fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include <linux/sort.h>
  34 #include "internal.h"
  35 #include "pnfs.h"
  36 #include "iostat.h"
  37 #include "nfs4trace.h"
  38 #include "delegation.h"
  39 #include "nfs42.h"
  40 #include "nfs4_fs.h"
  41
  42 #define NFSDBG_FACILITY         NFSDBG_PNFS
  43 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  44
  45 /* Locking:
  46  *
  47  * pnfs_spinlock:
  48  *      protects pnfs_modules_tbl.
  49  */
  50 static DEFINE_SPINLOCK(pnfs_spinlock);
  51
  52 /*
  53  * pnfs_modules_tbl holds all pnfs modules
  54  */
  55 static LIST_HEAD(pnfs_modules_tbl);
  56
  57 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  58 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  59                 struct list_head *free_me,
  60                 const struct pnfs_layout_range *range,
  61                 u32 seq);
  62 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  63                                 struct list_head *tmp_list);
  64
  65 /* Return the registered pnfs layout driver module matching given id */
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver_locked(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  72                 if (local->id == id)
  73                         goto out;
  74         local = NULL;
  75 out:
  76         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  77         return local;
  78 }
  79
  80 static struct pnfs_layoutdriver_type *
  81 find_pnfs_driver(u32 id)
  82 {
  83         struct pnfs_layoutdriver_type *local;
  84
  85         spin_lock(&pnfs_spinlock);
  86         local = find_pnfs_driver_locked(id);
  87         if (local != NULL && !try_module_get(local->owner)) {
  88                 dprintk("%s: Could not grab reference on module\n", __func__);
  89                 local = NULL;
  90         }
  91         spin_unlock(&pnfs_spinlock);
  92         return local;
  93 }
  94
  95 void
  96 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  97 {
  98         if (nfss->pnfs_curr_ld) {
  99                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
 100                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 101                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 102                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 103                         nfs4_deviceid_purge_client(nfss->nfs_client);
 104                 module_put(nfss->pnfs_curr_ld->owner);
 105         }
 106         nfss->pnfs_curr_ld = NULL;
 107 }
 108
 109 /*
 110  * When the server sends a list of layout types, we choose one in the order
 111  * given in the list below.
 112  *
 113  * FIXME: should this list be configurable in some fashion? module param?
 114  *        mount option? something else?
 115  */
 116 static const u32 ld_prefs[] = {
 117         LAYOUT_SCSI,
 118         LAYOUT_BLOCK_VOLUME,
 119         LAYOUT_OSD2_OBJECTS,
 120         LAYOUT_FLEX_FILES,
 121         LAYOUT_NFSV4_1_FILES,
 122         0
 123 };
 124
 125 static int
 126 ld_cmp(const void *e1, const void *e2)
 127 {
 128         u32 ld1 = *((u32 *)e1);
 129         u32 ld2 = *((u32 *)e2);
 130         int i;
 131
 132         for (i = 0; ld_prefs[i] != 0; i++) {
 133                 if (ld1 == ld_prefs[i])
 134                         return -1;
 135
 136                 if (ld2 == ld_prefs[i])
 137                         return 1;
 138         }
 139         return 0;
 140 }
 141
 142 /*
 143  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 144  * Currently only one pNFS layout driver per filesystem is supported.
 145  *
 146  * @ids array of layout types supported by MDS.
 147  */
 148 void
 149 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 150                       struct nfs_fsinfo *fsinfo)
 151 {
 152         struct pnfs_layoutdriver_type *ld_type = NULL;
 153         u32 id;
 154         int i;
 155
 156         if (fsinfo->nlayouttypes == 0)
 157                 goto out_no_driver;
 158         if (!(server->nfs_client->cl_exchange_flags &
 159                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 160                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 161                         __func__, server->nfs_client->cl_exchange_flags);
 162                 goto out_no_driver;
 163         }
 164
 165         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 166                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 167
 168         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 169                 id = fsinfo->layouttype[i];
 170                 ld_type = find_pnfs_driver(id);
 171                 if (!ld_type) {
 172                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 173                                         id);
 174                         ld_type = find_pnfs_driver(id);
 175                 }
 176                 if (ld_type)
 177                         break;
 178         }
 179
 180         if (!ld_type) {
 181                 dprintk("%s: No pNFS module found!\n", __func__);
 182                 goto out_no_driver;
 183         }
 184
 185         server->pnfs_curr_ld = ld_type;
 186         if (ld_type->set_layoutdriver
 187             && ld_type->set_layoutdriver(server, mntfh)) {
 188                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 189                         "driver %u.\n", __func__, id);
 190                 module_put(ld_type->owner);
 191                 goto out_no_driver;
 192         }
 193         /* Bump the MDS count */
 194         atomic_inc(&server->nfs_client->cl_mds_count);
 195
 196         dprintk("%s: pNFS module for %u set\n", __func__, id);
 197         return;
 198
 199 out_no_driver:
 200         dprintk("%s: Using NFSv4 I/O\n", __func__);
 201         server->pnfs_curr_ld = NULL;
 202 }
 203
 204 int
 205 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 206 {
 207         int status = -EINVAL;
 208         struct pnfs_layoutdriver_type *tmp;
 209
 210         if (ld_type->id == 0) {
 211                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 212                 return status;
 213         }
 214         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 215                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 216                        "alloc_lseg and free_lseg.\n", __func__);
 217                 return status;
 218         }
 219
 220         spin_lock(&pnfs_spinlock);
 221         tmp = find_pnfs_driver_locked(ld_type->id);
 222         if (!tmp) {
 223                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 224                 status = 0;
 225                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 226                         ld_type->name);
 227         } else {
 228                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 229                         __func__, ld_type->id);
 230         }
 231         spin_unlock(&pnfs_spinlock);
 232
 233         return status;
 234 }
 235 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 236
 237 void
 238 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 239 {
 240         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 241         spin_lock(&pnfs_spinlock);
 242         list_del(&ld_type->pnfs_tblid);
 243         spin_unlock(&pnfs_spinlock);
 244 }
 245 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 246
 247 /*
 248  * pNFS client layout cache
 249  */
 250
 251 /* Need to hold i_lock if caller does not already hold reference */
 252 void
 253 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 254 {
 255         refcount_inc(&lo->plh_refcount);
 256 }
 257
 258 static struct pnfs_layout_hdr *
 259 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 260 {
 261         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 262         return ld->alloc_layout_hdr(ino, gfp_flags);
 263 }
 264
 265 static void
 266 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 267 {
 268         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 269         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 270
 271         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
 272                 struct nfs_client *clp = server->nfs_client;
 273
 274                 spin_lock(&clp->cl_lock);
 275                 list_del_rcu(&lo->plh_layouts);
 276                 spin_unlock(&clp->cl_lock);
 277         }
 278         put_cred(lo->plh_lc_cred);
 279         return ld->free_layout_hdr(lo);
 280 }
 281
 282 static void
 283 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 284 {
 285         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 286         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 287         nfsi->layout = NULL;
 288         /* Reset MDS Threshold I/O counters */
 289         nfsi->write_io = 0;
 290         nfsi->read_io = 0;
 291 }
 292
 293 void
 294 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 295 {
 296         struct inode *inode;
 297         unsigned long i_state;
 298
 299         if (!lo)
 300                 return;
 301         inode = lo->plh_inode;
 302         pnfs_layoutreturn_before_put_layout_hdr(lo);
 303
 304         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 305                 if (!list_empty(&lo->plh_segs))
 306                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 307                 pnfs_detach_layout_hdr(lo);
 308                 i_state = inode->i_state;
 309                 spin_unlock(&inode->i_lock);
 310                 pnfs_free_layout_hdr(lo);
 311                 /* Notify pnfs_destroy_layout_final() that we're done */
 312                 if (i_state & (I_FREEING | I_CLEAR))
 313                         wake_up_var(lo);
 314         }
 315 }
 316
 317 static struct inode *
 318 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
 319 {
 320         struct inode *inode = igrab(lo->plh_inode);
 321         if (inode)
 322                 return inode;
 323         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
 324         return NULL;
 325 }
 326
 327 /*
 328  * Compare 2 layout stateid sequence ids, to see which is newer,
 329  * taking into account wraparound issues.
 330  */
 331 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 332 {
 333         return (s32)(s1 - s2) > 0;
 334 }
 335
 336 static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
 337 {
 338         if (pnfs_seqid_is_newer(newseq, lo->plh_barrier))
 339                 lo->plh_barrier = newseq;
 340 }
 341
 342 static void
 343 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 344                          u32 seq)
 345 {
 346         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 347                 iomode = IOMODE_ANY;
 348         lo->plh_return_iomode = iomode;
 349         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 350         if (seq != 0) {
 351                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 352                 lo->plh_return_seq = seq;
 353                 pnfs_barrier_update(lo, seq);
 354         }
 355 }
 356
 357 static void
 358 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 359 {
 360         struct pnfs_layout_segment *lseg;
 361         lo->plh_return_iomode = 0;
 362         lo->plh_return_seq = 0;
 363         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 364         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 365                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 366                         continue;
 367                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 368         }
 369 }
 370
 371 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 372 {
 373         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 374         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 375         smp_mb__after_atomic();
 376         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 377         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 378 }
 379
 380 static void
 381 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 382                 struct list_head *free_me)
 383 {
 384         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 385         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 386         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 387                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 388         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 389                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 390 }
 391
 392 /*
 393  * Update the seqid of a layout stateid after receiving
 394  * NFS4ERR_OLD_STATEID
 395  */
 396 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
 397                 struct pnfs_layout_range *dst_range,
 398                 struct inode *inode)
 399 {
 400         struct pnfs_layout_hdr *lo;
 401         struct pnfs_layout_range range = {
 402                 .iomode = IOMODE_ANY,
 403                 .offset = 0,
 404                 .length = NFS4_MAX_UINT64,
 405         };
 406         bool ret = false;
 407         LIST_HEAD(head);
 408         int err;
 409
 410         spin_lock(&inode->i_lock);
 411         lo = NFS_I(inode)->layout;
 412         if (lo &&  pnfs_layout_is_valid(lo) &&
 413             nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 414                 /* Is our call using the most recent seqid? If so, bump it */
 415                 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
 416                         nfs4_stateid_seqid_inc(dst);
 417                         ret = true;
 418                         goto out;
 419                 }
 420                 /* Try to update the seqid to the most recent */
 421                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 422                 if (err != -EBUSY) {
 423                         dst->seqid = lo->plh_stateid.seqid;
 424                         *dst_range = range;
 425                         ret = true;
 426                 }
 427         }
 428 out:
 429         spin_unlock(&inode->i_lock);
 430         pnfs_free_lseg_list(&head);
 431         return ret;
 432 }
 433
 434 /*
 435  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 436  *
 437  * In order to continue using the pnfs_layout_hdr, a full recovery
 438  * is required.
 439  * Note that caller must hold inode->i_lock.
 440  */
 441 int
 442 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 443                 struct list_head *lseg_list)
 444 {
 445         struct pnfs_layout_range range = {
 446                 .iomode = IOMODE_ANY,
 447                 .offset = 0,
 448                 .length = NFS4_MAX_UINT64,
 449         };
 450         struct pnfs_layout_segment *lseg, *next;
 451
 452         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 453         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 454                 pnfs_clear_lseg_state(lseg, lseg_list);
 455         pnfs_clear_layoutreturn_info(lo);
 456         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 457         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 458             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 459                 pnfs_clear_layoutreturn_waitbit(lo);
 460         return !list_empty(&lo->plh_segs);
 461 }
 462
 463 static int
 464 pnfs_iomode_to_fail_bit(u32 iomode)
 465 {
 466         return iomode == IOMODE_RW ?
