1/* 2 * Copyright (c) 2000-2005 Silicon Graphics, Inc. 3 * All Rights Reserved. 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License as 7 * published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it would be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 * GNU General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public License 15 * along with this program; if not, write the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 17 */ 18#include "xfs.h" 19#include "xfs_fs.h" 20#include "xfs_bit.h" 21#include "xfs_log.h" 22#include "xfs_inum.h" 23#include "xfs_trans.h" 24#include "xfs_sb.h" 25#include "xfs_ag.h" 26#include "xfs_alloc.h" 27#include "xfs_quota.h" 28#include "xfs_mount.h" 29#include "xfs_bmap_btree.h" 30#include "xfs_ialloc_btree.h" 31#include "xfs_dinode.h" 32#include "xfs_inode.h" 33#include "xfs_ialloc.h" 34#include "xfs_itable.h" 35#include "xfs_rtalloc.h" 36#include "xfs_error.h" 37#include "xfs_bmap.h" 38#include "xfs_attr.h" 39#include "xfs_buf_item.h" 40#include "xfs_trans_space.h" 41#include "xfs_utils.h" 42#include "xfs_qm.h" 43#include "xfs_trace.h" 44 45/* 46 * The global quota manager. There is only one of these for the entire 47 * system, _not_ one per file system. XQM keeps track of the overall 48 * quota functionality, including maintaining the freelist and hash 49 * tables of dquots. 50 */ 51struct mutex xfs_Gqm_lock; 52struct xfs_qm *xfs_Gqm; 53uint ndquot; 54 55kmem_zone_t *qm_dqzone; 56kmem_zone_t *qm_dqtrxzone; 57 58static cred_t xfs_zerocr; 59 60STATIC void xfs_qm_list_init(xfs_dqlist_t *, char *, int); 61STATIC void xfs_qm_list_destroy(xfs_dqlist_t *); 62 63STATIC int xfs_qm_init_quotainos(xfs_mount_t *); 64STATIC int xfs_qm_init_quotainfo(xfs_mount_t *); 65STATIC int xfs_qm_shake(struct shrinker *, int, gfp_t); 66 67static struct shrinker xfs_qm_shaker = { 68 .shrink = xfs_qm_shake, 69 .seeks = DEFAULT_SEEKS, 70}; 71 72#ifdef DEBUG 73extern struct mutex qcheck_lock; 74#endif 75 76#ifdef QUOTADEBUG 77static void 78xfs_qm_dquot_list_print( 79 struct xfs_mount *mp) 80{ 81 xfs_dquot_t *dqp; 82 int i = 0; 83 84 list_for_each_entry(dqp, &mp->m_quotainfo->qi_dqlist_lock, qi_mplist) { 85 cmn_err(CE_DEBUG, " %d. \"%d (%s)\" " 86 "bcnt = %lld, icnt = %lld, refs = %d", 87 i++, be32_to_cpu(dqp->q_core.d_id), 88 DQFLAGTO_TYPESTR(dqp), 89 (long long)be64_to_cpu(dqp->q_core.d_bcount), 90 (long long)be64_to_cpu(dqp->q_core.d_icount), 91 dqp->q_nrefs); 92 } 93} 94#else 95static void xfs_qm_dquot_list_print(struct xfs_mount *mp) { } 96#endif 97 98/* 99 * Initialize the XQM structure. 100 * Note that there is not one quota manager per file system. 101 */ 102STATIC struct xfs_qm * 103xfs_Gqm_init(void) 104{ 105 xfs_dqhash_t *udqhash, *gdqhash; 106 xfs_qm_t *xqm; 107 size_t hsize; 108 uint i; 109 110 /* 111 * Initialize the dquot hash tables. 112 */ 113 udqhash = kmem_zalloc_greedy(&hsize, 114 XFS_QM_HASHSIZE_LOW * sizeof(xfs_dqhash_t), 115 XFS_QM_HASHSIZE_HIGH * sizeof(xfs_dqhash_t)); 116 if (!udqhash) 117 goto out; 118 119 gdqhash = kmem_zalloc_large(hsize); 120 if (!gdqhash) 121 goto out_free_udqhash; 122 123 hsize /= sizeof(xfs_dqhash_t); 124 ndquot = hsize << 8; 125 126 xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP); 127 xqm->qm_dqhashmask = hsize - 1; 128 xqm->qm_usr_dqhtable = udqhash; 129 xqm->qm_grp_dqhtable = gdqhash; 130 ASSERT(xqm->qm_usr_dqhtable != NULL); 131 ASSERT(xqm->qm_grp_dqhtable != NULL); 132 133 for (i = 0; i < hsize; i++) { 134 xfs_qm_list_init(&(xqm->qm_usr_dqhtable[i]), "uxdqh", i); 135 xfs_qm_list_init(&(xqm->qm_grp_dqhtable[i]), "gxdqh", i); 136 } 137 138 /* 139 * Freelist of all dquots of all file systems 140 */ 141 INIT_LIST_HEAD(&xqm->qm_dqfrlist); 142 xqm->qm_dqfrlist_cnt = 0; 143 mutex_init(&xqm->qm_dqfrlist_lock); 144 145 /* 146 * dquot zone. we register our own low-memory callback. 147 */ 148 if (!qm_dqzone) { 149 xqm->qm_dqzone = kmem_zone_init(sizeof(xfs_dquot_t), 150 "xfs_dquots"); 151 qm_dqzone = xqm->qm_dqzone; 152 } else 153 xqm->qm_dqzone = qm_dqzone; 154 155 register_shrinker(&xfs_qm_shaker); 156 157 /* 158 * The t_dqinfo portion of transactions. 159 */ 160 if (!qm_dqtrxzone) { 161 xqm->qm_dqtrxzone = kmem_zone_init(sizeof(xfs_dquot_acct_t), 162 "xfs_dqtrx"); 163 qm_dqtrxzone = xqm->qm_dqtrxzone; 164 } else 165 xqm->qm_dqtrxzone = qm_dqtrxzone; 166 167 atomic_set(&xqm->qm_totaldquots, 0); 168 xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO; 169 xqm->qm_nrefs = 0; 170#ifdef DEBUG 171 mutex_init(&qcheck_lock); 172#endif 173 return xqm; 174 175 out_free_udqhash: 176 kmem_free_large(udqhash); 177 out: 178 return NULL; 179} 180 181/* 182 * Destroy the global quota manager when its reference count goes to zero. 183 */ 184STATIC void 185xfs_qm_destroy( 186 struct xfs_qm *xqm) 187{ 188 struct xfs_dquot *dqp, *n; 189 int hsize, i; 190 191 ASSERT(xqm != NULL); 192 ASSERT(xqm->qm_nrefs == 0); 193 unregister_shrinker(&xfs_qm_shaker); 194 hsize = xqm->qm_dqhashmask + 1; 195 for (i = 0; i < hsize; i++) { 196 xfs_qm_list_destroy(&(xqm->qm_usr_dqhtable[i])); 197 xfs_qm_list_destroy(&(xqm->qm_grp_dqhtable[i])); 198 } 199 kmem_free_large(xqm->qm_usr_dqhtable); 200 kmem_free_large(xqm->qm_grp_dqhtable); 201 xqm->qm_usr_dqhtable = NULL; 202 xqm->qm_grp_dqhtable = NULL; 203 xqm->qm_dqhashmask = 0; 204 205 /* frlist cleanup */ 206 mutex_lock(&xqm->qm_dqfrlist_lock); 207 list_for_each_entry_safe(dqp, n, &xqm->qm_dqfrlist, q_freelist) { 208 xfs_dqlock(dqp); 209#ifdef QUOTADEBUG 210 cmn_err(CE_DEBUG, "FREELIST destroy 0x%p", dqp); 211#endif 212 list_del_init(&dqp->q_freelist); 213 xfs_Gqm->qm_dqfrlist_cnt--; 214 xfs_dqunlock(dqp); 215 xfs_qm_dqdestroy(dqp); 216 } 217 mutex_unlock(&xqm->qm_dqfrlist_lock); 218 mutex_destroy(&xqm->qm_dqfrlist_lock); 219#ifdef DEBUG 220 mutex_destroy(&qcheck_lock); 221#endif 222 kmem_free(xqm); 223} 224 225/* 226 * Called at mount time to let XQM know that another file system is 227 * starting quotas. This isn't crucial information as the individual mount 228 * structures are pretty independent, but it helps the XQM keep a 229 * global view of what's going on. 230 */ 231/* ARGSUSED */ 232STATIC int 233xfs_qm_hold_quotafs_ref( 234 struct xfs_mount *mp) 235{ 236 /* 237 * Need to lock the xfs_Gqm structure for things like this. For example, 238 * the structure could disappear between the entry to this routine and 239 * a HOLD operation if not locked. 240 */ 241 mutex_lock(&xfs_Gqm_lock); 242 243 if (!xfs_Gqm) { 244 xfs_Gqm = xfs_Gqm_init(); 245 if (!xfs_Gqm) { 246 mutex_unlock(&xfs_Gqm_lock); 247 return ENOMEM; 248 } 249 } 250 251 /* 252 * We can keep a list of all filesystems with quotas mounted for 253 * debugging and statistical purposes, but ... 254 * Just take a reference and get out. 255 */ 256 xfs_Gqm->qm_nrefs++; 257 mutex_unlock(&xfs_Gqm_lock); 258 259 return 0; 260} 261 262 263/* 264 * Release the reference that a filesystem took at mount time, 265 * so that we know when we need to destroy the entire quota manager. 266 */ 267/* ARGSUSED */ 268STATIC void 269xfs_qm_rele_quotafs_ref( 270 struct xfs_mount *mp) 271{ 272 xfs_dquot_t *dqp, *n; 273 274 ASSERT(xfs_Gqm); 275 ASSERT(xfs_Gqm->qm_nrefs > 0); 276 277 /* 278 * Go thru the freelist and destroy all inactive dquots. 279 */ 280 mutex_lock(&xfs_Gqm->qm_dqfrlist_lock); 281 282 list_for_each_entry_safe(dqp, n, &xfs_Gqm->qm_dqfrlist, q_freelist) { 283 xfs_dqlock(dqp); 284 if (dqp->dq_flags & XFS_DQ_INACTIVE) { 285 ASSERT(dqp->q_mount == NULL); 286 ASSERT(! XFS_DQ_IS_DIRTY(dqp)); 287 ASSERT(list_empty(&dqp->q_hashlist)); 288 ASSERT(list_empty(&dqp->q_mplist)); 289 list_del_init(&dqp->q_freelist); 290 xfs_Gqm->qm_dqfrlist_cnt--; 291 xfs_dqunlock(dqp); 292 xfs_qm_dqdestroy(dqp); 293 } else { 294 xfs_dqunlock(dqp); 295 } 296 } 297 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock); 298 299 /* 300 * Destroy the entire XQM. If somebody mounts with quotaon, this'll 301 * be restarted. 302 */ 303 mutex_lock(&xfs_Gqm_lock); 304 if (--xfs_Gqm->qm_nrefs == 0) { 305 xfs_qm_destroy(xfs_Gqm); 306 xfs_Gqm = NULL; 307 } 308 mutex_unlock(&xfs_Gqm_lock); 309} 310 311/* 312 * Just destroy the quotainfo structure. 313 */ 314void 315xfs_qm_unmount( 316 struct xfs_mount *mp) 317{ 318 if (mp->m_quotainfo) { 319 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL); 320 xfs_qm_destroy_quotainfo(mp); 321 } 322} 323 324 325/* 326 * This is called from xfs_mountfs to start quotas and initialize all 327 * necessary data structures like quotainfo. This is also responsible for 328 * running a quotacheck as necessary. We are guaranteed that the superblock 329 * is consistently read in at this point. 330 * 331 * If we fail here, the mount will continue with quota turned off. We don't 332 * need to inidicate success or failure at all. 333 */ 334void 335xfs_qm_mount_quotas( 336 xfs_mount_t *mp) 337{ 338 int error = 0; 339 uint sbf; 340 341 /* 342 * If quotas on realtime volumes is not supported, we disable 343 * quotas immediately. 