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 <[email protected]>
   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 (!list_empty(&lo->plh_layouts)) {
 272                 struct nfs_client *clp = server->nfs_client;
 273
 274                 spin_lock(&clp->cl_lock);
 275                 list_del_init(&lo->plh_layouts);
 276                 spin_unlock(&clp->cl_lock);
 277         }
 278         put_rpccred(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
 298         if (!lo)
 299                 return;
 300         inode = lo->plh_inode;
 301         pnfs_layoutreturn_before_put_layout_hdr(lo);
 302
 303         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 304                 if (!list_empty(&lo->plh_segs))
 305                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 306                 pnfs_detach_layout_hdr(lo);
 307                 spin_unlock(&inode->i_lock);
 308                 pnfs_free_layout_hdr(lo);
 309         }
 310 }
 311
 312 static void
 313 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 314                          u32 seq)
 315 {
 316         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 317                 iomode = IOMODE_ANY;
 318         lo->plh_return_iomode = iomode;
 319         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 320         if (seq != 0) {
 321                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 322                 lo->plh_return_seq = seq;
 323         }
 324 }
 325
 326 static void
 327 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 328 {
 329         struct pnfs_layout_segment *lseg;
 330         lo->plh_return_iomode = 0;
 331         lo->plh_return_seq = 0;
 332         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 333         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 334                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 335                         continue;
 336                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 337         }
 338 }
 339
 340 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 341 {
 342         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 343         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 344         smp_mb__after_atomic();
 345         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 346         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 347 }
 348
 349 static void
 350 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 351                 struct list_head *free_me)
 352 {
 353         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 354         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 355         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 356                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 357         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 358                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 359 }
 360
 361 /*
 362  * Update the seqid of a layout stateid
 363  */
 364 bool nfs4_layoutreturn_refresh_stateid(nfs4_stateid *dst,
 365                 struct pnfs_layout_range *dst_range,
 366                 struct inode *inode)
 367 {
 368         struct pnfs_layout_hdr *lo;
 369         struct pnfs_layout_range range = {
 370                 .iomode = IOMODE_ANY,
 371                 .offset = 0,
 372                 .length = NFS4_MAX_UINT64,
 373         };
 374         bool ret = false;
 375         LIST_HEAD(head);
 376         int err;
 377
 378         spin_lock(&inode->i_lock);
 379         lo = NFS_I(inode)->layout;
 380         if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
 381                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
 382                 if (err != -EBUSY) {
 383                         dst->seqid = lo->plh_stateid.seqid;
 384                         *dst_range = range;
 385                         ret = true;
 386                 }
 387         }
 388         spin_unlock(&inode->i_lock);
 389         pnfs_free_lseg_list(&head);
 390         return ret;
 391 }
 392
 393 /*
 394  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 395  *
 396  * In order to continue using the pnfs_layout_hdr, a full recovery
 397  * is required.
 398  * Note that caller must hold inode->i_lock.
 399  */
 400 int
 401 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 402                 struct list_head *lseg_list)
 403 {
 404         struct pnfs_layout_range range = {
 405                 .iomode = IOMODE_ANY,
 406                 .offset = 0,
 407                 .length = NFS4_MAX_UINT64,
 408         };
 409         struct pnfs_layout_segment *lseg, *next;
 410
 411         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 412         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 413                 pnfs_clear_lseg_state(lseg, lseg_list);
 414         pnfs_clear_layoutreturn_info(lo);
 415         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 416         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 417             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 418                 pnfs_clear_layoutreturn_waitbit(lo);
 419         return !list_empty(&lo->plh_segs);
 420 }
 421
 422 static int
 423 pnfs_iomode_to_fail_bit(u32 iomode)
 424 {
 425         return iomode == IOMODE_RW ?
 426                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 427 }
 428
 429 static void
 430 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 431 {
 432         lo->plh_retry_timestamp = jiffies;
 433         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 434                 refcount_inc(&lo->plh_refcount);
 435 }
 436
 437 static void
 438 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 439 {
 440         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 441                 refcount_dec(&lo->plh_refcount);
 442 }
 443
 444 static void
 445 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 446 {
 447         struct inode *inode = lo->plh_inode;
 448         struct pnfs_layout_range range = {
 449                 .iomode = iomode,
 450                 .offset = 0,
 451                 .length = NFS4_MAX_UINT64,
 452         };
 453         LIST_HEAD(head);
 454
 455         spin_lock(&inode->i_lock);
 456         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 457         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 458         spin_unlock(&inode->i_lock);
 459         pnfs_free_lseg_list(&head);
 460         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 461                         iomode == IOMODE_RW ?  "RW" : "READ");
 462 }
 463
 464 static bool
 465 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 466 {
 467         unsigned long start, end;
 468         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 469
 470         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 471                 return false;
 472         end = jiffies;
 473         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 474         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 475                 /* It is time to retry the failed layoutgets */
 476                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 477                 return false;
 478         }
 479         return true;
 480 }
 481
 482 static void
 483 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 484                 const struct pnfs_layout_range *range,
 485                 const nfs4_stateid *stateid)
 486 {
 487         INIT_LIST_HEAD(&lseg->pls_list);
 488         INIT_LIST_HEAD(&lseg->pls_lc_list);
 489         refcount_set(&lseg->pls_refcount, 1);
 490         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 491         lseg->pls_layout = lo;
 492         lseg->pls_range = *range;
 493         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 494 }
 495
 496 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 497 {
 498         if (lseg != NULL) {
 499                 struct inode *inode = lseg->pls_layout->plh_inode;
 500                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 501         }
 502 }
 503
 504 static void
 505 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 506                 struct pnfs_layout_segment *lseg)
 507 {
 508         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 509         list_del_init(&lseg->pls_list);
 510         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 511         refcount_dec(&lo->plh_refcount);
 512         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 513                 return;
 514         if (list_empty(&lo->plh_segs) &&
 515             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 516             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 517                 if (atomic_read(&lo->plh_outstanding) == 0)
 518                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 519                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 520         }
 521 }
 522
 523 static bool
 524 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 525                 struct pnfs_layout_segment *lseg)
 526 {
 527         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 528             pnfs_layout_is_valid(lo)) {
 529                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 530                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 531                 return true;
 532         }
 533         return false;
 534 }
 535
 536 void
 537 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 538 {
 539         struct pnfs_layout_hdr *lo;
 540         struct inode *inode;
 541
 542         if (!lseg)
 543                 return;
 544
 545         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 546                 refcount_read(&lseg->pls_refcount),
 547                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 548
 549         lo = lseg->pls_layout;
 550         inode = lo->plh_inode;
 551
 552         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 553                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 554                         spin_unlock(&inode->i_lock);
 555                         return;
 556                 }
 557                 pnfs_get_layout_hdr(lo);
 558                 pnfs_layout_remove_lseg(lo, lseg);
 559                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 560                         lseg = NULL;
 561                 spin_unlock(&inode->i_lock);
 562                 pnfs_free_lseg(lseg);
 563                 pnfs_put_layout_hdr(lo);
 564         }
 565 }
 566 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 567
 568 /*
 569  * is l2 fully contained in l1?
 570  *   start1                             end1
 571  *   [----------------------------------)
 572  *           start2           end2
 573  *           [----------------)
 574  */
 575 static bool
 576 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 577                  const struct pnfs_layout_range *l2)
 578 {
 579         u64 start1 = l1->offset;
 580         u64 end1 = pnfs_end_offset(start1, l1->length);
 581         u64 start2 = l2->offset;
 582         u64 end2 = pnfs_end_offset(start2, l2->length);
 583
 584         return (start1 <= start2) && (end1 >= end2);
 585 }
 586
 587 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 588                 struct list_head *tmp_list)
 589 {
 590         if (!refcount_dec_and_test(&lseg->pls_refcount))
 591                 return false;
 592         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 593         list_add(&lseg->pls_list, tmp_list);
 594         return true;
 595 }
 596
 597 /* Returns 1 if lseg is removed from list, 0 otherwise */
 598 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 599                              struct list_head *tmp_list)
 600 {
 601         int rv = 0;
 602
 603         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 604                 /* Remove the reference keeping the lseg in the
 605                  * list.  It will now be removed when all
 606                  * outstanding io is finished.
