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 = be32_to_cpu(lo->plh_stateid.seqid);
 970         u32 newseq = be32_to_cpu(new->seqid);
 971
 972         if (!pnfs_layout_is_valid(lo)) {
 973                 pnfs_set_layout_cred(lo, cred);
 974                 nfs4_stateid_copy(&lo->plh_stateid, new);
 975                 lo->plh_barrier = newseq;
 976                 pnfs_clear_layoutreturn_info(lo);
 977                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 978                 return;
 979         }
 980
 981         if (pnfs_seqid_is_newer(newseq, oldseq))
 982                 nfs4_stateid_copy(&lo->plh_stateid, new);
 983
 984         if (update_barrier) {
 985                 pnfs_barrier_update(lo, newseq);
 986                 return;
 987         }
 988         /*
 989          * Because of wraparound, we want to keep the barrier
 990          * "close" to the current seqids. We really only want to
 991          * get here from a layoutget call.
 992          */
 993         if (atomic_read(&lo->plh_outstanding) == 1)
 994                  pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
 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         pnfs_put_layout_hdr(lgp->lo);
1132         put_nfs_open_context(lgp->args.ctx);
1133         kfree(lgp);
1134 }
1135
1136 static void pnfs_clear_layoutcommit(struct inode *inode,
1137                 struct list_head *head)
1138 {
1139         struct nfs_inode *nfsi = NFS_I(inode);
1140         struct pnfs_layout_segment *lseg, *tmp;
1141
1142         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1143                 return;
1144         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1145                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1146                         continue;
1147                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1148         }
1149 }
1150
1151 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1152                 const nfs4_stateid *arg_stateid,
1153                 const struct pnfs_layout_range *range,
1154                 const nfs4_stateid *stateid)
1155 {
1156         struct inode *inode = lo->plh_inode;
1157         LIST_HEAD(freeme);
1158
1159         spin_lock(&inode->i_lock);
1160         if (!pnfs_layout_is_valid(lo) ||
1161             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1162                 goto out_unlock;
1163         if (stateid) {
1164                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1165
1166                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1167                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1168                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1169         } else
1170                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1171 out_unlock:
1172         pnfs_clear_layoutreturn_waitbit(lo);
1173         spin_unlock(&inode->i_lock);
1174         pnfs_free_lseg_list(&freeme);
1175
1176 }
1177
1178 static bool
1179 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1180                 nfs4_stateid *stateid,
1181                 const struct cred **cred,
1182                 enum pnfs_iomode *iomode)
1183 {
1184         /* Serialise LAYOUTGET/LAYOUTRETURN */
1185         if (atomic_read(&lo->plh_outstanding) != 0)
1186                 return false;
1187         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1188                 return false;
1189         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1190         pnfs_get_layout_hdr(lo);
1191         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1192         *cred = get_cred(lo->plh_lc_cred);
1193         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1194                 if (lo->plh_return_seq != 0)
1195                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1196                 if (iomode != NULL)
1197                         *iomode = lo->plh_return_iomode;
1198                 pnfs_clear_layoutreturn_info(lo);
1199         } else if (iomode != NULL)
1200                 *iomode = IOMODE_ANY;
1201         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1202         return true;
1203 }
1204
1205 static void
1206 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1207                 struct pnfs_layout_hdr *lo,
1208                 const nfs4_stateid *stateid,
1209                 enum pnfs_iomode iomode)
1210 {
1211         struct inode *inode = lo->plh_inode;
1212
1213         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1214         args->inode = inode;
1215         args->range.iomode = iomode;
1216         args->range.offset = 0;
1217         args->range.length = NFS4_MAX_UINT64;
1218         args->layout = lo;
1219         nfs4_stateid_copy(&args->stateid, stateid);
1220 }
1221
1222 static int
1223 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1224                        const nfs4_stateid *stateid,
1225                        const struct cred **pcred,
1226                        enum pnfs_iomode iomode,
1227                        bool sync)
1228 {
1229         struct inode *ino = lo->plh_inode;
1230         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1231         struct nfs4_layoutreturn *lrp;
1232         const struct cred *cred = *pcred;
1233         int status = 0;
1234
1235         *pcred = NULL;
1236         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1237         if (unlikely(lrp == NULL)) {
1238                 status = -ENOMEM;
1239                 spin_lock(&ino->i_lock);
1240                 pnfs_clear_layoutreturn_waitbit(lo);
1241                 spin_unlock(&ino->i_lock);
1242                 put_cred(cred);
1243                 pnfs_put_layout_hdr(lo);
1244                 goto out;
1245         }
1246
1247         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1248         lrp->args.ld_private = &lrp->ld_private;
1249         lrp->clp = NFS_SERVER(ino)->nfs_client;
1250         lrp->cred = cred;
1251         if (ld->prepare_layoutreturn)
1252                 ld->prepare_layoutreturn(&lrp->args);
1253
1254         status = nfs4_proc_layoutreturn(lrp, sync);
1255 out:
1256         dprintk("<-- %s status: %d\n", __func__, status);
1257         return status;
1258 }
1259
1260 static bool
1261 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1262                                 enum pnfs_iomode iomode,
1263                                 u32 seq)
1264 {
1265         struct pnfs_layout_range recall_range = {
1266                 .length = NFS4_MAX_UINT64,
1267                 .iomode = iomode,
1268         };
1269         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1270                                                &recall_range, seq) != -EBUSY;
1271 }
1272
1273 /* Return true if layoutreturn is needed */
1274 static bool
1275 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1276 {
1277         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1278                 return false;
1279         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1280                                                lo->plh_return_seq);
1281 }
1282
1283 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1284 {
1285         struct inode *inode= lo->plh_inode;
1286
1287         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1288                 return;
1289         spin_lock(&inode->i_lock);
1290         if (pnfs_layout_need_return(lo)) {
1291                 const struct cred *cred;
1292                 nfs4_stateid stateid;
1293                 enum pnfs_iomode iomode;
1294                 bool send;
1295
1296                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1297                 spin_unlock(&inode->i_lock);
1298                 if (send) {
1299                         /* Send an async layoutreturn so we dont deadlock */
1300                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1301                 }
1302         } else
1303                 spin_unlock(&inode->i_lock);
1304 }
1305
1306 /*
1307  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1308  * when the layout segment list is empty.
1309  *
1310  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1311  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1312  * deviceid is marked invalid.
