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