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_pagf_init(mp, NULL, ag, trylock); 130 if (err) { 131 xfs_perag_put(pag); 132 if (err != -EAGAIN) 133 return err; 134 /* Couldn't lock the AGF, skip this AG. */ 135 continue; 136 } 137 } 138 139 /* Keep track of the AG with the most free blocks. */ 140 if (pag->pagf_freeblks > maxfree) { 141 maxfree = pag->pagf_freeblks; 142 max_ag = ag; 143 } 144 145 /* 146 * The AG reference count does two things: it enforces mutual 147 * exclusion when examining the suitability of an AG in this 148 * loop, and it guards against two filestreams being established 149 * in the same AG as each other. 150 */ 151 if (xfs_filestream_get_ag(mp, ag) > 1) { 152 xfs_filestream_put_ag(mp, ag); 153 goto next_ag; 154 } 155 156 longest = xfs_alloc_longest_free_extent(pag, 157 xfs_alloc_min_freelist(mp, pag), 158 xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE)); 159 if (((minlen && longest >= minlen) || 160 (!minlen && pag->pagf_freeblks >= minfree)) && 161 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) || 162 (flags & XFS_PICK_LOWSPACE))) { 163 164 /* Break out, retaining the reference on the AG. */ 165 free = pag->pagf_freeblks; 166 xfs_perag_put(pag); 167 *agp = ag; 168 break; 169 } 170 171 /* Drop the reference on this AG, it's not usable. */ 172 xfs_filestream_put_ag(mp, ag); 173 next_ag: 174 xfs_perag_put(pag); 175 /* Move to the next AG, wrapping to AG 0 if necessary. */ 176 if (++ag >= mp->m_sb.sb_agcount) 177 ag = 0; 178 179 /* If a full pass of the AGs hasn't been done yet, continue. */ 180 if (ag != startag) 181 continue; 182 183 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */ 184 if (trylock != 0) { 185 trylock = 0; 186 continue; 187 } 188 189 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */ 190 if (!(flags & XFS_PICK_LOWSPACE)) { 191 flags |= XFS_PICK_LOWSPACE; 192 continue; 193 } 194 195 /* 196 * Take the AG with the most free space, regardless of whether 197 * it's already in use by another filestream. 198 */ 199 if (max_ag != NULLAGNUMBER) { 200 xfs_filestream_get_ag(mp, max_ag); 201 free = maxfree; 202 *agp = max_ag; 203 break; 204 } 205 206 /* take AG 0 if none matched */ 207 trace_xfs_filestream_pick(ip, *agp, free, nscan); 208 *agp = 0; 209 return 0; 210 } 211 212 trace_xfs_filestream_pick(ip, *agp, free, nscan); 213 214 if (*agp == NULLAGNUMBER) 215 return 0; 216 217 err = -ENOMEM; 218 item = kmem_alloc(sizeof(*item), KM_MAYFAIL); 219 if (!item) 220 goto out_put_ag; 221 222 item->ag = *agp; 223 224 err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru); 225 if (err) { 226 if (err == -EEXIST) 227 err = 0; 228 goto out_free_item; 229 } 230 231 return 0; 232 233 out_free_item: 234 kmem_free(item); 235 out_put_ag: 236 xfs_filestream_put_ag(mp, *agp); 237 return err; 238 } 239 240 static struct xfs_inode * 241 xfs_filestream_get_parent( 242 struct xfs_inode *ip) 243 { 244 struct inode *inode = VFS_I(ip), *dir = NULL; 245 struct dentry *dentry, *parent; 246 247 dentry = d_find_alias(inode); 248 if (!dentry) 249 goto out; 250 251 parent = dget_parent(dentry); 252 if (!parent) 253 goto out_dput; 254 255 dir = igrab(d_inode(parent)); 256 dput(parent); 257 258 out_dput: 259 dput(dentry); 260 out: 261 return dir ? XFS_I(dir) : NULL; 262 } 263 264 /* 265 * Find the right allocation group for a file, either by finding an 266 * existing file stream or creating a new one. 267 * 268 * Returns NULLAGNUMBER in case of an error. 