 607                  */
 608                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 609                         refcount_read(&lseg->pls_refcount));
 610                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 611                         rv = 1;
 612         }
 613         return rv;
 614 }
 615
 616 /*
 617  * Compare 2 layout stateid sequence ids, to see which is newer,
 618  * taking into account wraparound issues.
 619  */
 620 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 621 {
 622         return (s32)(s1 - s2) > 0;
 623 }
 624
 625 static bool
 626 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 627                  const struct pnfs_layout_range *recall_range)
 628 {
 629         return (recall_range->iomode == IOMODE_ANY ||
 630                 lseg_range->iomode == recall_range->iomode) &&
 631                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 632 }
 633
 634 static bool
 635 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 636                 const struct pnfs_layout_range *recall_range,
 637                 u32 seq)
 638 {
 639         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 640                 return false;
 641         if (recall_range == NULL)
 642                 return true;
 643         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 644 }
 645
 646 /**
 647  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 648  * @lo: layout header containing the lsegs
 649  * @tmp_list: list head where doomed lsegs should go
 650  * @recall_range: optional recall range argument to match (may be NULL)
 651  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 652  *
 653  * Walk the list of lsegs in the layout header, and tear down any that should
 654  * be destroyed. If "recall_range" is specified then the segment must match
 655  * that range. If "seq" is non-zero, then only match segments that were handed
 656  * out at or before that sequence.
 657  *
 658  * Returns number of matching invalid lsegs remaining in list after scanning
 659  * it and purging them.
 660  */
 661 int
 662 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 663                             struct list_head *tmp_list,
 664                             const struct pnfs_layout_range *recall_range,
 665                             u32 seq)
 666 {
 667         struct pnfs_layout_segment *lseg, *next;
 668         int remaining = 0;
 669
 670         dprintk("%s:Begin lo %p\n", __func__, lo);
 671
 672         if (list_empty(&lo->plh_segs))
 673                 return 0;
 674         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 675                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 676                         dprintk("%s: freeing lseg %p iomode %d seq %u "
 677                                 "offset %llu length %llu\n", __func__,
 678                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 679                                 lseg->pls_range.offset, lseg->pls_range.length);
 680                         if (!mark_lseg_invalid(lseg, tmp_list))
 681                                 remaining++;
 682                 }
 683         dprintk("%s:Return %i\n", __func__, remaining);
 684         return remaining;
 685 }
 686
 687 static void
 688 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 689                 struct list_head *free_me,
 690                 const struct pnfs_layout_range *range,
 691                 u32 seq)
 692 {
 693         struct pnfs_layout_segment *lseg, *next;
 694
 695         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 696                 if (pnfs_match_lseg_recall(lseg, range, seq))
 697                         list_move_tail(&lseg->pls_list, free_me);
 698         }
 699 }
 700
 701 /* note free_me must contain lsegs from a single layout_hdr */
 702 void
 703 pnfs_free_lseg_list(struct list_head *free_me)
 704 {
 705         struct pnfs_layout_segment *lseg, *tmp;
 706
 707         if (list_empty(free_me))
 708                 return;
 709
 710         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 711                 list_del(&lseg->pls_list);
 712                 pnfs_free_lseg(lseg);
 713         }
 714 }
 715
 716 void
 717 pnfs_destroy_layout(struct nfs_inode *nfsi)
 718 {
 719         struct pnfs_layout_hdr *lo;
 720         LIST_HEAD(tmp_list);
 721
 722         spin_lock(&nfsi->vfs_inode.i_lock);
 723         lo = nfsi->layout;
 724         if (lo) {
 725                 pnfs_get_layout_hdr(lo);
 726                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 727                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 728                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 729                 spin_unlock(&nfsi->vfs_inode.i_lock);
 730                 pnfs_free_lseg_list(&tmp_list);
 731                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 732                 pnfs_put_layout_hdr(lo);
 733         } else
 734                 spin_unlock(&nfsi->vfs_inode.i_lock);
 735 }
 736 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 737
 738 static bool
 739 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 740                 struct list_head *layout_list)
 741 {
 742         struct pnfs_layout_hdr *lo;
 743         bool ret = false;
 744
 745         spin_lock(&inode->i_lock);
 746         lo = NFS_I(inode)->layout;
 747         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 748                 pnfs_get_layout_hdr(lo);
 749                 list_add(&lo->plh_bulk_destroy, layout_list);
 750                 ret = true;
 751         }
 752         spin_unlock(&inode->i_lock);
 753         return ret;
 754 }
 755
 756 /* Caller must hold rcu_read_lock and clp->cl_lock */
 757 static int
 758 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 759                 struct nfs_server *server,
 760                 struct list_head *layout_list)
 761 {
 762         struct pnfs_layout_hdr *lo, *next;
 763         struct inode *inode;
 764
 765         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 766                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 767                         continue;
 768                 inode = igrab(lo->plh_inode);
 769                 if (inode == NULL)
 770                         continue;
 771                 list_del_init(&lo->plh_layouts);
 772                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 773                         continue;
 774                 rcu_read_unlock();
 775                 spin_unlock(&clp->cl_lock);
 776                 iput(inode);
 777                 spin_lock(&clp->cl_lock);
 778                 rcu_read_lock();
 779                 return -EAGAIN;
 780         }
 781         return 0;
 782 }
 783
 784 static int
 785 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 786                 bool is_bulk_recall)
 787 {
 788         struct pnfs_layout_hdr *lo;
 789         struct inode *inode;
 790         LIST_HEAD(lseg_list);
 791         int ret = 0;
 792
 793         while (!list_empty(layout_list)) {
 794                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 795                                 plh_bulk_destroy);
 796                 dprintk("%s freeing layout for inode %lu\n", __func__,
 797                         lo->plh_inode->i_ino);
 798                 inode = lo->plh_inode;
 799
 800                 pnfs_layoutcommit_inode(inode, false);
 801
 802                 spin_lock(&inode->i_lock);
 803                 list_del_init(&lo->plh_bulk_destroy);
 804                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 805                         if (is_bulk_recall)
 806                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 807                         ret = -EAGAIN;
 808                 }
 809                 spin_unlock(&inode->i_lock);
 810                 pnfs_free_lseg_list(&lseg_list);
 811                 /* Free all lsegs that are attached to commit buckets */
 812                 nfs_commit_inode(inode, 0);
 813                 pnfs_put_layout_hdr(lo);
 814                 iput(inode);
 815         }
 816         return ret;
 817 }
 818
 819 int
 820 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 821                 struct nfs_fsid *fsid,
 822                 bool is_recall)
 823 {
 824         struct nfs_server *server;
 825         LIST_HEAD(layout_list);
 826
 827         spin_lock(&clp->cl_lock);
 828         rcu_read_lock();
 829 restart:
 830         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 831                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 832                         continue;
 833                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 834                                 server,
 835                                 &layout_list) != 0)
 836                         goto restart;
 837         }
 838         rcu_read_unlock();
 839         spin_unlock(&clp->cl_lock);
 840
 841         if (list_empty(&layout_list))
 842                 return 0;
 843         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 844 }
 845
 846 int
 847 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 848                 bool is_recall)
 849 {
 850         struct nfs_server *server;
 851         LIST_HEAD(layout_list);
 852
 853         spin_lock(&clp->cl_lock);
 854         rcu_read_lock();
 855 restart:
 856         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 857                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 858                                         server,
 859                                         &layout_list) != 0)
 860                         goto restart;
 861         }
 862         rcu_read_unlock();
 863         spin_unlock(&clp->cl_lock);
 864
 865         if (list_empty(&layout_list))
 866                 return 0;
 867         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 868 }
 869
 870 /*
 871  * Called by the state manger to remove all layouts established under an
 872  * expired lease.