 467                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 468 }
 469
 470 static void
 471 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 472 {
 473         lo->plh_retry_timestamp = jiffies;
 474         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 475                 refcount_inc(&lo->plh_refcount);
 476 }
 477
 478 static void
 479 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 480 {
 481         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 482                 refcount_dec(&lo->plh_refcount);
 483 }
 484
 485 static void
 486 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 487 {
 488         struct inode *inode = lo->plh_inode;
 489         struct pnfs_layout_range range = {
 490                 .iomode = iomode,
 491                 .offset = 0,
 492                 .length = NFS4_MAX_UINT64,
 493         };
 494         LIST_HEAD(head);
 495
 496         spin_lock(&inode->i_lock);
 497         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 498         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 499         spin_unlock(&inode->i_lock);
 500         pnfs_free_lseg_list(&head);
 501         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 502                         iomode == IOMODE_RW ?  "RW" : "READ");
 503 }
 504
 505 static bool
 506 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 507 {
 508         unsigned long start, end;
 509         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 510
 511         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 512                 return false;
 513         end = jiffies;
 514         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 515         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 516                 /* It is time to retry the failed layoutgets */
 517                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 518                 return false;
 519         }
 520         return true;
 521 }
 522
 523 static void
 524 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 525                 const struct pnfs_layout_range *range,
 526                 const nfs4_stateid *stateid)
 527 {
 528         INIT_LIST_HEAD(&lseg->pls_list);
 529         INIT_LIST_HEAD(&lseg->pls_lc_list);
 530         INIT_LIST_HEAD(&lseg->pls_commits);
 531         refcount_set(&lseg->pls_refcount, 1);
 532         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 533         lseg->pls_layout = lo;
 534         lseg->pls_range = *range;
 535         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 536 }
 537
 538 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 539 {
 540         if (lseg != NULL) {
 541                 struct inode *inode = lseg->pls_layout->plh_inode;
 542                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 543         }
 544 }
 545
 546 static void
 547 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 548                 struct pnfs_layout_segment *lseg)
 549 {
 550         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 551         list_del_init(&lseg->pls_list);
 552         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 553         refcount_dec(&lo->plh_refcount);
 554         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 555                 return;
 556         if (list_empty(&lo->plh_segs) &&
 557             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 558             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 559                 if (atomic_read(&lo->plh_outstanding) == 0)
 560                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 561                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 562         }
 563 }
 564
 565 static bool
 566 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 567                 struct pnfs_layout_segment *lseg)
 568 {
 569         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 570             pnfs_layout_is_valid(lo)) {
 571                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 572                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 573                 return true;
 574         }
 575         return false;
 576 }
 577
 578 void
 579 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 580 {
 581         struct pnfs_layout_hdr *lo;
 582         struct inode *inode;
 583
 584         if (!lseg)
 585                 return;
 586
 587         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 588                 refcount_read(&lseg->pls_refcount),
 589                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 590
 591         lo = lseg->pls_layout;
 592         inode = lo->plh_inode;
 593
 594         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 595                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 596                         spin_unlock(&inode->i_lock);
 597                         return;
 598                 }
 599                 pnfs_get_layout_hdr(lo);
 600                 pnfs_layout_remove_lseg(lo, lseg);
 601                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 602                         lseg = NULL;
 603                 spin_unlock(&inode->i_lock);
 604                 pnfs_free_lseg(lseg);
 605                 pnfs_put_layout_hdr(lo);
 606         }
 607 }
 608 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 609
 610 /*
 611  * is l2 fully contained in l1?
 612  *   start1                             end1
 613  *   [----------------------------------)
 614  *           start2           end2
 615  *           [----------------)
 616  */
 617 static bool
 618 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 619                  const struct pnfs_layout_range *l2)
 620 {
 621         u64 start1 = l1->offset;
 622         u64 end1 = pnfs_end_offset(start1, l1->length);
 623         u64 start2 = l2->offset;
 624         u64 end2 = pnfs_end_offset(start2, l2->length);
 625
 626         return (start1 <= start2) && (end1 >= end2);
 627 }
 628
 629 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 630                 struct list_head *tmp_list)
 631 {
 632         if (!refcount_dec_and_test(&lseg->pls_refcount))
 633                 return false;
 634         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 635         list_add(&lseg->pls_list, tmp_list);
 636         return true;
 637 }
 638
 639 /* Returns 1 if lseg is removed from list, 0 otherwise */
 640 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 641                              struct list_head *tmp_list)
 642 {
 643         int rv = 0;
 644
 645         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 646                 /* Remove the reference keeping the lseg in the
 647                  * list.  It will now be removed when all
 648                  * outstanding io is finished.
 649                  */
 650                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 651                         refcount_read(&lseg->pls_refcount));
 652                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 653                         rv = 1;
 654         }
 655         return rv;
 656 }
 657
 658 static bool
 659 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 660                  const struct pnfs_layout_range *recall_range)
 661 {
 662         return (recall_range->iomode == IOMODE_ANY ||
 663                 lseg_range->iomode == recall_range->iomode) &&
 664                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 665 }
 666
 667 static bool
 668 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 669                 const struct pnfs_layout_range *recall_range,
 670                 u32 seq)
 671 {
 672         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 673                 return false;
 674         if (recall_range == NULL)
 675                 return true;
 676         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 677 }
 678
 679 /**
 680  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 681  * @lo: layout header containing the lsegs
 682  * @tmp_list: list head where doomed lsegs should go
 683  * @recall_range: optional recall range argument to match (may be NULL)
 684  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 685  *
 686  * Walk the list of lsegs in the layout header, and tear down any that should
 687  * be destroyed. If "recall_range" is specified then the segment must match
 688  * that range. If "seq" is non-zero, then only match segments that were handed
 689  * out at or before that sequence.
 690  *
 691  * Returns number of matching invalid lsegs remaining in list after scanning
 692  * it and purging them.
 693  */
 694 int
 695 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 696                             struct list_head *tmp_list,
 697                             const struct pnfs_layout_range *recall_range,
 698                             u32 seq)
 699 {
 700         struct pnfs_layout_segment *lseg, *next;
 701         int remaining = 0;
 702
 703         dprintk("%s:Begin lo %p\n", __func__, lo);
 704
 705         if (list_empty(&lo->plh_segs))
 706                 return 0;
 707         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 708                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 709                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 710                                 "offset %llu length %llu\n", __func__,
 711                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 712                                 lseg->pls_range.offset, lseg->pls_range.length);
 713                         if (!mark_lseg_invalid(lseg, tmp_list))
 714                                 remaining++;
 715                 }
 716         dprintk("%s:Return %i\n", __func__, remaining);
 717         return remaining;
 718 }
 719
 720 static void
 721 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 722                 struct list_head *free_me,
 723                 const struct pnfs_layout_range *range,
 724                 u32 seq)
 725 {
 726         struct pnfs_layout_segment *lseg, *next;
 727
 728         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 729                 if (pnfs_match_lseg_recall(lseg, range, seq))
 730                         list_move_tail(&lseg->pls_list, free_me);
 731         }
 732 }
 733
 734 /* note free_me must contain lsegs from a single layout_hdr */
 735 void
 736 pnfs_free_lseg_list(struct list_head *free_me)
 737 {
 738         struct pnfs_layout_segment *lseg, *tmp;
 739
 740         if (list_empty(free_me))
 741                 return;
 742
 743         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 744                 list_del(&lseg->pls_list);
 745                 pnfs_free_lseg(lseg);
 746         }
 747 }
 748
 749 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
 750 {
 751         struct pnfs_layout_hdr *lo;
 752         LIST_HEAD(tmp_list);
 753
 754         spin_lock(&nfsi->vfs_inode.i_lock);
 755         lo = nfsi->layout;
 756         if (lo) {
 757                 pnfs_get_layout_hdr(lo);
 758                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 759                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 760                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 761                 spin_unlock(&nfsi->vfs_inode.i_lock);
 762                 pnfs_free_lseg_list(&tmp_list);
 763                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 764                 pnfs_put_layout_hdr(lo);
 765         } else
 766                 spin_unlock(&nfsi->vfs_inode.i_lock);
 767         return lo;
 768 }
 769
 770 void pnfs_destroy_layout(struct nfs_inode *nfsi)
 771 {
 772         __pnfs_destroy_layout(nfsi);
 773 }
 774 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 775
 776 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
 777                                 struct pnfs_layout_hdr *lo)
 778 {
 779         bool ret;
 780
 781         spin_lock(&nfsi->vfs_inode.i_lock);
 782         ret = nfsi->layout != lo;
 783         spin_unlock(&nfsi->vfs_inode.i_lock);
 784         return ret;
 785 }
 786
 787 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
 788 {
 789         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
 790
 791         if (lo)
 792                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
 793 }
 794
 795 static bool
 796 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 797                 struct list_head *layout_list)
 798 {
 799         struct pnfs_layout_hdr *lo;
 800         bool ret = false;
 801
 802         spin_lock(&inode->i_lock);
 803         lo = NFS_I(inode)->layout;
 804         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 805                 pnfs_get_layout_hdr(lo);
 806                 list_add(&lo->plh_bulk_destroy, layout_list);
 807                 ret = true;
 808         }
 809         spin_unlock(&inode->i_lock);
 810         return ret;
 811 }
 812
 813 /* Caller must hold rcu_read_lock and clp->cl_lock */
 814 static int
 815 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 816                 struct nfs_server *server,
 817                 struct list_head *layout_list)
 818         __must_hold(&clp->cl_lock)
 819         __must_hold(RCU)
 820 {
 821         struct pnfs_layout_hdr *lo, *next;
 822         struct inode *inode;
 823
 824         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 825                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
 826                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
 827                     !list_empty(&lo->plh_bulk_destroy))
 828                         continue;
 829                 /* If the sb is being destroyed, just bail */
 830                 if (!nfs_sb_active(server->super))
 831                         break;
 832                 inode = pnfs_grab_inode_layout_hdr(lo);
 833                 if (inode != NULL) {
 834                         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
 835                                 list_del_rcu(&lo->plh_layouts);
 836                         if (pnfs_layout_add_bulk_destroy_list(inode,
 837                                                 layout_list))
 838                                 continue;
 839                         rcu_read_unlock();
 840                         spin_unlock(&clp->cl_lock);
 841                         iput(inode);
 842                 } else {
 843                         rcu_read_unlock();
 844                         spin_unlock(&clp->cl_lock);
 845                 }
 846                 nfs_sb_deactive(server->super);
 847                 spin_lock(&clp->cl_lock);
 848                 rcu_read_lock();
 849                 return -EAGAIN;
 850         }
 851         return 0;
 852 }
 853
 854 static int
 855 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 856                 bool is_bulk_recall)
 857 {
 858         struct pnfs_layout_hdr *lo;
 859         struct inode *inode;
 860         LIST_HEAD(lseg_list);
 861         int ret = 0;
 862
 863         while (!list_empty(layout_list)) {
 864                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 865                                 plh_bulk_destroy);
 866                 dprintk("%s freeing layout for inode %lu\n", __func__,
 867                         lo->plh_inode->i_ino);
 868                 inode = lo->plh_inode;
 869
 870                 pnfs_layoutcommit_inode(inode, false);
 871
 872                 spin_lock(&inode->i_lock);
 873                 list_del_init(&lo->plh_bulk_destroy);
 874                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 875                         if (is_bulk_recall)
 876                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 877                         ret = -EAGAIN;
 878                 }
 879                 spin_unlock(&inode->i_lock);
 880                 pnfs_free_lseg_list(&lseg_list);
 881                 /* Free all lsegs that are attached to commit buckets */
 882                 nfs_commit_inode(inode, 0);
 883                 pnfs_put_layout_hdr(lo);
 884                 nfs_iput_and_deactive(inode);
 885         }
 886         return ret;
 887 }
 888
 889 int
 890 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 891                 struct nfs_fsid *fsid,
 892                 bool is_recall)
 893 {
 894         struct nfs_server *server;
 895         LIST_HEAD(layout_list);
 896
 897         spin_lock(&clp->cl_lock);
 898         rcu_read_lock();
 899 restart:
 900         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 901                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 902                         continue;
 903                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 904                                 server,
 905                                 &layout_list) != 0)
 906                         goto restart;
 907         }
 908         rcu_read_unlock();
 909         spin_unlock(&clp->cl_lock);
 910
 911         if (list_empty(&layout_list))
 912                 return 0;
 913         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 914 }
 915
 916 int
 917 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 918                 bool is_recall)
 919 {
 920         struct nfs_server *server;
 921         LIST_HEAD(layout_list);
 922
 923         spin_lock(&clp->cl_lock);
 924         rcu_read_lock();
 925 restart:
 926         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 927                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 928                                         server,
 929                                         &layout_list) != 0)
 930                         goto restart;
 931         }
 932         rcu_read_unlock();
 933         spin_unlock(&clp->cl_lock);
 934
 935         if (list_empty(&layout_list))
 936                 return 0;
 937         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 938 }
 939
 940 /*
 941  * Called by the state manager to remove all layouts established under an
 942  * expired lease.