1313  */
1314 int
1315 _pnfs_return_layout(struct inode *ino)
1316 {
1317         struct pnfs_layout_hdr *lo = NULL;
1318         struct nfs_inode *nfsi = NFS_I(ino);
1319         struct pnfs_layout_range range = {
1320                 .iomode         = IOMODE_ANY,
1321                 .offset         = 0,
1322                 .length         = NFS4_MAX_UINT64,
1323         };
1324         LIST_HEAD(tmp_list);
1325         const struct cred *cred;
1326         nfs4_stateid stateid;
1327         int status = 0;
1328         bool send, valid_layout;
1329
1330         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1331
1332         spin_lock(&ino->i_lock);
1333         lo = nfsi->layout;
1334         if (!lo) {
1335                 spin_unlock(&ino->i_lock);
1336                 dprintk("NFS: %s no layout to return\n", __func__);
1337                 goto out;
1338         }
1339         /* Reference matched in nfs4_layoutreturn_release */
1340         pnfs_get_layout_hdr(lo);
1341         /* Is there an outstanding layoutreturn ? */
1342         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1343                 spin_unlock(&ino->i_lock);
1344                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1345                                         TASK_UNINTERRUPTIBLE))
1346                         goto out_put_layout_hdr;
1347                 spin_lock(&ino->i_lock);
1348         }
1349         valid_layout = pnfs_layout_is_valid(lo);
1350         pnfs_clear_layoutcommit(ino, &tmp_list);
1351         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1352
1353         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1354                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1355
1356         /* Don't send a LAYOUTRETURN if list was initially empty */
1357         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1358                         !valid_layout) {
1359                 spin_unlock(&ino->i_lock);
1360                 dprintk("NFS: %s no layout segments to return\n", __func__);
1361                 goto out_wait_layoutreturn;
1362         }
1363
1364         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1365         spin_unlock(&ino->i_lock);
1366         if (send)
1367                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1368 out_wait_layoutreturn:
1369         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1370 out_put_layout_hdr:
1371         pnfs_free_lseg_list(&tmp_list);
1372         pnfs_put_layout_hdr(lo);
1373 out:
1374         dprintk("<-- %s status: %d\n", __func__, status);
1375         return status;
1376 }
1377
1378 int
1379 pnfs_commit_and_return_layout(struct inode *inode)
1380 {
1381         struct pnfs_layout_hdr *lo;
1382         int ret;
1383
1384         spin_lock(&inode->i_lock);
1385         lo = NFS_I(inode)->layout;
1386         if (lo == NULL) {
1387                 spin_unlock(&inode->i_lock);
1388                 return 0;
1389         }
1390         pnfs_get_layout_hdr(lo);
1391         /* Block new layoutgets and read/write to ds */
1392         lo->plh_block_lgets++;
1393         spin_unlock(&inode->i_lock);
1394         filemap_fdatawait(inode->i_mapping);
1395         ret = pnfs_layoutcommit_inode(inode, true);
1396         if (ret == 0)
1397                 ret = _pnfs_return_layout(inode);
1398         spin_lock(&inode->i_lock);
1399         lo->plh_block_lgets--;
1400         spin_unlock(&inode->i_lock);
1401         pnfs_put_layout_hdr(lo);
1402         return ret;
1403 }
1404
1405 bool pnfs_roc(struct inode *ino,
1406                 struct nfs4_layoutreturn_args *args,
1407                 struct nfs4_layoutreturn_res *res,
1408                 const struct cred *cred)
1409 {
1410         struct nfs_inode *nfsi = NFS_I(ino);
1411         struct nfs_open_context *ctx;
1412         struct nfs4_state *state;
1413         struct pnfs_layout_hdr *lo;
1414         struct pnfs_layout_segment *lseg, *next;
1415         const struct cred *lc_cred;
1416         nfs4_stateid stateid;
1417         enum pnfs_iomode iomode = 0;
1418         bool layoutreturn = false, roc = false;
1419         bool skip_read = false;
1420
1421         if (!nfs_have_layout(ino))
1422                 return false;
1423 retry:
1424         rcu_read_lock();
1425         spin_lock(&ino->i_lock);
1426         lo = nfsi->layout;
1427         if (!lo || !pnfs_layout_is_valid(lo) ||
1428             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1429                 lo = NULL;
1430                 goto out_noroc;
1431         }
1432         pnfs_get_layout_hdr(lo);
1433         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1434                 spin_unlock(&ino->i_lock);
1435                 rcu_read_unlock();
1436                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1437                                 TASK_UNINTERRUPTIBLE);
1438                 pnfs_put_layout_hdr(lo);
1439                 goto retry;
1440         }
1441
1442         /* no roc if we hold a delegation */
1443         if (nfs4_check_delegation(ino, FMODE_READ)) {
1444                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1445                         goto out_noroc;
1446                 skip_read = true;
1447         }
1448
1449         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1450                 state = ctx->state;
1451                 if (state == NULL)
1452                         continue;
1453                 /* Don't return layout if there is open file state */
1454                 if (state->state & FMODE_WRITE)
1455                         goto out_noroc;
1456                 if (state->state & FMODE_READ)
1457                         skip_read = true;
1458         }
1459
1460
1461         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1462                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1463                         continue;
1464                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1465                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1466                         continue;
1467                 /*
1468                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1469                  * doesn't invalidate the layout for us.
1470                  */
1471                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1472                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1473                         continue;
1474                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1475         }
1476
1477         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1478                 goto out_noroc;
1479
1480         /* ROC in two conditions:
1481          * 1. there are ROC lsegs
1482          * 2. we don't send layoutreturn
1483          */
1484         /* lo ref dropped in pnfs_roc_release() */
1485         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1486         /* If the creds don't match, we can't compound the layoutreturn */
1487         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1488                 goto out_noroc;
1489
1490         roc = layoutreturn;
1491         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1492         res->lrs_present = 0;
1493         layoutreturn = false;
1494         put_cred(lc_cred);
1495
1496 out_noroc:
1497         spin_unlock(&ino->i_lock);
1498         rcu_read_unlock();
1499         pnfs_layoutcommit_inode(ino, true);
1500         if (roc) {
1501                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1502                 if (ld->prepare_layoutreturn)
1503                         ld->prepare_layoutreturn(args);
1504                 pnfs_put_layout_hdr(lo);
1505                 return true;
1506         }
1507         if (layoutreturn)
1508                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1509         pnfs_put_layout_hdr(lo);
1510         return false;
1511 }
1512
1513 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1514                   struct nfs4_layoutreturn_res **respp, int *ret)
1515 {
1516         struct nfs4_layoutreturn_args *arg = *argpp;
1517         int retval = -EAGAIN;
1518
1519         if (!arg)
1520                 return 0;
1521         /* Handle Layoutreturn errors */
1522         switch (*ret) {
1523         case 0:
1524                 retval = 0;
1525                 break;
1526         case -NFS4ERR_NOMATCHING_LAYOUT:
1527                 /* Was there an RPC level error? If not, retry */
1528                 if (task->tk_rpc_status == 0)
1529                         break;
1530                 /* If the call was not sent, let caller handle it */
1531                 if (!RPC_WAS_SENT(task))
1532                         return 0;
1533                 /*
1534                  * Otherwise, assume the call succeeded and
1535                  * that we need to release the layout
1536                  */
1537                 *ret = 0;
1538                 (*respp)->lrs_present = 0;
1539                 retval = 0;
1540                 break;
1541         case -NFS4ERR_DELAY:
1542                 /* Let the caller handle the retry */
1543                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1544                 return 0;
1545         case -NFS4ERR_OLD_STATEID:
1546                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1547                                                      &arg->range, arg->inode))
1548                         break;
1549                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1550                 return -EAGAIN;
1551         }
1552         *argpp = NULL;
1553         *respp = NULL;
1554         return retval;
1555 }
1556
1557 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1558                 struct nfs4_layoutreturn_res *res,
1559                 int ret)
1560 {
1561         struct pnfs_layout_hdr *lo = args->layout;
1562         struct inode *inode = args->inode;
1563         const nfs4_stateid *res_stateid = NULL;
1564         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1565
1566         switch (ret) {
1567         case -NFS4ERR_NOMATCHING_LAYOUT:
1568                 spin_lock(&inode->i_lock);
1569                 if (pnfs_layout_is_valid(lo) &&
1570                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1571                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1572                 pnfs_clear_layoutreturn_waitbit(lo);
1573                 spin_unlock(&inode->i_lock);
1574                 break;
1575         case 0:
1576                 if (res->lrs_present)
1577                         res_stateid = &res->stateid;
1578                 fallthrough;
1579         default:
1580                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1581                                              res_stateid);
1582         }
1583         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1584         if (ld_private && ld_private->ops && ld_private->ops->free)
1585                 ld_private->ops->free(ld_private);
1586         pnfs_put_layout_hdr(lo);
1587 }
1588
1589 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1590 {
1591         struct nfs_inode *nfsi = NFS_I(ino);
1592         struct pnfs_layout_hdr *lo;
1593         bool sleep = false;
1594
1595         /* we might not have grabbed lo reference. so need to check under
1596          * i_lock */
1597         spin_lock(&ino->i_lock);
1598         lo = nfsi->layout;
1599         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1600                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1601                 sleep = true;
1602         }
1603         spin_unlock(&ino->i_lock);
1604         return sleep;
1605 }
1606
1607 /*
1608  * Compare two layout segments for sorting into layout cache.
1609  * We want to preferentially return RW over RO layouts, so ensure those
1610  * are seen first.