344 */ 345 if (mp->m_sb.sb_rextents) { 346 cmn_err(CE_NOTE, 347 "Cannot turn on quotas for realtime filesystem %s", 348 mp->m_fsname); 349 mp->m_qflags = 0; 350 goto write_changes; 351 } 352 353 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 354 355 /* 356 * Allocate the quotainfo structure inside the mount struct, and 357 * create quotainode(s), and change/rev superblock if necessary. 358 */ 359 error = xfs_qm_init_quotainfo(mp); 360 if (error) { 361 /* 362 * We must turn off quotas. 363 */ 364 ASSERT(mp->m_quotainfo == NULL); 365 mp->m_qflags = 0; 366 goto write_changes; 367 } 368 /* 369 * If any of the quotas are not consistent, do a quotacheck. 370 */ 371 if (XFS_QM_NEED_QUOTACHECK(mp)) { 372 error = xfs_qm_quotacheck(mp); 373 if (error) { 374 /* Quotacheck failed and disabled quotas. */ 375 return; 376 } 377 } 378 /* 379 * If one type of quotas is off, then it will lose its 380 * quotachecked status, since we won't be doing accounting for 381 * that type anymore. 382 */ 383 if (!XFS_IS_UQUOTA_ON(mp)) 384 mp->m_qflags &= ~XFS_UQUOTA_CHKD; 385 if (!(XFS_IS_GQUOTA_ON(mp) || XFS_IS_PQUOTA_ON(mp))) 386 mp->m_qflags &= ~XFS_OQUOTA_CHKD; 387 388 write_changes: 389 spin_lock(&mp->m_sb_lock); 390 sbf = mp->m_sb.sb_qflags; 391 mp->m_sb.sb_qflags = mp->m_qflags & XFS_MOUNT_QUOTA_ALL; 392 spin_unlock(&mp->m_sb_lock); 393 394 if (sbf != (mp->m_qflags & XFS_MOUNT_QUOTA_ALL)) { 395 if (xfs_qm_write_sb_changes(mp, XFS_SB_QFLAGS)) { 396 /* 397 * We could only have been turning quotas off. 398 * We aren't in very good shape actually because 399 * the incore structures are convinced that quotas are 400 * off, but the on disk superblock doesn't know that ! 401 */ 402 ASSERT(!(XFS_IS_QUOTA_RUNNING(mp))); 403 xfs_fs_cmn_err(CE_ALERT, mp, 404 "XFS mount_quotas: Superblock update failed!"); 405 } 406 } 407 408 if (error) { 409 xfs_fs_cmn_err(CE_WARN, mp, 410 "Failed to initialize disk quotas."); 411 return; 412 } 413 414#ifdef QUOTADEBUG 415 if (XFS_IS_QUOTA_ON(mp)) 416 xfs_qm_internalqcheck(mp); 417#endif 418} 419 420/* 421 * Called from the vfsops layer. 422 */ 423void 424xfs_qm_unmount_quotas( 425 xfs_mount_t *mp) 426{ 427 /* 428 * Release the dquots that root inode, et al might be holding, 429 * before we flush quotas and blow away the quotainfo structure. 430 */ 431 ASSERT(mp->m_rootip); 432 xfs_qm_dqdetach(mp->m_rootip); 433 if (mp->m_rbmip) 434 xfs_qm_dqdetach(mp->m_rbmip); 435 if (mp->m_rsumip) 436 xfs_qm_dqdetach(mp->m_rsumip); 437 438 /* 439 * Release the quota inodes. 440 */ 441 if (mp->m_quotainfo) { 442 if (mp->m_quotainfo->qi_uquotaip) { 443 IRELE(mp->m_quotainfo->qi_uquotaip); 444 mp->m_quotainfo->qi_uquotaip = NULL; 445 } 446 if (mp->m_quotainfo->qi_gquotaip) { 447 IRELE(mp->m_quotainfo->qi_gquotaip); 448 mp->m_quotainfo->qi_gquotaip = NULL; 449 } 450 } 451} 452 453/* 454 * Flush all dquots of the given file system to disk. The dquots are 455 * _not_ purged from memory here, just their data written to disk. 456 */ 457STATIC int 458xfs_qm_dqflush_all( 459 struct xfs_mount *mp, 460 int sync_mode) 461{ 462 struct xfs_quotainfo *q = mp->m_quotainfo; 463 int recl; 464 struct xfs_dquot *dqp; 465 int niters; 466 int error; 467 468 if (!q) 469 return 0; 470 niters = 0; 471again: 472 mutex_lock(&q->qi_dqlist_lock); 473 list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) { 474 xfs_dqlock(dqp); 475 if (! XFS_DQ_IS_DIRTY(dqp)) { 476 xfs_dqunlock(dqp); 477 continue; 478 } 479 480 recl = q->qi_dqreclaims; 481 if (!xfs_dqflock_nowait(dqp)) { 482 /* 483 * If we can't grab the flush lock then check 484 * to see if the dquot has been flushed delayed 485 * write. If so, grab its buffer and send it 486 * out immediately. We'll be able to acquire 487 * the flush lock when the I/O completes. 488 */ 489 xfs_qm_dqflock_pushbuf_wait(dqp); 490 } 491 /* 492 * Let go of the mplist lock. We don't want to hold it 493 * across a disk write. 494 */ 495 mutex_unlock(&q->qi_dqlist_lock); 496 error = xfs_qm_dqflush(dqp, sync_mode); 497 xfs_dqunlock(dqp); 498 if (error) 499 return error; 500 501 mutex_lock(&q->qi_dqlist_lock); 502 if (recl != q->qi_dqreclaims) { 503 mutex_unlock(&q->qi_dqlist_lock); 504 goto again; 505 } 506 } 507 508 mutex_unlock(&q->qi_dqlist_lock); 509 /* return ! busy */ 510 return 0; 511} 512/* 513 * Release the group dquot pointers the user dquots may be 514 * carrying around as a hint. mplist is locked on entry and exit. 515 */ 516STATIC void 517xfs_qm_detach_gdquots( 518 struct xfs_mount *mp) 519{ 520 struct xfs_quotainfo *q = mp->m_quotainfo; 521 struct xfs_dquot *dqp, *gdqp; 522 int nrecl; 523 524 again: 525 ASSERT(mutex_is_locked(&q->qi_dqlist_lock)); 526 list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) { 527 xfs_dqlock(dqp); 528 if ((gdqp = dqp->q_gdquot)) { 529 xfs_dqlock(gdqp); 530 dqp->q_gdquot = NULL; 531 } 532 xfs_dqunlock(dqp); 533 534 if (gdqp) { 535 /* 536 * Can't hold the mplist lock across a dqput. 537 * XXXmust convert to marker based iterations here. 538 */ 539 nrecl = q->qi_dqreclaims; 540 mutex_unlock(&q->qi_dqlist_lock); 541 xfs_qm_dqput(gdqp); 542 543 mutex_lock(&q->qi_dqlist_lock); 544 if (nrecl != q->qi_dqreclaims) 545 goto again; 546 } 547 } 548} 549 550/* 551 * Go through all the incore dquots of this file system and take them 552 * off the mplist and hashlist, if the dquot type matches the dqtype 553 * parameter. This is used when turning off quota accounting for 554 * users and/or groups, as well as when the filesystem is unmounting. 555 */ 556STATIC int 557xfs_qm_dqpurge_int( 558 struct xfs_mount *mp, 559 uint flags) 560{ 561 struct xfs_quotainfo *q = mp->m_quotainfo; 562 struct xfs_dquot *dqp, *n; 563 uint dqtype; 564 int nrecl; 565 int nmisses; 566 567 if (!q) 568 return 0; 569 570 dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0; 571 dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0; 572 dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0; 573 574 mutex_lock(&q->qi_dqlist_lock); 575 576 /* 577 * In the first pass through all incore dquots of this filesystem, 578 * we release the group dquot pointers the user dquots may be 579 * carrying around as a hint. We need to do this irrespective of 580 * what's being turned off. 581 */ 582 xfs_qm_detach_gdquots(mp); 583 584 again: 585 nmisses = 0; 586 ASSERT(mutex_is_locked(&q->qi_dqlist_lock)); 587 /* 588 * Try to get rid of all of the unwanted dquots. The idea is to 589 * get them off mplist and hashlist, but leave them on freelist. 590 */ 591 list_for_each_entry_safe(dqp, n, &q->qi_dqlist, q_mplist) { 592 /* 593 * It's OK to look at the type without taking dqlock here. 594 * We're holding the mplist lock here, and that's needed for 595 * a dqreclaim. 596 */ 597 if ((dqp->dq_flags & dqtype) == 0) 598 continue; 599 600 if (!mutex_trylock(&dqp->q_hash->qh_lock)) { 601 nrecl = q->qi_dqreclaims; 602 mutex_unlock(&q->qi_dqlist_lock); 603 mutex_lock(&dqp->q_hash->qh_lock); 604 mutex_lock(&q->qi_dqlist_lock); 605 606 /* 607 * XXXTheoretically, we can get into a very long 608 * ping pong game here. 609 * No one can be adding dquots to the mplist at 610 * this point, but somebody might be taking things off. 611 */ 612 if (nrecl != q->qi_dqreclaims) { 613 mutex_unlock(&dqp->q_hash->qh_lock); 614 goto again; 615 } 616 } 617 618 /* 619 * Take the dquot off the mplist and hashlist. It may remain on 620 * freelist in INACTIVE state. 621 */ 622 nmisses += xfs_qm_dqpurge(dqp); 623 } 624 mutex_unlock(&q->qi_dqlist_lock); 625 return nmisses; 626} 627 628int 629xfs_qm_dqpurge_all( 630 xfs_mount_t *mp, 631 uint flags) 632{ 633 int ndquots; 634 635 /* 636 * Purge the dquot cache. 637 * None of the dquots should really be busy at this point. 638 */ 639 if (mp->m_quotainfo) { 640 while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) { 641 delay(ndquots * 10); 642 } 643 } 644 return 0; 645} 646 647STATIC int 648xfs_qm_dqattach_one( 649 xfs_inode_t *ip, 650 xfs_dqid_t id, 651 uint type, 652 uint doalloc, 653 xfs_dquot_t *udqhint, /* hint */ 654 xfs_dquot_t **IO_idqpp) 655{ 656 xfs_dquot_t *dqp; 657 int error; 658 659 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 660 error = 0; 661 662 /* 663 * See if we already have it in the inode itself. IO_idqpp is 664 * &i_udquot or &i_gdquot. This made the code look weird, but 665 * made the logic a lot simpler. 666 */ 667 dqp = *IO_idqpp; 668 if (dqp) { 669 trace_xfs_dqattach_found(dqp); 670 return 0; 671 } 672 673 /* 674 * udqhint is the i_udquot field in inode, and is non-NULL only 675 * when the type arg is group/project. Its purpose is to save a 676 * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside 677 * the user dquot. 