 873  */
 874 void
 875 pnfs_destroy_all_layouts(struct nfs_client *clp)
 876 {
 877         nfs4_deviceid_mark_client_invalid(clp);
 878         nfs4_deviceid_purge_client(clp);
 879
 880         pnfs_destroy_layouts_byclid(clp, false);
 881 }
 882
 883 /* update lo->plh_stateid with new if is more recent */
 884 void
 885 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 886                         bool update_barrier)
 887 {
 888         u32 oldseq, newseq, new_barrier = 0;
 889
 890         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 891         newseq = be32_to_cpu(new->seqid);
 892
 893         if (!pnfs_layout_is_valid(lo)) {
 894                 nfs4_stateid_copy(&lo->plh_stateid, new);
 895                 lo->plh_barrier = newseq;
 896                 pnfs_clear_layoutreturn_info(lo);
 897                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 898                 return;
 899         }
 900         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 901                 nfs4_stateid_copy(&lo->plh_stateid, new);
 902                 /*
 903                  * Because of wraparound, we want to keep the barrier
 904                  * "close" to the current seqids.
 905                  */
 906                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 907         }
 908         if (update_barrier)
 909                 new_barrier = be32_to_cpu(new->seqid);
 910         else if (new_barrier == 0)
 911                 return;
 912         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 913                 lo->plh_barrier = new_barrier;
 914 }
 915
 916 static bool
 917 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 918                 const nfs4_stateid *stateid)
 919 {
 920         u32 seqid = be32_to_cpu(stateid->seqid);
 921
 922         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 923 }
 924
 925 /* lget is set to 1 if called from inside send_layoutget call chain */
 926 static bool
 927 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 928 {
 929         return lo->plh_block_lgets ||
 930                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 931 }
 932
 933 static struct nfs_server *
 934 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
 935 {
 936         struct nfs_server *server;
 937
 938         if (inode) {
 939                 server = NFS_SERVER(inode);
 940         } else {
 941                 struct dentry *parent_dir = dget_parent(ctx->dentry);
 942                 server = NFS_SERVER(parent_dir->d_inode);
 943                 dput(parent_dir);
 944         }
 945         return server;
 946 }
 947
 948 static void nfs4_free_pages(struct page **pages, size_t size)
 949 {
 950         int i;
 951
 952         if (!pages)
 953                 return;
 954
 955         for (i = 0; i < size; i++) {
 956                 if (!pages[i])
 957                         break;
 958                 __free_page(pages[i]);
 959         }
 960         kfree(pages);
 961 }
 962
 963 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
 964 {
 965         struct page **pages;
 966         int i;
 967
 968         pages = kcalloc(size, sizeof(struct page *), gfp_flags);
 969         if (!pages) {
 970                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
 971                 return NULL;
 972         }
 973
 974         for (i = 0; i < size; i++) {
 975                 pages[i] = alloc_page(gfp_flags);
 976                 if (!pages[i]) {
 977                         dprintk("%s: failed to allocate page\n", __func__);
 978                         nfs4_free_pages(pages, size);
 979                         return NULL;
 980                 }
 981         }
 982
 983         return pages;
 984 }
 985
 986 static struct nfs4_layoutget *
 987 pnfs_alloc_init_layoutget_args(struct inode *ino,
 988            struct nfs_open_context *ctx,
 989            const nfs4_stateid *stateid,
 990            const struct pnfs_layout_range *range,
 991            gfp_t gfp_flags)
 992 {
 993         struct nfs_server *server = pnfs_find_server(ino, ctx);
 994         size_t max_pages = max_response_pages(server);
 995         struct nfs4_layoutget *lgp;
 996
 997         dprintk("--> %s\n", __func__);
 998
 999         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1000         if (lgp == NULL)
1001                 return NULL;
1002
1003         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1004         if (!lgp->args.layout.pages) {
1005                 kfree(lgp);
1006                 return NULL;
1007         }
1008         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1009         lgp->res.layoutp = &lgp->args.layout;
1010
1011         /* Don't confuse uninitialised result and success */
1012         lgp->res.status = -NFS4ERR_DELAY;
1013
1014         lgp->args.minlength = PAGE_SIZE;
1015         if (lgp->args.minlength > range->length)
1016                 lgp->args.minlength = range->length;
1017         if (ino) {
1018                 loff_t i_size = i_size_read(ino);
1019
1020                 if (range->iomode == IOMODE_READ) {
1021                         if (range->offset >= i_size)
1022                                 lgp->args.minlength = 0;
1023                         else if (i_size - range->offset < lgp->args.minlength)
1024                                 lgp->args.minlength = i_size - range->offset;
1025                 }
1026         }
1027         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1028         pnfs_copy_range(&lgp->args.range, range);
1029         lgp->args.type = server->pnfs_curr_ld->id;
1030         lgp->args.inode = ino;
1031         lgp->args.ctx = get_nfs_open_context(ctx);
1032         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1033         lgp->gfp_flags = gfp_flags;
1034         lgp->cred = get_rpccred(ctx->cred);
1035         return lgp;
1036 }
1037
1038 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1039 {
1040         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1041
1042         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1043         if (lgp->args.inode)
1044                 pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1045         put_rpccred(lgp->cred);
1046         put_nfs_open_context(lgp->args.ctx);
1047         kfree(lgp);
1048 }
1049
1050 static void pnfs_clear_layoutcommit(struct inode *inode,
1051                 struct list_head *head)
1052 {
1053         struct nfs_inode *nfsi = NFS_I(inode);
1054         struct pnfs_layout_segment *lseg, *tmp;
1055
1056         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1057                 return;
1058         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1059                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1060                         continue;
1061                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1062         }
1063 }
1064
1065 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1066                 const nfs4_stateid *arg_stateid,
1067                 const struct pnfs_layout_range *range,
1068                 const nfs4_stateid *stateid)
1069 {
1070         struct inode *inode = lo->plh_inode;
1071         LIST_HEAD(freeme);
1072
1073         spin_lock(&inode->i_lock);
1074         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1075             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1076                 goto out_unlock;
1077         if (stateid) {
1078                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1079
1080                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1081                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1082                 pnfs_set_layout_stateid(lo, stateid, true);
1083         } else
1084                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1085 out_unlock:
1086         pnfs_clear_layoutreturn_waitbit(lo);
1087         spin_unlock(&inode->i_lock);
1088         pnfs_free_lseg_list(&freeme);
1089
1090 }
1091
1092 static bool
1093 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1094                 nfs4_stateid *stateid,
1095                 enum pnfs_iomode *iomode)
1096 {
1097         /* Serialise LAYOUTGET/LAYOUTRETURN */
1098         if (atomic_read(&lo->plh_outstanding) != 0)
1099                 return false;
1100         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1101                 return false;
1102         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1103         pnfs_get_layout_hdr(lo);
1104         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1105                 if (stateid != NULL) {
1106                         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1107                         if (lo->plh_return_seq != 0)
1108                                 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1109                 }
1110                 if (iomode != NULL)
1111                         *iomode = lo->plh_return_iomode;
1112                 pnfs_clear_layoutreturn_info(lo);
1113                 return true;
1114         }
1115         if (stateid != NULL)
1116                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1117         if (iomode != NULL)
1118                 *iomode = IOMODE_ANY;
1119         return true;
1120 }
1121
1122 static void
1123 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1124                 struct pnfs_layout_hdr *lo,
1125                 const nfs4_stateid *stateid,
1126                 enum pnfs_iomode iomode)
1127 {
1128         struct inode *inode = lo->plh_inode;
1129
1130         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1131         args->inode = inode;
1132         args->range.iomode = iomode;
1133         args->range.offset = 0;
1134         args->range.length = NFS4_MAX_UINT64;
1135         args->layout = lo;
1136         nfs4_stateid_copy(&args->stateid, stateid);
1137 }
1138
1139 static int
1140 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1141                        enum pnfs_iomode iomode, bool sync)
1142 {
1143         struct inode *ino = lo->plh_inode;
1144         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1145         struct nfs4_layoutreturn *lrp;
1146         int status = 0;
1147
1148         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1149         if (unlikely(lrp == NULL)) {
1150                 status = -ENOMEM;
1151                 spin_lock(&ino->i_lock);
1152                 pnfs_clear_layoutreturn_waitbit(lo);
1153                 spin_unlock(&ino->i_lock);
1154                 pnfs_put_layout_hdr(lo);
1155                 goto out;
1156         }
1157
1158         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1159         lrp->args.ld_private = &lrp->ld_private;
1160         lrp->clp = NFS_SERVER(ino)->nfs_client;
1161         lrp->cred = lo->plh_lc_cred;
1162         if (ld->prepare_layoutreturn)
1163                 ld->prepare_layoutreturn(&lrp->args);
1164
1165         status = nfs4_proc_layoutreturn(lrp, sync);
1166 out:
1167         dprintk("<-- %s status: %d\n", __func__, status);
1168         return status;
1169 }
1170
1171 /* Return true if layoutreturn is needed */
1172 static bool
1173 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1174 {
1175         struct pnfs_layout_segment *s;
1176         enum pnfs_iomode iomode;
1177         u32 seq;
1178
1179         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1180                 return false;
1181
1182         seq = lo->plh_return_seq;
1183         iomode = lo->plh_return_iomode;
1184
1185         /* Defer layoutreturn until all recalled lsegs are done */
1186         list_for_each_entry(s, &lo->plh_segs, pls_list) {
1187                 if (seq && pnfs_seqid_is_newer(s->pls_seq, seq))
1188                         continue;
1189                 if (iomode != IOMODE_ANY && s->pls_range.iomode != iomode)
1190                         continue;
1191                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1192                         return false;
1193         }
1194
1195         return true;
1196 }
1197
1198 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1199 {
1200         struct inode *inode= lo->plh_inode;
1201
1202         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1203                 return;
1204         spin_lock(&inode->i_lock);
1205         if (pnfs_layout_need_return(lo)) {
1206                 nfs4_stateid stateid;
1207                 enum pnfs_iomode iomode;
1208                 bool send;
1209
1210                 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1211                 spin_unlock(&inode->i_lock);
1212                 if (send) {
1213                         /* Send an async layoutreturn so we dont deadlock */
1214                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1215                 }
1216         } else
1217                 spin_unlock(&inode->i_lock);
1218 }
1219
1220 /*
1221  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1222  * when the layout segment list is empty.