1611  */
1612 static s64
1613 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1614            const struct pnfs_layout_range *l2)
1615 {
1616         s64 d;
1617
1618         /* high offset > low offset */
1619         d = l1->offset - l2->offset;
1620         if (d)
1621                 return d;
1622
1623         /* short length > long length */
1624         d = l2->length - l1->length;
1625         if (d)
1626                 return d;
1627
1628         /* read > read/write */
1629         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1630 }
1631
1632 static bool
1633 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1634                 const struct pnfs_layout_range *l2)
1635 {
1636         return pnfs_lseg_range_cmp(l1, l2) > 0;
1637 }
1638
1639 static bool
1640 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1641                 struct pnfs_layout_segment *old)
1642 {
1643         return false;
1644 }
1645
1646 void
1647 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1648                    struct pnfs_layout_segment *lseg,
1649                    bool (*is_after)(const struct pnfs_layout_range *,
1650                            const struct pnfs_layout_range *),
1651                    bool (*do_merge)(struct pnfs_layout_segment *,
1652                            struct pnfs_layout_segment *),
1653                    struct list_head *free_me)
1654 {
1655         struct pnfs_layout_segment *lp, *tmp;
1656
1657         dprintk("%s:Begin\n", __func__);
1658
1659         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1660                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1661                         continue;
1662                 if (do_merge(lseg, lp)) {
1663                         mark_lseg_invalid(lp, free_me);
1664                         continue;
1665                 }
1666                 if (is_after(&lseg->pls_range, &lp->pls_range))
1667                         continue;
1668                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1669                 dprintk("%s: inserted lseg %p "
1670                         "iomode %d offset %llu length %llu before "
1671                         "lp %p iomode %d offset %llu length %llu\n",
1672                         __func__, lseg, lseg->pls_range.iomode,
1673                         lseg->pls_range.offset, lseg->pls_range.length,
1674                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1675                         lp->pls_range.length);
1676                 goto out;
1677         }
1678         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1679         dprintk("%s: inserted lseg %p "
1680                 "iomode %d offset %llu length %llu at tail\n",
1681                 __func__, lseg, lseg->pls_range.iomode,
1682                 lseg->pls_range.offset, lseg->pls_range.length);
1683 out:
1684         pnfs_get_layout_hdr(lo);
1685
1686         dprintk("%s:Return\n", __func__);
1687 }
1688 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1689
1690 static void
1691 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1692                    struct pnfs_layout_segment *lseg,
1693                    struct list_head *free_me)
1694 {
1695         struct inode *inode = lo->plh_inode;
1696         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1697
1698         if (ld->add_lseg != NULL)
1699                 ld->add_lseg(lo, lseg, free_me);
1700         else
1701                 pnfs_generic_layout_insert_lseg(lo, lseg,
1702                                 pnfs_lseg_range_is_after,
1703                                 pnfs_lseg_no_merge,
1704                                 free_me);
1705 }
1706
1707 static struct pnfs_layout_hdr *
1708 alloc_init_layout_hdr(struct inode *ino,
1709                       struct nfs_open_context *ctx,
1710                       gfp_t gfp_flags)
1711 {
1712         struct pnfs_layout_hdr *lo;
1713
1714         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1715         if (!lo)
1716                 return NULL;
1717         refcount_set(&lo->plh_refcount, 1);
1718         INIT_LIST_HEAD(&lo->plh_layouts);
1719         INIT_LIST_HEAD(&lo->plh_segs);
1720         INIT_LIST_HEAD(&lo->plh_return_segs);
1721         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1722         lo->plh_inode = ino;
1723         lo->plh_lc_cred = get_cred(ctx->cred);
1724         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1725         return lo;
1726 }
1727
1728 static struct pnfs_layout_hdr *
1729 pnfs_find_alloc_layout(struct inode *ino,
1730                        struct nfs_open_context *ctx,
1731                        gfp_t gfp_flags)
1732         __releases(&ino->i_lock)
1733         __acquires(&ino->i_lock)
1734 {
1735         struct nfs_inode *nfsi = NFS_I(ino);
1736         struct pnfs_layout_hdr *new = NULL;
1737
1738         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1739
1740         if (nfsi->layout != NULL)
1741                 goto out_existing;
1742         spin_unlock(&ino->i_lock);
1743         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1744         spin_lock(&ino->i_lock);
1745
1746         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1747                 nfsi->layout = new;
1748                 return new;
1749         } else if (new != NULL)
1750                 pnfs_free_layout_hdr(new);
1751 out_existing:
1752         pnfs_get_layout_hdr(nfsi->layout);
1753         return nfsi->layout;
1754 }
1755
1756 /*
1757  * iomode matching rules:
1758  * iomode       lseg    strict match
1759  *                      iomode
1760  * -----        -----   ------ -----
1761  * ANY          READ    N/A    true
1762  * ANY          RW      N/A    true
1763  * RW           READ    N/A    false
1764  * RW           RW      N/A    true
1765  * READ         READ    N/A    true
1766  * READ         RW      true   false
1767  * READ         RW      false  true
1768  */
1769 static bool
1770 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1771                  const struct pnfs_layout_range *range,
1772                  bool strict_iomode)
1773 {
1774         struct pnfs_layout_range range1;
1775
1776         if ((range->iomode == IOMODE_RW &&
1777              ls_range->iomode != IOMODE_RW) ||
1778             (range->iomode != ls_range->iomode &&
1779              strict_iomode) ||
1780             !pnfs_lseg_range_intersecting(ls_range, range))
1781                 return false;
1782
1783         /* range1 covers only the first byte in the range */
1784         range1 = *range;
1785         range1.length = 1;
1786         return pnfs_lseg_range_contained(ls_range, &range1);
1787 }
1788
1789 /*
1790  * lookup range in layout
1791  */
1792 static struct pnfs_layout_segment *
1793 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1794                 struct pnfs_layout_range *range,
1795                 bool strict_iomode)
1796 {
1797         struct pnfs_layout_segment *lseg, *ret = NULL;
1798
1799         dprintk("%s:Begin\n", __func__);
1800
1801         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1802                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1803                     pnfs_lseg_range_match(&lseg->pls_range, range,
1804                                           strict_iomode)) {
1805                         ret = pnfs_get_lseg(lseg);
1806                         break;
1807                 }
1808         }
1809
1810         dprintk("%s:Return lseg %p ref %d\n",
1811                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1812         return ret;
1813 }
1814
1815 /*
1816  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1817  * to the MDS or over pNFS
1818  *
1819  * The nfs_inode read_io and write_io fields are cumulative counters reset
1820  * when there are no layout segments. Note that in pnfs_update_layout iomode
1821  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1822  * WRITE request.
1823  *
1824  * A return of true means use MDS I/O.
1825  *
1826  * From rfc 5661:
1827  * If a file's size is smaller than the file size threshold, data accesses
1828  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1829  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1830  * server.  If both file size and I/O size are provided, the client SHOULD
1831  * reach or exceed  both thresholds before sending its read or write
1832  * requests to the data server.
1833  */
1834 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1835                                      struct inode *ino, int iomode)
1836 {
1837         struct nfs4_threshold *t = ctx->mdsthreshold;
1838         struct nfs_inode *nfsi = NFS_I(ino);
1839         loff_t fsize = i_size_read(ino);
1840         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1841
1842         if (t == NULL)
1843                 return ret;
1844
1845         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1846                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1847
1848         switch (iomode) {
1849         case IOMODE_READ:
1850                 if (t->bm & THRESHOLD_RD) {
1851                         dprintk("%s fsize %llu\n", __func__, fsize);
1852                         size_set = true;
1853                         if (fsize < t->rd_sz)
1854                                 size = true;
1855                 }
1856                 if (t->bm & THRESHOLD_RD_IO) {
1857                         dprintk("%s nfsi->read_io %llu\n", __func__,
1858                                 nfsi->read_io);
1859                         io_set = true;
1860                         if (nfsi->read_io < t->rd_io_sz)
1861                                 io = true;
1862                 }
1863                 break;
1864         case IOMODE_RW:
1865                 if (t->bm & THRESHOLD_WR) {
1866                         dprintk("%s fsize %llu\n", __func__, fsize);
1867                         size_set = true;
1868                         if (fsize < t->wr_sz)
1869                                 size = true;
1870                 }
1871                 if (t->bm & THRESHOLD_WR_IO) {
1872                         dprintk("%s nfsi->write_io %llu\n", __func__,
1873                                 nfsi->write_io);
1874                         io_set = true;
1875                         if (nfsi->write_io < t->wr_io_sz)
1876                                 io = true;
1877                 }
1878                 break;
1879         }
1880         if (size_set && io_set) {
1881                 if (size && io)
1882                         ret = true;
1883         } else if (size || io)
1884                 ret = true;
1885
1886         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1887         return ret;
1888 }
1889
1890 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1891 {
1892         /*
1893          * send layoutcommit as it can hold up layoutreturn due to lseg
1894          * reference
1895          */
1896         pnfs_layoutcommit_inode(lo->plh_inode, false);
1897         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1898                                    nfs_wait_bit_killable,
1899                                    TASK_KILLABLE);
1900 }
1901
1902 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1903 {
1904         atomic_inc(&lo->plh_outstanding);
1905 }
1906
1907 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1908 {
1909         if (atomic_dec_and_test(&lo->plh_outstanding))
1910                 wake_up_var(&lo->plh_outstanding);
1911 }
1912
1913 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1914 {
1915         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1916 }
1917
1918 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1919 {
1920         unsigned long *bitlock = &lo->plh_flags;
1921
1922         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1923         smp_mb__after_atomic();
1924         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1925 }
1926
1927 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1928                                 struct nfs_server *server)
1929 {
1930         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1931                 struct nfs_client *clp = server->nfs_client;
1932
1933                 /* The lo must be on the clp list if there is any
1934                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1935                  */
1936                 spin_lock(&clp->cl_lock);
1937                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1938                 spin_unlock(&clp->cl_lock);
1939         }
1940 }
1941
1942 /*
1943  * Layout segment is retreived from the server if not cached.
