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_defer.h" 26 #include "xfs_inode.h" 27 #include "xfs_bmap.h" 28 #include "xfs_bmap_util.h" 29 #include "xfs_alloc.h" 30 #include "xfs_mru_cache.h" 31 #include "xfs_filestream.h" 32 #include "xfs_trace.h" 33 34 struct xfs_fstrm_item { 35 struct xfs_mru_cache_elem mru; 36 struct xfs_inode *ip; 37 xfs_agnumber_t ag; /* AG in use for this directory */ 38 }; 39 40 enum xfs_fstrm_alloc { 41 XFS_PICK_USERDATA = 1, 42 XFS_PICK_LOWSPACE = 2, 43 }; 44 45 /* 46 * Allocation group filestream associations are tracked with per-ag atomic 47 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a 48 * particular AG already has active filestreams associated with it. The mount 49 * point's m_peraglock is used to protect these counters from per-ag array 50 * re-allocation during a growfs operation. When xfs_growfs_data_private() is 51 * about to reallocate the array, it calls xfs_filestream_flush() with the 52 * m_peraglock held in write mode. 53 * 54 * Since xfs_mru_cache_flush() guarantees that all the free functions for all 55 * the cache elements have finished executing before it returns, it's safe for 56 * the free functions to use the atomic counters without m_peraglock protection. 57 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about 58 * whether it was called with the m_peraglock held in read mode, write mode or 59 * not held at all. The race condition this addresses is the following: 60 * 61 * - The work queue scheduler fires and pulls a filestream directory cache 62 * element off the LRU end of the cache for deletion, then gets pre-empted. 63 * - A growfs operation grabs the m_peraglock in write mode, flushes all the 64 * remaining items from the cache and reallocates the mount point's per-ag 65 * array, resetting all the counters to zero. 66 * - The work queue thread resumes and calls the free function for the element 67 * it started cleaning up earlier. In the process it decrements the 68 * filestreams counter for an AG that now has no references. 69 * 70 * With a shrinkfs feature, the above scenario could panic the system. 71 * 72 * All other uses of the following macros should be protected by either the 73 * m_peraglock held in read mode, or the cache's internal locking exposed by the 74 * interval between a call to xfs_mru_cache_lookup() and a call to 75 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode 76 * when new elements are added to the cache. 77 * 78 * Combined, these locking rules ensure that no associations will ever exist in 79 * the cache that reference per-ag array elements that have since been 80 * reallocated. 81 */ 82 int 83 xfs_filestream_peek_ag( 84 xfs_mount_t *mp, 85 xfs_agnumber_t agno) 86 { 87 struct xfs_perag *pag; 88 int ret; 89 90 pag = xfs_perag_get(mp, agno); 91 ret = atomic_read(&pag->pagf_fstrms); 92 xfs_perag_put(pag); 93 return ret; 94 } 95 96 static int 97 xfs_filestream_get_ag( 98 xfs_mount_t *mp, 99 xfs_agnumber_t agno) 100 { 101 struct xfs_perag *pag; 102 int ret; 103 104 pag = xfs_perag_get(mp, agno); 105 ret = atomic_inc_return(&pag->pagf_fstrms); 106 xfs_perag_put(pag); 107 return ret; 108 } 109 110 static void 111 xfs_filestream_put_ag( 112 xfs_mount_t *mp, 113 xfs_agnumber_t agno) 114 { 115 struct xfs_perag *pag; 116 117 pag = xfs_perag_get(mp, agno); 118 atomic_dec(&pag->pagf_fstrms); 119 xfs_perag_put(pag); 120 } 121 122 static void 123 xfs_fstrm_free_func( 124 struct xfs_mru_cache_elem *mru) 125 { 126 struct xfs_fstrm_item *item = 127 container_of(mru, struct xfs_fstrm_item, mru); 128 129 xfs_filestream_put_ag(item->ip->i_mount, item->ag); 130 131 trace_xfs_filestream_free(item->ip, item->ag); 132 133 kmem_free(item); 134 } 135 