1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2006-2007 Silicon Graphics, Inc. 4 * Copyright (c) 2014 Christoph Hellwig. 5 * All Rights Reserved. 6 */ 7 #include "xfs.h" 8 #include "xfs_shared.h" 9 #include "xfs_format.h" 10 #include "xfs_log_format.h" 11 #include "xfs_trans_resv.h" 12 #include "xfs_mount.h" 13 #include "xfs_inode.h" 14 #include "xfs_bmap.h" 15 #include "xfs_alloc.h" 16 #include "xfs_mru_cache.h" 17 #include "xfs_trace.h" 18 #include "xfs_ag.h" 19 #include "xfs_ag_resv.h" 20 #include "xfs_trans.h" 21 #include "xfs_filestream.h" 22 23 struct xfs_fstrm_item { 24 struct xfs_mru_cache_elem mru; 25 xfs_agnumber_t ag; /* AG in use for this directory */ 26 }; 27 28 enum xfs_fstrm_alloc { 29 XFS_PICK_USERDATA = 1, 30 XFS_PICK_LOWSPACE = 2, 31 }; 32 33 /* 34 * Allocation group filestream associations are tracked with per-ag atomic 35 * counters. These counters allow xfs_filestream_pick_ag() to tell whether a 36 * particular AG already has active filestreams associated with it. 37 */ 38 int 39 xfs_filestream_peek_ag( 40 xfs_mount_t *mp, 41 xfs_agnumber_t agno) 42 { 43 struct xfs_perag *pag; 44 int ret; 45 46 pag = xfs_perag_get(mp, agno); 47 ret = atomic_read(&pag->pagf_fstrms); 48 xfs_perag_put(pag); 49 return ret; 50 } 51 52 static int 53 xfs_filestream_get_ag( 54 xfs_mount_t *mp, 55 xfs_agnumber_t agno) 56 { 57 struct xfs_perag *pag; 58 int ret; 59 60 pag = xfs_perag_get(mp, agno); 61 ret = atomic_inc_return(&pag->pagf_fstrms); 62 xfs_perag_put(pag); 63 return ret; 64 } 65 66 static void 67 xfs_filestream_put_ag( 68 xfs_mount_t *mp, 69 xfs_agnumber_t agno) 70 { 71 struct xfs_perag *pag; 72 73 pag = xfs_perag_get(mp, agno); 74 atomic_dec(&pag->pagf_fstrms); 75 xfs_perag_put(pag); 76 } 77 78 static void 79 xfs_fstrm_free_func( 80 void *data, 81 struct xfs_mru_cache_elem *mru) 82 { 83 struct xfs_mount *mp = data; 84 struct xfs_fstrm_item *item = 85 container_of(mru, struct xfs_fstrm_item, mru); 86 87 xfs_filestream_put_ag(mp, item->ag); 88 trace_xfs_filestream_free(mp, mru->key, item->ag); 89 90 kmem_free(item); 91 } 92 93 /* 94 * Scan the AGs starting at startag looking for an AG that isn't in use and has 95 * at least minlen blocks free. 96 */ 97 static int 98 xfs_filestream_pick_ag( 99 struct xfs_inode *ip, 100 xfs_agnumber_t startag, 101 xfs_agnumber_t *agp, 102 int flags, 103 xfs_extlen_t minlen) 104 { 105 struct xfs_mount *mp = ip->i_mount; 106 struct xfs_fstrm_item *item; 107 struct xfs_perag *pag; 108 xfs_extlen_t longest, free = 0, minfree, maxfree = 0; 109 xfs_agnumber_t ag, max_ag = NULLAGNUMBER; 110 int err, trylock, nscan; 111 112 ASSERT(S_ISDIR(VFS_I(ip)->i_mode)); 113 114 /* 2% of an AG's blocks must be free for it to be chosen. */ 115 minfree = mp->m_sb.sb_agblocks / 50; 116 117 ag = startag; 118 *agp = NULLAGNUMBER; 119 120 /* For the first pass, don't sleep trying to init the per-AG. */ 121 trylock = XFS_ALLOC_FLAG_TRYLOCK; 122 123 for (nscan = 0; 1; nscan++) { 124 trace_xfs_filestream_scan(mp, ip->i_ino, ag); 125 126 pag = xfs_perag_get(mp, ag); 127 128 if (!pag->pagf_init) { 129 err = xfs_alloc_read_agf(pag, NULL, trylock, NULL); 130 if (err) { 131 if (err != -EAGAIN) { 132 xfs_perag_put(pag); 133 return err; 134 } 135 /* Couldn't lock the AGF, skip this AG. */ 136 goto next_ag; 137 } 138 } 139 140 /* Keep track of the AG with the most free blocks. */ 141 if (pag->pagf_freeblks > maxfree) { 142 maxfree = pag->pagf_freeblks; 143 max_ag = ag; 144 } 145 146 /* 147 * The AG reference count does two things: it enforces mutual 148 * exclusion when examining the suitability of an AG in this 149 * loop, and it guards against two filestreams being established 150 * in the same AG as each other. 