678 */ 679 if (udqhint) { 680 ASSERT(type == XFS_DQ_GROUP || type == XFS_DQ_PROJ); 681 xfs_dqlock(udqhint); 682 683 /* 684 * No need to take dqlock to look at the id. 685 * 686 * The ID can't change until it gets reclaimed, and it won't 687 * be reclaimed as long as we have a ref from inode and we 688 * hold the ilock. 689 */ 690 dqp = udqhint->q_gdquot; 691 if (dqp && be32_to_cpu(dqp->q_core.d_id) == id) { 692 xfs_dqlock(dqp); 693 XFS_DQHOLD(dqp); 694 ASSERT(*IO_idqpp == NULL); 695 *IO_idqpp = dqp; 696 697 xfs_dqunlock(dqp); 698 xfs_dqunlock(udqhint); 699 return 0; 700 } 701 702 /* 703 * We can't hold a dquot lock when we call the dqget code. 704 * We'll deadlock in no time, because of (not conforming to) 705 * lock ordering - the inodelock comes before any dquot lock, 706 * and we may drop and reacquire the ilock in xfs_qm_dqget(). 707 */ 708 xfs_dqunlock(udqhint); 709 } 710 711 /* 712 * Find the dquot from somewhere. This bumps the 713 * reference count of dquot and returns it locked. 714 * This can return ENOENT if dquot didn't exist on 715 * disk and we didn't ask it to allocate; 716 * ESRCH if quotas got turned off suddenly. 717 */ 718 error = xfs_qm_dqget(ip->i_mount, ip, id, type, XFS_QMOPT_DOWARN, &dqp); 719 if (error) 720 return error; 721 722 trace_xfs_dqattach_get(dqp); 723 724 /* 725 * dqget may have dropped and re-acquired the ilock, but it guarantees 726 * that the dquot returned is the one that should go in the inode. 727 */ 728 *IO_idqpp = dqp; 729 xfs_dqunlock(dqp); 730 return 0; 731} 732 733 734/* 735 * Given a udquot and gdquot, attach a ptr to the group dquot in the 736 * udquot as a hint for future lookups. The idea sounds simple, but the 737 * execution isn't, because the udquot might have a group dquot attached 738 * already and getting rid of that gets us into lock ordering constraints. 739 * The process is complicated more by the fact that the dquots may or may not 740 * be locked on entry. 741 */ 742STATIC void 743xfs_qm_dqattach_grouphint( 744 xfs_dquot_t *udq, 745 xfs_dquot_t *gdq) 746{ 747 xfs_dquot_t *tmp; 748 749 xfs_dqlock(udq); 750 751 if ((tmp = udq->q_gdquot)) { 752 if (tmp == gdq) { 753 xfs_dqunlock(udq); 754 return; 755 } 756 757 udq->q_gdquot = NULL; 758 /* 759 * We can't keep any dqlocks when calling dqrele, 760 * because the freelist lock comes before dqlocks. 761 */ 762 xfs_dqunlock(udq); 763 /* 764 * we took a hard reference once upon a time in dqget, 765 * so give it back when the udquot no longer points at it 766 * dqput() does the unlocking of the dquot. 767 */ 768 xfs_qm_dqrele(tmp); 769 770 xfs_dqlock(udq); 771 xfs_dqlock(gdq); 772 773 } else { 774 ASSERT(XFS_DQ_IS_LOCKED(udq)); 775 xfs_dqlock(gdq); 776 } 777 778 ASSERT(XFS_DQ_IS_LOCKED(udq)); 779 ASSERT(XFS_DQ_IS_LOCKED(gdq)); 780 /* 781 * Somebody could have attached a gdquot here, 782 * when we dropped the uqlock. If so, just do nothing. 783 */ 784 if (udq->q_gdquot == NULL) { 785 XFS_DQHOLD(gdq); 786 udq->q_gdquot = gdq; 787 } 788 789 xfs_dqunlock(gdq); 790 xfs_dqunlock(udq); 791} 792 793 794/* 795 * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON 796 * into account. 797 * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed. 798 * Inode may get unlocked and relocked in here, and the caller must deal with 799 * the consequences. 800 */ 801int 802xfs_qm_dqattach_locked( 803 xfs_inode_t *ip, 804 uint flags) 805{ 806 xfs_mount_t *mp = ip->i_mount; 807 uint nquotas = 0; 808 int error = 0; 809 810 if (!XFS_IS_QUOTA_RUNNING(mp) || 811 !XFS_IS_QUOTA_ON(mp) || 812 !XFS_NOT_DQATTACHED(mp, ip) || 813 ip->i_ino == mp->m_sb.sb_uquotino || 814 ip->i_ino == mp->m_sb.sb_gquotino) 815 return 0; 816 817 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 818 819 if (XFS_IS_UQUOTA_ON(mp)) { 820 error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER, 821 flags & XFS_QMOPT_DQALLOC, 822 NULL, &ip->i_udquot); 823 if (error) 824 goto done; 825 nquotas++; 826 } 827 828 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 829 if (XFS_IS_OQUOTA_ON(mp)) { 830 error = XFS_IS_GQUOTA_ON(mp) ? 831 xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP, 832 flags & XFS_QMOPT_DQALLOC, 833 ip->i_udquot, &ip->i_gdquot) : 834 xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ, 835 flags & XFS_QMOPT_DQALLOC, 836 ip->i_udquot, &ip->i_gdquot); 837 /* 838 * Don't worry about the udquot that we may have 839 * attached above. It'll get detached, if not already. 840 */ 841 if (error) 842 goto done; 843 nquotas++; 844 } 845 846 /* 847 * Attach this group quota to the user quota as a hint. 848 * This WON'T, in general, result in a thrash. 849 */ 850 if (nquotas == 2) { 851 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 852 ASSERT(ip->i_udquot); 853 ASSERT(ip->i_gdquot); 854 855 /* 856 * We may or may not have the i_udquot locked at this point, 857 * but this check is OK since we don't depend on the i_gdquot to 858 * be accurate 100% all the time. It is just a hint, and this 859 * will succeed in general. 860 */ 861 if (ip->i_udquot->q_gdquot == ip->i_gdquot) 862 goto done; 863 /* 864 * Attach i_gdquot to the gdquot hint inside the i_udquot. 865 */ 866 xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot); 867 } 868 869 done: 870#ifdef QUOTADEBUG 871 if (! error) { 872 if (XFS_IS_UQUOTA_ON(mp)) 873 ASSERT(ip->i_udquot); 874 if (XFS_IS_OQUOTA_ON(mp)) 875 ASSERT(ip->i_gdquot); 876 } 877 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 878#endif 879 return error; 880} 881 882int 883xfs_qm_dqattach( 884 struct xfs_inode *ip, 885 uint flags) 886{ 887 int error; 888 889 xfs_ilock(ip, XFS_ILOCK_EXCL); 890 error = xfs_qm_dqattach_locked(ip, flags); 891 xfs_iunlock(ip, XFS_ILOCK_EXCL); 892 893 return error; 894} 895 896/* 897 * Release dquots (and their references) if any. 898 * The inode should be locked EXCL except when this's called by 899 * xfs_ireclaim. 900 */ 901void 902xfs_qm_dqdetach( 903 xfs_inode_t *ip) 904{ 905 if (!(ip->i_udquot || ip->i_gdquot)) 906 return; 907 908 trace_xfs_dquot_dqdetach(ip); 909 910 ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino); 911 ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino); 912 if (ip->i_udquot) { 913 xfs_qm_dqrele(ip->i_udquot); 914 ip->i_udquot = NULL; 915 } 916 if (ip->i_gdquot) { 917 xfs_qm_dqrele(ip->i_gdquot); 918 ip->i_gdquot = NULL; 919 } 920} 921 922int 923xfs_qm_sync( 924 struct xfs_mount *mp, 925 int flags) 926{ 927 struct xfs_quotainfo *q = mp->m_quotainfo; 928 int recl, restarts; 929 struct xfs_dquot *dqp; 930 int error; 931 932 if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp)) 933 return 0; 934 935 restarts = 0; 936 937 again: 938 mutex_lock(&q->qi_dqlist_lock); 939 /* 940 * dqpurge_all() also takes the mplist lock and iterate thru all dquots 941 * in quotaoff. However, if the QUOTA_ACTIVE bits are not cleared 942 * when we have the mplist lock, we know that dquots will be consistent 943 * as long as we have it locked. 944 */ 945 if (!XFS_IS_QUOTA_ON(mp)) { 946 mutex_unlock(&q->qi_dqlist_lock); 947 return 0; 948 } 949 ASSERT(mutex_is_locked(&q->qi_dqlist_lock)); 950 list_for_each_entry(dqp, &q->qi_dqlist, q_mplist) { 951 /* 952 * If this is vfs_sync calling, then skip the dquots that 953 * don't 'seem' to be dirty. ie. don't acquire dqlock. 954 * This is very similar to what xfs_sync does with inodes. 955 */ 956 if (flags & SYNC_TRYLOCK) { 957 if (!XFS_DQ_IS_DIRTY(dqp)) 958 continue; 959 if (!xfs_qm_dqlock_nowait(dqp)) 960 continue; 961 } else { 962 xfs_dqlock(dqp); 963 } 964 965 /* 966 * Now, find out for sure if this dquot is dirty or not. 967 */ 968 if (! XFS_DQ_IS_DIRTY(dqp)) { 969 xfs_dqunlock(dqp); 970 continue; 971 } 972 973 recl = q->qi_dqreclaims; 974 if (!xfs_dqflock_nowait(dqp)) { 975 if (flags & SYNC_TRYLOCK) { 976 xfs_dqunlock(dqp); 977 continue; 978 } 979 /* 980 * If we can't grab the flush lock then if the caller 981 * really wanted us to give this our best shot, so 982 * see if we can give a push to the buffer before we wait 983 * on the flush lock. At this point, we know that 984 * even though the dquot is being flushed, 985 * it has (new) dirty data. 986 */ 987 xfs_qm_dqflock_pushbuf_wait(dqp); 988 } 989 /* 990 * Let go of the mplist lock. We don't want to hold it 991 * across a disk write 992 */ 993 mutex_unlock(&q->qi_dqlist_lock); 994 error = xfs_qm_dqflush(dqp, flags); 995 xfs_dqunlock(dqp); 996 if (error && XFS_FORCED_SHUTDOWN(mp)) 997 return 0; /* Need to prevent umount failure */ 998 else if (error) 999 return error; 1000 1001 mutex_lock(&q->qi_dqlist_lock); 1002 if (recl != q->qi_dqreclaims) { 1003 if (++restarts >= XFS_QM_SYNC_MAX_RESTARTS) 1004 break; 1005 1006 mutex_unlock(&q->qi_dqlist_lock); 1007 goto again; 1008 } 1009 } 1010 1011 mutex_unlock(&q->qi_dqlist_lock); 1012 return 0; 1013} 1014 1015/* 1016 * The hash chains and the mplist use the same xfs_dqhash structure as 1017 * their list head, but we can take the mplist qh_lock and one of the 1018 * hash qh_locks at the same time without any problem as they aren't 1019 * related. 