fs/xfs/xfs_filestream.c

   1 /*
   2  * Copyright (c) 2006-2007 Silicon Graphics, Inc.
   3  * Copyright (c) 2014 Christoph Hellwig.
   4  * All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it would be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write the Free Software Foundation,
  17  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  18  */
  19 #include "xfs.h"
  20 #include "xfs_format.h"
  21 #include "xfs_log_format.h"
  22 #include "xfs_trans_resv.h"
  23 #include "xfs_ag.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_mount.h"
  26 #include "xfs_inum.h"
  27 #include "xfs_inode.h"
  28 #include "xfs_bmap.h"
  29 #include "xfs_bmap_util.h"
  30 #include "xfs_alloc.h"
  31 #include "xfs_mru_cache.h"
  32 #include "xfs_dinode.h"
  33 #include "xfs_filestream.h"
  34 #include "xfs_trace.h"
  35
  36 #define TRACE_AG_SCAN(mp, ag, ag2)
  37 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
  38 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
  39 #define TRACE_FREE(mp, ip, pip, ag, cnt)
  40 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
  41
  42 static kmem_zone_t *item_zone;
  43
  44 struct xfs_fstrm_item {
  45         struct xfs_mru_cache_elem       mru;
  46         struct xfs_inode                *ip;
  47         xfs_agnumber_t                  ag; /* AG in use for this directory */
  48 };
  49
  50 enum xfs_fstrm_alloc {
  51         XFS_PICK_USERDATA = 1,
  52         XFS_PICK_LOWSPACE = 2,
  53 };
  54
  55 /*
  56  * Allocation group filestream associations are tracked with per-ag atomic
  57  * counters.  These counters allow xfs_filestream_pick_ag() to tell whether a
  58  * particular AG already has active filestreams associated with it. The mount
  59  * point's m_peraglock is used to protect these counters from per-ag array
  60  * re-allocation during a growfs operation.  When xfs_growfs_data_private() is
  61  * about to reallocate the array, it calls xfs_filestream_flush() with the
  62  * m_peraglock held in write mode.
  63  *
  64  * Since xfs_mru_cache_flush() guarantees that all the free functions for all
  65  * the cache elements have finished executing before it returns, it's safe for
  66  * the free functions to use the atomic counters without m_peraglock protection.
  67  * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
  68  * whether it was called with the m_peraglock held in read mode, write mode or
  69  * not held at all.  The race condition this addresses is the following:
  70  *
  71  *  - The work queue scheduler fires and pulls a filestream directory cache
  72  *    element off the LRU end of the cache for deletion, then gets pre-empted.
  73  *  - A growfs operation grabs the m_peraglock in write mode, flushes all the
  74  *    remaining items from the cache and reallocates the mount point's per-ag
  75  *    array, resetting all the counters to zero.
  76  *  - The work queue thread resumes and calls the free function for the element
  77  *    it started cleaning up earlier.  In the process it decrements the
  78  *    filestreams counter for an AG that now has no references.
  79  *
  80  * With a shrinkfs feature, the above scenario could panic the system.
  81  *
  82  * All other uses of the following macros should be protected by either the
  83  * m_peraglock held in read mode, or the cache's internal locking exposed by the
  84  * interval between a call to xfs_mru_cache_lookup() and a call to
  85  * xfs_mru_cache_done().  In addition, the m_peraglock must be held in read mode
  86  * when new elements are added to the cache.
  87  *
  88  * Combined, these locking rules ensure that no associations will ever exist in
  89  * the cache that reference per-ag array elements that have since been
  90  * reallocated.
