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_filestream.h" 33 #include "xfs_trace.h" 34 35 struct xfs_fstrm_item { 36 struct xfs_mru_cache_elem mru; 37 struct xfs_inode *ip; 38 xfs_agnumber_t ag; /* AG in use for this directory */ 39 }; 40 41 enum xfs_fstrm_alloc { 42 XFS_PICK_USERDATA = 1, 43 XFS_PICK_LOWSPACE = 2, 44 }; 45 46 /* 47 * Allocation group filestream associations are tracked with per-ag atomic 48 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a 49 * particular AG already has active filestreams associated with it. The mount 50 * point's m_peraglock is used to protect these counters from per-ag array 51 * re-allocation during a growfs operation. When xfs_growfs_data_private() is 52 * about to reallocate the array, it calls xfs_filestream_flush() with the 53 * m_peraglock held in write mode. 54 * 55 * Since xfs_mru_cache_flush() guarantees that all the free functions for all 56 * the cache elements have finished executing before it returns, it's safe for 57 * the free functions to use the atomic counters without m_peraglock protection. 58 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about 59 * whether it was called with the m_peraglock held in read mode, write mode or 60 * not held at all. The race condition this addresses is the following: 61 * 62 * - The work queue scheduler fires and pulls a filestream directory cache 63 * element off the LRU end of the cache for deletion, then gets pre-empted. 64 * - A growfs operation grabs the m_peraglock in write mode, flushes all the 65 * remaining items from the cache and reallocates the mount point's per-ag 66 * array, resetting all the counters to zero. 67 * - The work queue thread resumes and calls the free function for the element 68 * it started cleaning up earlier. In the process it decrements the 69 * filestreams counter for an AG that now has no references. 70 * 71 * With a shrinkfs feature, the above scenario could panic the system. 72 * 73 * All other uses of the following macros should be protected by either the 74 * m_peraglock held in read mode, or the cache's internal locking exposed by the 75 * interval between a call to xfs_mru_cache_lookup() and a call to 76 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode 77 * when new elements are added to the cache. 78 * 79 * Combined, these locking rules ensure that no associations will ever exist in 80 * the cache that reference per-ag array elements that have since been 81 * reallocated. 82 */ 83 int 84 xfs_filestream_peek_ag( 85 xfs_mount_t *mp, 86 xfs_agnumber_t agno) 87 { 88 struct xfs_perag *pag; 89 int ret; 90 91 pag = xfs_perag_get(mp, agno); 92 ret = atomic_read(&pag->pagf_fstrms); 93 xfs_perag_put(pag); 94 return ret; 95 } 96 97 static int 98 xfs_filestream_get_ag( 99 xfs_mount_t *mp, 100 xfs_agnumber_t agno) 101 { 102 struct xfs_perag *pag; 103 int ret; 104 105 pag = xfs_perag_get(mp, agno); 106 ret = atomic_inc_return(&pag->pagf_fstrms); 107 xfs_perag_put(pag); 108 return ret; 109 } 110 111 static void 112 xfs_filestream_put_ag( 113 xfs_mount_t *mp, 114 xfs_agnumber_t agno) 115 { 116 struct xfs_perag *pag; 117 118 pag = xfs_perag_get(mp, agno); 119 atomic_dec(&pag->pagf_fstrms); 120 xfs_perag_put(pag); 121 } 122 123 static void 124 xfs_fstrm_free_func( 125 struct xfs_mru_cache_elem *mru) 126 { 127 struct xfs_fstrm_item *item = 128 container_of(mru, struct xfs_fstrm_item, mru); 129 130 xfs_filestream_put_ag(item->ip->i_mount, item->ag); 131 132 trace_xfs_filestream_free(item->ip, item->ag); 133 134 kmem_free(item); 135 } 136 137 /* 138 * Scan the AGs starting at startag looking for an AG that isn't in use and has 139 * at least minlen blocks free. 140 */ 141 static int 142 xfs_filestream_pick_ag( 