1944  * The appropriate layout segment is referenced and returned to the caller.
1945  */
1946 struct pnfs_layout_segment *
1947 pnfs_update_layout(struct inode *ino,
1948                    struct nfs_open_context *ctx,
1949                    loff_t pos,
1950                    u64 count,
1951                    enum pnfs_iomode iomode,
1952                    bool strict_iomode,
1953                    gfp_t gfp_flags)
1954 {
1955         struct pnfs_layout_range arg = {
1956                 .iomode = iomode,
1957                 .offset = pos,
1958                 .length = count,
1959         };
1960         unsigned pg_offset;
1961         struct nfs_server *server = NFS_SERVER(ino);
1962         struct nfs_client *clp = server->nfs_client;
1963         struct pnfs_layout_hdr *lo = NULL;
1964         struct pnfs_layout_segment *lseg = NULL;
1965         struct nfs4_layoutget *lgp;
1966         nfs4_stateid stateid;
1967         long timeout = 0;
1968         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1969         bool first;
1970
1971         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1972                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1973                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1974                 goto out;
1975         }
1976
1977         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1978                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1979                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1980                 goto out;
1981         }
1982
1983 lookup_again:
1984         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1985         if (IS_ERR(lseg))
1986                 goto out;
1987         first = false;
1988         spin_lock(&ino->i_lock);
1989         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1990         if (lo == NULL) {
1991                 spin_unlock(&ino->i_lock);
1992                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1993                                  PNFS_UPDATE_LAYOUT_NOMEM);
1994                 goto out;
1995         }
1996
1997         /* Do we even need to bother with this? */
1998         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1999                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2000                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2001                 dprintk("%s matches recall, use MDS\n", __func__);
2002                 goto out_unlock;
2003         }
2004
2005         /* if LAYOUTGET already failed once we don't try again */
2006         if (pnfs_layout_io_test_failed(lo, iomode)) {
2007                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2008                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2009                 goto out_unlock;
2010         }
2011
2012         /*
2013          * If the layout segment list is empty, but there are outstanding
2014          * layoutget calls, then they might be subject to a layoutrecall.
2015          */
2016         if ((list_empty(&lo->plh_segs) || !pnfs_layout_is_valid(lo)) &&
2017             atomic_read(&lo->plh_outstanding) != 0) {
2018                 spin_unlock(&ino->i_lock);
2019                 lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
2020                                         !atomic_read(&lo->plh_outstanding)));
2021                 if (IS_ERR(lseg))
2022                         goto out_put_layout_hdr;
2023                 pnfs_put_layout_hdr(lo);
2024                 goto lookup_again;
2025         }
2026
2027         /*
2028          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2029          * for LAYOUTRETURN.
2030          */
2031         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2032                 spin_unlock(&ino->i_lock);
2033                 dprintk("%s wait for layoutreturn\n", __func__);
2034                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2035                 if (!IS_ERR(lseg)) {
2036                         pnfs_put_layout_hdr(lo);
2037                         dprintk("%s retrying\n", __func__);
2038                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2039                                                  lseg,
2040                                                  PNFS_UPDATE_LAYOUT_RETRY);
2041                         goto lookup_again;
2042                 }
2043                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2044                                          PNFS_UPDATE_LAYOUT_RETURN);
2045                 goto out_put_layout_hdr;
2046         }
2047
2048         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2049         if (lseg) {
2050                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2051                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2052                 goto out_unlock;
2053         }
2054
2055         /*
2056          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2057          * stateid, or it has been invalidated, then we must use the open
2058          * stateid.
2059          */
2060         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2061                 int status;
2062
2063                 /*
2064                  * The first layoutget for the file. Need to serialize per
2065                  * RFC 5661 Errata 3208.
2066                  */
2067                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2068                                      &lo->plh_flags)) {
2069                         spin_unlock(&ino->i_lock);
2070                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2071                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2072                                                 TASK_KILLABLE));
2073                         if (IS_ERR(lseg))
2074                                 goto out_put_layout_hdr;
2075                         pnfs_put_layout_hdr(lo);
2076                         dprintk("%s retrying\n", __func__);
2077                         goto lookup_again;
2078                 }
2079
2080                 spin_unlock(&ino->i_lock);
2081                 first = true;
2082                 status = nfs4_select_rw_stateid(ctx->state,
2083                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2084                                         NULL, &stateid, NULL);
2085                 if (status != 0) {
2086                         lseg = ERR_PTR(status);
2087                         trace_pnfs_update_layout(ino, pos, count,
2088                                         iomode, lo, lseg,
2089                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2090                         nfs4_schedule_stateid_recovery(server, ctx->state);
2091                         pnfs_clear_first_layoutget(lo);
2092                         pnfs_put_layout_hdr(lo);
2093                         goto lookup_again;
2094                 }
2095                 spin_lock(&ino->i_lock);
2096         } else {
2097                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2098         }
2099
2100         if (pnfs_layoutgets_blocked(lo)) {
2101                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2102                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2103                 goto out_unlock;
2104         }
2105         nfs_layoutget_begin(lo);
2106         spin_unlock(&ino->i_lock);
2107
2108         _add_to_server_list(lo, server);
2109
2110         pg_offset = arg.offset & ~PAGE_MASK;
2111         if (pg_offset) {
2112                 arg.offset -= pg_offset;
2113                 arg.length += pg_offset;
2114         }
2115         if (arg.length != NFS4_MAX_UINT64)
2116                 arg.length = PAGE_ALIGN(arg.length);
2117
2118         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2119         if (!lgp) {
2120                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2121                                          PNFS_UPDATE_LAYOUT_NOMEM);
2122                 nfs_layoutget_end(lo);
2123                 goto out_put_layout_hdr;
2124         }
2125
2126         lgp->lo = lo;
2127         pnfs_get_layout_hdr(lo);
2128
2129         lseg = nfs4_proc_layoutget(lgp, &timeout);
2130         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2131                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2132         nfs_layoutget_end(lo);
2133         if (IS_ERR(lseg)) {
2134                 switch(PTR_ERR(lseg)) {
2135                 case -EBUSY:
2136                         if (time_after(jiffies, giveup))
2137                                 lseg = NULL;
2138                         break;
2139                 case -ERECALLCONFLICT:
2140                 case -EAGAIN:
2141                         break;
2142                 default:
2143                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2144                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2145                                 lseg = NULL;
2146                         }
2147                         goto out_put_layout_hdr;
2148                 }
2149                 if (lseg) {
2150                         if (first)
2151                                 pnfs_clear_first_layoutget(lo);
2152                         trace_pnfs_update_layout(ino, pos, count,
2153                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2154                         pnfs_put_layout_hdr(lo);
2155                         goto lookup_again;
2156                 }
2157         } else {
2158                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2159         }
2160
2161 out_put_layout_hdr:
2162         if (first)
2163                 pnfs_clear_first_layoutget(lo);
2164         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2165                                  PNFS_UPDATE_LAYOUT_EXIT);
2166         pnfs_put_layout_hdr(lo);
2167 out:
2168         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2169                         "(%s, offset: %llu, length: %llu)\n",
2170                         __func__, ino->i_sb->s_id,
2171                         (unsigned long long)NFS_FILEID(ino),
2172                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2173                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2174                         (unsigned long long)pos,
2175                         (unsigned long long)count);
2176         return lseg;
2177 out_unlock:
2178         spin_unlock(&ino->i_lock);
2179         goto out_put_layout_hdr;
2180 }
