unless
 932          * a bulk scan is able to roll them up.
 933          */
 934         if (radix < HAMMER2_FREEMAP_BLOCK_RADIX) {
 935                 count = 1;
 936                 if (how == HAMMER2_FREEMAP_DOREALFREE)
 937                         how = HAMMER2_FREEMAP_DOMAYFREE;
 938         } else {
 939                 count = 1 << (radix - HAMMER2_FREEMAP_BLOCK_RADIX);
 940         }
 941
 942         /*
 943          * [re]load the bmap and bitmap pointers.  Each bmap entry covers
 944          * a 2MB swath.  The bmap itself (LEVEL1) covers 2GB.
 945          *
 946          * Be sure to reset the linear iterator to ensure that the adjustment
 947          * is not ignored.
 948          */
 949 again:
 950         bmap = &chain->data->bmdata[(int)(data_off >> HAMMER2_SEGRADIX) &
 951                                     (HAMMER2_FREEMAP_COUNT - 1)];
 952         bitmap = &bmap->bitmap[(int)(data_off >> (HAMMER2_SEGRADIX - 3)) & 7];
 953
 954         if (modified)
 955                 bmap->linear = 0;
 956
 957         while (count) {
 958                 KKASSERT(bmmask11);
 959                 if (how == HAMMER2_FREEMAP_DORECOVER) {
 960                         /*
 961                          * Recovery request, mark as allocated.
 962                          */
 963                         if ((*bitmap & bmmask11) != bmmask11) {
 964                                 if (modified == 0) {
 965                                         hammer2_chain_modify(trans, chain, 0);
 966                                         modified = 1;
 967                                         goto again;
 968                                 }
 969                                 if ((*bitmap & bmmask11) == bmmask00) {
 970                                         bmap->avail -=
 971                                                 HAMMER2_FREEMAP_BLOCK_SIZE;
 972                                 }
 973                                 if (bmap->class == 0)
 974                                         bmap->class = class;
 975                                 *bitmap |= bmmask11;
 976                                 if (hammer2_debug & 0x0040) {
 977                                         kprintf("hammer2_freemap_recover: "
 978                                                 "fixup type=%02x "
 979                                                 "block=%016jx/%zd\n",
 980                                                 bref->type, data_off, bytes);
 981                                 }
 982                         } else {
 983                                 /*
 984                                 kprintf("hammer2_freemap_recover:  good "
 985                                         "type=%02x block=%016jx/%zd\n",
 986                                         bref->type, data_off, bytes);
 987                                 */
 988                         }
 989                 }
 990 #if 0
 991                 /*
 992                  * XXX this stuff doesn't work, avail is miscalculated and
 993                  * code 10 means something else now.
 994                  */
 995                 else if ((*bitmap & bmmask11) == bmmask11) {
 996                         /*
 997                          * Mayfree/Realfree request and bitmap is currently
 998                          * marked as being fully allocated.
 999                          */
1000                         if (!modified) {
1001                                 hammer2_chain_modify(trans, chain, 0);
1002                                 modified = 1;
1003                                 goto again;
1004                         }
1005                         if (how == HAMMER2_FREEMAP_DOREALFREE)
1006                                 *bitmap &= ~bmmask11;
1007                         else
1008                                 *bitmap = (*bitmap & ~bmmask11) | bmmask10;
1009                 } else if ((*bitmap & bmmask11) == bmmask10) {
1010                         /*
1011                          * Mayfree/Realfree request and bitmap is currently
1012                          * marked as being possibly freeable.
1013                          */
1014                         if (how == HAMMER2_FREEMAP_DOREALFREE) {
1015                                 if (!modified) {
1016                                         hammer2_chain_modify(trans, chain, 0);
1017                                         modified = 1;
1018                                         goto again;
1019                                 }
1020                                 *bitmap &= ~bmmask11;
1021                         }
1022                 } else {
1023                         /*
1024                          * 01 - Not implemented, currently illegal state
1025                          * 00 - Not allocated at all, illegal free.