143 struct xfs_inode *ip, 144 xfs_agnumber_t startag, 145 xfs_agnumber_t *agp, 146 int flags, 147 xfs_extlen_t minlen) 148 { 149 struct xfs_mount *mp = ip->i_mount; 150 struct xfs_fstrm_item *item; 151 struct xfs_perag *pag; 152 xfs_extlen_t longest, free = 0, minfree, maxfree = 0; 153 xfs_agnumber_t ag, max_ag = NULLAGNUMBER; 154 int err, trylock, nscan; 155 156 ASSERT(S_ISDIR(ip->i_d.di_mode)); 157 158 /* 2% of an AG's blocks must be free for it to be chosen. */ 159 minfree = mp->m_sb.sb_agblocks / 50; 160 161 ag = startag; 162 *agp = NULLAGNUMBER; 163 164 /* For the first pass, don't sleep trying to init the per-AG. */ 165 trylock = XFS_ALLOC_FLAG_TRYLOCK; 166 167 for (nscan = 0; 1; nscan++) { 168 trace_xfs_filestream_scan(ip, ag); 169 170 pag = xfs_perag_get(mp, ag); 171 172 if (!pag->pagf_init) { 173 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); 174 if (err && !trylock) { 175 xfs_perag_put(pag); 176 return err; 177 } 178 } 179 180 /* Might fail sometimes during the 1st pass with trylock set. */ 181 if (!pag->pagf_init) 182 goto next_ag; 183 184 /* Keep track of the AG with the most free blocks. */ 185 if (pag->pagf_freeblks > maxfree) { 186 maxfree = pag->pagf_freeblks; 187 max_ag = ag; 188 } 189 190 /* 191 * The AG reference count does two things: it enforces mutual 192 * exclusion when examining the suitability of an AG in this 193 * loop, and it guards against two filestreams being established 194 * in the same AG as each other. 195 */ 196 if (xfs_filestream_get_ag(mp, ag) > 1) { 197 xfs_filestream_put_ag(mp, ag); 198 goto next_ag; 199 } 200 201 longest = xfs_alloc_longest_free_extent(mp, pag); 202 if (((minlen && longest >= minlen) || 203 (!minlen && pag->pagf_freeblks >= minfree)) && 204 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || 205 (flags & XFS_PICK_LOWSPACE))) { 206 207 /* Break out, retaining the reference on the AG. */ 208 free = pag->pagf_freeblks; 209 xfs_perag_put(pag); 210 *agp = ag; 211 break; 212 } 213 214 /* Drop the reference on this AG, it's not usable. */ 215 xfs_filestream_put_ag(mp, ag); 216 next_ag: 217 xfs_perag_put(pag); 218 /* Move to the next AG, wrapping to AG 0 if necessary. */ 219 if (++ag >= mp->m_sb.sb_agcount) 220 ag = 0; 221 222 /* If a full pass of the AGs hasn't been done yet, continue. */ 223 if (ag != startag) 224 continue; 225 226 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ 227 if (trylock != 0) { 228 trylock = 0; 229 continue; 230 } 231 232 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ 233 if (!(flags & XFS_PICK_LOWSPACE)) { 234 flags |= XFS_PICK_LOWSPACE; 235 continue; 236 } 237 238 /* 239 * Take the AG with the most free space, regardless of whether 240 * it's already in use by another filestream. 241 */ 242 if (max_ag != NULLAGNUMBER) { 243 xfs_filestream_get_ag(mp, max_ag); 244 free = maxfree; 245 *agp = max_ag; 246 break; 247 } 248 249 /* take AG 0 if none matched */ 250 trace_xfs_filestream_pick(ip, *agp, free, nscan); 251 *agp = 0; 252 return 0; 253 } 254 255 trace_xfs_filestream_pick(ip, *agp, free, nscan); 256 257 if (*agp == NULLAGNUMBER) 258 return 0; 259 260 err = -ENOMEM; 261 item = kmem_alloc(sizeof(*item), KM_MAYFAIL); 262 if (!item) 263 goto out_put_ag; 264 265 item->ag = *agp; 266 item->ip = ip; 267 268 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); 269 if (err) { 270 if (err == -EEXIST) 271 err = 0; 272 goto out_free_item; 273 } 274 275 return 0; 276 277 out_free_item: 278 kmem_free(item); 279 out_put_ag: 280 xfs_filestream_put_ag(mp, *agp); 281 return err; 282 } 283 284 static struct xfs_inode * 285 xfs_filestream_get_parent( 286 struct xfs_inode *ip) 287 { 288 struct inode *inode = VFS_I(ip), *dir = NULL; 289 struct dentry *dentry, *parent; 290 291 dentry = d_find_alias(inode); 292 if (!dentry) 293 goto out; 294 295 parent = dget_parent(dentry); 296 if (!parent) 297 goto out_dput; 298 299 dir = igrab(parent->d_inode); 300 dput(parent); 301 302 out_dput: 303 dput(dentry); 304 out: 305 return dir ? XFS_I(dir) : NULL; 306 } 307 308 /* 309 * Find the right allocation group for a file, either by finding an 310 * existing file stream or creating a new one. 311 * 312 * Returns NULLAGNUMBER in case of an error. 