2181 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2182
2183 static bool
2184 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2185 {
2186         switch (range->iomode) {
2187         case IOMODE_READ:
2188         case IOMODE_RW:
2189                 break;
2190         default:
2191                 return false;
2192         }
2193         if (range->offset == NFS4_MAX_UINT64)
2194                 return false;
2195         if (range->length == 0)
2196                 return false;
2197         if (range->length != NFS4_MAX_UINT64 &&
2198             range->length > NFS4_MAX_UINT64 - range->offset)
2199                 return false;
2200         return true;
2201 }
2202
2203 static struct pnfs_layout_hdr *
2204 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2205 {
2206         struct pnfs_layout_hdr *lo;
2207
2208         spin_lock(&ino->i_lock);
2209         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2210         if (!lo)
2211                 goto out_unlock;
2212         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2213                 goto out_unlock;
2214         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2215                 goto out_unlock;
2216         if (pnfs_layoutgets_blocked(lo))
2217                 goto out_unlock;
2218         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2219                 goto out_unlock;
2220         nfs_layoutget_begin(lo);
2221         spin_unlock(&ino->i_lock);
2222         _add_to_server_list(lo, NFS_SERVER(ino));
2223         return lo;
2224
2225 out_unlock:
2226         spin_unlock(&ino->i_lock);
2227         pnfs_put_layout_hdr(lo);
2228         return NULL;
2229 }
2230
2231 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2232                                      struct nfs_open_context *ctx)
2233 {
2234         struct inode *ino = data->dentry->d_inode;
2235         struct pnfs_layout_range rng = {
2236                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2237                           IOMODE_RW: IOMODE_READ,
2238                 .offset = 0,
2239                 .length = NFS4_MAX_UINT64,
2240         };
2241         struct nfs4_layoutget *lgp;
2242         struct pnfs_layout_hdr *lo;
2243
2244         /* Heuristic: don't send layoutget if we have cached data */
2245         if (rng.iomode == IOMODE_READ &&
2246            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2247                 return;
2248
2249         lo = _pnfs_grab_empty_layout(ino, ctx);
2250         if (!lo)
2251                 return;
2252         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2253                                              &rng, GFP_KERNEL);
2254         if (!lgp) {
2255                 pnfs_clear_first_layoutget(lo);
2256                 nfs_layoutget_end(lo);
2257                 pnfs_put_layout_hdr(lo);
2258                 return;
2259         }
2260         lgp->lo = lo;
2261         data->lgp = lgp;
2262         data->o_arg.lg_args = &lgp->args;
2263         data->o_res.lg_res = &lgp->res;
2264 }
2265
2266 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2267                                      struct nfs_open_context *ctx)
2268 {
2269         struct inode *ino = data->dentry->d_inode;
2270         struct pnfs_layout_range rng = {
2271                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2272                           IOMODE_RW: IOMODE_READ,
2273                 .offset = 0,
2274                 .length = NFS4_MAX_UINT64,
2275         };
2276         struct nfs4_layoutget *lgp;
2277
2278         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2279                                              &rng, GFP_KERNEL);
2280         if (!lgp)
2281                 return;
2282         data->lgp = lgp;
2283         data->o_arg.lg_args = &lgp->args;
2284         data->o_res.lg_res = &lgp->res;
2285 }
2286
2287 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2288                          struct nfs_open_context *ctx)
2289 {
2290         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2291
2292         if (!(pnfs_enabled_sb(server) &&
2293               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2294                 return;
2295         /* Could check on max_ops, but currently hardcoded high enough */
2296         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2297                 return;
2298         if (data->lgp)
2299                 return;
2300         if (data->state)
2301                 _lgopen_prepare_attached(data, ctx);
2302         else
2303                 _lgopen_prepare_floating(data, ctx);
2304 }
2305
2306 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2307                        struct nfs_open_context *ctx)
2308 {
2309         struct pnfs_layout_hdr *lo;
2310         struct pnfs_layout_segment *lseg;
2311         struct nfs_server *srv = NFS_SERVER(ino);
2312         u32 iomode;
2313
2314         if (!lgp)
2315                 return;
2316         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2317         if (lgp->res.status) {
2318                 switch (lgp->res.status) {
2319                 default:
2320                         break;
2321                 /*
2322                  * Halt lgopen attempts if the server doesn't recognise
2323                  * the "current stateid" value, the layout type, or the
2324                  * layoutget operation as being valid.
2325                  * Also if it complains about too many ops in the compound
2326                  * or of the request/reply being too big.
2327                  */
2328                 case -NFS4ERR_BAD_STATEID:
2329                 case -NFS4ERR_NOTSUPP:
2330                 case -NFS4ERR_REP_TOO_BIG:
2331                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2332                 case -NFS4ERR_REQ_TOO_BIG:
2333                 case -NFS4ERR_TOO_MANY_OPS:
2334                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2335                         srv->caps &= ~NFS_CAP_LGOPEN;
2336                 }
2337                 return;
2338         }
2339         if (!lgp->lo) {
2340                 lo = _pnfs_grab_empty_layout(ino, ctx);
2341                 if (!lo)
2342                         return;
2343                 lgp->lo = lo;
2344         } else
2345                 lo = lgp->lo;
2346
2347         lseg = pnfs_layout_process(lgp);
2348         if (!IS_ERR(lseg)) {
2349                 iomode = lgp->args.range.iomode;
2350                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2351                 pnfs_put_lseg(lseg);
2352         }
2353 }
2354
2355 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2356 {
2357         if (lgp != NULL) {
2358                 if (lgp->lo) {
2359                         pnfs_clear_first_layoutget(lgp->lo);
2360                         nfs_layoutget_end(lgp->lo);
2361                 }
2362                 pnfs_layoutget_free(lgp);
2363         }
2364 }
2365
2366 struct pnfs_layout_segment *
2367 pnfs_layout_process(struct nfs4_layoutget *lgp)
2368 {
2369         struct pnfs_layout_hdr *lo = lgp->lo;
2370         struct nfs4_layoutget_res *res = &lgp->res;
2371         struct pnfs_layout_segment *lseg;
2372         struct inode *ino = lo->plh_inode;
2373         LIST_HEAD(free_me);
2374
2375         if (!pnfs_sanity_check_layout_range(&res->range))
2376                 return ERR_PTR(-EINVAL);
2377
2378         /* Inject layout blob into I/O device driver */
2379         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2380         if (IS_ERR_OR_NULL(lseg)) {
2381                 if (!lseg)
2382                         lseg = ERR_PTR(-ENOMEM);
2383
2384                 dprintk("%s: Could not allocate layout: error %ld\n",
2385                        __func__, PTR_ERR(lseg));
2386                 return lseg;
2387         }
2388
2389         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2390
2391         spin_lock(&ino->i_lock);
2392         if (pnfs_layoutgets_blocked(lo)) {
2393                 dprintk("%s forget reply due to state\n", __func__);
2394                 goto out_forget;
2395         }
2396
2397         if (!pnfs_layout_is_valid(lo) && !pnfs_is_first_layoutget(lo))
2398                 goto out_forget;
2399
2400         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2401                 /* existing state ID, make sure the sequence number matches. */
2402                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2403                         if (!pnfs_layout_is_valid(lo))
2404                                 lo->plh_barrier = 0;
2405                         dprintk("%s forget reply due to sequence\n", __func__);
2406                         goto out_forget;
2407                 }
2408                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2409         } else if (pnfs_layout_is_valid(lo)) {
2410                 /*
2411                  * We got an entirely new state ID.  Mark all segments for the
2412                  * inode invalid, and retry the layoutget
2413                  */
2414                 struct pnfs_layout_range range = {
2415                         .iomode = IOMODE_ANY,
2416                         .length = NFS4_MAX_UINT64,
2417                 };
2418                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2419                 goto out_forget;
2420         } else {
2421                 /* We have a completely new layout */
2422                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2423         }
2424
2425         pnfs_get_lseg(lseg);
2426         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2427
2428
2429         if (res->return_on_close)
2430                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2431
2432         spin_unlock(&ino->i_lock);
2433         pnfs_free_lseg_list(&free_me);
2434         return lseg;
2435
2436 out_forget:
2437         spin_unlock(&ino->i_lock);
2438         lseg->pls_layout = lo;
2439         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2440         return ERR_PTR(-EAGAIN);
2441 }
2442
2443 /**
2444  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2445  * @lo: pointer to layout header
2446  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2447  * @return_range: describe layout segment ranges to be returned
2448  * @seq: stateid seqid to match
2449  *
2450  * This function is mainly intended for use by layoutrecall. It attempts
2451  * to free the layout segment immediately, or else to mark it for return
2452  * as soon as its reference count drops to zero.