151 */ 152 if (xfs_filestream_get_ag(mp, ag) > 1) { 153 xfs_filestream_put_ag(mp, ag); 154 goto next_ag; 155 } 156 157 longest = xfs_alloc_longest_free_extent(pag, 158 xfs_alloc_min_freelist(mp, pag), 159 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE)); 160 if (((minlen && longest >= minlen) || 161 (!minlen && pag->pagf_freeblks >= minfree)) && 162 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || 163 (flags & XFS_PICK_LOWSPACE))) { 164 165 /* Break out, retaining the reference on the AG. */ 166 free = pag->pagf_freeblks; 167 xfs_perag_put(pag); 168 *agp = ag; 169 break; 170 } 171 172 /* Drop the reference on this AG, it's not usable. */ 173 xfs_filestream_put_ag(mp, ag); 174 next_ag: 175 xfs_perag_put(pag); 176 /* Move to the next AG, wrapping to AG 0 if necessary. */ 177 if (++ag >= mp->m_sb.sb_agcount) 178 ag = 0; 179 180 /* If a full pass of the AGs hasn't been done yet, continue. */ 181 if (ag != startag) 182 continue; 183 184 /* Allow sleeping in xfs_alloc_read_agf() on the 2nd pass. */ 185 if (trylock != 0) { 186 trylock = 0; 187 continue; 188 } 189 190 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ 191 if (!(flags & XFS_PICK_LOWSPACE)) { 192 flags |= XFS_PICK_LOWSPACE; 193 continue; 194 } 195 196 /* 197 * Take the AG with the most free space, regardless of whether 198 * it's already in use by another filestream. 199 */ 200 if (max_ag != NULLAGNUMBER) { 201 xfs_filestream_get_ag(mp, max_ag); 202 free = maxfree; 203 *agp = max_ag; 204 break; 205 } 206 207 /* take AG 0 if none matched */ 208 trace_xfs_filestream_pick(ip, *agp, free, nscan); 209 *agp = 0; 210 return 0; 211 } 212 213 trace_xfs_filestream_pick(ip, *agp, free, nscan); 214 215 if (*agp == NULLAGNUMBER) 216 return 0; 217 218 err = -ENOMEM; 219 item = kmem_alloc(sizeof(*item), KM_MAYFAIL); 220 if (!item) 221 goto out_put_ag; 222 223 item->ag = *agp; 224 225 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); 226 if (err) { 227 if (err == -EEXIST) 228 err = 0; 229 goto out_free_item; 230 } 231 232 return 0; 233 234 out_free_item: 235 kmem_free(item); 236 out_put_ag: 237 xfs_filestream_put_ag(mp, *agp); 238 return err; 239 } 240 241 static struct xfs_inode * 242 xfs_filestream_get_parent( 243 struct xfs_inode *ip) 244 { 245 struct inode *inode = VFS_I(ip), *dir = NULL; 246 struct dentry *dentry, *parent; 247 248 dentry = d_find_alias(inode); 249 if (!dentry) 250 goto out; 251 252 parent = dget_parent(dentry); 253 if (!parent) 254 goto out_dput; 255 256 dir = igrab(d_inode(parent)); 257 dput(parent); 258 259 out_dput: 260 dput(dentry); 261 out: 262 return dir ? XFS_I(dir) : NULL; 263 } 264 265 /* 266 * Find the right allocation group for a file, either by finding an 267 * existing file stream or creating a new one. 268 * 269 * Returns NULLAGNUMBER in case of an error. 