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