 943  */
 944 void
 945 pnfs_destroy_all_layouts(struct nfs_client *clp)
 946 {
 947         nfs4_deviceid_mark_client_invalid(clp);
 948         nfs4_deviceid_purge_client(clp);
 949
 950         pnfs_destroy_layouts_byclid(clp, false);
 951 }
 952
 953 static void
 954 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
 955 {
 956         const struct cred *old;
 957
 958         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
 959                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
 960                 put_cred(old);
 961         }
 962 }
 963
 964 /* update lo->plh_stateid with new if is more recent */
 965 void
 966 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 967                         const struct cred *cred, bool update_barrier)
 968 {
 969         u32 oldseq, newseq, new_barrier = 0;
 970
 971         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 972         newseq = be32_to_cpu(new->seqid);
 973
 974         if (!pnfs_layout_is_valid(lo)) {
 975                 pnfs_set_layout_cred(lo, cred);
 976                 nfs4_stateid_copy(&lo->plh_stateid, new);
 977                 lo->plh_barrier = newseq;
 978                 pnfs_clear_layoutreturn_info(lo);
 979                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 980                 return;
 981         }
 982         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 983                 nfs4_stateid_copy(&lo->plh_stateid, new);
 984                 /*
 985                  * Because of wraparound, we want to keep the barrier
 986                  * "close" to the current seqids.
 987                  */
 988                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 989         }
 990         if (update_barrier)
 991                 new_barrier = be32_to_cpu(new->seqid);
 992         else if (new_barrier == 0)
 993                 return;
 994         pnfs_barrier_update(lo, new_barrier);
 995 }
 996
 997 static bool
 998 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 999                 const nfs4_stateid *stateid)
1000 {
1001         u32 seqid = be32_to_cpu(stateid->seqid);
1002
1003         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier) && lo->plh_barrier;
1004 }
1005
1006 /* lget is set to 1 if called from inside send_layoutget call chain */
1007 static bool
1008 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1009 {
1010         return lo->plh_block_lgets ||
1011                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1012 }
1013
1014 static struct nfs_server *
1015 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1016 {
1017         struct nfs_server *server;
1018
1019         if (inode) {
1020                 server = NFS_SERVER(inode);
1021         } else {
1022                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1023                 server = NFS_SERVER(parent_dir->d_inode);
1024                 dput(parent_dir);
1025         }
1026         return server;
1027 }
1028
1029 static void nfs4_free_pages(struct page **pages, size_t size)
1030 {
1031         int i;
1032
1033         if (!pages)
1034                 return;
1035
1036         for (i = 0; i < size; i++) {
1037                 if (!pages[i])
1038                         break;
1039                 __free_page(pages[i]);
1040         }
1041         kfree(pages);
1042 }
1043
1044 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1045 {
1046         struct page **pages;
1047         int i;
1048
1049         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1050         if (!pages) {
1051                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1052                 return NULL;
1053         }
1054
1055         for (i = 0; i < size; i++) {
1056                 pages[i] = alloc_page(gfp_flags);
1057                 if (!pages[i]) {
1058                         dprintk("%s: failed to allocate page\n", __func__);
1059                         nfs4_free_pages(pages, i);
1060                         return NULL;
1061                 }
1062         }
1063
1064         return pages;
1065 }
1066
1067 static struct nfs4_layoutget *
1068 pnfs_alloc_init_layoutget_args(struct inode *ino,
1069            struct nfs_open_context *ctx,
1070            const nfs4_stateid *stateid,
1071            const struct pnfs_layout_range *range,
1072            gfp_t gfp_flags)
1073 {
1074         struct nfs_server *server = pnfs_find_server(ino, ctx);
1075         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1076         size_t max_pages = max_response_pages(server);
1077         struct nfs4_layoutget *lgp;
1078
1079         dprintk("--> %s\n", __func__);
1080
1081         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1082         if (lgp == NULL)
1083                 return NULL;
1084
1085         if (max_reply_sz) {
1086                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1087                 if (npages < max_pages)
1088                         max_pages = npages;
1089         }
1090
1091         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1092         if (!lgp->args.layout.pages) {
1093                 kfree(lgp);
1094                 return NULL;
1095         }
1096         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1097         lgp->res.layoutp = &lgp->args.layout;
1098
1099         /* Don't confuse uninitialised result and success */
1100         lgp->res.status = -NFS4ERR_DELAY;
1101
1102         lgp->args.minlength = PAGE_SIZE;
1103         if (lgp->args.minlength > range->length)
1104                 lgp->args.minlength = range->length;
1105         if (ino) {
1106                 loff_t i_size = i_size_read(ino);
1107
1108                 if (range->iomode == IOMODE_READ) {
1109                         if (range->offset >= i_size)
1110                                 lgp->args.minlength = 0;
1111                         else if (i_size - range->offset < lgp->args.minlength)
1112                                 lgp->args.minlength = i_size - range->offset;
1113                 }
1114         }
1115         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1116         pnfs_copy_range(&lgp->args.range, range);
1117         lgp->args.type = server->pnfs_curr_ld->id;
1118         lgp->args.inode = ino;
1119         lgp->args.ctx = get_nfs_open_context(ctx);
1120         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1121         lgp->gfp_flags = gfp_flags;
1122         lgp->cred = ctx->cred;
1123         return lgp;
1124 }
1125
1126 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1127 {
1128         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1129
1130         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1131         if (lgp->args.inode)
1132                 pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1133         put_nfs_open_context(lgp->args.ctx);
1134         kfree(lgp);
1135 }
1136
1137 static void pnfs_clear_layoutcommit(struct inode *inode,
1138                 struct list_head *head)
1139 {
1140         struct nfs_inode *nfsi = NFS_I(inode);
1141         struct pnfs_layout_segment *lseg, *tmp;
1142
1143         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1144                 return;
1145         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1146                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1147                         continue;
1148                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1149         }
1150 }
1151
1152 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1153                 const nfs4_stateid *arg_stateid,
1154                 const struct pnfs_layout_range *range,
1155                 const nfs4_stateid *stateid)
1156 {
1157         struct inode *inode = lo->plh_inode;
1158         LIST_HEAD(freeme);
1159
1160         spin_lock(&inode->i_lock);
1161         if (!pnfs_layout_is_valid(lo) ||
1162             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1163                 goto out_unlock;
1164         if (stateid) {
1165                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1166
1167                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1168                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1169                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1170         } else
1171                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1172 out_unlock:
1173         pnfs_clear_layoutreturn_waitbit(lo);
1174         spin_unlock(&inode->i_lock);
1175         pnfs_free_lseg_list(&freeme);
1176
1177 }
1178
1179 static bool
1180 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1181                 nfs4_stateid *stateid,
1182                 const struct cred **cred,
1183                 enum pnfs_iomode *iomode)
1184 {
1185         /* Serialise LAYOUTGET/LAYOUTRETURN */
1186         if (atomic_read(&lo->plh_outstanding) != 0)
1187                 return false;
1188         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1189                 return false;
1190         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1191         pnfs_get_layout_hdr(lo);
1192         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1193         *cred = get_cred(lo->plh_lc_cred);
1194         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1195                 if (lo->plh_return_seq != 0)
1196                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1197                 if (iomode != NULL)
1198                         *iomode = lo->plh_return_iomode;
1199                 pnfs_clear_layoutreturn_info(lo);
1200         } else if (iomode != NULL)
1201                 *iomode = IOMODE_ANY;
1202         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1203         return true;
1204 }
1205
1206 static void
1207 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1208                 struct pnfs_layout_hdr *lo,
1209                 const nfs4_stateid *stateid,
1210                 enum pnfs_iomode iomode)
1211 {
1212         struct inode *inode = lo->plh_inode;
1213
1214         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1215         args->inode = inode;
1216         args->range.iomode = iomode;
1217         args->range.offset = 0;
1218         args->range.length = NFS4_MAX_UINT64;
1219         args->layout = lo;
1220         nfs4_stateid_copy(&args->stateid, stateid);
1221 }
1222
1223 static int
1224 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1225                        const nfs4_stateid *stateid,
1226                        const struct cred **pcred,
1227                        enum pnfs_iomode iomode,
1228                        bool sync)
1229 {
1230         struct inode *ino = lo->plh_inode;
1231         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1232         struct nfs4_layoutreturn *lrp;
1233         const struct cred *cred = *pcred;
1234         int status = 0;
1235
1236         *pcred = NULL;
1237         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1238         if (unlikely(lrp == NULL)) {
1239                 status = -ENOMEM;
1240                 spin_lock(&ino->i_lock);
1241                 pnfs_clear_layoutreturn_waitbit(lo);
1242                 spin_unlock(&ino->i_lock);
1243                 put_cred(cred);
1244                 pnfs_put_layout_hdr(lo);
1245                 goto out;
1246         }
1247
1248         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1249         lrp->args.ld_private = &lrp->ld_private;
1250         lrp->clp = NFS_SERVER(ino)->nfs_client;
1251         lrp->cred = cred;
1252         if (ld->prepare_layoutreturn)
1253                 ld->prepare_layoutreturn(&lrp->args);
1254
1255         status = nfs4_proc_layoutreturn(lrp, sync);
1256 out:
1257         dprintk("<-- %s status: %d\n", __func__, status);
1258         return status;
1259 }
1260
1261 static bool
1262 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1263                                 enum pnfs_iomode iomode,
1264                                 u32 seq)
1265 {
1266         struct pnfs_layout_range recall_range = {
1267                 .length = NFS4_MAX_UINT64,
1268                 .iomode = iomode,
1269         };
1270         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1271                                                &recall_range, seq) != -EBUSY;
1272 }
1273
1274 /* Return true if layoutreturn is needed */
1275 static bool
1276 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1277 {
1278         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1279                 return false;
1280         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1281                                                lo->plh_return_seq);
1282 }
1283
1284 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1285 {
1286         struct inode *inode= lo->plh_inode;
1287
1288         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1289                 return;
1290         spin_lock(&inode->i_lock);
1291         if (pnfs_layout_need_return(lo)) {
1292                 const struct cred *cred;
1293                 nfs4_stateid stateid;
1294                 enum pnfs_iomode iomode;
1295                 bool send;
1296
1297                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1298                 spin_unlock(&inode->i_lock);
1299                 if (send) {
1300                         /* Send an async layoutreturn so we dont deadlock */
1301                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1302                 }
1303         } else
1304                 spin_unlock(&inode->i_lock);
1305 }
1306
1307 /*
1308  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1309  * when the layout segment list is empty.
1310  *
1311  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1312  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1313  * deviceid is marked invalid.