1223  *
1224  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1225  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1226  * deviceid is marked invalid.
1227  */
1228 int
1229 _pnfs_return_layout(struct inode *ino)
1230 {
1231         struct pnfs_layout_hdr *lo = NULL;
1232         struct nfs_inode *nfsi = NFS_I(ino);
1233         LIST_HEAD(tmp_list);
1234         nfs4_stateid stateid;
1235         int status = 0;
1236         bool send, valid_layout;
1237
1238         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1239
1240         spin_lock(&ino->i_lock);
1241         lo = nfsi->layout;
1242         if (!lo) {
1243                 spin_unlock(&ino->i_lock);
1244                 dprintk("NFS: %s no layout to return\n", __func__);
1245                 goto out;
1246         }
1247         /* Reference matched in nfs4_layoutreturn_release */
1248         pnfs_get_layout_hdr(lo);
1249         /* Is there an outstanding layoutreturn ? */
1250         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1251                 spin_unlock(&ino->i_lock);
1252                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1253                                         TASK_UNINTERRUPTIBLE))
1254                         goto out_put_layout_hdr;
1255                 spin_lock(&ino->i_lock);
1256         }
1257         valid_layout = pnfs_layout_is_valid(lo);
1258         pnfs_clear_layoutcommit(ino, &tmp_list);
1259         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1260
1261         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1262                 struct pnfs_layout_range range = {
1263                         .iomode         = IOMODE_ANY,
1264                         .offset         = 0,
1265                         .length         = NFS4_MAX_UINT64,
1266                 };
1267                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1268         }
1269
1270         /* Don't send a LAYOUTRETURN if list was initially empty */
1271         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1272                         !valid_layout) {
1273                 spin_unlock(&ino->i_lock);
1274                 dprintk("NFS: %s no layout segments to return\n", __func__);
1275                 goto out_put_layout_hdr;
1276         }
1277
1278         send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1279         spin_unlock(&ino->i_lock);
1280         if (send)
1281                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1282 out_put_layout_hdr:
1283         pnfs_free_lseg_list(&tmp_list);
1284         pnfs_put_layout_hdr(lo);
1285 out:
1286         dprintk("<-- %s status: %d\n", __func__, status);
1287         return status;
1288 }
1289
1290 int
1291 pnfs_commit_and_return_layout(struct inode *inode)
1292 {
1293         struct pnfs_layout_hdr *lo;
1294         int ret;
1295
1296         spin_lock(&inode->i_lock);
1297         lo = NFS_I(inode)->layout;
1298         if (lo == NULL) {
1299                 spin_unlock(&inode->i_lock);
1300                 return 0;
1301         }
1302         pnfs_get_layout_hdr(lo);
1303         /* Block new layoutgets and read/write to ds */
1304         lo->plh_block_lgets++;
1305         spin_unlock(&inode->i_lock);
1306         filemap_fdatawait(inode->i_mapping);
1307         ret = pnfs_layoutcommit_inode(inode, true);
1308         if (ret == 0)
1309                 ret = _pnfs_return_layout(inode);
1310         spin_lock(&inode->i_lock);
1311         lo->plh_block_lgets--;
1312         spin_unlock(&inode->i_lock);
1313         pnfs_put_layout_hdr(lo);
1314         return ret;
1315 }
1316
1317 bool pnfs_roc(struct inode *ino,
1318                 struct nfs4_layoutreturn_args *args,
1319                 struct nfs4_layoutreturn_res *res,
1320                 const struct rpc_cred *cred)
1321 {
1322         struct nfs_inode *nfsi = NFS_I(ino);
1323         struct nfs_open_context *ctx;
1324         struct nfs4_state *state;
1325         struct pnfs_layout_hdr *lo;
1326         struct pnfs_layout_segment *lseg, *next;
1327         nfs4_stateid stateid;
1328         enum pnfs_iomode iomode = 0;
1329         bool layoutreturn = false, roc = false;
1330         bool skip_read = false;
1331
1332         if (!nfs_have_layout(ino))
1333                 return false;
1334 retry:
1335         spin_lock(&ino->i_lock);
1336         lo = nfsi->layout;
1337         if (!lo || !pnfs_layout_is_valid(lo) ||
1338             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1339                 lo = NULL;
1340                 goto out_noroc;
1341         }
1342         pnfs_get_layout_hdr(lo);
1343         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1344                 spin_unlock(&ino->i_lock);
1345                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1346                                 TASK_UNINTERRUPTIBLE);
1347                 pnfs_put_layout_hdr(lo);
1348                 goto retry;
1349         }
1350
1351         /* no roc if we hold a delegation */
1352         if (nfs4_check_delegation(ino, FMODE_READ)) {
1353                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1354                         goto out_noroc;
1355                 skip_read = true;
1356         }
1357
1358         list_for_each_entry(ctx, &nfsi->open_files, list) {
1359                 state = ctx->state;
1360                 if (state == NULL)
1361                         continue;
1362                 /* Don't return layout if there is open file state */
1363                 if (state->state & FMODE_WRITE)
1364                         goto out_noroc;
1365                 if (state->state & FMODE_READ)
1366                         skip_read = true;
1367         }
1368
1369
1370         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1371                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1372                         continue;
1373                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1374                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1375                         continue;
1376                 /*
1377                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1378                  * doesn't invalidate the layout for us.