1026                          */
1027                         panic("hammer2_freemap_adjust: "
1028                               "Illegal state %08x(%08x)",
1029                               *bitmap, *bitmap & bmmask11);
1030                 }
1031 #endif
1032                 --count;
1033                 bmmask01 <<= 2;
1034                 bmmask10 <<= 2;
1035                 bmmask11 <<= 2;
1036         }
1037         if (how == HAMMER2_FREEMAP_DOREALFREE && modified) {
1038                 bmap->avail += 1 << radix;
1039                 KKASSERT(bmap->avail <= HAMMER2_SEGSIZE);
1040                 if (bmap->avail == HAMMER2_SEGSIZE &&
1041                     bmap->bitmap[0] == 0 &&
1042                     bmap->bitmap[1] == 0 &&
1043                     bmap->bitmap[2] == 0 &&
1044                     bmap->bitmap[3] == 0 &&
1045                     bmap->bitmap[4] == 0 &&
1046                     bmap->bitmap[5] == 0 &&
1047                     bmap->bitmap[6] == 0 &&
1048                     bmap->bitmap[7] == 0) {
1049                         key = H2FMBASE(data_off, HAMMER2_FREEMAP_LEVEL0_RADIX);
1050                         kprintf("Freeseg %016jx\n", (intmax_t)key);
1051                         bmap->class = 0;
1052                 }
1053         }
1054
1055         /*
1056          * chain->bref.check.freemap.bigmask (XXX)
1057          *
1058          * Setting bigmask is a hint to the allocation code that there might
1059          * be something allocatable.  We also set this in recovery... it
1060          * doesn't hurt and we might want to use the hint for other validation
1061          * operations later on.
1062          */
1063         if (modified)
1064                 chain->bref.check.freemap.bigmask |= 1 << radix;
1065
1066         hammer2_chain_unlock(chain);
1067         hammer2_chain_drop(chain);
1068 done:
1069         hammer2_chain_unlock(parent);
1070         hammer2_chain_drop(parent);
1071 }
1072
1073 /*
1074  * Validate the freemap, in three stages.
1075  *
1076  * stage-1      ALLOCATED     -> POSSIBLY FREE
1077  *              POSSIBLY FREE -> POSSIBLY FREE (type corrected)
1078  *
1079  *      This transitions bitmap entries from ALLOCATED to POSSIBLY FREE.
1080  *      The POSSIBLY FREE state does not mean that a block is actually free
1081  *      and may be transitioned back to ALLOCATED in stage-2.
1082  *
1083  *      This is typically done during normal filesystem operations when
1084  *      something is deleted or a block is replaced.
1085  *
1086  *      This is done by bulkfree in-bulk after a memory-bounded meta-data
1087  *      scan to try to determine what might be freeable.
1088  *
1089  *      This can be done unconditionally through a freemap scan when the
1090  *      intention is to brute-force recover the proper state of the freemap.
1091  *
1092  * stage-2      POSSIBLY FREE -> ALLOCATED      (scan metadata topology)
1093  *
1094  *      This is done by bulkfree during a meta-data scan to ensure that
1095  *      all blocks still actually allocated by the filesystem are marked
1096  *      as such.
1097  *
1098  *      NOTE! Live filesystem transitions to POSSIBLY FREE can occur while
1099  *            the bulkfree stage-2 and stage-3 is running.  The live filesystem
1100  *            will use the alternative POSSIBLY FREE type (2) to prevent
1101  *            stage-3 from improperly transitioning unvetted possibly-free
1102  *            blocks to FREE.
1103  *
1104  * stage-3      POSSIBLY FREE (type 1) -> FREE  (scan freemap)
1105  *
1106  *      This is done by bulkfree to finalize POSSIBLY FREE states.
1107  *
1108  */