1314  */
1315 int
1316 _pnfs_return_layout(struct inode *ino)
1317 {
1318         struct pnfs_layout_hdr *lo = NULL;
1319         struct nfs_inode *nfsi = NFS_I(ino);
1320         struct pnfs_layout_range range = {
1321                 .iomode         = IOMODE_ANY,
1322                 .offset         = 0,
1323                 .length         = NFS4_MAX_UINT64,
1324         };
1325         LIST_HEAD(tmp_list);
1326         const struct cred *cred;
1327         nfs4_stateid stateid;
1328         int status = 0;
1329         bool send, valid_layout;
1330
1331         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1332
1333         spin_lock(&ino->i_lock);
1334         lo = nfsi->layout;
1335         if (!lo) {
1336                 spin_unlock(&ino->i_lock);
1337                 dprintk("NFS: %s no layout to return\n", __func__);
1338                 goto out;
1339         }
1340         /* Reference matched in nfs4_layoutreturn_release */
1341         pnfs_get_layout_hdr(lo);
1342         /* Is there an outstanding layoutreturn ? */
1343         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1344                 spin_unlock(&ino->i_lock);
1345                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1346                                         TASK_UNINTERRUPTIBLE))
1347                         goto out_put_layout_hdr;
1348                 spin_lock(&ino->i_lock);
1349         }
1350         valid_layout = pnfs_layout_is_valid(lo);
1351         pnfs_clear_layoutcommit(ino, &tmp_list);
1352         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1353
1354         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1355                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1356
1357         /* Don't send a LAYOUTRETURN if list was initially empty */
1358         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1359                         !valid_layout) {
1360                 spin_unlock(&ino->i_lock);
1361                 dprintk("NFS: %s no layout segments to return\n", __func__);
1362                 goto out_wait_layoutreturn;
1363         }
1364
1365         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1366         spin_unlock(&ino->i_lock);
1367         if (send)
1368                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1369 out_wait_layoutreturn:
1370         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1371 out_put_layout_hdr:
1372         pnfs_free_lseg_list(&tmp_list);
1373         pnfs_put_layout_hdr(lo);
1374 out:
1375         dprintk("<-- %s status: %d\n", __func__, status);
1376         return status;
1377 }
1378
1379 int
1380 pnfs_commit_and_return_layout(struct inode *inode)
1381 {
1382         struct pnfs_layout_hdr *lo;
1383         int ret;
1384
1385         spin_lock(&inode->i_lock);
1386         lo = NFS_I(inode)->layout;
1387         if (lo == NULL) {
1388                 spin_unlock(&inode->i_lock);
1389                 return 0;
1390         }
1391         pnfs_get_layout_hdr(lo);
1392         /* Block new layoutgets and read/write to ds */
1393         lo->plh_block_lgets++;
1394         spin_unlock(&inode->i_lock);
1395         filemap_fdatawait(inode->i_mapping);
1396         ret = pnfs_layoutcommit_inode(inode, true);
1397         if (ret == 0)
1398                 ret = _pnfs_return_layout(inode);
1399         spin_lock(&inode->i_lock);
1400         lo->plh_block_lgets--;
1401         spin_unlock(&inode->i_lock);
1402         pnfs_put_layout_hdr(lo);
1403         return ret;
1404 }
1405
1406 bool pnfs_roc(struct inode *ino,
1407                 struct nfs4_layoutreturn_args *args,
1408                 struct nfs4_layoutreturn_res *res,
1409                 const struct cred *cred)
1410 {
1411         struct nfs_inode *nfsi = NFS_I(ino);
1412         struct nfs_open_context *ctx;
1413         struct nfs4_state *state;
1414         struct pnfs_layout_hdr *lo;
1415         struct pnfs_layout_segment *lseg, *next;
1416         const struct cred *lc_cred;
1417         nfs4_stateid stateid;
1418         enum pnfs_iomode iomode = 0;
1419         bool layoutreturn = false, roc = false;
1420         bool skip_read = false;
1421
1422         if (!nfs_have_layout(ino))
1423                 return false;
1424 retry:
1425         rcu_read_lock();
1426         spin_lock(&ino->i_lock);
1427         lo = nfsi->layout;
1428         if (!lo || !pnfs_layout_is_valid(lo) ||
1429             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1430                 lo = NULL;
1431                 goto out_noroc;
1432         }
1433         pnfs_get_layout_hdr(lo);
1434         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1435                 spin_unlock(&ino->i_lock);
1436                 rcu_read_unlock();
1437                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1438                                 TASK_UNINTERRUPTIBLE);
1439                 pnfs_put_layout_hdr(lo);
1440                 goto retry;
1441         }
1442
1443         /* no roc if we hold a delegation */
1444         if (nfs4_check_delegation(ino, FMODE_READ)) {
1445                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1446                         goto out_noroc;
1447                 skip_read = true;
1448         }
1449
1450         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1451                 state = ctx->state;
1452                 if (state == NULL)
1453                         continue;
1454                 /* Don't return layout if there is open file state */
1455                 if (state->state & FMODE_WRITE)
1456                         goto out_noroc;
1457                 if (state->state & FMODE_READ)
1458                         skip_read = true;
1459         }
1460
1461
1462         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1463                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1464                         continue;
1465                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1466                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1467                         continue;
1468                 /*
1469                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1470                  * doesn't invalidate the layout for us.
1471                  */
1472                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1473                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1474                         continue;
1475                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1476         }
1477
1478         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1479                 goto out_noroc;
1480
1481         /* ROC in two conditions:
1482          * 1. there are ROC lsegs
1483          * 2. we don't send layoutreturn
1484          */
1485         /* lo ref dropped in pnfs_roc_release() */
1486         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1487         /* If the creds don't match, we can't compound the layoutreturn */
1488         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1489                 goto out_noroc;
1490
1491         roc = layoutreturn;
1492         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1493         res->lrs_present = 0;
1494         layoutreturn = false;
1495         put_cred(lc_cred);
1496
1497 out_noroc:
1498         spin_unlock(&ino->i_lock);
1499         rcu_read_unlock();
1500         pnfs_layoutcommit_inode(ino, true);
1501         if (roc) {
1502                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1503                 if (ld->prepare_layoutreturn)
1504                         ld->prepare_layoutreturn(args);
1505                 pnfs_put_layout_hdr(lo);
1506                 return true;
1507         }
1508         if (layoutreturn)
1509                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1510         pnfs_put_layout_hdr(lo);
1511         return false;
1512 }
1513
1514 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1515                   struct nfs4_layoutreturn_res **respp, int *ret)
1516 {
1517         struct nfs4_layoutreturn_args *arg = *argpp;
1518         int retval = -EAGAIN;
1519
1520         if (!arg)
1521                 return 0;
1522         /* Handle Layoutreturn errors */
1523         switch (*ret) {
1524         case 0:
1525                 retval = 0;
1526                 break;
1527         case -NFS4ERR_NOMATCHING_LAYOUT:
1528                 /* Was there an RPC level error? If not, retry */
1529                 if (task->tk_rpc_status == 0)
1530                         break;
1531                 /* If the call was not sent, let caller handle it */
1532                 if (!RPC_WAS_SENT(task))
1533                         return 0;
1534                 /*
1535                  * Otherwise, assume the call succeeded and
1536                  * that we need to release the layout
1537                  */
1538                 *ret = 0;
1539                 (*respp)->lrs_present = 0;
1540                 retval = 0;
1541                 break;
1542         case -NFS4ERR_DELAY:
1543                 /* Let the caller handle the retry */
1544                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1545                 return 0;
1546         case -NFS4ERR_OLD_STATEID:
1547                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1548                                                      &arg->range, arg->inode))
1549                         break;
1550                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1551                 return -EAGAIN;
1552         }
1553         *argpp = NULL;
1554         *respp = NULL;
1555         return retval;
1556 }
1557
1558 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1559                 struct nfs4_layoutreturn_res *res,
1560                 int ret)
1561 {
1562         struct pnfs_layout_hdr *lo = args->layout;
1563         struct inode *inode = args->inode;
1564         const nfs4_stateid *res_stateid = NULL;
1565         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1566
1567         switch (ret) {
1568         case -NFS4ERR_NOMATCHING_LAYOUT:
1569                 spin_lock(&inode->i_lock);
1570                 if (pnfs_layout_is_valid(lo) &&
1571                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1572                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1573                 pnfs_clear_layoutreturn_waitbit(lo);
1574                 spin_unlock(&inode->i_lock);
1575                 break;
1576         case 0:
1577                 if (res->lrs_present)
1578                         res_stateid = &res->stateid;
1579                 fallthrough;
1580         default:
1581                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1582                                              res_stateid);
1583         }
1584         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1585         if (ld_private && ld_private->ops && ld_private->ops->free)
1586                 ld_private->ops->free(ld_private);
1587         pnfs_put_layout_hdr(lo);
1588 }
1589
1590 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1591 {
1592         struct nfs_inode *nfsi = NFS_I(ino);
1593         struct pnfs_layout_hdr *lo;
1594         bool sleep = false;
1595
1596         /* we might not have grabbed lo reference. so need to check under
1597          * i_lock */
1598         spin_lock(&ino->i_lock);
1599         lo = nfsi->layout;
1600         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1601                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1602                 sleep = true;
1603         }
1604         spin_unlock(&ino->i_lock);
1605         return sleep;
1606 }
1607
1608 /*
1609  * Compare two layout segments for sorting into layout cache.
1610  * We want to preferentially return RW over RO layouts, so ensure those
1611  * are seen first.