1379                  */
1380                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1381                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1382                         continue;
1383                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1384         }
1385
1386         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1387                 goto out_noroc;
1388
1389         /* ROC in two conditions:
1390          * 1. there are ROC lsegs
1391          * 2. we don't send layoutreturn
1392          */
1393         /* lo ref dropped in pnfs_roc_release() */
1394         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1395         /* If the creds don't match, we can't compound the layoutreturn */
1396         if (!layoutreturn || cred != lo->plh_lc_cred)
1397                 goto out_noroc;
1398
1399         roc = layoutreturn;
1400         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1401         res->lrs_present = 0;
1402         layoutreturn = false;
1403
1404 out_noroc:
1405         spin_unlock(&ino->i_lock);
1406         pnfs_layoutcommit_inode(ino, true);
1407         if (roc) {
1408                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1409                 if (ld->prepare_layoutreturn)
1410                         ld->prepare_layoutreturn(args);
1411                 pnfs_put_layout_hdr(lo);
1412                 return true;
1413         }
1414         if (layoutreturn)
1415                 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1416         pnfs_put_layout_hdr(lo);
1417         return false;
1418 }
1419
1420 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1421                 struct nfs4_layoutreturn_res *res,
1422                 int ret)
1423 {
1424         struct pnfs_layout_hdr *lo = args->layout;
1425         const nfs4_stateid *arg_stateid = NULL;
1426         const nfs4_stateid *res_stateid = NULL;
1427         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1428
1429         if (ret == 0) {
1430                 arg_stateid = &args->stateid;
1431                 if (res->lrs_present)
1432                         res_stateid = &res->stateid;
1433         }
1434         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1435                         res_stateid);
1436         if (ld_private && ld_private->ops && ld_private->ops->free)
1437                 ld_private->ops->free(ld_private);
1438         pnfs_put_layout_hdr(lo);
1439         trace_nfs4_layoutreturn_on_close(args->inode, 0);
1440 }
1441
1442 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1443 {
1444         struct nfs_inode *nfsi = NFS_I(ino);
1445         struct pnfs_layout_hdr *lo;
1446         bool sleep = false;
1447
1448         /* we might not have grabbed lo reference. so need to check under
1449          * i_lock */
1450         spin_lock(&ino->i_lock);
1451         lo = nfsi->layout;
1452         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1453                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1454                 sleep = true;
1455         }
1456         spin_unlock(&ino->i_lock);
1457         return sleep;
1458 }
1459
1460 /*
1461  * Compare two layout segments for sorting into layout cache.
1462  * We want to preferentially return RW over RO layouts, so ensure those
1463  * are seen first.
1464  */
1465 static s64
1466 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1467            const struct pnfs_layout_range *l2)
1468 {
1469         s64 d;
1470
1471         /* high offset > low offset */
1472         d = l1->offset - l2->offset;
1473         if (d)
1474                 return d;
1475
1476         /* short length > long length */
1477         d = l2->length - l1->length;
1478         if (d)
1479                 return d;
1480
1481         /* read > read/write */
1482         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1483 }
1484
1485 static bool
1486 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1487                 const struct pnfs_layout_range *l2)
1488 {
1489         return pnfs_lseg_range_cmp(l1, l2) > 0;
1490 }
1491
1492 static bool
1493 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1494                 struct pnfs_layout_segment *old)
1495 {
1496         return false;
1497 }
1498
1499 void
1500 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1501                    struct pnfs_layout_segment *lseg,
1502                    bool (*is_after)(const struct pnfs_layout_range *,
1503                            const struct pnfs_layout_range *),
1504                    bool (*do_merge)(struct pnfs_layout_segment *,
1505                            struct pnfs_layout_segment *),
1506                    struct list_head *free_me)
1507 {
1508         struct pnfs_layout_segment *lp, *tmp;
1509
1510         dprintk("%s:Begin\n", __func__);
1511
1512         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1513                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1514                         continue;
1515                 if (do_merge(lseg, lp)) {
1516                         mark_lseg_invalid(lp, free_me);
1517                         continue;
1518                 }
1519                 if (is_after(&lseg->pls_range, &lp->pls_range))
1520                         continue;
1521                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1522                 dprintk("%s: inserted lseg %p "
1523                         "iomode %d offset %llu length %llu before "
1524                         "lp %p iomode %d offset %llu length %llu\n",
1525                         __func__, lseg, lseg->pls_range.iomode,
1526                         lseg->pls_range.offset, lseg->pls_range.length,
1527                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1528                         lp->pls_range.length);
1529                 goto out;
1530         }
1531         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1532         dprintk("%s: inserted lseg %p "
1533                 "iomode %d offset %llu length %llu at tail\n",
1534                 __func__, lseg, lseg->pls_range.iomode,
1535                 lseg->pls_range.offset, lseg->pls_range.length);
1536 out:
1537         pnfs_get_layout_hdr(lo);
1538
1539         dprintk("%s:Return\n", __func__);
1540 }
1541 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1542
1543 static void
1544 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1545                    struct pnfs_layout_segment *lseg,
1546                    struct list_head *free_me)
1547 {
1548         struct inode *inode = lo->plh_inode;
1549         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1550
1551         if (ld->add_lseg != NULL)
1552                 ld->add_lseg(lo, lseg, free_me);
1553         else
1554                 pnfs_generic_layout_insert_lseg(lo, lseg,
1555                                 pnfs_lseg_range_is_after,
1556                                 pnfs_lseg_no_merge,
1557                                 free_me);
1558 }
1559
1560 static struct pnfs_layout_hdr *
1561 alloc_init_layout_hdr(struct inode *ino,
1562                       struct nfs_open_context *ctx,
1563                       gfp_t gfp_flags)
1564 {
1565         struct pnfs_layout_hdr *lo;
1566
1567         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1568         if (!lo)
1569                 return NULL;
1570         refcount_set(&lo->plh_refcount, 1);
1571         INIT_LIST_HEAD(&lo->plh_layouts);
1572         INIT_LIST_HEAD(&lo->plh_segs);
1573         INIT_LIST_HEAD(&lo->plh_return_segs);
1574         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1575         lo->plh_inode = ino;
1576         lo->plh_lc_cred = get_rpccred(ctx->cred);
1577         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1578         return lo;
1579 }
1580
1581 static struct pnfs_layout_hdr *
1582 pnfs_find_alloc_layout(struct inode *ino,
1583                        struct nfs_open_context *ctx,
1584                        gfp_t gfp_flags)
1585         __releases(&ino->i_lock)
1586         __acquires(&ino->i_lock)
1587 {
1588         struct nfs_inode *nfsi = NFS_I(ino);
1589         struct pnfs_layout_hdr *new = NULL;
1590
1591         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1592
1593         if (nfsi->layout != NULL)
1594                 goto out_existing;
1595         spin_unlock(&ino->i_lock);
1596         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1597         spin_lock(&ino->i_lock);
1598
1599         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1600                 nfsi->layout = new;
1601                 return new;
1602         } else if (new != NULL)
1603                 pnfs_free_layout_hdr(new);
1604 out_existing:
1605         pnfs_get_layout_hdr(nfsi->layout);
1606         return nfsi->layout;
1607 }
1608
1609 /*
1610  * iomode matching rules:
1611  * iomode       lseg    strict match
1612  *                      iomode
1613  * -----        -----   ------ -----
1614  * ANY          READ    N/A    true
1615  * ANY          RW      N/A    true
1616  * RW           READ    N/A    false
1617  * RW           RW      N/A    true
1618  * READ         READ    N/A    true
1619  * READ         RW      true   false
1620  * READ         RW      false  true
1621  */
1622 static bool
1623 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1624                  const struct pnfs_layout_range *range,
1625                  bool strict_iomode)
1626 {
1627         struct pnfs_layout_range range1;
1628
1629         if ((range->iomode == IOMODE_RW &&
1630              ls_range->iomode != IOMODE_RW) ||
1631             (range->iomode != ls_range->iomode &&
1632              strict_iomode) ||
1633             !pnfs_lseg_range_intersecting(ls_range, range))
1634                 return false;
1635
1636         /* range1 covers only the first byte in the range */
1637         range1 = *range;
1638         range1.length = 1;
1639         return pnfs_lseg_range_contained(ls_range, &range1);
1640 }
1641
1642 /*
1643  * lookup range in layout
1644  */
1645 static struct pnfs_layout_segment *
1646 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1647                 struct pnfs_layout_range *range,
1648                 bool strict_iomode)
1649 {
1650         struct pnfs_layout_segment *lseg, *ret = NULL;
1651
1652         dprintk("%s:Begin\n", __func__);
1653
1654         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1655                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1656                     pnfs_lseg_range_match(&lseg->pls_range, range,
1657                                           strict_iomode)) {
1658                         ret = pnfs_get_lseg(lseg);
1659                         break;
1660                 }
1661         }
1662
1663         dprintk("%s:Return lseg %p ref %d\n",
1664                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1665         return ret;
1666 }
1667
1668 /*
1669  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1670  * to the MDS or over pNFS
1671  *
1672  * The nfs_inode read_io and write_io fields are cumulative counters reset
1673  * when there are no layout segments. Note that in pnfs_update_layout iomode
1674  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1675  * WRITE request.
1676  *
1677  * A return of true means use MDS I/O.