2453  *
2454  * Returns
2455  * - 0: a layoutreturn needs to be scheduled.
2456  * - EBUSY: there are layout segment that are still in use.
2457  * - ENOENT: there are no layout segments that need to be returned.
2458  */
2459 int
2460 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2461                                 struct list_head *tmp_list,
2462                                 const struct pnfs_layout_range *return_range,
2463                                 u32 seq)
2464 {
2465         struct pnfs_layout_segment *lseg, *next;
2466         int remaining = 0;
2467
2468         dprintk("%s:Begin lo %p\n", __func__, lo);
2469
2470         assert_spin_locked(&lo->plh_inode->i_lock);
2471
2472         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2473                 tmp_list = &lo->plh_return_segs;
2474
2475         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2476                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2477                         dprintk("%s: marking lseg %p iomode %d "
2478                                 "offset %llu length %llu\n", __func__,
2479                                 lseg, lseg->pls_range.iomode,
2480                                 lseg->pls_range.offset,
2481                                 lseg->pls_range.length);
2482                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2483                                 tmp_list = &lo->plh_return_segs;
2484                         if (mark_lseg_invalid(lseg, tmp_list))
2485                                 continue;
2486                         remaining++;
2487                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2488                 }
2489
2490         if (remaining) {
2491                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2492                 return -EBUSY;
2493         }
2494
2495         if (!list_empty(&lo->plh_return_segs)) {
2496                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2497                 return 0;
2498         }
2499
2500         return -ENOENT;
2501 }
2502
2503 static void
2504 pnfs_mark_layout_for_return(struct inode *inode,
2505                             const struct pnfs_layout_range *range)
2506 {
2507         struct pnfs_layout_hdr *lo;
2508         bool return_now = false;
2509
2510         spin_lock(&inode->i_lock);
2511         lo = NFS_I(inode)->layout;
2512         if (!pnfs_layout_is_valid(lo)) {
2513                 spin_unlock(&inode->i_lock);
2514                 return;
2515         }
2516         pnfs_set_plh_return_info(lo, range->iomode, 0);
2517         /*
2518          * mark all matching lsegs so that we are sure to have no live
2519          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2520          * for how it works.
2521          */
2522         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2523                 const struct cred *cred;
2524                 nfs4_stateid stateid;
2525                 enum pnfs_iomode iomode;
2526
2527                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2528                 spin_unlock(&inode->i_lock);
2529                 if (return_now)
2530                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2531         } else {
2532                 spin_unlock(&inode->i_lock);
2533                 nfs_commit_inode(inode, 0);
2534         }
2535 }
2536
2537 void pnfs_error_mark_layout_for_return(struct inode *inode,
2538                                        struct pnfs_layout_segment *lseg)
2539 {
2540         struct pnfs_layout_range range = {
2541                 .iomode = lseg->pls_range.iomode,
2542                 .offset = 0,
2543                 .length = NFS4_MAX_UINT64,
2544         };
2545
2546         pnfs_mark_layout_for_return(inode, &range);
2547 }
2548 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2549
2550 static bool
2551 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2552 {
2553         return pnfs_layout_is_valid(lo) &&
2554                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2555                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2556 }
2557
2558 static struct pnfs_layout_segment *
2559 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2560                      const struct pnfs_layout_range *range,
2561                      enum pnfs_iomode iomode)
2562 {
2563         struct pnfs_layout_segment *lseg;
2564
2565         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2566                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2567                         continue;
2568                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2569                         continue;
2570                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2571                         continue;
2572                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2573                         return lseg;
2574         }
2575         return NULL;
2576 }
2577
2578 /* Find open file states whose mode matches that of the range */
2579 static bool
2580 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2581                                  const struct pnfs_layout_range *range)
2582 {
2583         struct list_head *head;
2584         struct nfs_open_context *ctx;
2585         fmode_t mode = 0;
2586
2587         if (!pnfs_layout_can_be_returned(lo) ||
2588             !pnfs_find_first_lseg(lo, range, range->iomode))
2589                 return false;
2590
2591         head = &NFS_I(lo->plh_inode)->open_files;
2592         list_for_each_entry_rcu(ctx, head, list) {
2593                 if (ctx->state)
2594                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2595         }
2596
2597         switch (range->iomode) {
2598         default:
2599                 break;
2600         case IOMODE_READ:
2601                 mode &= ~FMODE_WRITE;
2602                 break;
2603         case IOMODE_RW:
2604                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2605                         mode &= ~FMODE_READ;
2606         }
2607         return mode == 0;
2608 }
2609
2610 static int
2611 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2612 {
2613         const struct pnfs_layout_range *range = data;
2614         struct pnfs_layout_hdr *lo;
2615         struct inode *inode;
2616 restart:
2617         rcu_read_lock();
2618         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2619                 if (!pnfs_layout_can_be_returned(lo) ||
2620                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2621                         continue;
2622                 inode = lo->plh_inode;
2623                 spin_lock(&inode->i_lock);
2624                 if (!pnfs_should_return_unused_layout(lo, range)) {
2625                         spin_unlock(&inode->i_lock);
2626                         continue;
2627                 }
2628                 spin_unlock(&inode->i_lock);
2629                 inode = pnfs_grab_inode_layout_hdr(lo);
2630                 if (!inode)
2631                         continue;
2632                 rcu_read_unlock();
2633                 pnfs_mark_layout_for_return(inode, range);
2634                 iput(inode);
2635                 cond_resched();
2636                 goto restart;
2637         }
2638         rcu_read_unlock();
2639         return 0;
2640 }
2641
2642 void
2643 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2644                                  enum pnfs_iomode iomode)
2645 {
2646         struct pnfs_layout_range range = {
2647                 .iomode = iomode,
2648                 .offset = 0,
2649                 .length = NFS4_MAX_UINT64,
2650         };
2651
2652         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2653                         &range);
2654 }
2655
2656 void
2657 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2658 {
2659         if (pgio->pg_lseg == NULL ||
2660             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2661                 return;
2662         pnfs_put_lseg(pgio->pg_lseg);
2663         pgio->pg_lseg = NULL;
2664 }
2665 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2666
2667 /*
2668  * Check for any intersection between the request and the pgio->pg_lseg,
2669  * and if none, put this pgio->pg_lseg away.
2670  */
2671 void
2672 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2673 {
2674         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2675                 pnfs_put_lseg(pgio->pg_lseg);
2676                 pgio->pg_lseg = NULL;
2677         }
2678 }
2679 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2680
2681 void
2682 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2683 {
2684         u64 rd_size;
2685
2686         pnfs_generic_pg_check_layout(pgio);
2687         pnfs_generic_pg_check_range(pgio, req);
2688         if (pgio->pg_lseg == NULL) {
2689                 if (pgio->pg_dreq == NULL)
2690                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2691                 else
2692                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2693
2694                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2695                                                    nfs_req_openctx(req),
2696                                                    req_offset(req),
2697                                                    rd_size,
2698                                                    IOMODE_READ,
2699                                                    false,
2700                                                    GFP_KERNEL);
2701                 if (IS_ERR(pgio->pg_lseg)) {
2702                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2703                         pgio->pg_lseg = NULL;
2704                         return;
2705                 }
2706         }
2707         /* If no lseg, fall back to read through mds */
2708         if (pgio->pg_lseg == NULL)
2709                 nfs_pageio_reset_read_mds(pgio);
2710
2711 }
2712 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2713
2714 void
2715 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2716                            struct nfs_page *req, u64 wb_size)
2717 {
2718         pnfs_generic_pg_check_layout(pgio);
2719         pnfs_generic_pg_check_range(pgio, req);
2720         if (pgio->pg_lseg == NULL) {
2721                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2722                                                    nfs_req_openctx(req),
2723                                                    req_offset(req),
2724                                                    wb_size,
2725                                                    IOMODE_RW,
2726                                                    false,
2727                                                    GFP_KERNEL);
2728                 if (IS_ERR(pgio->pg_lseg)) {
2729                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2730                         pgio->pg_lseg = NULL;
2731                         return;
2732                 }
2733         }
2734         /* If no lseg, fall back to write through mds */
2735         if (pgio->pg_lseg == NULL)
2736                 nfs_pageio_reset_write_mds(pgio);
2737 }
2738 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2739
2740 void
2741 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2742 {
2743         if (desc->pg_lseg) {
2744                 pnfs_put_lseg(desc->pg_lseg);
2745                 desc->pg_lseg = NULL;
2746         }
2747 }
2748 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2749
2750 /*
2751  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2752  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2753  */
2754 size_t
2755 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2756                      struct nfs_page *prev, struct nfs_page *req)
2757 {
2758         unsigned int size;
2759         u64 seg_end, req_start, seg_left;
2760
2761         size = nfs_generic_pg_test(pgio, prev, req);
2762         if (!size)
2763                 return 0;
2764
2765         /*
2766          * 'size' contains the number of bytes left in the current page (up
2767          * to the original size asked for in @req->wb_bytes).