136 /* 137 * Scan the AGs starting at startag looking for an AG that isn't in use and has 138 * at least minlen blocks free. 139 */ 140 static int 141 xfs_filestream_pick_ag( 142 struct xfs_inode *ip, 143 xfs_agnumber_t startag, 144 xfs_agnumber_t *agp, 145 int flags, 146 xfs_extlen_t minlen) 147 { 148 struct xfs_mount *mp = ip->i_mount; 149 struct xfs_fstrm_item *item; 150 struct xfs_perag *pag; 151 xfs_extlen_t longest, free = 0, minfree, maxfree = 0; 152 xfs_agnumber_t ag, max_ag = NULLAGNUMBER; 153 int err, trylock, nscan; 154 155 ASSERT(S_ISDIR(VFS_I(ip)->i_mode)); 156 157 /* 2% of an AG's blocks must be free for it to be chosen. */ 158 minfree = mp->m_sb.sb_agblocks / 50; 159 160 ag = startag; 161 *agp = NULLAGNUMBER; 162 163 /* For the first pass, don't sleep trying to init the per-AG. */ 164 trylock = XFS_ALLOC_FLAG_TRYLOCK; 165 166 for (nscan = 0; 1; nscan++) { 167 trace_xfs_filestream_scan(ip, ag); 168 169 pag = xfs_perag_get(mp, ag); 170 171 if (!pag->pagf_init) { 172 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock); 173 if (err && !trylock) { 174 xfs_perag_put(pag); 175 return err; 176 } 177 } 178 179 /* Might fail sometimes during the 1st pass with trylock set. */ 180 if (!pag->pagf_init) 181 goto next_ag; 182 183 /* Keep track of the AG with the most free blocks. */ 184 if (pag->pagf_freeblks > maxfree) { 185 maxfree = pag->pagf_freeblks; 186 max_ag = ag; 187 } 188 189 /* 190 * The AG reference count does two things: it enforces mutual 191 * exclusion when examining the suitability of an AG in this 192 * loop, and it guards against two filestreams being established 193 * in the same AG as each other. 194 */ 195 if (xfs_filestream_get_ag(mp, ag) > 1) { 196 xfs_filestream_put_ag(mp, ag); 197 goto next_ag; 198 } 199 200 longest = xfs_alloc_longest_free_extent(mp, pag, 201 xfs_alloc_min_freelist(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(d_inode(parent)); 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(VFS_I(ip)->i_mode)); 324 325 pip = xfs_filestream_get_parent(ip); 326 if (!pip) 327 return NULLAGNUMBER; 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 = 0; 373 int err = 0; 374 struct xfs_mru_cache_elem *mru; 375 376 *agp = NULLAGNUMBER; 377 378 pip = xfs_filestream_get_parent(ip); 379 if (!pip) 380 goto exit; 381 382 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); 383 if (mru) { 384 struct xfs_fstrm_item *item = 385 container_of(mru, struct xfs_fstrm_item, mru); 386 startag = (item->ag + 1) % mp->m_sb.sb_agcount; 387 } 388 389 if (xfs_alloc_is_userdata(ap->datatype)) 390 flags |= XFS_PICK_USERDATA; 391 if (ap->dfops->dop_low) 392 flags |= XFS_PICK_LOWSPACE; 393 394 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); 395 396 /* 397 * Only free the item here so we skip over the old AG earlier. 398 */ 399 if (mru) 400 xfs_fstrm_free_func(mru); 401 402 IRELE(pip); 403 exit: 404 if (*agp == NULLAGNUMBER) 405 *agp = 0; 406 return err; 407 } 408 409 void 410 xfs_filestream_deassociate( 411 struct xfs_inode *ip) 412 { 413 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); 414 } 415 416 int 417 xfs_filestream_mount( 418 xfs_mount_t *mp) 419 { 420 /* 421 * The filestream timer tunable is currently fixed within the range of 422 * one second to four minutes, with five seconds being the default. The 423 * group count is somewhat arbitrary, but it'd be nice to adhere to the 424 * timer tunable to within about 10 percent. This requires at least 10 425 * groups. 426 */ 427 return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10, 428 10, xfs_fstrm_free_func); 429 } 430 431 void 432 xfs_filestream_unmount( 433 xfs_mount_t *mp) 434 { 435 xfs_mru_cache_destroy(mp->m_filestream); 436 } 437