1020 */ 1021static struct lock_class_key xfs_quota_mplist_class; 1022 1023/* 1024 * This initializes all the quota information that's kept in the 1025 * mount structure 1026 */ 1027STATIC int 1028xfs_qm_init_quotainfo( 1029 xfs_mount_t *mp) 1030{ 1031 xfs_quotainfo_t *qinf; 1032 int error; 1033 xfs_dquot_t *dqp; 1034 1035 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 1036 1037 /* 1038 * Tell XQM that we exist as soon as possible. 1039 */ 1040 if ((error = xfs_qm_hold_quotafs_ref(mp))) { 1041 return error; 1042 } 1043 1044 qinf = mp->m_quotainfo = kmem_zalloc(sizeof(xfs_quotainfo_t), KM_SLEEP); 1045 1046 /* 1047 * See if quotainodes are setup, and if not, allocate them, 1048 * and change the superblock accordingly. 1049 */ 1050 if ((error = xfs_qm_init_quotainos(mp))) { 1051 kmem_free(qinf); 1052 mp->m_quotainfo = NULL; 1053 return error; 1054 } 1055 1056 INIT_LIST_HEAD(&qinf->qi_dqlist); 1057 mutex_init(&qinf->qi_dqlist_lock); 1058 lockdep_set_class(&qinf->qi_dqlist_lock, &xfs_quota_mplist_class); 1059 1060 qinf->qi_dqreclaims = 0; 1061 1062 /* mutex used to serialize quotaoffs */ 1063 mutex_init(&qinf->qi_quotaofflock); 1064 1065 /* Precalc some constants */ 1066 qinf->qi_dqchunklen = XFS_FSB_TO_BB(mp, XFS_DQUOT_CLUSTER_SIZE_FSB); 1067 ASSERT(qinf->qi_dqchunklen); 1068 qinf->qi_dqperchunk = BBTOB(qinf->qi_dqchunklen); 1069 do_div(qinf->qi_dqperchunk, sizeof(xfs_dqblk_t)); 1070 1071 mp->m_qflags |= (mp->m_sb.sb_qflags & XFS_ALL_QUOTA_CHKD); 1072 1073 /* 1074 * We try to get the limits from the superuser's limits fields. 1075 * This is quite hacky, but it is standard quota practice. 1076 * We look at the USR dquot with id == 0 first, but if user quotas 1077 * are not enabled we goto the GRP dquot with id == 0. 1078 * We don't really care to keep separate default limits for user 1079 * and group quotas, at least not at this point. 1080 */ 1081 error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)0, 1082 XFS_IS_UQUOTA_RUNNING(mp) ? XFS_DQ_USER : 1083 (XFS_IS_GQUOTA_RUNNING(mp) ? XFS_DQ_GROUP : 1084 XFS_DQ_PROJ), 1085 XFS_QMOPT_DQSUSER|XFS_QMOPT_DOWARN, 1086 &dqp); 1087 if (! error) { 1088 xfs_disk_dquot_t *ddqp = &dqp->q_core; 1089 1090 /* 1091 * The warnings and timers set the grace period given to 1092 * a user or group before he or she can not perform any 1093 * more writing. If it is zero, a default is used. 1094 */ 1095 qinf->qi_btimelimit = ddqp->d_btimer ? 1096 be32_to_cpu(ddqp->d_btimer) : XFS_QM_BTIMELIMIT; 1097 qinf->qi_itimelimit = ddqp->d_itimer ? 1098 be32_to_cpu(ddqp->d_itimer) : XFS_QM_ITIMELIMIT; 1099 qinf->qi_rtbtimelimit = ddqp->d_rtbtimer ? 1100 be32_to_cpu(ddqp->d_rtbtimer) : XFS_QM_RTBTIMELIMIT; 1101 qinf->qi_bwarnlimit = ddqp->d_bwarns ? 1102 be16_to_cpu(ddqp->d_bwarns) : XFS_QM_BWARNLIMIT; 1103 qinf->qi_iwarnlimit = ddqp->d_iwarns ? 1104 be16_to_cpu(ddqp->d_iwarns) : XFS_QM_IWARNLIMIT; 1105 qinf->qi_rtbwarnlimit = ddqp->d_rtbwarns ? 1106 be16_to_cpu(ddqp->d_rtbwarns) : XFS_QM_RTBWARNLIMIT; 1107 qinf->qi_bhardlimit = be64_to_cpu(ddqp->d_blk_hardlimit); 1108 qinf->qi_bsoftlimit = be64_to_cpu(ddqp->d_blk_softlimit); 1109 qinf->qi_ihardlimit = be64_to_cpu(ddqp->d_ino_hardlimit); 1110 qinf->qi_isoftlimit = be64_to_cpu(ddqp->d_ino_softlimit); 1111 qinf->qi_rtbhardlimit = be64_to_cpu(ddqp->d_rtb_hardlimit); 1112 qinf->qi_rtbsoftlimit = be64_to_cpu(ddqp->d_rtb_softlimit); 1113 1114 /* 1115 * We sent the XFS_QMOPT_DQSUSER flag to dqget because 1116 * we don't want this dquot cached. We haven't done a 1117 * quotacheck yet, and quotacheck doesn't like incore dquots. 1118 */ 1119 xfs_qm_dqdestroy(dqp); 1120 } else { 1121 qinf->qi_btimelimit = XFS_QM_BTIMELIMIT; 1122 qinf->qi_itimelimit = XFS_QM_ITIMELIMIT; 1123 qinf->qi_rtbtimelimit = XFS_QM_RTBTIMELIMIT; 1124 qinf->qi_bwarnlimit = XFS_QM_BWARNLIMIT; 1125 qinf->qi_iwarnlimit = XFS_QM_IWARNLIMIT; 1126 qinf->qi_rtbwarnlimit = XFS_QM_RTBWARNLIMIT; 1127 } 1128 1129 return 0; 1130} 1131 1132 1133/* 1134 * Gets called when unmounting a filesystem or when all quotas get 1135 * turned off. 1136 * This purges the quota inodes, destroys locks and frees itself. 1137 */ 1138void 1139xfs_qm_destroy_quotainfo( 1140 xfs_mount_t *mp) 1141{ 1142 xfs_quotainfo_t *qi; 1143 1144 qi = mp->m_quotainfo; 1145 ASSERT(qi != NULL); 1146 ASSERT(xfs_Gqm != NULL); 1147 1148 /* 1149 * Release the reference that XQM kept, so that we know 1150 * when the XQM structure should be freed. We cannot assume 1151 * that xfs_Gqm is non-null after this point. 1152 */ 1153 xfs_qm_rele_quotafs_ref(mp); 1154 1155 ASSERT(list_empty(&qi->qi_dqlist)); 1156 mutex_destroy(&qi->qi_dqlist_lock); 1157 1158 if (qi->qi_uquotaip) { 1159 IRELE(qi->qi_uquotaip); 1160 qi->qi_uquotaip = NULL; /* paranoia */ 1161 } 1162 if (qi->qi_gquotaip) { 1163 IRELE(qi->qi_gquotaip); 1164 qi->qi_gquotaip = NULL; 1165 } 1166 mutex_destroy(&qi->qi_quotaofflock); 1167 kmem_free(qi); 1168 mp->m_quotainfo = NULL; 1169} 1170 1171 1172 1173/* ------------------- PRIVATE STATIC FUNCTIONS ----------------------- */ 1174 1175/* ARGSUSED */ 1176STATIC void 1177xfs_qm_list_init( 1178 xfs_dqlist_t *list, 1179 char *str, 1180 int n) 1181{ 1182 mutex_init(&list->qh_lock); 1183 INIT_LIST_HEAD(&list->qh_list); 1184 list->qh_version = 0; 1185 list->qh_nelems = 0; 1186} 1187 1188STATIC void 1189xfs_qm_list_destroy( 1190 xfs_dqlist_t *list) 1191{ 1192 mutex_destroy(&(list->qh_lock)); 1193} 1194 1195 1196/* 1197 * Stripped down version of dqattach. This doesn't attach, or even look at the 1198 * dquots attached to the inode. The rationale is that there won't be any 1199 * attached at the time this is called from quotacheck. 1200 */ 1201STATIC int 1202xfs_qm_dqget_noattach( 1203 xfs_inode_t *ip, 1204 xfs_dquot_t **O_udqpp, 1205 xfs_dquot_t **O_gdqpp) 1206{ 1207 int error; 1208 xfs_mount_t *mp; 1209 xfs_dquot_t *udqp, *gdqp; 1210 1211 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 1212 mp = ip->i_mount; 1213 udqp = NULL; 1214 gdqp = NULL; 1215 1216 if (XFS_IS_UQUOTA_ON(mp)) { 1217 ASSERT(ip->i_udquot == NULL); 1218 /* 1219 * We want the dquot allocated if it doesn't exist. 1220 */ 1221 if ((error = xfs_qm_dqget(mp, ip, ip->i_d.di_uid, XFS_DQ_USER, 1222 XFS_QMOPT_DQALLOC | XFS_QMOPT_DOWARN, 1223 &udqp))) { 1224 /* 1225 * Shouldn't be able to turn off quotas here. 1226 */ 1227 ASSERT(error != ESRCH); 1228 ASSERT(error != ENOENT); 1229 return error; 1230 } 1231 ASSERT(udqp); 1232 } 1233 1234 if (XFS_IS_OQUOTA_ON(mp)) { 1235 ASSERT(ip->i_gdquot == NULL); 1236 if (udqp) 1237 xfs_dqunlock(udqp); 1238 error = XFS_IS_GQUOTA_ON(mp) ? 1239 xfs_qm_dqget(mp, ip, 1240 ip->i_d.di_gid, XFS_DQ_GROUP, 1241 XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN, 1242 &gdqp) : 1243 xfs_qm_dqget(mp, ip, 1244 ip->i_d.di_projid, XFS_DQ_PROJ, 1245 XFS_QMOPT_DQALLOC|XFS_QMOPT_DOWARN, 1246 &gdqp); 1247 if (error) { 1248 if (udqp) 1249 xfs_qm_dqrele(udqp); 1250 ASSERT(error != ESRCH); 1251 ASSERT(error != ENOENT); 1252 return error; 1253 } 1254 ASSERT(gdqp); 1255 1256 /* Reacquire the locks in the right order */ 1257 if (udqp) { 1258 if (! xfs_qm_dqlock_nowait(udqp)) { 1259 xfs_dqunlock(gdqp); 1260 xfs_dqlock(udqp); 1261 xfs_dqlock(gdqp); 1262 } 1263 } 1264 } 1265 1266 *O_udqpp = udqp; 1267 *O_gdqpp = gdqp; 1268 1269#ifdef QUOTADEBUG 1270 if (udqp) ASSERT(XFS_DQ_IS_LOCKED(udqp)); 1271 if (gdqp) ASSERT(XFS_DQ_IS_LOCKED(gdqp)); 1272#endif 1273 return 0; 1274} 1275 1276/* 1277 * Create an inode and return with a reference already taken, but unlocked 1278 * This is how we create quota inodes 1279 */ 1280STATIC int 1281xfs_qm_qino_alloc( 1282 xfs_mount_t *mp, 1283 xfs_inode_t **ip, 1284 __int64_t sbfields, 1285 uint flags) 1286{ 1287 xfs_trans_t *tp; 1288 int error; 1289 int committed; 1290 1291 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_QINOCREATE); 1292 if ((error = xfs_trans_reserve(tp, 1293 XFS_QM_QINOCREATE_SPACE_RES(mp), 1294 XFS_CREATE_LOG_RES(mp), 0, 1295 XFS_TRANS_PERM_LOG_RES, 1296 XFS_CREATE_LOG_COUNT))) { 1297 xfs_trans_cancel(tp, 0); 1298 return error; 1299 } 1300 1301 if ((error = xfs_dir_ialloc(&tp, NULL, S_IFREG, 1, 0, 1302 &xfs_zerocr, 0, 1, ip, &committed))) { 1303 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | 1304 XFS_TRANS_ABORT); 1305 return error; 1306 } 1307 1308 /* 1309 * Keep an extra reference to this quota inode. This inode is 1310 * locked exclusively and joined to the transaction already. 1311 */ 1312 ASSERT(xfs_isilocked(*ip, XFS_ILOCK_EXCL)); 1313 IHOLD(*ip); 1314 1315 /* 1316 * Make the changes in the superblock, and log those too. 1317 * sbfields arg may contain fields other than *QUOTINO; 1318 * VERSIONNUM for example. 