1612  */
1613 static s64
1614 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1615            const struct pnfs_layout_range *l2)
1616 {
1617         s64 d;
1618
1619         /* high offset > low offset */
1620         d = l1->offset - l2->offset;
1621         if (d)
1622                 return d;
1623
1624         /* short length > long length */
1625         d = l2->length - l1->length;
1626         if (d)
1627                 return d;
1628
1629         /* read > read/write */
1630         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1631 }
1632
1633 static bool
1634 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1635                 const struct pnfs_layout_range *l2)
1636 {
1637         return pnfs_lseg_range_cmp(l1, l2) > 0;
1638 }
1639
1640 static bool
1641 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1642                 struct pnfs_layout_segment *old)
1643 {
1644         return false;
1645 }
1646
1647 void
1648 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1649                    struct pnfs_layout_segment *lseg,
1650                    bool (*is_after)(const struct pnfs_layout_range *,
1651                            const struct pnfs_layout_range *),
1652                    bool (*do_merge)(struct pnfs_layout_segment *,
1653                            struct pnfs_layout_segment *),
1654                    struct list_head *free_me)
1655 {
1656         struct pnfs_layout_segment *lp, *tmp;
1657
1658         dprintk("%s:Begin\n", __func__);
1659
1660         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1661                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1662                         continue;
1663                 if (do_merge(lseg, lp)) {
1664                         mark_lseg_invalid(lp, free_me);
1665                         continue;
1666                 }
1667                 if (is_after(&lseg->pls_range, &lp->pls_range))
1668                         continue;
1669                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1670                 dprintk("%s: inserted lseg %p "
1671                         "iomode %d offset %llu length %llu before "
1672                         "lp %p iomode %d offset %llu length %llu\n",
1673                         __func__, lseg, lseg->pls_range.iomode,
1674                         lseg->pls_range.offset, lseg->pls_range.length,
1675                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1676                         lp->pls_range.length);
1677                 goto out;
1678         }
1679         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1680         dprintk("%s: inserted lseg %p "
1681                 "iomode %d offset %llu length %llu at tail\n",
1682                 __func__, lseg, lseg->pls_range.iomode,
1683                 lseg->pls_range.offset, lseg->pls_range.length);
1684 out:
1685         pnfs_get_layout_hdr(lo);
1686
1687         dprintk("%s:Return\n", __func__);
1688 }
1689 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1690
1691 static void
1692 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1693                    struct pnfs_layout_segment *lseg,
1694                    struct list_head *free_me)
1695 {
1696         struct inode *inode = lo->plh_inode;
1697         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1698
1699         if (ld->add_lseg != NULL)
1700                 ld->add_lseg(lo, lseg, free_me);
1701         else
1702                 pnfs_generic_layout_insert_lseg(lo, lseg,
1703                                 pnfs_lseg_range_is_after,
1704                                 pnfs_lseg_no_merge,
1705                                 free_me);
1706 }
1707
1708 static struct pnfs_layout_hdr *
1709 alloc_init_layout_hdr(struct inode *ino,
1710                       struct nfs_open_context *ctx,
1711                       gfp_t gfp_flags)
1712 {
1713         struct pnfs_layout_hdr *lo;
1714
1715         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1716         if (!lo)
1717                 return NULL;
1718         refcount_set(&lo->plh_refcount, 1);
1719         INIT_LIST_HEAD(&lo->plh_layouts);
1720         INIT_LIST_HEAD(&lo->plh_segs);
1721         INIT_LIST_HEAD(&lo->plh_return_segs);
1722         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1723         lo->plh_inode = ino;
1724         lo->plh_lc_cred = get_cred(ctx->cred);
1725         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1726         return lo;
1727 }
1728
1729 static struct pnfs_layout_hdr *
1730 pnfs_find_alloc_layout(struct inode *ino,
1731                        struct nfs_open_context *ctx,
1732                        gfp_t gfp_flags)
1733         __releases(&ino->i_lock)
1734         __acquires(&ino->i_lock)
1735 {
1736         struct nfs_inode *nfsi = NFS_I(ino);
1737         struct pnfs_layout_hdr *new = NULL;
1738
1739         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1740
1741         if (nfsi->layout != NULL)
1742                 goto out_existing;
1743         spin_unlock(&ino->i_lock);
1744         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1745         spin_lock(&ino->i_lock);
1746
1747         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1748                 nfsi->layout = new;
1749                 return new;
1750         } else if (new != NULL)
1751                 pnfs_free_layout_hdr(new);
1752 out_existing:
1753         pnfs_get_layout_hdr(nfsi->layout);
1754         return nfsi->layout;
1755 }
1756
1757 /*
1758  * iomode matching rules:
1759  * iomode       lseg    strict match
1760  *                      iomode
1761  * -----        -----   ------ -----
1762  * ANY          READ    N/A    true
1763  * ANY          RW      N/A    true
1764  * RW           READ    N/A    false
1765  * RW           RW      N/A    true
1766  * READ         READ    N/A    true
1767  * READ         RW      true   false
1768  * READ         RW      false  true
1769  */
1770 static bool
1771 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1772                  const struct pnfs_layout_range *range,
1773                  bool strict_iomode)
1774 {
1775         struct pnfs_layout_range range1;
1776
1777         if ((range->iomode == IOMODE_RW &&
1778              ls_range->iomode != IOMODE_RW) ||
1779             (range->iomode != ls_range->iomode &&
1780              strict_iomode) ||
1781             !pnfs_lseg_range_intersecting(ls_range, range))
1782                 return false;
1783
1784         /* range1 covers only the first byte in the range */
1785         range1 = *range;
1786         range1.length = 1;
1787         return pnfs_lseg_range_contained(ls_range, &range1);
1788 }
1789
1790 /*
1791  * lookup range in layout
1792  */
1793 static struct pnfs_layout_segment *
1794 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1795                 struct pnfs_layout_range *range,
1796                 bool strict_iomode)
1797 {
1798         struct pnfs_layout_segment *lseg, *ret = NULL;
1799
1800         dprintk("%s:Begin\n", __func__);
1801
1802         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1803                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1804                     pnfs_lseg_range_match(&lseg->pls_range, range,
1805                                           strict_iomode)) {
1806                         ret = pnfs_get_lseg(lseg);
1807                         break;
1808                 }
1809         }
1810
1811         dprintk("%s:Return lseg %p ref %d\n",
1812                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1813         return ret;
1814 }
1815
1816 /*
1817  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1818  * to the MDS or over pNFS
1819  *
1820  * The nfs_inode read_io and write_io fields are cumulative counters reset
1821  * when there are no layout segments. Note that in pnfs_update_layout iomode
1822  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1823  * WRITE request.
1824  *
1825  * A return of true means use MDS I/O.
1826  *
1827  * From rfc 5661:
1828  * If a file's size is smaller than the file size threshold, data accesses
1829  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1830  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1831  * server.  If both file size and I/O size are provided, the client SHOULD
1832  * reach or exceed  both thresholds before sending its read or write
1833  * requests to the data server.
1834  */
1835 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1836                                      struct inode *ino, int iomode)
1837 {
1838         struct nfs4_threshold *t = ctx->mdsthreshold;
1839         struct nfs_inode *nfsi = NFS_I(ino);
1840         loff_t fsize = i_size_read(ino);
1841         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1842
1843         if (t == NULL)
1844                 return ret;
1845
1846         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1847                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1848
1849         switch (iomode) {
1850         case IOMODE_READ:
1851                 if (t->bm & THRESHOLD_RD) {
1852                         dprintk("%s fsize %llu\n", __func__, fsize);
1853                         size_set = true;
1854                         if (fsize < t->rd_sz)
1855                                 size = true;
1856                 }
1857                 if (t->bm & THRESHOLD_RD_IO) {
1858                         dprintk("%s nfsi->read_io %llu\n", __func__,
1859                                 nfsi->read_io);
1860                         io_set = true;
1861                         if (nfsi->read_io < t->rd_io_sz)
1862                                 io = true;
1863                 }
1864                 break;
1865         case IOMODE_RW:
1866                 if (t->bm & THRESHOLD_WR) {
1867                         dprintk("%s fsize %llu\n", __func__, fsize);
1868                         size_set = true;
1869                         if (fsize < t->wr_sz)
1870                                 size = true;
1871                 }
1872                 if (t->bm & THRESHOLD_WR_IO) {
1873                         dprintk("%s nfsi->write_io %llu\n", __func__,
1874                                 nfsi->write_io);
1875                         io_set = true;
1876                         if (nfsi->write_io < t->wr_io_sz)
1877                                 io = true;
1878                 }
1879                 break;
1880         }
1881         if (size_set && io_set) {
1882                 if (size && io)
1883                         ret = true;
1884         } else if (size || io)
1885                 ret = true;
1886
1887         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1888         return ret;
1889 }
1890
1891 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1892 {
1893         /*
1894          * send layoutcommit as it can hold up layoutreturn due to lseg
1895          * reference
1896          */
1897         pnfs_layoutcommit_inode(lo->plh_inode, false);
1898         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1899                                    nfs_wait_bit_killable,
1900                                    TASK_KILLABLE);
1901 }
1902
1903 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1904 {
1905         atomic_inc(&lo->plh_outstanding);
1906 }
1907
1908 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1909 {
1910         if (atomic_dec_and_test(&lo->plh_outstanding))
1911                 wake_up_var(&lo->plh_outstanding);
1912 }
1913
1914 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1915 {
1916         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1917 }
1918
1919 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1920 {
1921         unsigned long *bitlock = &lo->plh_flags;
1922
1923         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1924         smp_mb__after_atomic();
1925         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1926 }
1927
1928 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1929                                 struct nfs_server *server)
1930 {
1931         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1932                 struct nfs_client *clp = server->nfs_client;
1933
1934                 /* The lo must be on the clp list if there is any
1935                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1936                  */
1937                 spin_lock(&clp->cl_lock);
1938                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1939                 spin_unlock(&clp->cl_lock);
1940         }
1941 }
1942
1943 /*
1944  * Layout segment is retreived from the server if not cached.
1945  * The appropriate layout segment is referenced and returned to the caller.
1946  */
1947 struct pnfs_layout_segment *
1948 pnfs_update_layout(struct inode *ino,
1949                    struct nfs_open_context *ctx,
1950                    loff_t pos,
1951                    u64 count,
1952                    enum pnfs_iomode iomode,
1953                    bool strict_iomode,
1954                    gfp_t gfp_flags)
1955 {
1956         struct pnfs_layout_range arg = {
1957                 .iomode = iomode,
1958                 .offset = pos,
1959                 .length = count,
1960         };
1961         unsigned pg_offset;
1962         struct nfs_server *server = NFS_SERVER(ino);
1963         struct nfs_client *clp = server->nfs_client;
1964         struct pnfs_layout_hdr *lo = NULL;
1965         struct pnfs_layout_segment *lseg = NULL;
1966         struct nfs4_layoutget *lgp;
1967         nfs4_stateid stateid;
1968         long timeout = 0;
1969         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1970         bool first;
1971
1972         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1973                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1974                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1975                 goto out;
1976         }
1977
1978         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1979                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1980                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1981                 goto out;
1982         }
1983
1984 lookup_again:
1985         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1986         if (IS_ERR(lseg))
1987                 goto out;
1988         first = false;
1989         spin_lock(&ino->i_lock);
1990         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1991         if (lo == NULL) {
1992                 spin_unlock(&ino->i_lock);
1993                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1994                                  PNFS_UPDATE_LAYOUT_NOMEM);
1995                 goto out;
1996         }
1997
1998         /* Do we even need to bother with this? */
1999         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2000                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2001                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2002                 dprintk("%s matches recall, use MDS\n", __func__);
2003                 goto out_unlock;
2004         }
2005
2006         /* if LAYOUTGET already failed once we don't try again */
2007         if (pnfs_layout_io_test_failed(lo, iomode)) {
2008                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2009                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2010                 goto out_unlock;
2011         }
2012
2013         /*
2014          * If the layout segment list is empty, but there are outstanding
2015          * layoutget calls, then they might be subject to a layoutrecall.
2016          */
2017         if (list_empty(&lo->plh_segs) &&
2018             atomic_read(&lo->plh_outstanding) != 0) {
2019                 spin_unlock(&ino->i_lock);
2020                 lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
2021                                         !atomic_read(&lo->plh_outstanding)));
2022                 if (IS_ERR(lseg))
2023                         goto out_put_layout_hdr;
2024                 pnfs_put_layout_hdr(lo);
2025                 goto lookup_again;
2026         }
2027
2028         /*
2029          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2030          * for LAYOUTRETURN.
2031          */
2032         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2033                 spin_unlock(&ino->i_lock);
2034                 dprintk("%s wait for layoutreturn\n", __func__);
2035                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2036                 if (!IS_ERR(lseg)) {
2037                         pnfs_put_layout_hdr(lo);
2038                         dprintk("%s retrying\n", __func__);
2039                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2040                                                  lseg,
2041                                                  PNFS_UPDATE_LAYOUT_RETRY);
2042                         goto lookup_again;
2043                 }
2044                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2045                                          PNFS_UPDATE_LAYOUT_RETURN);
2046                 goto out_put_layout_hdr;
2047         }
2048
2049         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2050         if (lseg) {
2051                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2052                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2053                 goto out_unlock;
2054         }
2055
2056         /*
2057          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2058          * stateid, or it has been invalidated, then we must use the open
2059          * stateid.
2060          */
2061         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2062                 int status;
2063
2064                 /*
2065                  * The first layoutget for the file. Need to serialize per
2066                  * RFC 5661 Errata 3208.