270 */ 271 xfs_agnumber_t 272 xfs_filestream_lookup_ag( 273 struct xfs_inode *ip) 274 { 275 struct xfs_mount *mp = ip->i_mount; 276 struct xfs_inode *pip = NULL; 277 xfs_agnumber_t startag, ag = NULLAGNUMBER; 278 struct xfs_mru_cache_elem *mru; 279 280 ASSERT(S_ISREG(VFS_I(ip)->i_mode)); 281 282 pip = xfs_filestream_get_parent(ip); 283 if (!pip) 284 return NULLAGNUMBER; 285 286 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); 287 if (mru) { 288 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag; 289 xfs_mru_cache_done(mp->m_filestream); 290 291 trace_xfs_filestream_lookup(mp, ip->i_ino, ag); 292 goto out; 293 } 294 295 /* 296 * Set the starting AG using the rotor for inode32, otherwise 297 * use the directory inode's AG. 298 */ 299 if (xfs_is_inode32(mp)) { 300 xfs_agnumber_t rotorstep = xfs_rotorstep; 301 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; 302 mp->m_agfrotor = (mp->m_agfrotor + 1) % 303 (mp->m_sb.sb_agcount * rotorstep); 304 } else 305 startag = XFS_INO_TO_AGNO(mp, pip->i_ino); 306 307 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0)) 308 ag = NULLAGNUMBER; 309 out: 310 xfs_irele(pip); 311 return ag; 312 } 313 314 /* 315 * Pick a new allocation group for the current file and its file stream. 316 * 317 * This is called when the allocator can't find a suitable extent in the 318 * current AG, and we have to move the stream into a new AG with more space. 319 */ 320 int 321 xfs_filestream_new_ag( 322 struct xfs_bmalloca *ap, 323 xfs_agnumber_t *agp) 324 { 325 struct xfs_inode *ip = ap->ip, *pip; 326 struct xfs_mount *mp = ip->i_mount; 327 xfs_extlen_t minlen = ap->length; 328 xfs_agnumber_t startag = 0; 329 int flags = 0; 330 int err = 0; 331 struct xfs_mru_cache_elem *mru; 332 333 *agp = NULLAGNUMBER; 334 335 pip = xfs_filestream_get_parent(ip); 336 if (!pip) 337 goto exit; 338 339 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); 340 if (mru) { 341 struct xfs_fstrm_item *item = 342 container_of(mru, struct xfs_fstrm_item, mru); 343 startag = (item->ag + 1) % mp->m_sb.sb_agcount; 344 } 345 346 if (ap->datatype & XFS_ALLOC_USERDATA) 347 flags |= XFS_PICK_USERDATA; 348 if (ap->tp->t_flags & XFS_TRANS_LOWMODE) 349 flags |= XFS_PICK_LOWSPACE; 350 351 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); 352 353 /* 354 * Only free the item here so we skip over the old AG earlier. 355 */ 356 if (mru) 357 xfs_fstrm_free_func(mp, mru); 358 359 xfs_irele(pip); 360 exit: 361 if (*agp == NULLAGNUMBER) 362 *agp = 0; 363 return err; 364 } 365 366 void 367 xfs_filestream_deassociate( 368 struct xfs_inode *ip) 369 { 370 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); 371 } 372 373 int 374 xfs_filestream_mount( 375 xfs_mount_t *mp) 376 { 377 /* 378 * The filestream timer tunable is currently fixed within the range of 379 * one second to four minutes, with five seconds being the default. The 380 * group count is somewhat arbitrary, but it'd be nice to adhere to the 381 * timer tunable to within about 10 percent. This requires at least 10 382 * groups. 383 */ 384 return xfs_mru_cache_create(&mp->m_filestream, mp, 385 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func); 386 } 387 388 void 389 xfs_filestream_unmount( 390 xfs_mount_t *mp) 391 { 392 xfs_mru_cache_destroy(mp->m_filestream); 393 } 394