1319 */ 1320 spin_lock(&mp->m_sb_lock); 1321 if (flags & XFS_QMOPT_SBVERSION) { 1322 ASSERT(!xfs_sb_version_hasquota(&mp->m_sb)); 1323 ASSERT((sbfields & (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO | 1324 XFS_SB_GQUOTINO | XFS_SB_QFLAGS)) == 1325 (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO | 1326 XFS_SB_GQUOTINO | XFS_SB_QFLAGS)); 1327 1328 xfs_sb_version_addquota(&mp->m_sb); 1329 mp->m_sb.sb_uquotino = NULLFSINO; 1330 mp->m_sb.sb_gquotino = NULLFSINO; 1331 1332 /* qflags will get updated _after_ quotacheck */ 1333 mp->m_sb.sb_qflags = 0; 1334 } 1335 if (flags & XFS_QMOPT_UQUOTA) 1336 mp->m_sb.sb_uquotino = (*ip)->i_ino; 1337 else 1338 mp->m_sb.sb_gquotino = (*ip)->i_ino; 1339 spin_unlock(&mp->m_sb_lock); 1340 xfs_mod_sb(tp, sbfields); 1341 1342 if ((error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES))) { 1343 xfs_fs_cmn_err(CE_ALERT, mp, "XFS qino_alloc failed!"); 1344 return error; 1345 } 1346 return 0; 1347} 1348 1349 1350STATIC void 1351xfs_qm_reset_dqcounts( 1352 xfs_mount_t *mp, 1353 xfs_buf_t *bp, 1354 xfs_dqid_t id, 1355 uint type) 1356{ 1357 xfs_disk_dquot_t *ddq; 1358 int j; 1359 1360 trace_xfs_reset_dqcounts(bp, _RET_IP_); 1361 1362 /* 1363 * Reset all counters and timers. They'll be 1364 * started afresh by xfs_qm_quotacheck. 1365 */ 1366#ifdef DEBUG 1367 j = XFS_FSB_TO_B(mp, XFS_DQUOT_CLUSTER_SIZE_FSB); 1368 do_div(j, sizeof(xfs_dqblk_t)); 1369 ASSERT(mp->m_quotainfo->qi_dqperchunk == j); 1370#endif 1371 ddq = (xfs_disk_dquot_t *)XFS_BUF_PTR(bp); 1372 for (j = 0; j < mp->m_quotainfo->qi_dqperchunk; j++) { 1373 /* 1374 * Do a sanity check, and if needed, repair the dqblk. Don't 1375 * output any warnings because it's perfectly possible to 1376 * find uninitialised dquot blks. See comment in xfs_qm_dqcheck. 1377 */ 1378 (void) xfs_qm_dqcheck(ddq, id+j, type, XFS_QMOPT_DQREPAIR, 1379 "xfs_quotacheck"); 1380 ddq->d_bcount = 0; 1381 ddq->d_icount = 0; 1382 ddq->d_rtbcount = 0; 1383 ddq->d_btimer = 0; 1384 ddq->d_itimer = 0; 1385 ddq->d_rtbtimer = 0; 1386 ddq->d_bwarns = 0; 1387 ddq->d_iwarns = 0; 1388 ddq->d_rtbwarns = 0; 1389 ddq = (xfs_disk_dquot_t *) ((xfs_dqblk_t *)ddq + 1); 1390 } 1391} 1392 1393STATIC int 1394xfs_qm_dqiter_bufs( 1395 xfs_mount_t *mp, 1396 xfs_dqid_t firstid, 1397 xfs_fsblock_t bno, 1398 xfs_filblks_t blkcnt, 1399 uint flags) 1400{ 1401 xfs_buf_t *bp; 1402 int error; 1403 int notcommitted; 1404 int incr; 1405 int type; 1406 1407 ASSERT(blkcnt > 0); 1408 notcommitted = 0; 1409 incr = (blkcnt > XFS_QM_MAX_DQCLUSTER_LOGSZ) ? 1410 XFS_QM_MAX_DQCLUSTER_LOGSZ : blkcnt; 1411 type = flags & XFS_QMOPT_UQUOTA ? XFS_DQ_USER : 1412 (flags & XFS_QMOPT_PQUOTA ? XFS_DQ_PROJ : XFS_DQ_GROUP); 1413 error = 0; 1414 1415 /* 1416 * Blkcnt arg can be a very big number, and might even be 1417 * larger than the log itself. So, we have to break it up into 1418 * manageable-sized transactions. 1419 * Note that we don't start a permanent transaction here; we might 1420 * not be able to get a log reservation for the whole thing up front, 1421 * and we don't really care to either, because we just discard 1422 * everything if we were to crash in the middle of this loop. 1423 */ 1424 while (blkcnt--) { 1425 error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, 1426 XFS_FSB_TO_DADDR(mp, bno), 1427 mp->m_quotainfo->qi_dqchunklen, 0, &bp); 1428 if (error) 1429 break; 1430 1431 xfs_qm_reset_dqcounts(mp, bp, firstid, type); 1432 xfs_bdwrite(mp, bp); 1433 /* 1434 * goto the next block. 1435 */ 1436 bno++; 1437 firstid += mp->m_quotainfo->qi_dqperchunk; 1438 } 1439 return error; 1440} 1441 1442/* 1443 * Iterate over all allocated USR/GRP/PRJ dquots in the system, calling a 1444 * caller supplied function for every chunk of dquots that we find. 1445 */ 1446STATIC int 1447xfs_qm_dqiterate( 1448 xfs_mount_t *mp, 1449 xfs_inode_t *qip, 1450 uint flags) 1451{ 1452 xfs_bmbt_irec_t *map; 1453 int i, nmaps; /* number of map entries */ 1454 int error; /* return value */ 1455 xfs_fileoff_t lblkno; 1456 xfs_filblks_t maxlblkcnt; 1457 xfs_dqid_t firstid; 1458 xfs_fsblock_t rablkno; 1459 xfs_filblks_t rablkcnt; 1460 1461 error = 0; 1462 /* 1463 * This looks racy, but we can't keep an inode lock across a 1464 * trans_reserve. But, this gets called during quotacheck, and that 1465 * happens only at mount time which is single threaded. 1466 */ 1467 if (qip->i_d.di_nblocks == 0) 1468 return 0; 1469 1470 map = kmem_alloc(XFS_DQITER_MAP_SIZE * sizeof(*map), KM_SLEEP); 1471 1472 lblkno = 0; 1473 maxlblkcnt = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp)); 1474 do { 1475 nmaps = XFS_DQITER_MAP_SIZE; 1476 /* 1477 * We aren't changing the inode itself. Just changing 1478 * some of its data. No new blocks are added here, and 1479 * the inode is never added to the transaction. 1480 */ 1481 xfs_ilock(qip, XFS_ILOCK_SHARED); 1482 error = xfs_bmapi(NULL, qip, lblkno, 1483 maxlblkcnt - lblkno, 1484 XFS_BMAPI_METADATA, 1485 NULL, 1486 0, map, &nmaps, NULL); 1487 xfs_iunlock(qip, XFS_ILOCK_SHARED); 1488 if (error) 1489 break; 1490 1491 ASSERT(nmaps <= XFS_DQITER_MAP_SIZE); 1492 for (i = 0; i < nmaps; i++) { 1493 ASSERT(map[i].br_startblock != DELAYSTARTBLOCK); 1494 ASSERT(map[i].br_blockcount); 1495 1496 1497 lblkno += map[i].br_blockcount; 1498 1499 if (map[i].br_startblock == HOLESTARTBLOCK) 1500 continue; 1501 1502 firstid = (xfs_dqid_t) map[i].br_startoff * 1503 mp->m_quotainfo->qi_dqperchunk; 1504 /* 1505 * Do a read-ahead on the next extent. 1506 */ 1507 if ((i+1 < nmaps) && 1508 (map[i+1].br_startblock != HOLESTARTBLOCK)) { 1509 rablkcnt = map[i+1].br_blockcount; 1510 rablkno = map[i+1].br_startblock; 1511 while (rablkcnt--) { 1512 xfs_baread(mp->m_ddev_targp, 1513 XFS_FSB_TO_DADDR(mp, rablkno), 1514 mp->m_quotainfo->qi_dqchunklen); 1515 rablkno++; 1516 } 1517 } 1518 /* 1519 * Iterate thru all the blks in the extent and 1520 * reset the counters of all the dquots inside them. 1521 */ 1522 if ((error = xfs_qm_dqiter_bufs(mp, 1523 firstid, 1524 map[i].br_startblock, 1525 map[i].br_blockcount, 1526 flags))) { 1527 break; 1528 } 1529 } 1530 1531 if (error) 1532 break; 1533 } while (nmaps > 0); 1534 1535 kmem_free(map); 1536 1537 return error; 1538} 1539 1540/* 1541 * Called by dqusage_adjust in doing a quotacheck. 1542 * Given the inode, and a dquot (either USR or GRP, doesn't matter), 1543 * this updates its incore copy as well as the buffer copy. This is 1544 * so that once the quotacheck is done, we can just log all the buffers, 1545 * as opposed to logging numerous updates to individual dquots. 1546 */ 1547STATIC void 1548xfs_qm_quotacheck_dqadjust( 1549 xfs_dquot_t *dqp, 1550 xfs_qcnt_t nblks, 1551 xfs_qcnt_t rtblks) 1552{ 1553 ASSERT(XFS_DQ_IS_LOCKED(dqp)); 1554 1555 trace_xfs_dqadjust(dqp); 1556 1557 /* 1558 * Adjust the inode count and the block count to reflect this inode's 1559 * resource usage. 1560 */ 1561 be64_add_cpu(&dqp->q_core.d_icount, 1); 1562 dqp->q_res_icount++; 1563 if (nblks) { 1564 be64_add_cpu(&dqp->q_core.d_bcount, nblks); 1565 dqp->q_res_bcount += nblks; 1566 } 1567 if (rtblks) { 1568 be64_add_cpu(&dqp->q_core.d_rtbcount, rtblks); 1569 dqp->q_res_rtbcount += rtblks; 1570 } 1571 1572 /* 1573 * Set default limits, adjust timers (since we changed usages) 1574 * 1575 * There are no timers for the default values set in the root dquot. 1576 */ 1577 if (dqp->q_core.d_id) { 1578 xfs_qm_adjust_dqlimits(dqp->q_mount, &dqp->q_core); 1579 xfs_qm_adjust_dqtimers(dqp->q_mount, &dqp->q_core); 1580 } 1581 1582 dqp->dq_flags |= XFS_DQ_DIRTY; 1583} 1584 1585STATIC int 1586xfs_qm_get_rtblks( 1587 xfs_inode_t *ip, 1588 xfs_qcnt_t *O_rtblks) 1589{ 1590 xfs_filblks_t rtblks; /* total rt blks */ 1591 xfs_extnum_t idx; /* extent record index */ 1592 xfs_ifork_t *ifp; /* inode fork pointer */ 1593 xfs_extnum_t nextents; /* number of extent entries */ 1594 int error; 1595 1596 ASSERT(XFS_IS_REALTIME_INODE(ip)); 1597 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK); 1598 if (!(ifp->if_flags & XFS_IFEXTENTS)) { 1599 if ((error = xfs_iread_extents(NULL, ip, XFS_DATA_FORK))) 1600 return error; 1601 } 1602 rtblks = 0; 1603 nextents = ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t); 1604 for (idx = 0; idx < nextents; idx++) 1605 rtblks += xfs_bmbt_get_blockcount(xfs_iext_get_ext(ifp, idx)); 1606 *O_rtblks = (xfs_qcnt_t)rtblks; 1607 return 0; 1608} 1609 1610/* 1611 * callback routine supplied to bulkstat(). Given an inumber, find its 1612 * dquots and update them to account for resources taken by that inode. 