269 */ 270 xfs_agnumber_t 271 xfs_filestream_lookup_ag( 272 struct xfs_inode *ip) 273 { 274 struct xfs_mount *mp = ip->i_mount; 275 struct xfs_inode *pip = NULL; 276 xfs_agnumber_t startag, ag = NULLAGNUMBER; 277 struct xfs_mru_cache_elem *mru; 278 279 ASSERT(S_ISREG(VFS_I(ip)->i_mode)); 280 281 pip = xfs_filestream_get_parent(ip); 282 if (!pip) 283 return NULLAGNUMBER; 284 285 mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino); 286 if (mru) { 287 ag = container_of(mru, struct xfs_fstrm_item, mru)->ag; 288 xfs_mru_cache_done(mp->m_filestream); 289 290 trace_xfs_filestream_lookup(mp, ip->i_ino, ag); 291 goto out; 292 } 293 294 /* 295 * Set the starting AG using the rotor for inode32, otherwise 296 * use the directory inode's AG. 297 */ 298 if (xfs_is_inode32(mp)) { 299 xfs_agnumber_t rotorstep = xfs_rotorstep; 300 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount; 301 mp->m_agfrotor = (mp->m_agfrotor + 1) % 302 (mp->m_sb.sb_agcount * rotorstep); 303 } else 304 startag = XFS_INO_TO_AGNO(mp, pip->i_ino); 305 306 if (xfs_filestream_pick_ag(pip, startag, &ag, 0, 0)) 307 ag = NULLAGNUMBER; 308 out: 309 xfs_irele(pip); 310 return ag; 311 } 312 313 /* 314 * Pick a new allocation group for the current file and its file stream. 315 * 316 * This is called when the allocator can't find a suitable extent in the 317 * current AG, and we have to move the stream into a new AG with more space. 318 */ 319 int 320 xfs_filestream_new_ag( 321 struct xfs_bmalloca *ap, 322 xfs_agnumber_t *agp) 323 { 324 struct xfs_inode *ip = ap->ip, *pip; 325 struct xfs_mount *mp = ip->i_mount; 326 xfs_extlen_t minlen = ap->length; 327 xfs_agnumber_t startag = 0; 328 int flags = 0; 329 int err = 0; 330 struct xfs_mru_cache_elem *mru; 331 332 *agp = NULLAGNUMBER; 333 334 pip = xfs_filestream_get_parent(ip); 335 if (!pip) 336 goto exit; 337 338 mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino); 339 if (mru) { 340 struct xfs_fstrm_item *item = 341 container_of(mru, struct xfs_fstrm_item, mru); 342 startag = (item->ag + 1) % mp->m_sb.sb_agcount; 343 } 344 345 if (ap->datatype & XFS_ALLOC_USERDATA) 346 flags |= XFS_PICK_USERDATA; 347 if (ap->tp->t_flags & XFS_TRANS_LOWMODE) 348 flags |= XFS_PICK_LOWSPACE; 349 350 err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen); 351 352 /* 353 * Only free the item here so we skip over the old AG earlier. 354 */ 355 if (mru) 356 xfs_fstrm_free_func(mp, mru); 357 358 xfs_irele(pip); 359 exit: 360 if (*agp == NULLAGNUMBER) 361 *agp = 0; 362 return err; 363 } 364 365 void 366 xfs_filestream_deassociate( 367 struct xfs_inode *ip) 368 { 369 xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino); 370 } 371 372 int 373 xfs_filestream_mount( 374 xfs_mount_t *mp) 375 { 376 /* 377 * The filestream timer tunable is currently fixed within the range of 378 * one second to four minutes, with five seconds being the default. The 379 * group count is somewhat arbitrary, but it'd be nice to adhere to the 380 * timer tunable to within about 10 percent. This requires at least 10 381 * groups. 382 */ 383 return xfs_mru_cache_create(&mp->m_filestream, mp, 384 xfs_fstrm_centisecs * 10, 10, xfs_fstrm_free_func); 385 } 386 387 void 388 xfs_filestream_unmount( 389 xfs_mount_t *mp) 390 { 391 xfs_mru_cache_destroy(mp->m_filestream); 392 } 393