  91  */
  92 static int
  93 xfs_filestream_peek_ag(
  94         xfs_mount_t     *mp,
  95         xfs_agnumber_t  agno)
  96 {
  97         struct xfs_perag *pag;
  98         int             ret;
  99
 100         pag = xfs_perag_get(mp, agno);
 101         ret = atomic_read(&pag->pagf_fstrms);
 102         xfs_perag_put(pag);
 103         return ret;
 104 }
 105
 106 static int
 107 xfs_filestream_get_ag(
 108         xfs_mount_t     *mp,
 109         xfs_agnumber_t  agno)
 110 {
 111         struct xfs_perag *pag;
 112         int             ret;
 113
 114         pag = xfs_perag_get(mp, agno);
 115         ret = atomic_inc_return(&pag->pagf_fstrms);
 116         xfs_perag_put(pag);
 117         return ret;
 118 }
 119
 120 static void
 121 xfs_filestream_put_ag(
 122         xfs_mount_t     *mp,
 123         xfs_agnumber_t  agno)
 124 {
 125         struct xfs_perag *pag;
 126
 127         pag = xfs_perag_get(mp, agno);
 128         atomic_dec(&pag->pagf_fstrms);
 129         xfs_perag_put(pag);
 130 }
 131
 132 static void
 133 xfs_fstrm_free_func(
 134         struct xfs_mru_cache_elem *mru)
 135 {
 136         struct xfs_fstrm_item   *item =
 137                 container_of(mru, struct xfs_fstrm_item, mru);
 138
 139         xfs_filestream_put_ag(item->ip->i_mount, item->ag);
 140
 141         TRACE_FREE(mp, ip, NULL, item->ag,
 142                    xfs_filestream_peek_ag(mp, item->ag));
 143
 144         kmem_zone_free(item_zone, item);
 145 }
 146
 147 /*
 148  * Scan the AGs starting at startag looking for an AG that isn't in use and has
 149  * at least minlen blocks free.
 150  */
 151 static int
 152 xfs_filestream_pick_ag(
 153         struct xfs_inode        *ip,
 154         xfs_agnumber_t          startag,
 155         xfs_agnumber_t          *agp,
 156         int                     flags,
 157         xfs_extlen_t            minlen)
 158 {
 159         struct xfs_mount        *mp = ip->i_mount;
 160         struct xfs_fstrm_item   *item;
 161         struct xfs_perag        *pag;
 162         xfs_extlen_t            longest, free, minfree, maxfree = 0;
 163         xfs_agnumber_t          ag, max_ag = NULLAGNUMBER;
 164         int                     streams, max_streams;
 165         int                     err, trylock, nscan;
 166
 167         ASSERT(S_ISDIR(ip->i_d.di_mode));
 168
 169         /* 2% of an AG's blocks must be free for it to be chosen. */
 170         minfree = mp->m_sb.sb_agblocks / 50;
 171
 172         ag = startag;
 173         *agp = NULLAGNUMBER;
 174
 175         /* For the first pass, don't sleep trying to init the per-AG. */
 176         trylock = XFS_ALLOC_FLAG_TRYLOCK;
 177
 178         for (nscan = 0; 1; nscan++) {
 179                 pag = xfs_perag_get(mp, ag);
 180                 TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms));
 181
 182                 if (!pag->pagf_init) {
 183                         err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
 184                         if (err && !trylock) {
 185                                 xfs_perag_put(pag);
 186                                 return err;
 187                         }
 188                 }
 189
 190                 /* Might fail sometimes during the 1st pass with trylock set. */
 191                 if (!pag->pagf_init)
 192                         goto next_ag;
 193
 194                 /* Keep track of the AG with the most free blocks. */
 195                 if (pag->pagf_freeblks > maxfree) {
 196                         maxfree = pag->pagf_freeblks;
 197                         max_streams = atomic_read(&pag->pagf_fstrms);
 198                         max_ag = ag;
 199                 }
 200
 201                 /*
 202                  * The AG reference count does two things: it enforces mutual
 203                  * exclusion when examining the suitability of an AG in this
 204                  * loop, and it guards against two filestreams being established
 205                  * in the same AG as each other.