1678  *
1679  * From rfc 5661:
1680  * If a file's size is smaller than the file size threshold, data accesses
1681  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1682  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1683  * server.  If both file size and I/O size are provided, the client SHOULD
1684  * reach or exceed  both thresholds before sending its read or write
1685  * requests to the data server.
1686  */
1687 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1688                                      struct inode *ino, int iomode)
1689 {
1690         struct nfs4_threshold *t = ctx->mdsthreshold;
1691         struct nfs_inode *nfsi = NFS_I(ino);
1692         loff_t fsize = i_size_read(ino);
1693         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1694
1695         if (t == NULL)
1696                 return ret;
1697
1698         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1699                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1700
1701         switch (iomode) {
1702         case IOMODE_READ:
1703                 if (t->bm & THRESHOLD_RD) {
1704                         dprintk("%s fsize %llu\n", __func__, fsize);
1705                         size_set = true;
1706                         if (fsize < t->rd_sz)
1707                                 size = true;
1708                 }
1709                 if (t->bm & THRESHOLD_RD_IO) {
1710                         dprintk("%s nfsi->read_io %llu\n", __func__,
1711                                 nfsi->read_io);
1712                         io_set = true;
1713                         if (nfsi->read_io < t->rd_io_sz)
1714                                 io = true;
1715                 }
1716                 break;
1717         case IOMODE_RW:
1718                 if (t->bm & THRESHOLD_WR) {
1719                         dprintk("%s fsize %llu\n", __func__, fsize);
1720                         size_set = true;
1721                         if (fsize < t->wr_sz)
1722                                 size = true;
1723                 }
1724                 if (t->bm & THRESHOLD_WR_IO) {
1725                         dprintk("%s nfsi->write_io %llu\n", __func__,
1726                                 nfsi->write_io);
1727                         io_set = true;
1728                         if (nfsi->write_io < t->wr_io_sz)
1729                                 io = true;
1730                 }
1731                 break;
1732         }
1733         if (size_set && io_set) {
1734                 if (size && io)
1735                         ret = true;
1736         } else if (size || io)
1737                 ret = true;
1738
1739         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1740         return ret;
1741 }
1742
1743 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1744 {
1745         /*
1746          * send layoutcommit as it can hold up layoutreturn due to lseg
1747          * reference
1748          */
1749         pnfs_layoutcommit_inode(lo->plh_inode, false);
1750         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1751                                    nfs_wait_bit_killable,
1752                                    TASK_UNINTERRUPTIBLE);
1753 }
1754
1755 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1756 {
1757         atomic_inc(&lo->plh_outstanding);
1758 }
1759
1760 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1761 {
1762         if (atomic_dec_and_test(&lo->plh_outstanding))
1763                 wake_up_var(&lo->plh_outstanding);
1764 }
1765
1766 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1767 {
1768         unsigned long *bitlock = &lo->plh_flags;
1769
1770         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1771         smp_mb__after_atomic();
1772         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1773 }
1774
1775 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1776                                 struct nfs_server *server)
1777 {
1778         if (list_empty(&lo->plh_layouts)) {
1779                 struct nfs_client *clp = server->nfs_client;
1780
1781                 /* The lo must be on the clp list if there is any
1782                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1783                  */
1784                 spin_lock(&clp->cl_lock);
1785                 if (list_empty(&lo->plh_layouts))
1786                         list_add_tail(&lo->plh_layouts, &server->layouts);
1787                 spin_unlock(&clp->cl_lock);
1788         }
1789 }
1790
1791 /*
1792  * Layout segment is retreived from the server if not cached.
1793  * The appropriate layout segment is referenced and returned to the caller.
1794  */
1795 struct pnfs_layout_segment *
1796 pnfs_update_layout(struct inode *ino,
1797                    struct nfs_open_context *ctx,
1798                    loff_t pos,
1799                    u64 count,
1800                    enum pnfs_iomode iomode,
1801                    bool strict_iomode,
1802                    gfp_t gfp_flags)
1803 {
1804         struct pnfs_layout_range arg = {
1805                 .iomode = iomode,
1806                 .offset = pos,
1807                 .length = count,
1808         };
1809         unsigned pg_offset;
1810         struct nfs_server *server = NFS_SERVER(ino);
1811         struct nfs_client *clp = server->nfs_client;
1812         struct pnfs_layout_hdr *lo = NULL;
1813         struct pnfs_layout_segment *lseg = NULL;
1814         struct nfs4_layoutget *lgp;
1815         nfs4_stateid stateid;
1816         long timeout = 0;
1817         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1818         bool first;
1819
1820         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1821                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1822                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1823                 goto out;
1824         }
1825
1826         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1827                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1828                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1829                 goto out;
1830         }
1831
1832 lookup_again:
1833         nfs4_client_recover_expired_lease(clp);
1834         first = false;
1835         spin_lock(&ino->i_lock);
1836         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1837         if (lo == NULL) {
1838                 spin_unlock(&ino->i_lock);
1839                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1840                                  PNFS_UPDATE_LAYOUT_NOMEM);
1841                 goto out;
1842         }
1843
1844         /* Do we even need to bother with this? */
1845         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1846                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1847                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1848                 dprintk("%s matches recall, use MDS\n", __func__);
1849                 goto out_unlock;
1850         }
1851
1852         /* if LAYOUTGET already failed once we don't try again */
1853         if (pnfs_layout_io_test_failed(lo, iomode)) {
1854                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1855                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1856                 goto out_unlock;
1857         }
1858
1859         /*
1860          * If the layout segment list is empty, but there are outstanding
1861          * layoutget calls, then they might be subject to a layoutrecall.
1862          */
1863         if (list_empty(&lo->plh_segs) &&
1864             atomic_read(&lo->plh_outstanding) != 0) {
1865                 spin_unlock(&ino->i_lock);
1866                 if (wait_var_event_killable(&lo->plh_outstanding,
1867                                         atomic_read(&lo->plh_outstanding) == 0
1868                                         || !list_empty(&lo->plh_segs)))
1869                         goto out_put_layout_hdr;
1870                 pnfs_put_layout_hdr(lo);
1871                 goto lookup_again;
1872         }
1873
1874         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1875         if (lseg) {
1876                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1877                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1878                 goto out_unlock;
1879         }
1880
1881         if (!nfs4_valid_open_stateid(ctx->state)) {
1882                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1883                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1884                 goto out_unlock;
1885         }
1886
1887         /*
1888          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1889          * stateid, or it has been invalidated, then we must use the open
1890          * stateid.
1891          */
1892         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1893
1894                 /*
1895                  * The first layoutget for the file. Need to serialize per
1896                  * RFC 5661 Errata 3208.
1897                  */
1898                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1899                                      &lo->plh_flags)) {
1900                         spin_unlock(&ino->i_lock);
1901                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1902                                     TASK_UNINTERRUPTIBLE);
1903                         pnfs_put_layout_hdr(lo);
1904                         dprintk("%s retrying\n", __func__);
1905                         goto lookup_again;
1906                 }
1907
1908                 first = true;
1909                 if (nfs4_select_rw_stateid(ctx->state,
1910                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1911                                         NULL, &stateid, NULL) != 0) {
1912                         trace_pnfs_update_layout(ino, pos, count,
1913                                         iomode, lo, lseg,
1914                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1915                         goto out_unlock;
1916                 }
1917         } else {
1918                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1919         }
1920
1921         /*
1922          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1923          * for LAYOUTRETURN even if first is true.