313 */ 314 xfs_agnumber_t 315 xfs_filestream_lookup_ag( 316 struct xfs_inode *ip) 317 { 318 struct xfs_mount *mp = ip->i_mount; 319 struct xfs_inode *pip = NULL; 320 xfs_agnumber_t startag, ag = NULLAGNUMBER; 321 struct xfs_mru_cache_elem *mru; 322 323 ASSERT(S_ISREG(ip->i_d.di_mode)); 324 325 pip = xfs_filestream_get_parent(ip); 326 if (!pip) 327 goto out; 328 329 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); 330 if (mru) { 331 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag; 332 xfs_mru_cache_done(mp->m_filestream); 333 334 trace_xfs_filestream_lookup(ip, ag); 335 goto out; 336 } 337 338 /* 339 * Set the starting AG using the rotor for inode32, otherwise 340 * use the directory inode's AG. 341 */ 342 if (mp->m_flags & XFS_MOUNT_32BITINODES) { 343 xfs_agnumber_t rotorstep = xfs_rotorstep; 344 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; 345 mp->m_agfrotor = (mp->m_agfrotor + 1) % 346 (mp->m_sb.sb_agcount * rotorstep); 347 } else 348 startag = XFS_INO_TO_AGNO(mp, pip->i_ino); 349 350 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0)) 351 ag = NULLAGNUMBER; 352 out: 353 IRELE(pip); 354 return ag; 355 } 356 357 /* 358 * Pick a new allocation group for the current file and its file stream. 359 * 360 * This is called when the allocator can't find a suitable extent in the 361 * current AG, and we have to move the stream into a new AG with more space. 362 */ 363 int 364 xfs_filestream_new_ag( 365 struct xfs_bmalloca *ap, 366 xfs_agnumber_t *agp) 367 { 368 struct xfs_inode *ip = ap->ip, *pip; 369 struct xfs_mount *mp = ip->i_mount; 370 xfs_extlen_t minlen = ap->length; 371 xfs_agnumber_t startag = 0; 372 int flags, err = 0; 373 struct xfs_mru_cache_elem *mru; 374 375 *agp = NULLAGNUMBER; 376 377 pip = xfs_filestream_get_parent(ip); 378 if (!pip) 379 goto exit; 380 381 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); 382 if (mru) { 383 struct xfs_fstrm_item *item = 384 container_of(mru, struct xfs_fstrm_item, mru); 385 startag = (item->ag + 1) % mp->m_sb.sb_agcount; 386 } 387 388 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | 389 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0); 390 391 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); 392 393 /* 394 * Only free the item here so we skip over the old AG earlier. 395 */ 396 if (mru) 397 xfs_fstrm_free_func(mru); 398 399 IRELE(pip); 400 exit: 401 if (*agp == NULLAGNUMBER) 402 *agp = 0; 403 return err; 404 } 405 406 void 407 xfs_filestream_deassociate( 408 struct xfs_inode *ip) 409 { 410 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); 411 } 412 413 int 414 xfs_filestream_mount( 415 xfs_mount_t *mp) 416 { 417 /* 418 * The filestream timer tunable is currently fixed within the range of 419 * one second to four minutes, with five seconds being the default. The 420 * group count is somewhat arbitrary, but it'd be nice to adhere to the 421 * timer tunable to within about 10 percent. This requires at least 10 422 * groups. 423 */ 424 return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10, 425 10, xfs_fstrm_free_func); 426 } 427 428 void 429 xfs_filestream_unmount( 430 xfs_mount_t *mp) 431 { 432 xfs_mru_cache_destroy(mp->m_filestream); 433 } 434