1613 */ 1614/* ARGSUSED */ 1615STATIC int 1616xfs_qm_dqusage_adjust( 1617 xfs_mount_t *mp, /* mount point for filesystem */ 1618 xfs_ino_t ino, /* inode number to get data for */ 1619 void __user *buffer, /* not used */ 1620 int ubsize, /* not used */ 1621 int *ubused, /* not used */ 1622 int *res) /* result code value */ 1623{ 1624 xfs_inode_t *ip; 1625 xfs_dquot_t *udqp, *gdqp; 1626 xfs_qcnt_t nblks, rtblks; 1627 int error; 1628 1629 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 1630 1631 /* 1632 * rootino must have its resources accounted for, not so with the quota 1633 * inodes. 1634 */ 1635 if (ino == mp->m_sb.sb_uquotino || ino == mp->m_sb.sb_gquotino) { 1636 *res = BULKSTAT_RV_NOTHING; 1637 return XFS_ERROR(EINVAL); 1638 } 1639 1640 /* 1641 * We don't _need_ to take the ilock EXCL. However, the xfs_qm_dqget 1642 * interface expects the inode to be exclusively locked because that's 1643 * the case in all other instances. It's OK that we do this because 1644 * quotacheck is done only at mount time. 1645 */ 1646 if ((error = xfs_iget(mp, NULL, ino, 0, XFS_ILOCK_EXCL, &ip))) { 1647 *res = BULKSTAT_RV_NOTHING; 1648 return error; 1649 } 1650 1651 /* 1652 * Obtain the locked dquots. In case of an error (eg. allocation 1653 * fails for ENOSPC), we return the negative of the error number 1654 * to bulkstat, so that it can get propagated to quotacheck() and 1655 * making us disable quotas for the file system. 1656 */ 1657 if ((error = xfs_qm_dqget_noattach(ip, &udqp, &gdqp))) { 1658 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1659 IRELE(ip); 1660 *res = BULKSTAT_RV_GIVEUP; 1661 return error; 1662 } 1663 1664 rtblks = 0; 1665 if (! XFS_IS_REALTIME_INODE(ip)) { 1666 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks; 1667 } else { 1668 /* 1669 * Walk thru the extent list and count the realtime blocks. 1670 */ 1671 if ((error = xfs_qm_get_rtblks(ip, &rtblks))) { 1672 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1673 IRELE(ip); 1674 if (udqp) 1675 xfs_qm_dqput(udqp); 1676 if (gdqp) 1677 xfs_qm_dqput(gdqp); 1678 *res = BULKSTAT_RV_GIVEUP; 1679 return error; 1680 } 1681 nblks = (xfs_qcnt_t)ip->i_d.di_nblocks - rtblks; 1682 } 1683 ASSERT(ip->i_delayed_blks == 0); 1684 1685 /* 1686 * We can't release the inode while holding its dquot locks. 1687 * The inode can go into inactive and might try to acquire the dquotlocks. 1688 * So, just unlock here and do a vn_rele at the end. 1689 */ 1690 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1691 1692 /* 1693 * Add the (disk blocks and inode) resources occupied by this 1694 * inode to its dquots. We do this adjustment in the incore dquot, 1695 * and also copy the changes to its buffer. 1696 * We don't care about putting these changes in a transaction 1697 * envelope because if we crash in the middle of a 'quotacheck' 1698 * we have to start from the beginning anyway. 1699 * Once we're done, we'll log all the dquot bufs. 1700 * 1701 * The *QUOTA_ON checks below may look pretty racy, but quotachecks 1702 * and quotaoffs don't race. (Quotachecks happen at mount time only). 1703 */ 1704 if (XFS_IS_UQUOTA_ON(mp)) { 1705 ASSERT(udqp); 1706 xfs_qm_quotacheck_dqadjust(udqp, nblks, rtblks); 1707 xfs_qm_dqput(udqp); 1708 } 1709 if (XFS_IS_OQUOTA_ON(mp)) { 1710 ASSERT(gdqp); 1711 xfs_qm_quotacheck_dqadjust(gdqp, nblks, rtblks); 1712 xfs_qm_dqput(gdqp); 1713 } 1714 /* 1715 * Now release the inode. This will send it to 'inactive', and 1716 * possibly even free blocks. 1717 */ 1718 IRELE(ip); 1719 1720 /* 1721 * Goto next inode. 1722 */ 1723 *res = BULKSTAT_RV_DIDONE; 1724 return 0; 1725} 1726 1727/* 1728 * Walk thru all the filesystem inodes and construct a consistent view 1729 * of the disk quota world. If the quotacheck fails, disable quotas. 1730 */ 1731int 1732xfs_qm_quotacheck( 1733 xfs_mount_t *mp) 1734{ 1735 int done, count, error; 1736 xfs_ino_t lastino; 1737 size_t structsz; 1738 xfs_inode_t *uip, *gip; 1739 uint flags; 1740 1741 count = INT_MAX; 1742 structsz = 1; 1743 lastino = 0; 1744 flags = 0; 1745 1746 ASSERT(mp->m_quotainfo->qi_uquotaip || mp->m_quotainfo->qi_gquotaip); 1747 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 1748 1749 /* 1750 * There should be no cached dquots. The (simplistic) quotacheck 1751 * algorithm doesn't like that. 1752 */ 1753 ASSERT(list_empty(&mp->m_quotainfo->qi_dqlist)); 1754 1755 cmn_err(CE_NOTE, "XFS quotacheck %s: Please wait.", mp->m_fsname); 1756 1757 /* 1758 * First we go thru all the dquots on disk, USR and GRP/PRJ, and reset 1759 * their counters to zero. We need a clean slate. 1760 * We don't log our changes till later. 1761 */ 1762 uip = mp->m_quotainfo->qi_uquotaip; 1763 if (uip) { 1764 error = xfs_qm_dqiterate(mp, uip, XFS_QMOPT_UQUOTA); 1765 if (error) 1766 goto error_return; 1767 flags |= XFS_UQUOTA_CHKD; 1768 } 1769 1770 gip = mp->m_quotainfo->qi_gquotaip; 1771 if (gip) { 1772 error = xfs_qm_dqiterate(mp, gip, XFS_IS_GQUOTA_ON(mp) ? 1773 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA); 1774 if (error) 1775 goto error_return; 1776 flags |= XFS_OQUOTA_CHKD; 1777 } 1778 1779 do { 1780 /* 1781 * Iterate thru all the inodes in the file system, 1782 * adjusting the corresponding dquot counters in core. 1783 */ 1784 error = xfs_bulkstat(mp, &lastino, &count, 1785 xfs_qm_dqusage_adjust, 1786 structsz, NULL, &done); 1787 if (error) 1788 break; 1789 1790 } while (!done); 1791 1792 /* 1793 * We've made all the changes that we need to make incore. 1794 * Flush them down to disk buffers if everything was updated 1795 * successfully. 1796 */ 1797 if (!error) 1798 error = xfs_qm_dqflush_all(mp, 0); 1799 1800 /* 1801 * We can get this error if we couldn't do a dquot allocation inside 1802 * xfs_qm_dqusage_adjust (via bulkstat). We don't care about the 1803 * dirty dquots that might be cached, we just want to get rid of them 1804 * and turn quotaoff. The dquots won't be attached to any of the inodes 1805 * at this point (because we intentionally didn't in dqget_noattach). 1806 */ 1807 if (error) { 1808 xfs_qm_dqpurge_all(mp, XFS_QMOPT_QUOTALL); 1809 goto error_return; 1810 } 1811 1812 /* 1813 * We didn't log anything, because if we crashed, we'll have to 1814 * start the quotacheck from scratch anyway. However, we must make 1815 * sure that our dquot changes are secure before we put the 1816 * quotacheck'd stamp on the superblock. So, here we do a synchronous 1817 * flush. 1818 */ 1819 XFS_bflush(mp->m_ddev_targp); 1820 1821 /* 1822 * If one type of quotas is off, then it will lose its 1823 * quotachecked status, since we won't be doing accounting for 1824 * that type anymore. 1825 */ 1826 mp->m_qflags &= ~(XFS_OQUOTA_CHKD | XFS_UQUOTA_CHKD); 1827 mp->m_qflags |= flags; 1828 1829 xfs_qm_dquot_list_print(mp); 1830 1831 error_return: 1832 if (error) { 1833 cmn_err(CE_WARN, "XFS quotacheck %s: Unsuccessful (Error %d): " 1834 "Disabling quotas.", 1835 mp->m_fsname, error); 1836 /* 1837 * We must turn off quotas. 1838 */ 1839 ASSERT(mp->m_quotainfo != NULL); 1840 ASSERT(xfs_Gqm != NULL); 1841 xfs_qm_destroy_quotainfo(mp); 1842 if (xfs_mount_reset_sbqflags(mp)) { 1843 cmn_err(CE_WARN, "XFS quotacheck %s: " 1844 "Failed to reset quota flags.", mp->m_fsname); 1845 } 1846 } else { 1847 cmn_err(CE_NOTE, "XFS quotacheck %s: Done.", mp->m_fsname); 1848 } 1849 return (error); 1850} 1851 1852/* 1853 * This is called after the superblock has been read in and we're ready to 1854 * iget the quota inodes. 1855 */ 1856STATIC int 1857xfs_qm_init_quotainos( 1858 xfs_mount_t *mp) 1859{ 1860 xfs_inode_t *uip, *gip; 1861 int error; 1862 __int64_t sbflags; 1863 uint flags; 1864 1865 ASSERT(mp->m_quotainfo); 1866 uip = gip = NULL; 1867 sbflags = 0; 1868 flags = 0; 1869 1870 /* 1871 * Get the uquota and gquota inodes 1872 */ 1873 if (xfs_sb_version_hasquota(&mp->m_sb)) { 1874 if (XFS_IS_UQUOTA_ON(mp) && 1875 mp->m_sb.sb_uquotino != NULLFSINO) { 1876 ASSERT(mp->m_sb.sb_uquotino > 0); 1877 if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_uquotino, 1878 0, 0, &uip))) 1879 return XFS_ERROR(error); 1880 } 1881 if (XFS_IS_OQUOTA_ON(mp) && 1882 mp->m_sb.sb_gquotino != NULLFSINO) { 1883 ASSERT(mp->m_sb.sb_gquotino > 0); 1884 if ((error = xfs_iget(mp, NULL, mp->m_sb.sb_gquotino, 1885 0, 0, &gip))) { 1886 if (uip) 1887 IRELE(uip); 1888 return XFS_ERROR(error); 1889 } 1890 } 1891 } else { 1892 flags |= XFS_QMOPT_SBVERSION; 1893 sbflags |= (XFS_SB_VERSIONNUM | XFS_SB_UQUOTINO | 1894 XFS_SB_GQUOTINO | XFS_SB_QFLAGS); 1895 } 1896 1897 /* 1898 * Create the two inodes, if they don't exist already. The changes 1899 * made above will get added to a transaction and logged in one of 1900 * the qino_alloc calls below. If the device is readonly, 1901 * temporarily switch to read-write to do this. 1902 */ 1903 if (XFS_IS_UQUOTA_ON(mp) && uip == NULL) { 1904 if ((error = xfs_qm_qino_alloc(mp, &uip, 1905 sbflags | XFS_SB_UQUOTINO, 1906 flags | XFS_QMOPT_UQUOTA))) 1907 return XFS_ERROR(error); 1908 1909 flags &= ~XFS_QMOPT_SBVERSION; 1910 } 1911 if (XFS_IS_OQUOTA_ON(mp) && gip == NULL) { 1912 flags |= (XFS_IS_GQUOTA_ON(mp) ? 