 206                  */
 207                 if (xfs_filestream_get_ag(mp, ag) > 1) {
 208                         xfs_filestream_put_ag(mp, ag);
 209                         goto next_ag;
 210                 }
 211
 212                 longest = xfs_alloc_longest_free_extent(mp, pag);
 213                 if (((minlen && longest >= minlen) ||
 214                      (!minlen && pag->pagf_freeblks >= minfree)) &&
 215                     (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
 216                      (flags & XFS_PICK_LOWSPACE))) {
 217
 218                         /* Break out, retaining the reference on the AG. */
 219                         free = pag->pagf_freeblks;
 220                         streams = atomic_read(&pag->pagf_fstrms);
 221                         xfs_perag_put(pag);
 222                         *agp = ag;
 223                         break;
 224                 }
 225
 226                 /* Drop the reference on this AG, it's not usable. */
 227                 xfs_filestream_put_ag(mp, ag);
 228 next_ag:
 229                 xfs_perag_put(pag);
 230                 /* Move to the next AG, wrapping to AG 0 if necessary. */
 231                 if (++ag >= mp->m_sb.sb_agcount)
 232                         ag = 0;
 233
 234                 /* If a full pass of the AGs hasn't been done yet, continue. */
 235                 if (ag != startag)
 236                         continue;
 237
 238                 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
 239                 if (trylock != 0) {
 240                         trylock = 0;
 241                         continue;
 242                 }
 243
 244                 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
 245                 if (!(flags & XFS_PICK_LOWSPACE)) {
 246                         flags |= XFS_PICK_LOWSPACE;
 247                         continue;
 248                 }
 249
 250                 /*
 251                  * Take the AG with the most free space, regardless of whether
 252                  * it's already in use by another filestream.
 253                  */
 254                 if (max_ag != NULLAGNUMBER) {
 255                         xfs_filestream_get_ag(mp, max_ag);
 256                         TRACE_AG_PICK1(mp, max_ag, maxfree);
 257                         streams = max_streams;
 258                         free = maxfree;
 259                         *agp = max_ag;
 260                         break;
 261                 }
 262
 263                 /* take AG 0 if none matched */
 264                 TRACE_AG_PICK1(mp, max_ag, maxfree);
 265                 *agp = 0;
 266                 return 0;
 267         }
 268
 269         TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
 270
 271         if (*agp == NULLAGNUMBER)
 272                 return 0;
 273
 274         err = ENOMEM;
 275         item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
 276         if (!item)
 277                 goto out_put_ag;
 278
 279         item->ag = *agp;
 280         item->ip = ip;
 281
 282         err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
 283         if (err) {
 284                 if (err == EEXIST)
 285                         err = 0;
 286                 goto out_free_item;
 287         }
 288
 289         return 0;
 290
 291 out_free_item:
 292         kmem_zone_free(item_zone, item);
 293 out_put_ag:
 294         xfs_filestream_put_ag(mp, *agp);
 295         return err;
 296 }
 297
 298 static struct xfs_inode *
 299 xfs_filestream_get_parent(
 300         struct xfs_inode        *ip)
 301 {
 302         struct inode            *inode = VFS_I(ip), *dir = NULL;
 303         struct dentry           *dentry, *parent;
 304
 305         dentry = d_find_alias(inode);
 306         if (!dentry)
 307                 goto out;
 308
 309         parent = dget_parent(dentry);
 310         if (!parent)
 311                 goto out_dput;
 312
 313         dir = igrab(parent->d_inode);
 314         dput(parent);
 315
 316 out_dput:
 317         dput(dentry);
 318 out:
 319         return dir ? XFS_I(dir) : NULL;
 320 }
 321
 322 /*
 323  * Return the AG of the filestream the file or directory belongs to, or
 324  * NULLAGNUMBER otherwise.
 325  */
 326 xfs_agnumber_t
 327 xfs_filestream_lookup_ag(
 328         struct xfs_inode        *ip)
 329 {
 330         struct xfs_mount        *mp = ip->i_mount;
 331         struct xfs_fstrm_item   *item;
 332         struct xfs_inode        *pip = NULL;
 333         xfs_agnumber_t          ag = NULLAGNUMBER;
 334         int                     ref = 0;
 335         struct xfs_mru_cache_elem *mru;
 336
 337         ASSERT(S_ISREG(ip->i_d.di_mode));
 338
 339         pip = xfs_filestream_get_parent(ip);
 340         if (!pip)
 341                 goto out;
 342
 343         mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
 344         if (!mru)
 345                 goto out;
 346
 347         item = container_of(mru, struct xfs_fstrm_item, mru);
 348
 349         ag = item->ag;
 350         xfs_mru_cache_done(mp->m_filestream);
 351
 352         ref = xfs_filestream_peek_ag(ip->i_mount, ag);
 353 out:
 354         TRACE_LOOKUP(mp, ip, pip, ag, ref);
 355         IRELE(pip);
 356         return ag;
 357 }
 358
 359 /*
 360  * Make sure a directory has a filestream associated with it.