1924          */
1925         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1926                 spin_unlock(&ino->i_lock);
1927                 dprintk("%s wait for layoutreturn\n", __func__);
1928                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1929                         if (first)
1930                                 pnfs_clear_first_layoutget(lo);
1931                         pnfs_put_layout_hdr(lo);
1932                         dprintk("%s retrying\n", __func__);
1933                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1934                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1935                         goto lookup_again;
1936                 }
1937                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1938                                 PNFS_UPDATE_LAYOUT_RETURN);
1939                 goto out_put_layout_hdr;
1940         }
1941
1942         if (pnfs_layoutgets_blocked(lo)) {
1943                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1944                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1945                 goto out_unlock;
1946         }
1947         nfs_layoutget_begin(lo);
1948         spin_unlock(&ino->i_lock);
1949
1950         _add_to_server_list(lo, server);
1951
1952         pg_offset = arg.offset & ~PAGE_MASK;
1953         if (pg_offset) {
1954                 arg.offset -= pg_offset;
1955                 arg.length += pg_offset;
1956         }
1957         if (arg.length != NFS4_MAX_UINT64)
1958                 arg.length = PAGE_ALIGN(arg.length);
1959
1960         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
1961         if (!lgp) {
1962                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
1963                                          PNFS_UPDATE_LAYOUT_NOMEM);
1964                 nfs_layoutget_end(lo);
1965                 goto out_put_layout_hdr;
1966         }
1967
1968         lseg = nfs4_proc_layoutget(lgp, &timeout);
1969         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1970                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1971         nfs_layoutget_end(lo);
1972         if (IS_ERR(lseg)) {
1973                 switch(PTR_ERR(lseg)) {
1974                 case -EBUSY:
1975                         if (time_after(jiffies, giveup))
1976                                 lseg = NULL;
1977                         break;
1978                 case -ERECALLCONFLICT:
1979                 case -EAGAIN:
1980                         break;
1981                 default:
1982                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1983                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1984                                 lseg = NULL;
1985                         }
1986                         goto out_put_layout_hdr;
1987                 }
1988                 if (lseg) {
1989                         if (first)
1990                                 pnfs_clear_first_layoutget(lo);
1991                         trace_pnfs_update_layout(ino, pos, count,
1992                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1993                         pnfs_put_layout_hdr(lo);
1994                         goto lookup_again;
1995                 }
1996         } else {
1997                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1998         }
1999
2000 out_put_layout_hdr:
2001         if (first)
2002                 pnfs_clear_first_layoutget(lo);
2003         pnfs_put_layout_hdr(lo);
2004 out:
2005         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2006                         "(%s, offset: %llu, length: %llu)\n",
2007                         __func__, ino->i_sb->s_id,
2008                         (unsigned long long)NFS_FILEID(ino),
2009                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2010                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2011                         (unsigned long long)pos,
2012                         (unsigned long long)count);
2013         return lseg;
2014 out_unlock:
2015         spin_unlock(&ino->i_lock);
2016         goto out_put_layout_hdr;
2017 }
2018 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2019
2020 static bool
2021 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2022 {
2023         switch (range->iomode) {
2024         case IOMODE_READ:
2025         case IOMODE_RW:
2026                 break;
2027         default:
2028                 return false;
2029         }
2030         if (range->offset == NFS4_MAX_UINT64)
2031                 return false;
2032         if (range->length == 0)
2033                 return false;
2034         if (range->length != NFS4_MAX_UINT64 &&
2035             range->length > NFS4_MAX_UINT64 - range->offset)
2036                 return false;
2037         return true;
2038 }
2039
2040 static struct pnfs_layout_hdr *
2041 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2042 {
2043         struct pnfs_layout_hdr *lo;
2044
2045         spin_lock(&ino->i_lock);
2046         lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2047         if (!lo)
2048                 goto out_unlock;
2049         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2050                 goto out_unlock;
2051         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2052                 goto out_unlock;
2053         if (pnfs_layoutgets_blocked(lo))
2054                 goto out_unlock;
2055         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2056                 goto out_unlock;
2057         nfs_layoutget_begin(lo);
2058         spin_unlock(&ino->i_lock);
2059         _add_to_server_list(lo, NFS_SERVER(ino));
2060         return lo;
2061
2062 out_unlock:
2063         spin_unlock(&ino->i_lock);
2064         pnfs_put_layout_hdr(lo);
2065         return NULL;
2066 }
2067
2068 extern const nfs4_stateid current_stateid;
2069
2070 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2071                                      struct nfs_open_context *ctx)
2072 {
2073         struct inode *ino = data->dentry->d_inode;
2074         struct pnfs_layout_range rng = {
2075                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2076                           IOMODE_RW: IOMODE_READ,
2077                 .offset = 0,
2078                 .length = NFS4_MAX_UINT64,
2079         };
2080         struct nfs4_layoutget *lgp;
2081         struct pnfs_layout_hdr *lo;
2082
2083         /* Heuristic: don't send layoutget if we have cached data */
2084         if (rng.iomode == IOMODE_READ &&
2085            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2086                 return;
2087
2088         lo = _pnfs_grab_empty_layout(ino, ctx);
2089         if (!lo)
2090                 return;
2091         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2092                                              &rng, GFP_KERNEL);
2093         if (!lgp) {
2094                 pnfs_clear_first_layoutget(lo);
2095                 pnfs_put_layout_hdr(lo);
2096                 return;
2097         }
2098         data->lgp = lgp;
2099         data->o_arg.lg_args = &lgp->args;
2100         data->o_res.lg_res = &lgp->res;
2101 }
2102
2103 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2104                                      struct nfs_open_context *ctx)
2105 {
2106         struct pnfs_layout_range rng = {
2107                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2108                           IOMODE_RW: IOMODE_READ,
2109                 .offset = 0,
2110                 .length = NFS4_MAX_UINT64,
2111         };
2112         struct nfs4_layoutget *lgp;
2113
2114         lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2115                                              &rng, GFP_KERNEL);
2116         if (!lgp)
2117                 return;
2118         data->lgp = lgp;
2119         data->o_arg.lg_args = &lgp->args;
2120         data->o_res.lg_res = &lgp->res;
2121 }
2122
2123 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2124                          struct nfs_open_context *ctx)
2125 {
2126         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2127
2128         if (!(pnfs_enabled_sb(server) &&
2129               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2130                 return;
2131         /* Could check on max_ops, but currently hardcoded high enough */
2132         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2133                 return;
2134         if (data->state)
2135                 _lgopen_prepare_attached(data, ctx);
2136         else
2137                 _lgopen_prepare_floating(data, ctx);
2138 }
2139
2140 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2141                        struct nfs_open_context *ctx)
2142 {
2143         struct pnfs_layout_hdr *lo;
2144         struct pnfs_layout_segment *lseg;
2145         struct nfs_server *srv = NFS_SERVER(ino);
2146         u32 iomode;
2147
2148         if (!lgp)
2149                 return;
2150         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2151         if (lgp->res.status) {
2152                 switch (lgp->res.status) {
2153                 default:
2154                         break;
2155                 /*
2156                  * Halt lgopen attempts if the server doesn't recognise
2157                  * the "current stateid" value, the layout type, or the
2158                  * layoutget operation as being valid.
2159                  * Also if it complains about too many ops in the compound
2160                  * or of the request/reply being too big.