1913 XFS_QMOPT_GQUOTA : XFS_QMOPT_PQUOTA); 1914 error = xfs_qm_qino_alloc(mp, &gip, 1915 sbflags | XFS_SB_GQUOTINO, flags); 1916 if (error) { 1917 if (uip) 1918 IRELE(uip); 1919 1920 return XFS_ERROR(error); 1921 } 1922 } 1923 1924 mp->m_quotainfo->qi_uquotaip = uip; 1925 mp->m_quotainfo->qi_gquotaip = gip; 1926 1927 return 0; 1928} 1929 1930 1931 1932/* 1933 * Just pop the least recently used dquot off the freelist and 1934 * recycle it. The returned dquot is locked. 1935 */ 1936STATIC xfs_dquot_t * 1937xfs_qm_dqreclaim_one(void) 1938{ 1939 xfs_dquot_t *dqpout; 1940 xfs_dquot_t *dqp; 1941 int restarts; 1942 1943 restarts = 0; 1944 dqpout = NULL; 1945 1946 /* lockorder: hashchainlock, freelistlock, mplistlock, dqlock, dqflock */ 1947startagain: 1948 mutex_lock(&xfs_Gqm->qm_dqfrlist_lock); 1949 1950 list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) { 1951 struct xfs_mount *mp = dqp->q_mount; 1952 xfs_dqlock(dqp); 1953 1954 /* 1955 * We are racing with dqlookup here. Naturally we don't 1956 * want to reclaim a dquot that lookup wants. We release the 1957 * freelist lock and start over, so that lookup will grab 1958 * both the dquot and the freelistlock. 1959 */ 1960 if (dqp->dq_flags & XFS_DQ_WANT) { 1961 ASSERT(! (dqp->dq_flags & XFS_DQ_INACTIVE)); 1962 1963 trace_xfs_dqreclaim_want(dqp); 1964 1965 xfs_dqunlock(dqp); 1966 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock); 1967 if (++restarts >= XFS_QM_RECLAIM_MAX_RESTARTS) 1968 return NULL; 1969 XQM_STATS_INC(xqmstats.xs_qm_dqwants); 1970 goto startagain; 1971 } 1972 1973 /* 1974 * If the dquot is inactive, we are assured that it is 1975 * not on the mplist or the hashlist, and that makes our 1976 * life easier. 1977 */ 1978 if (dqp->dq_flags & XFS_DQ_INACTIVE) { 1979 ASSERT(mp == NULL); 1980 ASSERT(! XFS_DQ_IS_DIRTY(dqp)); 1981 ASSERT(list_empty(&dqp->q_hashlist)); 1982 ASSERT(list_empty(&dqp->q_mplist)); 1983 list_del_init(&dqp->q_freelist); 1984 xfs_Gqm->qm_dqfrlist_cnt--; 1985 xfs_dqunlock(dqp); 1986 dqpout = dqp; 1987 XQM_STATS_INC(xqmstats.xs_qm_dqinact_reclaims); 1988 break; 1989 } 1990 1991 ASSERT(dqp->q_hash); 1992 ASSERT(!list_empty(&dqp->q_mplist)); 1993 1994 /* 1995 * Try to grab the flush lock. If this dquot is in the process of 1996 * getting flushed to disk, we don't want to reclaim it. 1997 */ 1998 if (!xfs_dqflock_nowait(dqp)) { 1999 xfs_dqunlock(dqp); 2000 continue; 2001 } 2002 2003 /* 2004 * We have the flush lock so we know that this is not in the 2005 * process of being flushed. So, if this is dirty, flush it 2006 * DELWRI so that we don't get a freelist infested with 2007 * dirty dquots. 2008 */ 2009 if (XFS_DQ_IS_DIRTY(dqp)) { 2010 int error; 2011 2012 trace_xfs_dqreclaim_dirty(dqp); 2013 2014 /* 2015 * We flush it delayed write, so don't bother 2016 * releasing the freelist lock. 2017 */ 2018 error = xfs_qm_dqflush(dqp, 0); 2019 if (error) { 2020 xfs_fs_cmn_err(CE_WARN, mp, 2021 "xfs_qm_dqreclaim: dquot %p flush failed", dqp); 2022 } 2023 xfs_dqunlock(dqp); /* dqflush unlocks dqflock */ 2024 continue; 2025 } 2026 2027 /* 2028 * We're trying to get the hashlock out of order. This races 2029 * with dqlookup; so, we giveup and goto the next dquot if 2030 * we couldn't get the hashlock. This way, we won't starve 2031 * a dqlookup process that holds the hashlock that is 2032 * waiting for the freelist lock. 2033 */ 2034 if (!mutex_trylock(&dqp->q_hash->qh_lock)) { 2035 restarts++; 2036 goto dqfunlock; 2037 } 2038 2039 /* 2040 * This races with dquot allocation code as well as dqflush_all 2041 * and reclaim code. So, if we failed to grab the mplist lock, 2042 * giveup everything and start over. 2043 */ 2044 if (!mutex_trylock(&mp->m_quotainfo->qi_dqlist_lock)) { 2045 restarts++; 2046 mutex_unlock(&dqp->q_hash->qh_lock); 2047 xfs_dqfunlock(dqp); 2048 xfs_dqunlock(dqp); 2049 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock); 2050 if (restarts++ >= XFS_QM_RECLAIM_MAX_RESTARTS) 2051 return NULL; 2052 goto startagain; 2053 } 2054 2055 ASSERT(dqp->q_nrefs == 0); 2056 list_del_init(&dqp->q_mplist); 2057 mp->m_quotainfo->qi_dquots--; 2058 mp->m_quotainfo->qi_dqreclaims++; 2059 list_del_init(&dqp->q_hashlist); 2060 dqp->q_hash->qh_version++; 2061 list_del_init(&dqp->q_freelist); 2062 xfs_Gqm->qm_dqfrlist_cnt--; 2063 dqpout = dqp; 2064 mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock); 2065 mutex_unlock(&dqp->q_hash->qh_lock); 2066dqfunlock: 2067 xfs_dqfunlock(dqp); 2068 xfs_dqunlock(dqp); 2069 if (dqpout) 2070 break; 2071 if (restarts >= XFS_QM_RECLAIM_MAX_RESTARTS) 2072 return NULL; 2073 } 2074 mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock); 2075 return dqpout; 2076} 2077 2078/* 2079 * Traverse the freelist of dquots and attempt to reclaim a maximum of 2080 * 'howmany' dquots. This operation races with dqlookup(), and attempts to 2081 * favor the lookup function ... 2082 */ 2083STATIC int 2084xfs_qm_shake_freelist( 2085 int howmany) 2086{ 2087 int nreclaimed = 0; 2088 xfs_dquot_t *dqp; 2089 2090 if (howmany <= 0) 2091 return 0; 2092 2093 while (nreclaimed < howmany) { 2094 dqp = xfs_qm_dqreclaim_one(); 2095 if (!dqp) 2096 return nreclaimed; 2097 xfs_qm_dqdestroy(dqp); 2098 nreclaimed++; 2099 } 2100 return nreclaimed; 2101} 2102 2103/* 2104 * The kmem_shake interface is invoked when memory is running low. 2105 */ 2106/* ARGSUSED */ 2107STATIC int 2108xfs_qm_shake( 2109 struct shrinker *shrink, 2110 int nr_to_scan, 2111 gfp_t gfp_mask) 2112{ 2113 int ndqused, nfree, n; 2114 2115 if (!kmem_shake_allow(gfp_mask)) 2116 return 0; 2117 if (!xfs_Gqm) 2118 return 0; 2119 2120 nfree = xfs_Gqm->qm_dqfrlist_cnt; /* free dquots */ 2121 /* incore dquots in all f/s's */ 2122 ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree; 2123 2124 ASSERT(ndqused >= 0); 2125 2126 if (nfree <= ndqused && nfree < ndquot) 2127 return 0; 2128 2129 ndqused *= xfs_Gqm->qm_dqfree_ratio; /* target # of free dquots */ 2130 n = nfree - ndqused - ndquot; /* # over target */ 2131 2132 return xfs_qm_shake_freelist(MAX(nfree, n)); 2133} 2134 2135 2136/*------------------------------------------------------------------*/ 2137 2138/* 2139 * Return a new incore dquot. Depending on the number of 2140 * dquots in the system, we either allocate a new one on the kernel heap, 2141 * or reclaim a free one. 2142 * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed 2143 * to reclaim an existing one from the freelist. 2144 */ 2145boolean_t 2146xfs_qm_dqalloc_incore( 2147 xfs_dquot_t **O_dqpp) 2148{ 2149 xfs_dquot_t *dqp; 2150 2151 /* 2152 * Check against high water mark to see if we want to pop 2153 * a nincompoop dquot off the freelist. 2154 */ 2155 if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) { 2156 /* 2157 * Try to recycle a dquot from the freelist. 2158 */ 2159 if ((dqp = xfs_qm_dqreclaim_one())) { 2160 XQM_STATS_INC(xqmstats.xs_qm_dqreclaims); 2161 memset(&dqp->q_core, 0, sizeof(dqp->q_core)); 2162 *O_dqpp = dqp; 2163 return B_FALSE; 2164 } 2165 XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses); 2166 } 2167 2168 /* 2169 * Allocate a brand new dquot on the kernel heap and return it 2170 * to the caller to initialize. 2171 */ 2172 ASSERT(xfs_Gqm->qm_dqzone != NULL); 2173 *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP); 2174 atomic_inc(&xfs_Gqm->qm_totaldquots); 2175 2176 return B_TRUE; 2177} 2178 2179 2180/* 2181 * Start a transaction and write the incore superblock changes to 2182 * disk. flags parameter indicates which fields have changed. 2183 */ 2184int 2185xfs_qm_write_sb_changes( 2186 xfs_mount_t *mp, 2187 __int64_t flags) 2188{ 2189 xfs_trans_t *tp; 2190 int error; 2191 2192#ifdef QUOTADEBUG 2193 cmn_err(CE_NOTE, "Writing superblock quota changes :%s", mp->m_fsname); 2194#endif 2195 tp = xfs_trans_alloc(mp, XFS_TRANS_QM_SBCHANGE); 2196 if ((error = xfs_trans_reserve(tp, 0, 2197 mp->m_sb.sb_sectsize + 128, 0, 2198 0, 2199 XFS_DEFAULT_LOG_COUNT))) { 2200 xfs_trans_cancel(tp, 0); 2201 return error; 2202 } 2203 2204 xfs_mod_sb(tp, flags); 2205 error = xfs_trans_commit(tp, 0); 2206 2207 return error; 2208} 2209 2210 2211/* --------------- utility functions for vnodeops ---------------- */ 2212 2213 2214/* 2215 * Given an inode, a uid and gid (from cred_t) make sure that we have 2216 * allocated relevant dquot(s) on disk, and that we won't exceed inode 2217 * quotas by creating this file. 2218 * This also attaches dquot(s) to the given inode after locking it, 2219 * and returns the dquots corresponding to the uid and/or gid. 2220 * 2221 * in : inode (unlocked) 2222 * out : udquot, gdquot with references taken and unlocked 2223 */ 2224int 2225xfs_qm_vop_dqalloc( 2226 struct xfs_inode *ip, 2227 uid_t uid, 2228 gid_t gid, 2229 prid_t prid, 2230 uint flags, 2231 struct xfs_dquot **O_udqpp, 2232 struct xfs_dquot **O_gdqpp) 2233{ 2234 struct xfs_mount *mp = ip->i_mount; 2235 struct xfs_dquot *uq, *gq; 2236 int error; 2237 uint lockflags; 2238 2239 if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp)) 2240 return 0; 2241 2242 lockflags = XFS_ILOCK_EXCL; 2243 xfs_ilock(ip, lockflags); 2244 2245 if ((flags & XFS_QMOPT_INHERIT) && XFS_INHERIT_GID(ip)) 2246 gid = ip->i_d.di_gid; 2247 2248 /* 2249 * Attach the dquot(s) to this inode, doing a dquot allocation 2250 * if necessary. The dquot(s) will not be locked. 