2067                  */
2068                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2069                                      &lo->plh_flags)) {
2070                         spin_unlock(&ino->i_lock);
2071                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2072                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2073                                                 TASK_KILLABLE));
2074                         if (IS_ERR(lseg))
2075                                 goto out_put_layout_hdr;
2076                         pnfs_put_layout_hdr(lo);
2077                         dprintk("%s retrying\n", __func__);
2078                         goto lookup_again;
2079                 }
2080
2081                 spin_unlock(&ino->i_lock);
2082                 first = true;
2083                 status = nfs4_select_rw_stateid(ctx->state,
2084                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2085                                         NULL, &stateid, NULL);
2086                 if (status != 0) {
2087                         lseg = ERR_PTR(status);
2088                         trace_pnfs_update_layout(ino, pos, count,
2089                                         iomode, lo, lseg,
2090                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2091                         nfs4_schedule_stateid_recovery(server, ctx->state);
2092                         pnfs_clear_first_layoutget(lo);
2093                         pnfs_put_layout_hdr(lo);
2094                         goto lookup_again;
2095                 }
2096                 spin_lock(&ino->i_lock);
2097         } else {
2098                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2099         }
2100
2101         if (pnfs_layoutgets_blocked(lo)) {
2102                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2103                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2104                 goto out_unlock;
2105         }
2106         nfs_layoutget_begin(lo);
2107         spin_unlock(&ino->i_lock);
2108
2109         _add_to_server_list(lo, server);
2110
2111         pg_offset = arg.offset & ~PAGE_MASK;
2112         if (pg_offset) {
2113                 arg.offset -= pg_offset;
2114                 arg.length += pg_offset;
2115         }
2116         if (arg.length != NFS4_MAX_UINT64)
2117                 arg.length = PAGE_ALIGN(arg.length);
2118
2119         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2120         if (!lgp) {
2121                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2122                                          PNFS_UPDATE_LAYOUT_NOMEM);
2123                 nfs_layoutget_end(lo);
2124                 goto out_put_layout_hdr;
2125         }
2126
2127         lseg = nfs4_proc_layoutget(lgp, &timeout);
2128         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2129                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2130         nfs_layoutget_end(lo);
2131         if (IS_ERR(lseg)) {
2132                 switch(PTR_ERR(lseg)) {
2133                 case -EBUSY:
2134                         if (time_after(jiffies, giveup))
2135                                 lseg = NULL;
2136                         break;
2137                 case -ERECALLCONFLICT:
2138                 case -EAGAIN:
2139                         break;
2140                 default:
2141                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2142                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2143                                 lseg = NULL;
2144                         }
2145                         goto out_put_layout_hdr;
2146                 }
2147                 if (lseg) {
2148                         if (first)
2149                                 pnfs_clear_first_layoutget(lo);
2150                         trace_pnfs_update_layout(ino, pos, count,
2151                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2152                         pnfs_put_layout_hdr(lo);
2153                         goto lookup_again;
2154                 }
2155         } else {
2156                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2157         }
2158
2159 out_put_layout_hdr:
2160         if (first)
2161                 pnfs_clear_first_layoutget(lo);
2162         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2163                                  PNFS_UPDATE_LAYOUT_EXIT);
2164         pnfs_put_layout_hdr(lo);
2165 out:
2166         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2167                         "(%s, offset: %llu, length: %llu)\n",
2168                         __func__, ino->i_sb->s_id,
2169                         (unsigned long long)NFS_FILEID(ino),
2170                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2171                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2172                         (unsigned long long)pos,
2173                         (unsigned long long)count);
2174         return lseg;
2175 out_unlock:
2176         spin_unlock(&ino->i_lock);
2177         goto out_put_layout_hdr;
2178 }
2179 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2180
2181 static bool
2182 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2183 {
2184         switch (range->iomode) {
2185         case IOMODE_READ:
2186         case IOMODE_RW:
2187                 break;
2188         default:
2189                 return false;
2190         }
2191         if (range->offset == NFS4_MAX_UINT64)
2192                 return false;
2193         if (range->length == 0)
2194                 return false;
2195         if (range->length != NFS4_MAX_UINT64 &&
2196             range->length > NFS4_MAX_UINT64 - range->offset)
2197                 return false;
2198         return true;
2199 }
2200
2201 static struct pnfs_layout_hdr *
2202 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2203 {
2204         struct pnfs_layout_hdr *lo;
2205
2206         spin_lock(&ino->i_lock);
2207         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2208         if (!lo)
2209                 goto out_unlock;
2210         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2211                 goto out_unlock;
2212         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2213                 goto out_unlock;
2214         if (pnfs_layoutgets_blocked(lo))
2215                 goto out_unlock;
2216         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2217                 goto out_unlock;
2218         nfs_layoutget_begin(lo);
2219         spin_unlock(&ino->i_lock);
2220         _add_to_server_list(lo, NFS_SERVER(ino));
2221         return lo;
2222
2223 out_unlock:
2224         spin_unlock(&ino->i_lock);
2225         pnfs_put_layout_hdr(lo);
2226         return NULL;
2227 }
2228
2229 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2230                                      struct nfs_open_context *ctx)
2231 {
2232         struct inode *ino = data->dentry->d_inode;
2233         struct pnfs_layout_range rng = {
2234                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2235                           IOMODE_RW: IOMODE_READ,
2236                 .offset = 0,
2237                 .length = NFS4_MAX_UINT64,
2238         };
2239         struct nfs4_layoutget *lgp;
2240         struct pnfs_layout_hdr *lo;
2241
2242         /* Heuristic: don't send layoutget if we have cached data */
2243         if (rng.iomode == IOMODE_READ &&
2244            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2245                 return;
2246
2247         lo = _pnfs_grab_empty_layout(ino, ctx);
2248         if (!lo)
2249                 return;
2250         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2251                                              &rng, GFP_KERNEL);
2252         if (!lgp) {
2253                 pnfs_clear_first_layoutget(lo);
2254                 nfs_layoutget_end(lo);
2255                 pnfs_put_layout_hdr(lo);
2256                 return;
2257         }
2258         data->lgp = lgp;
2259         data->o_arg.lg_args = &lgp->args;
2260         data->o_res.lg_res = &lgp->res;
2261 }
2262
2263 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2264                                      struct nfs_open_context *ctx)
2265 {
2266         struct pnfs_layout_range rng = {
2267                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2268                           IOMODE_RW: IOMODE_READ,
2269                 .offset = 0,
2270                 .length = NFS4_MAX_UINT64,
2271         };
2272         struct nfs4_layoutget *lgp;
2273
2274         lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2275                                              &rng, GFP_KERNEL);
2276         if (!lgp)
2277                 return;
2278         data->lgp = lgp;
2279         data->o_arg.lg_args = &lgp->args;
2280         data->o_res.lg_res = &lgp->res;
2281 }
2282
2283 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2284                          struct nfs_open_context *ctx)
2285 {
2286         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2287
2288         if (!(pnfs_enabled_sb(server) &&
2289               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2290                 return;
2291         /* Could check on max_ops, but currently hardcoded high enough */
2292         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2293                 return;
2294         if (data->state)
2295                 _lgopen_prepare_attached(data, ctx);
2296         else
2297                 _lgopen_prepare_floating(data, ctx);
2298 }
2299
2300 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2301                        struct nfs_open_context *ctx)
2302 {
2303         struct pnfs_layout_hdr *lo;
2304         struct pnfs_layout_segment *lseg;
2305         struct nfs_server *srv = NFS_SERVER(ino);
2306         u32 iomode;
2307
2308         if (!lgp)
2309                 return;
2310         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2311         if (lgp->res.status) {
2312                 switch (lgp->res.status) {
2313                 default:
2314                         break;
2315                 /*
2316                  * Halt lgopen attempts if the server doesn't recognise
2317                  * the "current stateid" value, the layout type, or the
2318                  * layoutget operation as being valid.
2319                  * Also if it complains about too many ops in the compound
2320                  * or of the request/reply being too big.
2321                  */
2322                 case -NFS4ERR_BAD_STATEID:
2323                 case -NFS4ERR_NOTSUPP:
2324                 case -NFS4ERR_REP_TOO_BIG:
2325                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2326                 case -NFS4ERR_REQ_TOO_BIG:
2327                 case -NFS4ERR_TOO_MANY_OPS:
2328                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2329                         srv->caps &= ~NFS_CAP_LGOPEN;
2330                 }
2331                 return;
2332         }
2333         if (!lgp->args.inode) {
2334                 lo = _pnfs_grab_empty_layout(ino, ctx);
2335                 if (!lo)
2336                         return;
2337                 lgp->args.inode = ino;
2338         } else
2339                 lo = NFS_I(lgp->args.inode)->layout;
2340
2341         lseg = pnfs_layout_process(lgp);
2342         if (!IS_ERR(lseg)) {
2343                 iomode = lgp->args.range.iomode;
2344                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2345                 pnfs_put_lseg(lseg);
2346         }
2347 }
2348
2349 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2350 {
2351         if (lgp != NULL) {
2352                 struct inode *inode = lgp->args.inode;
2353                 if (inode) {
2354                         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2355                         pnfs_clear_first_layoutget(lo);
2356                         nfs_layoutget_end(lo);
2357                 }
2358                 pnfs_layoutget_free(lgp);
2359         }
2360 }
2361
2362 struct pnfs_layout_segment *
2363 pnfs_layout_process(struct nfs4_layoutget *lgp)
2364 {
2365         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2366         struct nfs4_layoutget_res *res = &lgp->res;
2367         struct pnfs_layout_segment *lseg;
2368         struct inode *ino = lo->plh_inode;
2369         LIST_HEAD(free_me);
2370
2371         if (!pnfs_sanity_check_layout_range(&res->range))
2372                 return ERR_PTR(-EINVAL);
2373
2374         /* Inject layout blob into I/O device driver */
2375         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2376         if (IS_ERR_OR_NULL(lseg)) {
2377                 if (!lseg)
2378                         lseg = ERR_PTR(-ENOMEM);
2379
2380                 dprintk("%s: Could not allocate layout: error %ld\n",
2381                        __func__, PTR_ERR(lseg));
2382                 return lseg;
2383         }
2384
2385         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2386
2387         spin_lock(&ino->i_lock);
2388         if (pnfs_layoutgets_blocked(lo)) {
2389                 dprintk("%s forget reply due to state\n", __func__);
2390                 goto out_forget;
2391         }
2392
2393         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2394                 /* existing state ID, make sure the sequence number matches. */
2395                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2396                         if (!pnfs_layout_is_valid(lo) &&
2397                             pnfs_is_first_layoutget(lo))
2398                                 lo->plh_barrier = 0;
2399                         dprintk("%s forget reply due to sequence\n", __func__);
2400                         goto out_forget;
2401                 }
2402                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2403         } else if (pnfs_layout_is_valid(lo)) {
2404                 /*
2405                  * We got an entirely new state ID.  Mark all segments for the
2406                  * inode invalid, and retry the layoutget
2407                  */
2408                 struct pnfs_layout_range range = {
2409                         .iomode = IOMODE_ANY,
2410                         .length = NFS4_MAX_UINT64,
2411                 };
2412                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2413                 goto out_forget;
2414         } else {
2415                 /* We have a completely new layout */
2416                 if (!pnfs_is_first_layoutget(lo))
2417                         goto out_forget;
2418                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2419         }
2420
2421         pnfs_get_lseg(lseg);
2422         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2423
2424
2425         if (res->return_on_close)
2426                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2427
2428         spin_unlock(&ino->i_lock);
2429         pnfs_free_lseg_list(&free_me);
2430         return lseg;
2431
2432 out_forget:
2433         spin_unlock(&ino->i_lock);
2434         lseg->pls_layout = lo;
2435         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2436         return ERR_PTR(-EAGAIN);
2437 }
2438
2439 /**
2440  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2441  * @lo: pointer to layout header
2442  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2443  * @return_range: describe layout segment ranges to be returned
2444  * @seq: stateid seqid to match
2445  *
2446  * This function is mainly intended for use by layoutrecall. It attempts
2447  * to free the layout segment immediately, or else to mark it for return
2448  * as soon as its reference count drops to zero.
2449  *
2450  * Returns
2451  * - 0: a layoutreturn needs to be scheduled.
2452  * - EBUSY: there are layout segment that are still in use.
2453  * - ENOENT: there are no layout segments that need to be returned.