2768          *
2769          * Calculate how many bytes are left in the layout segment
2770          * and if there are less bytes than 'size', return that instead.
2771          *
2772          * Please also note that 'end_offset' is actually the offset of the
2773          * first byte that lies outside the pnfs_layout_range. FIXME?
2774          *
2775          */
2776         if (pgio->pg_lseg) {
2777                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2778                                      pgio->pg_lseg->pls_range.length);
2779                 req_start = req_offset(req);
2780
2781                 /* start of request is past the last byte of this segment */
2782                 if (req_start >= seg_end)
2783                         return 0;
2784
2785                 /* adjust 'size' iff there are fewer bytes left in the
2786                  * segment than what nfs_generic_pg_test returned */
2787                 seg_left = seg_end - req_start;
2788                 if (seg_left < size)
2789                         size = (unsigned int)seg_left;
2790         }
2791
2792         return size;
2793 }
2794 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2795
2796 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2797 {
2798         struct nfs_pageio_descriptor pgio;
2799
2800         /* Resend all requests through the MDS */
2801         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2802                               hdr->completion_ops);
2803         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2804         return nfs_pageio_resend(&pgio, hdr);
2805 }
2806 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2807
2808 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2809 {
2810
2811         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2812         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2813             PNFS_LAYOUTRET_ON_ERROR) {
2814                 pnfs_return_layout(hdr->inode);
2815         }
2816         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2817                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2818 }
2819
2820 /*
2821  * Called by non rpc-based layout drivers
2822  */
2823 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2824 {
2825         if (likely(!hdr->pnfs_error)) {
2826                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2827                                 hdr->mds_offset + hdr->res.count);
2828                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2829         }
2830         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2831         if (unlikely(hdr->pnfs_error))
2832                 pnfs_ld_handle_write_error(hdr);
2833         hdr->mds_ops->rpc_release(hdr);
2834 }
2835 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2836
2837 static void
2838 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2839                 struct nfs_pgio_header *hdr)
2840 {
2841         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2842
2843         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2844                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2845                 nfs_pageio_reset_write_mds(desc);
2846                 mirror->pg_recoalesce = 1;
2847         }
2848         hdr->completion_ops->completion(hdr);
2849 }
2850
2851 static enum pnfs_try_status
2852 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2853                         const struct rpc_call_ops *call_ops,
2854                         struct pnfs_layout_segment *lseg,
2855                         int how)
2856 {
2857         struct inode *inode = hdr->inode;
2858         enum pnfs_try_status trypnfs;
2859         struct nfs_server *nfss = NFS_SERVER(inode);
2860
2861         hdr->mds_ops = call_ops;
2862
2863         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2864                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2865         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2866         if (trypnfs != PNFS_NOT_ATTEMPTED)
2867                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2868         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2869         return trypnfs;
2870 }
2871
2872 static void
2873 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2874               struct nfs_pgio_header *hdr, int how)
2875 {
2876         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2877         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2878         enum pnfs_try_status trypnfs;
2879
2880         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2881         switch (trypnfs) {
2882         case PNFS_NOT_ATTEMPTED:
2883                 pnfs_write_through_mds(desc, hdr);
2884                 break;
2885         case PNFS_ATTEMPTED:
2886                 break;
2887         case PNFS_TRY_AGAIN:
2888                 /* cleanup hdr and prepare to redo pnfs */
2889                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2890                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2891                         list_splice_init(&hdr->pages, &mirror->pg_list);
2892                         mirror->pg_recoalesce = 1;
2893                 }
2894                 hdr->mds_ops->rpc_release(hdr);
2895         }
2896 }
2897
2898 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2899 {
2900         pnfs_put_lseg(hdr->lseg);
2901         nfs_pgio_header_free(hdr);
2902 }
2903
2904 int
2905 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2906 {
2907         struct nfs_pgio_header *hdr;
2908         int ret;
2909
2910         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2911         if (!hdr) {
2912                 desc->pg_error = -ENOMEM;
2913                 return desc->pg_error;
2914         }
2915         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2916
2917         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2918         ret = nfs_generic_pgio(desc, hdr);
2919         if (!ret)
2920                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2921
2922         return ret;
2923 }
2924 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2925
2926 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2927 {
2928         struct nfs_pageio_descriptor pgio;
2929
2930         /* Resend all requests through the MDS */
2931         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2932         return nfs_pageio_resend(&pgio, hdr);
2933 }
2934 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2935
2936 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2937 {
2938         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2939         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2940             PNFS_LAYOUTRET_ON_ERROR) {
2941                 pnfs_return_layout(hdr->inode);
2942         }
2943         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2944                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2945 }
2946
2947 /*
2948  * Called by non rpc-based layout drivers
2949  */
2950 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2951 {
2952         if (likely(!hdr->pnfs_error))
2953                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2954         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2955         if (unlikely(hdr->pnfs_error))
2956                 pnfs_ld_handle_read_error(hdr);
2957         hdr->mds_ops->rpc_release(hdr);
2958 }
2959 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2960
2961 static void
2962 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2963                 struct nfs_pgio_header *hdr)
2964 {
2965         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2966
2967         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2968                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2969                 nfs_pageio_reset_read_mds(desc);
2970                 mirror->pg_recoalesce = 1;
2971         }
2972         hdr->completion_ops->completion(hdr);
2973 }
2974
2975 /*
2976  * Call the appropriate parallel I/O subsystem read function.
2977  */
2978 static enum pnfs_try_status
2979 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2980                        const struct rpc_call_ops *call_ops,
2981                        struct pnfs_layout_segment *lseg)
2982 {
2983         struct inode *inode = hdr->inode;
2984         struct nfs_server *nfss = NFS_SERVER(inode);
2985         enum pnfs_try_status trypnfs;
2986
2987         hdr->mds_ops = call_ops;
2988
2989         dprintk("%s: Reading ino:%lu %u@%llu\n",
2990                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2991
2992         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2993         if (trypnfs != PNFS_NOT_ATTEMPTED)
2994                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2995         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2996         return trypnfs;
2997 }
2998
2999 /* Resend all requests through pnfs. */
3000 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3001                            unsigned int mirror_idx)
3002 {
3003         struct nfs_pageio_descriptor pgio;
3004
3005         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3006                 /* Prevent deadlocks with layoutreturn! */
3007                 pnfs_put_lseg(hdr->lseg);
3008                 hdr->lseg = NULL;
3009
3010                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3011                                         hdr->completion_ops);
3012                 pgio.pg_mirror_idx = mirror_idx;
3013                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3014         }
3015 }
3016 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3017
3018 static void
3019 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3020 {
3021         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3022         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3023         enum pnfs_try_status trypnfs;
3024
3025         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3026         switch (trypnfs) {
3027         case PNFS_NOT_ATTEMPTED:
3028                 pnfs_read_through_mds(desc, hdr);
3029                 break;
3030         case PNFS_ATTEMPTED:
3031                 break;
3032         case PNFS_TRY_AGAIN:
3033                 /* cleanup hdr and prepare to redo pnfs */
3034                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3035                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3036                         list_splice_init(&hdr->pages, &mirror->pg_list);
3037                         mirror->pg_recoalesce = 1;
3038                 }
3039                 hdr->mds_ops->rpc_release(hdr);
3040         }
3041 }
3042
3043 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3044 {
3045         pnfs_put_lseg(hdr->lseg);
3046         nfs_pgio_header_free(hdr);
3047 }
3048
3049 int
3050 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3051 {
3052         struct nfs_pgio_header *hdr;
3053         int ret;
3054
3055         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3056         if (!hdr) {
3057                 desc->pg_error = -ENOMEM;
3058                 return desc->pg_error;
3059         }
3060         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3061         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3062         ret = nfs_generic_pgio(desc, hdr);
3063         if (!ret)
3064                 pnfs_do_read(desc, hdr);
3065         return ret;
3066 }
3067 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3068
3069 static void pnfs_clear_layoutcommitting(struct inode *inode)
3070 {
3071         unsigned long *bitlock = &NFS_I(inode)->flags;
3072
3073         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3074         smp_mb__after_atomic();
3075         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3076 }
3077
3078 /*
3079  * There can be multiple RW segments.