2251 */ 2252 if (XFS_NOT_DQATTACHED(mp, ip)) { 2253 error = xfs_qm_dqattach_locked(ip, XFS_QMOPT_DQALLOC); 2254 if (error) { 2255 xfs_iunlock(ip, lockflags); 2256 return error; 2257 } 2258 } 2259 2260 uq = gq = NULL; 2261 if ((flags & XFS_QMOPT_UQUOTA) && XFS_IS_UQUOTA_ON(mp)) { 2262 if (ip->i_d.di_uid != uid) { 2263 /* 2264 * What we need is the dquot that has this uid, and 2265 * if we send the inode to dqget, the uid of the inode 2266 * takes priority over what's sent in the uid argument. 2267 * We must unlock inode here before calling dqget if 2268 * we're not sending the inode, because otherwise 2269 * we'll deadlock by doing trans_reserve while 2270 * holding ilock. 2271 */ 2272 xfs_iunlock(ip, lockflags); 2273 if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t) uid, 2274 XFS_DQ_USER, 2275 XFS_QMOPT_DQALLOC | 2276 XFS_QMOPT_DOWARN, 2277 &uq))) { 2278 ASSERT(error != ENOENT); 2279 return error; 2280 } 2281 /* 2282 * Get the ilock in the right order. 2283 */ 2284 xfs_dqunlock(uq); 2285 lockflags = XFS_ILOCK_SHARED; 2286 xfs_ilock(ip, lockflags); 2287 } else { 2288 /* 2289 * Take an extra reference, because we'll return 2290 * this to caller 2291 */ 2292 ASSERT(ip->i_udquot); 2293 uq = ip->i_udquot; 2294 xfs_dqlock(uq); 2295 XFS_DQHOLD(uq); 2296 xfs_dqunlock(uq); 2297 } 2298 } 2299 if ((flags & XFS_QMOPT_GQUOTA) && XFS_IS_GQUOTA_ON(mp)) { 2300 if (ip->i_d.di_gid != gid) { 2301 xfs_iunlock(ip, lockflags); 2302 if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)gid, 2303 XFS_DQ_GROUP, 2304 XFS_QMOPT_DQALLOC | 2305 XFS_QMOPT_DOWARN, 2306 &gq))) { 2307 if (uq) 2308 xfs_qm_dqrele(uq); 2309 ASSERT(error != ENOENT); 2310 return error; 2311 } 2312 xfs_dqunlock(gq); 2313 lockflags = XFS_ILOCK_SHARED; 2314 xfs_ilock(ip, lockflags); 2315 } else { 2316 ASSERT(ip->i_gdquot); 2317 gq = ip->i_gdquot; 2318 xfs_dqlock(gq); 2319 XFS_DQHOLD(gq); 2320 xfs_dqunlock(gq); 2321 } 2322 } else if ((flags & XFS_QMOPT_PQUOTA) && XFS_IS_PQUOTA_ON(mp)) { 2323 if (ip->i_d.di_projid != prid) { 2324 xfs_iunlock(ip, lockflags); 2325 if ((error = xfs_qm_dqget(mp, NULL, (xfs_dqid_t)prid, 2326 XFS_DQ_PROJ, 2327 XFS_QMOPT_DQALLOC | 2328 XFS_QMOPT_DOWARN, 2329 &gq))) { 2330 if (uq) 2331 xfs_qm_dqrele(uq); 2332 ASSERT(error != ENOENT); 2333 return (error); 2334 } 2335 xfs_dqunlock(gq); 2336 lockflags = XFS_ILOCK_SHARED; 2337 xfs_ilock(ip, lockflags); 2338 } else { 2339 ASSERT(ip->i_gdquot); 2340 gq = ip->i_gdquot; 2341 xfs_dqlock(gq); 2342 XFS_DQHOLD(gq); 2343 xfs_dqunlock(gq); 2344 } 2345 } 2346 if (uq) 2347 trace_xfs_dquot_dqalloc(ip); 2348 2349 xfs_iunlock(ip, lockflags); 2350 if (O_udqpp) 2351 *O_udqpp = uq; 2352 else if (uq) 2353 xfs_qm_dqrele(uq); 2354 if (O_gdqpp) 2355 *O_gdqpp = gq; 2356 else if (gq) 2357 xfs_qm_dqrele(gq); 2358 return 0; 2359} 2360 2361/* 2362 * Actually transfer ownership, and do dquot modifications. 2363 * These were already reserved. 2364 */ 2365xfs_dquot_t * 2366xfs_qm_vop_chown( 2367 xfs_trans_t *tp, 2368 xfs_inode_t *ip, 2369 xfs_dquot_t **IO_olddq, 2370 xfs_dquot_t *newdq) 2371{ 2372 xfs_dquot_t *prevdq; 2373 uint bfield = XFS_IS_REALTIME_INODE(ip) ? 2374 XFS_TRANS_DQ_RTBCOUNT : XFS_TRANS_DQ_BCOUNT; 2375 2376 2377 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 2378 ASSERT(XFS_IS_QUOTA_RUNNING(ip->i_mount)); 2379 2380 /* old dquot */ 2381 prevdq = *IO_olddq; 2382 ASSERT(prevdq); 2383 ASSERT(prevdq != newdq); 2384 2385 xfs_trans_mod_dquot(tp, prevdq, bfield, -(ip->i_d.di_nblocks)); 2386 xfs_trans_mod_dquot(tp, prevdq, XFS_TRANS_DQ_ICOUNT, -1); 2387 2388 /* the sparkling new dquot */ 2389 xfs_trans_mod_dquot(tp, newdq, bfield, ip->i_d.di_nblocks); 2390 xfs_trans_mod_dquot(tp, newdq, XFS_TRANS_DQ_ICOUNT, 1); 2391 2392 /* 2393 * Take an extra reference, because the inode 2394 * is going to keep this dquot pointer even 2395 * after the trans_commit. 2396 */ 2397 xfs_dqlock(newdq); 2398 XFS_DQHOLD(newdq); 2399 xfs_dqunlock(newdq); 2400 *IO_olddq = newdq; 2401 2402 return prevdq; 2403} 2404 2405/* 2406 * Quota reservations for setattr(AT_UID|AT_GID|AT_PROJID). 2407 */ 2408int 2409xfs_qm_vop_chown_reserve( 2410 xfs_trans_t *tp, 2411 xfs_inode_t *ip, 2412 xfs_dquot_t *udqp, 2413 xfs_dquot_t *gdqp, 2414 uint flags) 2415{ 2416 xfs_mount_t *mp = ip->i_mount; 2417 uint delblks, blkflags, prjflags = 0; 2418 xfs_dquot_t *unresudq, *unresgdq, *delblksudq, *delblksgdq; 2419 int error; 2420 2421 2422 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)); 2423 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 2424 2425 delblks = ip->i_delayed_blks; 2426 delblksudq = delblksgdq = unresudq = unresgdq = NULL; 2427 blkflags = XFS_IS_REALTIME_INODE(ip) ? 2428 XFS_QMOPT_RES_RTBLKS : XFS_QMOPT_RES_REGBLKS; 2429 2430 if (XFS_IS_UQUOTA_ON(mp) && udqp && 2431 ip->i_d.di_uid != (uid_t)be32_to_cpu(udqp->q_core.d_id)) { 2432 delblksudq = udqp; 2433 /* 2434 * If there are delayed allocation blocks, then we have to 2435 * unreserve those from the old dquot, and add them to the 2436 * new dquot. 2437 */ 2438 if (delblks) { 2439 ASSERT(ip->i_udquot); 2440 unresudq = ip->i_udquot; 2441 } 2442 } 2443 if (XFS_IS_OQUOTA_ON(ip->i_mount) && gdqp) { 2444 if (XFS_IS_PQUOTA_ON(ip->i_mount) && 2445 ip->i_d.di_projid != be32_to_cpu(gdqp->q_core.d_id)) 2446 prjflags = XFS_QMOPT_ENOSPC; 2447 2448 if (prjflags || 2449 (XFS_IS_GQUOTA_ON(ip->i_mount) && 2450 ip->i_d.di_gid != be32_to_cpu(gdqp->q_core.d_id))) { 2451 delblksgdq = gdqp; 2452 if (delblks) { 2453 ASSERT(ip->i_gdquot); 2454 unresgdq = ip->i_gdquot; 2455 } 2456 } 2457 } 2458 2459 if ((error = xfs_trans_reserve_quota_bydquots(tp, ip->i_mount, 2460 delblksudq, delblksgdq, ip->i_d.di_nblocks, 1, 2461 flags | blkflags | prjflags))) 2462 return (error); 2463 2464 /* 2465 * Do the delayed blks reservations/unreservations now. Since, these 2466 * are done without the help of a transaction, if a reservation fails 2467 * its previous reservations won't be automatically undone by trans 2468 * code. So, we have to do it manually here. 2469 */ 2470 if (delblks) { 2471 /* 2472 * Do the reservations first. Unreservation can't fail. 2473 */ 2474 ASSERT(delblksudq || delblksgdq); 2475 ASSERT(unresudq || unresgdq); 2476 if ((error = xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount, 2477 delblksudq, delblksgdq, (xfs_qcnt_t)delblks, 0, 2478 flags | blkflags | prjflags))) 2479 return (error); 2480 xfs_trans_reserve_quota_bydquots(NULL, ip->i_mount, 2481 unresudq, unresgdq, -((xfs_qcnt_t)delblks), 0, 2482 blkflags); 2483 } 2484 2485 return (0); 2486} 2487 2488int 2489xfs_qm_vop_rename_dqattach( 2490 struct xfs_inode **i_tab) 2491{ 2492 struct xfs_mount *mp = i_tab[0]->i_mount; 2493 int i; 2494 2495 if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp)) 2496 return 0; 2497 2498 for (i = 0; (i < 4 && i_tab[i]); i++) { 2499 struct xfs_inode *ip = i_tab[i]; 2500 int error; 2501 2502 /* 2503 * Watch out for duplicate entries in the table. 2504 */ 2505 if (i == 0 || ip != i_tab[i-1]) { 2506 if (XFS_NOT_DQATTACHED(mp, ip)) { 2507 error = xfs_qm_dqattach(ip, 0); 2508 if (error) 2509 return error; 2510 } 2511 } 2512 } 2513 return 0; 2514} 2515 2516void 2517xfs_qm_vop_create_dqattach( 2518 struct xfs_trans *tp, 2519 struct xfs_inode *ip, 2520 struct xfs_dquot *udqp, 2521 struct xfs_dquot *gdqp) 2522{ 2523 struct xfs_mount *mp = tp->t_mountp; 2524 2525 if (!XFS_IS_QUOTA_RUNNING(mp) || !XFS_IS_QUOTA_ON(mp)) 2526 return; 2527 2528 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL)); 2529 ASSERT(XFS_IS_QUOTA_RUNNING(mp)); 2530 2531 if (udqp) { 2532 xfs_dqlock(udqp); 2533 XFS_DQHOLD(udqp); 2534 xfs_dqunlock(udqp); 2535 ASSERT(ip->i_udquot == NULL); 2536 ip->i_udquot = udqp; 2537 ASSERT(XFS_IS_UQUOTA_ON(mp)); 2538 ASSERT(ip->i_d.di_uid == be32_to_cpu(udqp->q_core.d_id)); 2539 xfs_trans_mod_dquot(tp, udqp, XFS_TRANS_DQ_ICOUNT, 1); 2540 } 2541 if (gdqp) { 2542 xfs_dqlock(gdqp); 2543 XFS_DQHOLD(gdqp); 2544 xfs_dqunlock(gdqp); 2545 ASSERT(ip->i_gdquot == NULL); 2546 ip->i_gdquot = gdqp; 2547 ASSERT(XFS_IS_OQUOTA_ON(mp)); 2548 ASSERT((XFS_IS_GQUOTA_ON(mp) ? 2549 ip->i_d.di_gid : ip->i_d.di_projid) == 2550 be32_to_cpu(gdqp->q_core.d_id)); 2551 xfs_trans_mod_dquot(tp, gdqp, XFS_TRANS_DQ_ICOUNT, 1); 2552 } 2553} 2554