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 xfs_alloc_min_freelist(mp, pag)); 201 if (((minlen && longest >= minlen) || 202 (!minlen && pag->pagf_freeblks >= minfree)) && 203 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || 204 (flags & XFS_PICK_LOWSPACE))) { 205 206 /* Break out, retaining the reference on the AG. */ 207 free = pag->pagf_freeblks; 208 xfs_perag_put(pag); 209 *agp = ag; 210 break; 211 } 212 213 /* Drop the reference on this AG, it's not usable. */ 214 xfs_filestream_put_ag(mp, ag); 215 next_ag: 216 xfs_perag_put(pag); 217 /* Move to the next AG, wrapping to AG 0 if necessary. */ 218 if (++ag >= mp->m_sb.sb_agcount) 219 ag = 0; 220 221 /* If a full pass of the AGs hasn't been done yet, continue. */ 222 if (ag != startag) 223 continue; 224 225 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ 226 if (trylock != 0) { 227 trylock = 0; 228 continue; 229 } 230 231 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ 232 if (!(flags & XFS_PICK_LOWSPACE)) { 233 flags |= XFS_PICK_LOWSPACE; 234 continue; 235 } 236 237 /* 238 * Take the AG with the most free space, regardless of whether 239 * it's already in use by another filestream. 240 */ 241 if (max_ag != NULLAGNUMBER) { 242 xfs_filestream_get_ag(mp, max_ag); 243 free = maxfree; 244 *agp = max_ag; 245 break; 246 } 247 248 /* take AG 0 if none matched */ 249 trace_xfs_filestream_pick(ip, *agp, free, nscan); 250 *agp = 0; 251 return 0; 252 } 253 254 trace_xfs_filestream_pick(ip, *agp, free, nscan); 255 256 if (*agp == NULLAGNUMBER) 257 return 0; 258 259 err = -ENOMEM; 260 item = kmem_alloc(sizeof(*item), KM_MAYFAIL); 261 if (!item) 262 goto out_put_ag; 263 264 item->ag = *agp; 265 item->ip = ip; 266 267 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); 268 if (err) { 269 if (err == -EEXIST) 270 err = 0; 271 goto out_free_item; 272 } 273 274 return 0; 275 276 out_free_item: 277 kmem_free(item); 278 out_put_ag: 279 xfs_filestream_put_ag(mp, *agp); 280 return err; 281 } 282 283 static struct xfs_inode * 284 xfs_filestream_get_parent( 285 struct xfs_inode *ip) 286 { 287 struct inode *inode = VFS_I(ip), *dir = NULL; 288 struct dentry *dentry, *parent; 289 290 dentry = d_find_alias(inode); 291 if (!dentry) 292 goto out; 293 294 parent = dget_parent(dentry); 295 if (!parent) 296 goto out_dput; 297 298 dir = igrab(d_inode(parent)); 299 dput(parent); 300 301 out_dput: 302 dput(dentry); 303 out: 304 return dir ? XFS_I(dir) : NULL; 305 } 306 307 /* 308 * Find the right allocation group for a file, either by finding an 309 * existing file stream or creating a new one. 310 * 311 * Returns NULLAGNUMBER in case of an error. 312 */ 313 xfs_agnumber_t 314 xfs_filestream_lookup_ag( 315 struct xfs_inode *ip) 316 { 317 struct xfs_mount *mp = ip->i_mount; 318 struct xfs_inode *pip = NULL; 319 xfs_agnumber_t startag, ag = NULLAGNUMBER; 320 struct xfs_mru_cache_elem *mru; 321 322 ASSERT(S_ISREG(ip->i_d.di_mode)); 323 324 pip = xfs_filestream_get_parent(ip); 325 if (!pip) 326 return NULLAGNUMBER; 327 328 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); 329 if (mru) { 330 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag; 331 xfs_mru_cache_done(mp->m_filestream); 332 333 trace_xfs_filestream_lookup(ip, ag); 334 goto out; 335 } 336 337 /* 338 * Set the starting AG using the rotor for inode32, otherwise 339 * use the directory inode's AG. 340 */ 341 if (mp->m_flags & XFS_MOUNT_32BITINODES) { 342 xfs_agnumber_t rotorstep = xfs_rotorstep; 343 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; 344 mp->m_agfrotor = (mp->m_agfrotor + 1) % 345 (mp->m_sb.sb_agcount * rotorstep); 346 } else 347 startag = XFS_INO_TO_AGNO(mp, pip->i_ino); 348 349 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0)) 350 ag = NULLAGNUMBER; 351 out: 352 IRELE(pip); 353 return ag; 354 } 355 356 /* 357 * Pick a new allocation group for the current file and its file stream. 358 * 359 * This is called when the allocator can't find a suitable extent in the 360 * current AG, and we have to move the stream into a new AG with more space. 361 */ 362 int 363 xfs_filestream_new_ag( 364 struct xfs_bmalloca *ap, 365 xfs_agnumber_t *agp) 366 { 367 struct xfs_inode *ip = ap->ip, *pip; 368 struct xfs_mount *mp = ip->i_mount; 369 xfs_extlen_t minlen = ap->length; 370 xfs_agnumber_t startag = 0; 371 int flags, err = 0; 372 struct xfs_mru_cache_elem *mru; 373 374 *agp = NULLAGNUMBER; 375 376 pip = xfs_filestream_get_parent(ip); 377 if (!pip) 378 goto exit; 379 380 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); 381 if (mru) { 382 struct xfs_fstrm_item *item = 383 container_of(mru, struct xfs_fstrm_item, mru); 384 startag = (item->ag + 1) % mp->m_sb.sb_agcount; 385 } 386 387 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) | 388 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0); 389 390 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); 391 392 /* 393 * Only free the item here so we skip over the old AG earlier. 394 */ 395 if (mru) 396 xfs_fstrm_free_func(mru); 397 398 IRELE(pip); 399 exit: 400 if (*agp == NULLAGNUMBER) 401 *agp = 0; 402 return err; 403 } 404 405 void 406 xfs_filestream_deassociate( 407 struct xfs_inode *ip) 408 { 409 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); 410 } 411 412 int 413 xfs_filestream_mount( 414 xfs_mount_t *mp) 415 { 416 /* 417 * The filestream timer tunable is currently fixed within the range of 418 * one second to four minutes, with five seconds being the default. The 419 * group count is somewhat arbitrary, but it'd be nice to adhere to the 420 * timer tunable to within about 10 percent. This requires at least 10 421 * groups. 422 */ 423 return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10, 424 10, xfs_fstrm_free_func); 425 } 426 427 void 428 xfs_filestream_unmount( 429 xfs_mount_t *mp) 430 { 431 xfs_mru_cache_destroy(mp->m_filestream); 432 } 433