3080  */
3081 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3082 {
3083         struct pnfs_layout_segment *lseg;
3084
3085         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3086                 if (lseg->pls_range.iomode == IOMODE_RW &&
3087                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3088                         list_add(&lseg->pls_lc_list, listp);
3089         }
3090 }
3091
3092 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3093 {
3094         struct pnfs_layout_segment *lseg, *tmp;
3095
3096         /* Matched by references in pnfs_set_layoutcommit */
3097         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3098                 list_del_init(&lseg->pls_lc_list);
3099                 pnfs_put_lseg(lseg);
3100         }
3101
3102         pnfs_clear_layoutcommitting(inode);
3103 }
3104
3105 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3106 {
3107         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3108 }
3109 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3110
3111 void
3112 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3113                 loff_t end_pos)
3114 {
3115         struct nfs_inode *nfsi = NFS_I(inode);
3116         bool mark_as_dirty = false;
3117
3118         spin_lock(&inode->i_lock);
3119         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3120                 nfsi->layout->plh_lwb = end_pos;
3121                 mark_as_dirty = true;
3122                 dprintk("%s: Set layoutcommit for inode %lu ",
3123                         __func__, inode->i_ino);
3124         } else if (end_pos > nfsi->layout->plh_lwb)
3125                 nfsi->layout->plh_lwb = end_pos;
3126         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3127                 /* references matched in nfs4_layoutcommit_release */
3128                 pnfs_get_lseg(lseg);
3129         }
3130         spin_unlock(&inode->i_lock);
3131         dprintk("%s: lseg %p end_pos %llu\n",
3132                 __func__, lseg, nfsi->layout->plh_lwb);
3133
3134         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3135          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3136         if (mark_as_dirty)
3137                 mark_inode_dirty_sync(inode);
3138 }
3139 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3140
3141 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3142 {
3143         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3144
3145         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3146                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3147         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3148 }
3149
3150 /*
3151  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3152  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3153  * data to disk to allow the server to recover the data if it crashes.
3154  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3155  * is off, and a COMMIT is sent to a data server, or
3156  * if WRITEs to a data server return NFS_DATA_SYNC.
3157  */
3158 int
3159 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3160 {
3161         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3162         struct nfs4_layoutcommit_data *data;
3163         struct nfs_inode *nfsi = NFS_I(inode);
3164         loff_t end_pos;
3165         int status;
3166
3167         if (!pnfs_layoutcommit_outstanding(inode))
3168                 return 0;
3169
3170         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3171
3172         status = -EAGAIN;
3173         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3174                 if (!sync)
3175                         goto out;
3176                 status = wait_on_bit_lock_action(&nfsi->flags,
3177                                 NFS_INO_LAYOUTCOMMITTING,
3178                                 nfs_wait_bit_killable,
3179                                 TASK_KILLABLE);
3180                 if (status)
3181                         goto out;
3182         }
3183
3184         status = -ENOMEM;
3185         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3186         data = kzalloc(sizeof(*data), GFP_NOFS);
3187         if (!data)
3188                 goto clear_layoutcommitting;
3189
3190         status = 0;
3191         spin_lock(&inode->i_lock);
3192         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3193                 goto out_unlock;
3194
3195         INIT_LIST_HEAD(&data->lseg_list);
3196         pnfs_list_write_lseg(inode, &data->lseg_list);
3197
3198         end_pos = nfsi->layout->plh_lwb;
3199
3200         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3201         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3202         spin_unlock(&inode->i_lock);
3203
3204         data->args.inode = inode;
3205         nfs_fattr_init(&data->fattr);
3206         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3207         data->res.fattr = &data->fattr;
3208         if (end_pos != 0)
3209                 data->args.lastbytewritten = end_pos - 1;
3210         else
3211                 data->args.lastbytewritten = U64_MAX;
3212         data->res.server = NFS_SERVER(inode);
3213
3214         if (ld->prepare_layoutcommit) {
3215                 status = ld->prepare_layoutcommit(&data->args);
3216                 if (status) {
3217                         put_cred(data->cred);
3218                         spin_lock(&inode->i_lock);
3219                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3220                         if (end_pos > nfsi->layout->plh_lwb)
3221                                 nfsi->layout->plh_lwb = end_pos;
3222                         goto out_unlock;
3223                 }
3224         }
3225
3226
3227         status = nfs4_proc_layoutcommit(data, sync);
3228 out:
3229         if (status)
3230                 mark_inode_dirty_sync(inode);
3231         dprintk("<-- %s status %d\n", __func__, status);
3232         return status;
3233 out_unlock:
3234         spin_unlock(&inode->i_lock);
3235         kfree(data);
3236 clear_layoutcommitting:
3237         pnfs_clear_layoutcommitting(inode);
3238         goto out;
3239 }
3240 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3241
3242 int
3243 pnfs_generic_sync(struct inode *inode, bool datasync)
3244 {
3245         return pnfs_layoutcommit_inode(inode, true);
3246 }
3247 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3248
3249 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3250 {
3251         struct nfs4_threshold *thp;
3252
3253         thp = kzalloc(sizeof(*thp), GFP_NOFS);
3254         if (!thp) {
3255                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3256                 return NULL;
3257         }
3258         return thp;
3259 }
3260
3261 #if IS_ENABLED(CONFIG_NFS_V4_2)
3262 int
3263 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3264 {
3265         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3266         struct nfs_server *server = NFS_SERVER(inode);
3267         struct nfs_inode *nfsi = NFS_I(inode);
3268         struct nfs42_layoutstat_data *data;
3269         struct pnfs_layout_hdr *hdr;
3270         int status = 0;
3271
3272         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3273                 goto out;
3274
3275         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3276                 goto out;
3277
3278         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3279                 goto out;
3280
3281         spin_lock(&inode->i_lock);
3282         if (!NFS_I(inode)->layout) {
3283                 spin_unlock(&inode->i_lock);
3284                 goto out_clear_layoutstats;
3285         }
3286         hdr = NFS_I(inode)->layout;
3287         pnfs_get_layout_hdr(hdr);
3288         spin_unlock(&inode->i_lock);
3289
3290         data = kzalloc(sizeof(*data), gfp_flags);
3291         if (!data) {
3292                 status = -ENOMEM;
3293                 goto out_put;
3294         }
3295
3296         data->args.fh = NFS_FH(inode);
3297         data->args.inode = inode;
3298         status = ld->prepare_layoutstats(&data->args);
3299         if (status)
3300                 goto out_free;
3301
3302         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3303
3304 out:
3305         dprintk("%s returns %d\n", __func__, status);
3306         return status;
3307
3308 out_free:
3309         kfree(data);
3310 out_put:
3311         pnfs_put_layout_hdr(hdr);
3312 out_clear_layoutstats:
3313         smp_mb__before_atomic();
3314         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3315         smp_mb__after_atomic();
3316         goto out;
3317 }
3318 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3319 #endif
3320
3321 unsigned int layoutstats_timer;
3322 module_param(layoutstats_timer, uint, 0644);
3323 EXPORT_SYMBOL_GPL(layoutstats_timer);