 361  *
 362  * This is called when creating regular files in an directory that has
 363  * filestreams enabled, so that a stream is ready by the time we need it
 364  * in the allocator for the files inside the directory.
 365  */
 366 int
 367 xfs_filestream_associate(
 368         struct xfs_inode        *pip)
 369 {
 370         struct xfs_mount        *mp = pip->i_mount;
 371         struct xfs_mru_cache_elem *mru;
 372         xfs_agnumber_t          startag, ag;
 373
 374         ASSERT(S_ISDIR(pip->i_d.di_mode));
 375
 376         /*
 377          * If the directory already has a file stream associated we're done.
 378          */
 379         mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
 380         if (mru) {
 381                 xfs_mru_cache_done(mp->m_filestream);
 382                 return 0;
 383         }
 384
 385         /*
 386          * Set the starting AG using the rotor for inode32, otherwise
 387          * use the directory inode's AG.
 388          */
 389         if (mp->m_flags & XFS_MOUNT_32BITINODES) {
 390                 xfs_agnumber_t   rotorstep = xfs_rotorstep;
 391                 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
 392                 mp->m_agfrotor = (mp->m_agfrotor + 1) %
 393                                  (mp->m_sb.sb_agcount * rotorstep);
 394         } else
 395                 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
 396
 397         return xfs_filestream_pick_ag(pip, startag, &ag, 0, 0);
 398 }
 399
 400 /*
 401  * Pick a new allocation group for the current file and its file stream.
 402  *
 403  * This is called when the allocator can't find a suitable extent in the
 404  * current AG, and we have to move the stream into a new AG with more space.
 405  */
 406 int
 407 xfs_filestream_new_ag(
 408         struct xfs_bmalloca     *ap,
 409         xfs_agnumber_t          *agp)
 410 {
 411         struct xfs_inode        *ip = ap->ip, *pip;
 412         struct xfs_mount        *mp = ip->i_mount;
 413         xfs_extlen_t            minlen = ap->length;
 414         xfs_agnumber_t          startag = 0;
 415         int                     flags, err = 0;
 416         struct xfs_mru_cache_elem *mru;
 417
 418         *agp = NULLAGNUMBER;
 419
 420         pip = xfs_filestream_get_parent(ip);
 421         if (!pip)
 422                 goto exit;
 423
 424         mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
 425         if (mru) {
 426                 struct xfs_fstrm_item *item =
 427                         container_of(mru, struct xfs_fstrm_item, mru);
 428                 startag = (item->ag + 1) % mp->m_sb.sb_agcount;
 429         }
 430
 431         flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
 432                 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
 433
 434         err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
 435
 436         /*
 437          * Only free the item here so we skip over the old AG earlier.
 438          */
 439         if (mru)
 440                 xfs_fstrm_free_func(mru);
 441
 442         IRELE(pip);
 443 exit:
 444         if (*agp == NULLAGNUMBER)
 445                 *agp = 0;
 446         return err;
 447 }
 448
 449 void
 450 xfs_filestream_deassociate(
 451         struct xfs_inode        *ip)
 452 {
 453         xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
 454 }
 455
 456 int
 457 xfs_filestream_mount(
 458         xfs_mount_t     *mp)
 459 {
 460         /*
 461          * The filestream timer tunable is currently fixed within the range of
 462          * one second to four minutes, with five seconds being the default.  The
 463          * group count is somewhat arbitrary, but it'd be nice to adhere to the
 464          * timer tunable to within about 10 percent.  This requires at least 10
 465          * groups.
 466          */
 467         return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10,
 468                                     10, xfs_fstrm_free_func);
 469 }
 470
 471 void
 472 xfs_filestream_unmount(
 473         xfs_mount_t     *mp)
 474 {
 475         xfs_mru_cache_destroy(mp->m_filestream);
 476 }
 477
 478
 479 /* needs to return a positive errno for the init path */
 480 int
 481 xfs_filestream_init(void)
 482 {
 483         item_zone = kmem_zone_init(sizeof(struct xfs_fstrm_item), "fstrm_item");
 484         if (!item_zone)
 485                 return -ENOMEM;
 486         return 0;
 487 }
 488
 489 void
 490 xfs_filestream_uninit(void)
 491 {
 492         kmem_zone_destroy(item_zone);
 493 }