2454  */
2455 int
2456 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2457                                 struct list_head *tmp_list,
2458                                 const struct pnfs_layout_range *return_range,
2459                                 u32 seq)
2460 {
2461         struct pnfs_layout_segment *lseg, *next;
2462         int remaining = 0;
2463
2464         dprintk("%s:Begin lo %p\n", __func__, lo);
2465
2466         assert_spin_locked(&lo->plh_inode->i_lock);
2467
2468         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2469                 tmp_list = &lo->plh_return_segs;
2470
2471         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2472                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2473                         dprintk("%s: marking lseg %p iomode %d "
2474                                 "offset %llu length %llu\n", __func__,
2475                                 lseg, lseg->pls_range.iomode,
2476                                 lseg->pls_range.offset,
2477                                 lseg->pls_range.length);
2478                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2479                                 tmp_list = &lo->plh_return_segs;
2480                         if (mark_lseg_invalid(lseg, tmp_list))
2481                                 continue;
2482                         remaining++;
2483                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2484                 }
2485
2486         if (remaining) {
2487                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2488                 return -EBUSY;
2489         }
2490
2491         if (!list_empty(&lo->plh_return_segs)) {
2492                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2493                 return 0;
2494         }
2495
2496         return -ENOENT;
2497 }
2498
2499 static void
2500 pnfs_mark_layout_for_return(struct inode *inode,
2501                             const struct pnfs_layout_range *range)
2502 {
2503         struct pnfs_layout_hdr *lo;
2504         bool return_now = false;
2505
2506         spin_lock(&inode->i_lock);
2507         lo = NFS_I(inode)->layout;
2508         if (!pnfs_layout_is_valid(lo)) {
2509                 spin_unlock(&inode->i_lock);
2510                 return;
2511         }
2512         pnfs_set_plh_return_info(lo, range->iomode, 0);
2513         /*
2514          * mark all matching lsegs so that we are sure to have no live
2515          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2516          * for how it works.
2517          */
2518         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2519                 const struct cred *cred;
2520                 nfs4_stateid stateid;
2521                 enum pnfs_iomode iomode;
2522
2523                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2524                 spin_unlock(&inode->i_lock);
2525                 if (return_now)
2526                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2527         } else {
2528                 spin_unlock(&inode->i_lock);
2529                 nfs_commit_inode(inode, 0);
2530         }
2531 }
2532
2533 void pnfs_error_mark_layout_for_return(struct inode *inode,
2534                                        struct pnfs_layout_segment *lseg)
2535 {
2536         struct pnfs_layout_range range = {
2537                 .iomode = lseg->pls_range.iomode,
2538                 .offset = 0,
2539                 .length = NFS4_MAX_UINT64,
2540         };
2541
2542         pnfs_mark_layout_for_return(inode, &range);
2543 }
2544 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2545
2546 static bool
2547 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2548 {
2549         return pnfs_layout_is_valid(lo) &&
2550                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2551                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2552 }
2553
2554 static struct pnfs_layout_segment *
2555 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2556                      const struct pnfs_layout_range *range,
2557                      enum pnfs_iomode iomode)
2558 {
2559         struct pnfs_layout_segment *lseg;
2560
2561         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2562                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2563                         continue;
2564                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2565                         continue;
2566                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2567                         continue;
2568                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2569                         return lseg;
2570         }
2571         return NULL;
2572 }
2573
2574 /* Find open file states whose mode matches that of the range */
2575 static bool
2576 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2577                                  const struct pnfs_layout_range *range)
2578 {
2579         struct list_head *head;
2580         struct nfs_open_context *ctx;
2581         fmode_t mode = 0;
2582
2583         if (!pnfs_layout_can_be_returned(lo) ||
2584             !pnfs_find_first_lseg(lo, range, range->iomode))
2585                 return false;
2586
2587         head = &NFS_I(lo->plh_inode)->open_files;
2588         list_for_each_entry_rcu(ctx, head, list) {
2589                 if (ctx->state)
2590                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2591         }
2592
2593         switch (range->iomode) {
2594         default:
2595                 break;
2596         case IOMODE_READ:
2597                 mode &= ~FMODE_WRITE;
2598                 break;
2599         case IOMODE_RW:
2600                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2601                         mode &= ~FMODE_READ;
2602         }
2603         return mode == 0;
2604 }
2605
2606 static int
2607 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2608 {
2609         const struct pnfs_layout_range *range = data;
2610         struct pnfs_layout_hdr *lo;
2611         struct inode *inode;
2612 restart:
2613         rcu_read_lock();
2614         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2615                 if (!pnfs_layout_can_be_returned(lo) ||
2616                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2617                         continue;
2618                 inode = lo->plh_inode;
2619                 spin_lock(&inode->i_lock);
2620                 if (!pnfs_should_return_unused_layout(lo, range)) {
2621                         spin_unlock(&inode->i_lock);
2622                         continue;
2623                 }
2624                 spin_unlock(&inode->i_lock);
2625                 inode = pnfs_grab_inode_layout_hdr(lo);
2626                 if (!inode)
2627                         continue;
2628                 rcu_read_unlock();
2629                 pnfs_mark_layout_for_return(inode, range);
2630                 iput(inode);
2631                 cond_resched();
2632                 goto restart;
2633         }
2634         rcu_read_unlock();
2635         return 0;
2636 }
2637
2638 void
2639 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2640                                  enum pnfs_iomode iomode)
2641 {
2642         struct pnfs_layout_range range = {
2643                 .iomode = iomode,
2644                 .offset = 0,
2645                 .length = NFS4_MAX_UINT64,
2646         };
2647
2648         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2649                         &range);
2650 }
2651
2652 void
2653 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2654 {
2655         if (pgio->pg_lseg == NULL ||
2656             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2657                 return;
2658         pnfs_put_lseg(pgio->pg_lseg);
2659         pgio->pg_lseg = NULL;
2660 }
2661 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2662
2663 /*
2664  * Check for any intersection between the request and the pgio->pg_lseg,
2665  * and if none, put this pgio->pg_lseg away.
2666  */
2667 void
2668 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2669 {
2670         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2671                 pnfs_put_lseg(pgio->pg_lseg);
2672                 pgio->pg_lseg = NULL;
2673         }
2674 }
2675 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2676
2677 void
2678 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2679 {
2680         u64 rd_size;
2681
2682         pnfs_generic_pg_check_layout(pgio);
2683         pnfs_generic_pg_check_range(pgio, req);
2684         if (pgio->pg_lseg == NULL) {
2685                 if (pgio->pg_dreq == NULL)
2686                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2687                 else
2688                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2689
2690                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2691                                                    nfs_req_openctx(req),
2692                                                    req_offset(req),
2693                                                    rd_size,
2694                                                    IOMODE_READ,
2695                                                    false,
2696                                                    GFP_KERNEL);
2697                 if (IS_ERR(pgio->pg_lseg)) {
2698                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2699                         pgio->pg_lseg = NULL;
2700                         return;
2701                 }
2702         }
2703         /* If no lseg, fall back to read through mds */
2704         if (pgio->pg_lseg == NULL)
2705                 nfs_pageio_reset_read_mds(pgio);
2706
2707 }
2708 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2709
2710 void
2711 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2712                            struct nfs_page *req, u64 wb_size)
2713 {
2714         pnfs_generic_pg_check_layout(pgio);
2715         pnfs_generic_pg_check_range(pgio, req);
2716         if (pgio->pg_lseg == NULL) {
2717                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2718                                                    nfs_req_openctx(req),
2719                                                    req_offset(req),
2720                                                    wb_size,
2721                                                    IOMODE_RW,
2722                                                    false,
2723                                                    GFP_KERNEL);
2724                 if (IS_ERR(pgio->pg_lseg)) {
2725                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2726                         pgio->pg_lseg = NULL;
2727                         return;
2728                 }
2729         }
2730         /* If no lseg, fall back to write through mds */
2731         if (pgio->pg_lseg == NULL)
2732                 nfs_pageio_reset_write_mds(pgio);
2733 }
2734 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2735
2736 void
2737 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2738 {
2739         if (desc->pg_lseg) {
2740                 pnfs_put_lseg(desc->pg_lseg);
2741                 desc->pg_lseg = NULL;
2742         }
2743 }
2744 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2745
2746 /*
2747  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2748  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2749  */
2750 size_t
2751 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2752                      struct nfs_page *prev, struct nfs_page *req)
2753 {
2754         unsigned int size;
2755         u64 seg_end, req_start, seg_left;
2756
2757         size = nfs_generic_pg_test(pgio, prev, req);
2758         if (!size)
2759                 return 0;
2760
2761         /*
2762          * 'size' contains the number of bytes left in the current page (up
2763          * to the original size asked for in @req->wb_bytes).
2764          *
2765          * Calculate how many bytes are left in the layout segment
2766          * and if there are less bytes than 'size', return that instead.
2767          *
2768          * Please also note that 'end_offset' is actually the offset of the
2769          * first byte that lies outside the pnfs_layout_range. FIXME?
2770          *
2771          */
2772         if (pgio->pg_lseg) {
2773                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2774                                      pgio->pg_lseg->pls_range.length);
2775                 req_start = req_offset(req);
2776
2777                 /* start of request is past the last byte of this segment */
2778                 if (req_start >= seg_end)
2779                         return 0;
2780
2781                 /* adjust 'size' iff there are fewer bytes left in the
2782                  * segment than what nfs_generic_pg_test returned */
2783                 seg_left = seg_end - req_start;
2784                 if (seg_left < size)
2785                         size = (unsigned int)seg_left;
2786         }
2787
2788         return size;
2789 }
2790 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2791
2792 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2793 {
2794         struct nfs_pageio_descriptor pgio;
2795
2796         /* Resend all requests through the MDS */
2797         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2798                               hdr->completion_ops);
2799         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2800         return nfs_pageio_resend(&pgio, hdr);
2801 }
2802 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2803
2804 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2805 {
2806
2807         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2808         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2809             PNFS_LAYOUTRET_ON_ERROR) {
2810                 pnfs_return_layout(hdr->inode);
2811         }
2812         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2813                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2814 }
2815
2816 /*
2817  * Called by non rpc-based layout drivers
2818  */
2819 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2820 {
2821         if (likely(!hdr->pnfs_error)) {
2822                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2823                                 hdr->mds_offset + hdr->res.count);
2824                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2825         }
2826         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2827         if (unlikely(hdr->pnfs_error))
2828                 pnfs_ld_handle_write_error(hdr);
2829         hdr->mds_ops->rpc_release(hdr);
2830 }
2831 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2832
2833 static void
2834 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2835                 struct nfs_pgio_header *hdr)
2836 {
2837         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2838
2839         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2840                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2841                 nfs_pageio_reset_write_mds(desc);
2842                 mirror->pg_recoalesce = 1;
2843         }
2844         hdr->completion_ops->completion(hdr);
2845 }
2846
2847 static enum pnfs_try_status
2848 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2849                         const struct rpc_call_ops *call_ops,
2850                         struct pnfs_layout_segment *lseg,
2851                         int how)
2852 {
2853         struct inode *inode = hdr->inode;
2854         enum pnfs_try_status trypnfs;
2855         struct nfs_server *nfss = NFS_SERVER(inode);
2856
2857         hdr->mds_ops = call_ops;
2858
2859         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2860                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2861         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2862         if (trypnfs != PNFS_NOT_ATTEMPTED)
2863                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2864         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2865         return trypnfs;
2866 }
2867
2868 static void
2869 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2870               struct nfs_pgio_header *hdr, int how)
2871 {
2872         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2873         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2874         enum pnfs_try_status trypnfs;
2875
2876         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2877         switch (trypnfs) {
2878         case PNFS_NOT_ATTEMPTED:
2879                 pnfs_write_through_mds(desc, hdr);
2880                 break;
2881         case PNFS_ATTEMPTED:
2882                 break;
2883         case PNFS_TRY_AGAIN:
2884                 /* cleanup hdr and prepare to redo pnfs */
2885                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2886                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2887                         list_splice_init(&hdr->pages, &mirror->pg_list);
2888                         mirror->pg_recoalesce = 1;
2889                 }
2890                 hdr->mds_ops->rpc_release(hdr);
2891         }
2892 }
2893
2894 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2895 {
2896         pnfs_put_lseg(hdr->lseg);
2897         nfs_pgio_header_free(hdr);
2898 }
2899
2900 int
2901 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2902 {
2903         struct nfs_pgio_header *hdr;
2904         int ret;
2905
2906         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2907         if (!hdr) {
2908                 desc->pg_error = -ENOMEM;
2909                 return desc->pg_error;
2910         }
2911         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2912
2913         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2914         ret = nfs_generic_pgio(desc, hdr);
2915         if (!ret)
2916                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2917
2918         return ret;
2919 }
2920 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2921
2922 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2923 {
2924         struct nfs_pageio_descriptor pgio;
2925
2926         /* Resend all requests through the MDS */
2927         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2928         return nfs_pageio_resend(&pgio, hdr);
2929 }
2930 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2931
2932 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2933 {
2934         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2935         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2936             PNFS_LAYOUTRET_ON_ERROR) {
2937                 pnfs_return_layout(hdr->inode);
2938         }
2939         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2940                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2941 }
2942
2943 /*
2944  * Called by non rpc-based layout drivers
2945  */
2946 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2947 {
2948         if (likely(!hdr->pnfs_error))
2949                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2950         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2951         if (unlikely(hdr->pnfs_error))
2952                 pnfs_ld_handle_read_error(hdr);
2953         hdr->mds_ops->rpc_release(hdr);
2954 }
2955 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2956
2957 static void
2958 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2959                 struct nfs_pgio_header *hdr)
2960 {
2961         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2962
2963         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2964                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2965                 nfs_pageio_reset_read_mds(desc);
2966                 mirror->pg_recoalesce = 1;
2967         }
2968         hdr->completion_ops->completion(hdr);
2969 }
2970
2971 /*
2972  * Call the appropriate parallel I/O subsystem read function.
2973  */
2974 static enum pnfs_try_status
2975 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2976                        const struct rpc_call_ops *call_ops,
2977                        struct pnfs_layout_segment *lseg)
2978 {
2979         struct inode *inode = hdr->inode;
2980         struct nfs_server *nfss = NFS_SERVER(inode);
2981         enum pnfs_try_status trypnfs;
2982
2983         hdr->mds_ops = call_ops;
2984
2985         dprintk("%s: Reading ino:%lu %u@%llu\n",
2986                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2987
2988         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2989         if (trypnfs != PNFS_NOT_ATTEMPTED)
2990                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2991         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2992         return trypnfs;
2993 }
2994
2995 /* Resend all requests through pnfs. */
2996 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
2997                            unsigned int mirror_idx)
2998 {
2999         struct nfs_pageio_descriptor pgio;
3000
3001         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3002                 /* Prevent deadlocks with layoutreturn! */
3003                 pnfs_put_lseg(hdr->lseg);
3004                 hdr->lseg = NULL;
3005
3006                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3007                                         hdr->completion_ops);
3008                 pgio.pg_mirror_idx = mirror_idx;
3009                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3010         }
3011 }
3012 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3013
3014 static void
3015 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3016 {
3017         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3018         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3019         enum pnfs_try_status trypnfs;
3020
3021         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3022         switch (trypnfs) {
3023         case PNFS_NOT_ATTEMPTED:
3024                 pnfs_read_through_mds(desc, hdr);
3025                 break;
3026         case PNFS_ATTEMPTED:
3027                 break;
3028         case PNFS_TRY_AGAIN:
3029                 /* cleanup hdr and prepare to redo pnfs */
3030                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3031                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3032                         list_splice_init(&hdr->pages, &mirror->pg_list);
3033                         mirror->pg_recoalesce = 1;
3034                 }
3035                 hdr->mds_ops->rpc_release(hdr);
3036         }
3037 }
3038
3039 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3040 {
3041         pnfs_put_lseg(hdr->lseg);
3042         nfs_pgio_header_free(hdr);
3043 }
3044
3045 int
3046 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3047 {
3048         struct nfs_pgio_header *hdr;
3049         int ret;
3050
3051         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3052         if (!hdr) {
3053                 desc->pg_error = -ENOMEM;
3054                 return desc->pg_error;
3055         }
3056         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3057         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3058         ret = nfs_generic_pgio(desc, hdr);
3059         if (!ret)
3060                 pnfs_do_read(desc, hdr);
3061         return ret;
3062 }
3063 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3064
3065 static void pnfs_clear_layoutcommitting(struct inode *inode)
3066 {
3067         unsigned long *bitlock = &NFS_I(inode)->flags;
3068
3069         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3070         smp_mb__after_atomic();
3071         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3072 }
3073
3074 /*
3075  * There can be multiple RW segments.
3076  */
3077 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3078 {
3079         struct pnfs_layout_segment *lseg;
3080
3081         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3082                 if (lseg->pls_range.iomode == IOMODE_RW &&
3083                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3084                         list_add(&lseg->pls_lc_list, listp);
3085         }
3086 }
3087
3088 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3089 {
3090         struct pnfs_layout_segment *lseg, *tmp;
3091
3092         /* Matched by references in pnfs_set_layoutcommit */
3093         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3094                 list_del_init(&lseg->pls_lc_list);
3095                 pnfs_put_lseg(lseg);
3096         }
3097
3098         pnfs_clear_layoutcommitting(inode);
3099 }
3100
3101 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3102 {
3103         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3104 }
3105 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3106
3107 void
3108 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3109                 loff_t end_pos)
3110 {
3111         struct nfs_inode *nfsi = NFS_I(inode);
3112         bool mark_as_dirty = false;
3113
3114         spin_lock(&inode->i_lock);
3115         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3116                 nfsi->layout->plh_lwb = end_pos;
3117                 mark_as_dirty = true;
3118                 dprintk("%s: Set layoutcommit for inode %lu ",
3119                         __func__, inode->i_ino);
3120         } else if (end_pos > nfsi->layout->plh_lwb)
3121                 nfsi->layout->plh_lwb = end_pos;
3122         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3123                 /* references matched in nfs4_layoutcommit_release */
3124                 pnfs_get_lseg(lseg);
3125         }
3126         spin_unlock(&inode->i_lock);
3127         dprintk("%s: lseg %p end_pos %llu\n",
3128                 __func__, lseg, nfsi->layout->plh_lwb);
3129
3130         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3131          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3132         if (mark_as_dirty)
3133                 mark_inode_dirty_sync(inode);
3134 }
3135 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3136
3137 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3138 {
3139         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3140
3141         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3142                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3143         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3144 }
3145
3146 /*
3147  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3148  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3149  * data to disk to allow the server to recover the data if it crashes.
3150  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3151  * is off, and a COMMIT is sent to a data server, or
3152  * if WRITEs to a data server return NFS_DATA_SYNC.
3153  */
3154 int
3155 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3156 {
3157         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3158         struct nfs4_layoutcommit_data *data;
3159         struct nfs_inode *nfsi = NFS_I(inode);
3160         loff_t end_pos;
3161         int status;
3162
3163         if (!pnfs_layoutcommit_outstanding(inode))
3164                 return 0;
3165
3166         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3167
3168         status = -EAGAIN;
3169         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3170                 if (!sync)
3171                         goto out;
3172                 status = wait_on_bit_lock_action(&nfsi->flags,
3173                                 NFS_INO_LAYOUTCOMMITTING,
3174                                 nfs_wait_bit_killable,
3175                                 TASK_KILLABLE);
3176                 if (status)
3177                         goto out;
3178         }
3179
3180         status = -ENOMEM;
3181         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3182         data = kzalloc(sizeof(*data), GFP_NOFS);
3183         if (!data)
3184                 goto clear_layoutcommitting;
3185
3186         status = 0;
3187         spin_lock(&inode->i_lock);
3188         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3189                 goto out_unlock;
3190
3191         INIT_LIST_HEAD(&data->lseg_list);
3192         pnfs_list_write_lseg(inode, &data->lseg_list);
3193
3194         end_pos = nfsi->layout->plh_lwb;
3195
3196         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3197         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3198         spin_unlock(&inode->i_lock);
3199
3200         data->args.inode = inode;
3201         nfs_fattr_init(&data->fattr);
3202         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3203         data->res.fattr = &data->fattr;
3204         if (end_pos != 0)
3205                 data->args.lastbytewritten = end_pos - 1;
3206         else
3207                 data->args.lastbytewritten = U64_MAX;
3208         data->res.server = NFS_SERVER(inode);
3209
3210         if (ld->prepare_layoutcommit) {
3211                 status = ld->prepare_layoutcommit(&data->args);
3212                 if (status) {
3213                         put_cred(data->cred);
3214                         spin_lock(&inode->i_lock);
3215                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3216                         if (end_pos > nfsi->layout->plh_lwb)
3217                                 nfsi->layout->plh_lwb = end_pos;
3218                         goto out_unlock;
3219                 }
3220         }
3221
3222
3223         status = nfs4_proc_layoutcommit(data, sync);
3224 out:
3225         if (status)
3226                 mark_inode_dirty_sync(inode);
3227         dprintk("<-- %s status %d\n", __func__, status);
3228         return status;
3229 out_unlock:
3230         spin_unlock(&inode->i_lock);
3231         kfree(data);
3232 clear_layoutcommitting:
3233         pnfs_clear_layoutcommitting(inode);
3234         goto out;
3235 }
3236 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3237
3238 int
3239 pnfs_generic_sync(struct inode *inode, bool datasync)
3240 {
3241         return pnfs_layoutcommit_inode(inode, true);
3242 }
3243 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3244
3245 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3246 {
3247         struct nfs4_threshold *thp;
3248
3249         thp = kzalloc(sizeof(*thp), GFP_NOFS);
3250         if (!thp) {
3251                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3252                 return NULL;
3253         }
3254         return thp;
3255 }
3256
3257 #if IS_ENABLED(CONFIG_NFS_V4_2)
3258 int
3259 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3260 {
3261         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3262         struct nfs_server *server = NFS_SERVER(inode);
3263         struct nfs_inode *nfsi = NFS_I(inode);
3264         struct nfs42_layoutstat_data *data;
3265         struct pnfs_layout_hdr *hdr;
3266         int status = 0;
3267
3268         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3269                 goto out;
3270
3271         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3272                 goto out;
3273
3274         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3275                 goto out;
3276
3277         spin_lock(&inode->i_lock);
3278         if (!NFS_I(inode)->layout) {
3279                 spin_unlock(&inode->i_lock);
3280                 goto out_clear_layoutstats;
3281         }
3282         hdr = NFS_I(inode)->layout;
3283         pnfs_get_layout_hdr(hdr);
3284         spin_unlock(&inode->i_lock);
3285
3286         data = kzalloc(sizeof(*data), gfp_flags);
3287         if (!data) {
3288                 status = -ENOMEM;
3289                 goto out_put;
3290         }
3291
3292         data->args.fh = NFS_FH(inode);
3293         data->args.inode = inode;
3294         status = ld->prepare_layoutstats(&data->args);
3295         if (status)
3296                 goto out_free;
3297
3298         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3299
3300 out:
3301         dprintk("%s returns %d\n", __func__, status);
3302         return status;
3303
3304 out_free:
3305         kfree(data);
3306 out_put:
3307         pnfs_put_layout_hdr(hdr);
3308 out_clear_layoutstats:
3309         smp_mb__before_atomic();
3310         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3311         smp_mb__after_atomic();
3312         goto out;
3313 }
3314 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3315 #endif
3316
3317 unsigned int layoutstats_timer;
3318 module_param(layoutstats_timer, uint, 0644);
3319 EXPORT_SYMBOL_GPL(layoutstats_timer);