2161                  */
2162                 case -NFS4ERR_BAD_STATEID:
2163                 case -NFS4ERR_NOTSUPP:
2164                 case -NFS4ERR_REP_TOO_BIG:
2165                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2166                 case -NFS4ERR_REQ_TOO_BIG:
2167                 case -NFS4ERR_TOO_MANY_OPS:
2168                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2169                         srv->caps &= ~NFS_CAP_LGOPEN;
2170                 }
2171                 return;
2172         }
2173         if (!lgp->args.inode) {
2174                 lo = _pnfs_grab_empty_layout(ino, ctx);
2175                 if (!lo)
2176                         return;
2177                 lgp->args.inode = ino;
2178         } else
2179                 lo = NFS_I(lgp->args.inode)->layout;
2180
2181         lseg = pnfs_layout_process(lgp);
2182         if (!IS_ERR(lseg)) {
2183                 iomode = lgp->args.range.iomode;
2184                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2185                 pnfs_put_lseg(lseg);
2186         }
2187 }
2188
2189 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2190 {
2191         if (lgp != NULL) {
2192                 struct inode *inode = lgp->args.inode;
2193                 if (inode) {
2194                         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2195                         pnfs_clear_first_layoutget(lo);
2196                         nfs_layoutget_end(lo);
2197                 }
2198                 pnfs_layoutget_free(lgp);
2199         }
2200 }
2201
2202 struct pnfs_layout_segment *
2203 pnfs_layout_process(struct nfs4_layoutget *lgp)
2204 {
2205         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2206         struct nfs4_layoutget_res *res = &lgp->res;
2207         struct pnfs_layout_segment *lseg;
2208         struct inode *ino = lo->plh_inode;
2209         LIST_HEAD(free_me);
2210
2211         if (!pnfs_sanity_check_layout_range(&res->range))
2212                 return ERR_PTR(-EINVAL);
2213
2214         /* Inject layout blob into I/O device driver */
2215         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2216         if (IS_ERR_OR_NULL(lseg)) {
2217                 if (!lseg)
2218                         lseg = ERR_PTR(-ENOMEM);
2219
2220                 dprintk("%s: Could not allocate layout: error %ld\n",
2221                        __func__, PTR_ERR(lseg));
2222                 return lseg;
2223         }
2224
2225         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2226
2227         spin_lock(&ino->i_lock);
2228         if (pnfs_layoutgets_blocked(lo)) {
2229                 dprintk("%s forget reply due to state\n", __func__);
2230                 goto out_forget;
2231         }
2232
2233         if (!pnfs_layout_is_valid(lo)) {
2234                 /* We have a completely new layout */
2235                 pnfs_set_layout_stateid(lo, &res->stateid, true);
2236         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2237                 /* existing state ID, make sure the sequence number matches. */
2238                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2239                         dprintk("%s forget reply due to sequence\n", __func__);
2240                         goto out_forget;
2241                 }
2242                 pnfs_set_layout_stateid(lo, &res->stateid, false);
2243         } else {
2244                 /*
2245                  * We got an entirely new state ID.  Mark all segments for the
2246                  * inode invalid, and retry the layoutget
2247                  */
2248                 pnfs_mark_layout_stateid_invalid(lo, &free_me);
2249                 goto out_forget;
2250         }
2251
2252         pnfs_get_lseg(lseg);
2253         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2254
2255
2256         if (res->return_on_close)
2257                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2258
2259         spin_unlock(&ino->i_lock);
2260         pnfs_free_lseg_list(&free_me);
2261         return lseg;
2262
2263 out_forget:
2264         spin_unlock(&ino->i_lock);
2265         lseg->pls_layout = lo;
2266         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2267         return ERR_PTR(-EAGAIN);
2268 }
2269
2270 static int
2271 mark_lseg_invalid_or_return(struct pnfs_layout_segment *lseg,
2272                 struct list_head *tmp_list)
2273 {
2274         if (!mark_lseg_invalid(lseg, tmp_list))
2275                 return 0;
2276         pnfs_cache_lseg_for_layoutreturn(lseg->pls_layout, lseg);
2277         return 1;
2278 }
2279
2280 /**
2281  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2282  * @lo: pointer to layout header
2283  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2284  * @return_range: describe layout segment ranges to be returned
2285  * @seq: stateid seqid to match
2286  *
2287  * This function is mainly intended for use by layoutrecall. It attempts
2288  * to free the layout segment immediately, or else to mark it for return
2289  * as soon as its reference count drops to zero.
2290  *
2291  * Returns
2292  * - 0: a layoutreturn needs to be scheduled.
2293  * - EBUSY: there are layout segment that are still in use.
2294  * - ENOENT: there are no layout segments that need to be returned.
2295  */
2296 int
2297 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2298                                 struct list_head *tmp_list,
2299                                 const struct pnfs_layout_range *return_range,
2300                                 u32 seq)
2301 {
2302         struct pnfs_layout_segment *lseg, *next;
2303         int remaining = 0;
2304
2305         dprintk("%s:Begin lo %p\n", __func__, lo);
2306
2307         assert_spin_locked(&lo->plh_inode->i_lock);
2308
2309         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2310                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2311                         dprintk("%s: marking lseg %p iomode %d "
2312                                 "offset %llu length %llu\n", __func__,
2313                                 lseg, lseg->pls_range.iomode,
2314                                 lseg->pls_range.offset,
2315                                 lseg->pls_range.length);
2316                         if (mark_lseg_invalid_or_return(lseg, tmp_list))
2317                                 continue;
2318                         remaining++;
2319                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2320                 }
2321
2322         if (remaining) {
2323                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2324                 return -EBUSY;
2325         }
2326
2327         if (!list_empty(&lo->plh_return_segs)) {
2328                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2329                 return 0;
2330         }
2331
2332         return -ENOENT;
2333 }
2334
2335 void pnfs_error_mark_layout_for_return(struct inode *inode,
2336                                        struct pnfs_layout_segment *lseg)
2337 {
2338         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2339         struct pnfs_layout_range range = {
2340                 .iomode = lseg->pls_range.iomode,
2341                 .offset = 0,
2342                 .length = NFS4_MAX_UINT64,
2343         };
2344         bool return_now = false;
2345
2346         spin_lock(&inode->i_lock);
2347         if (!pnfs_layout_is_valid(lo)) {
2348                 spin_unlock(&inode->i_lock);
2349                 return;
2350         }
2351         pnfs_set_plh_return_info(lo, range.iomode, 0);
2352         /*
2353          * mark all matching lsegs so that we are sure to have no live
2354          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2355          * for how it works.
2356          */
2357         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0) != -EBUSY) {
2358                 nfs4_stateid stateid;
2359                 enum pnfs_iomode iomode;
2360
2361                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2362                 spin_unlock(&inode->i_lock);
2363                 if (return_now)
2364                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2365         } else {
2366                 spin_unlock(&inode->i_lock);
2367                 nfs_commit_inode(inode, 0);
2368         }
2369 }
2370 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2371
2372 void
2373 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2374 {
2375         if (pgio->pg_lseg == NULL ||
2376             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2377                 return;
2378         pnfs_put_lseg(pgio->pg_lseg);
2379         pgio->pg_lseg = NULL;
2380 }
2381 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2382
2383 /*
2384  * Check for any intersection between the request and the pgio->pg_lseg,
2385  * and if none, put this pgio->pg_lseg away.
2386  */
2387 static void
2388 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2389 {
2390         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2391                 pnfs_put_lseg(pgio->pg_lseg);
2392                 pgio->pg_lseg = NULL;
2393         }
2394 }
2395
2396 void
2397 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2398 {
2399         u64 rd_size = req->wb_bytes;
2400
2401         pnfs_generic_pg_check_layout(pgio);
2402         pnfs_generic_pg_check_range(pgio, req);
2403         if (pgio->pg_lseg == NULL) {
2404                 if (pgio->pg_dreq == NULL)
2405                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2406                 else
2407                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2408
2409                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2410                                                    req->wb_context,
2411                                                    req_offset(req),
2412                                                    rd_size,
2413                                                    IOMODE_READ,
2414                                                    false,
2415                                                    GFP_KERNEL);
2416                 if (IS_ERR(pgio->pg_lseg)) {
2417                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2418                         pgio->pg_lseg = NULL;
2419                         return;
2420                 }
2421         }
2422         /* If no lseg, fall back to read through mds */
2423         if (pgio->pg_lseg == NULL)
2424                 nfs_pageio_reset_read_mds(pgio);
2425
2426 }
2427 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2428
2429 void
2430 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2431                            struct nfs_page *req, u64 wb_size)
2432 {
2433         pnfs_generic_pg_check_layout(pgio);
2434         pnfs_generic_pg_check_range(pgio, req);
2435         if (pgio->pg_lseg == NULL) {
2436                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2437                                                    req->wb_context,
2438                                                    req_offset(req),
2439                                                    wb_size,
2440                                                    IOMODE_RW,
2441                                                    false,
2442                                                    GFP_NOFS);
2443                 if (IS_ERR(pgio->pg_lseg)) {
2444                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2445                         pgio->pg_lseg = NULL;
2446                         return;
2447                 }
2448         }
2449         /* If no lseg, fall back to write through mds */
2450         if (pgio->pg_lseg == NULL)
2451                 nfs_pageio_reset_write_mds(pgio);
2452 }
2453 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2454
2455 void
2456 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2457 {
2458         if (desc->pg_lseg) {
2459                 pnfs_put_lseg(desc->pg_lseg);
2460                 desc->pg_lseg = NULL;
2461         }
2462 }
2463 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2464
2465 /*
2466  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2467  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2468  */
2469 size_t
2470 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2471                      struct nfs_page *prev, struct nfs_page *req)
2472 {
2473         unsigned int size;
2474         u64 seg_end, req_start, seg_left;
2475
2476         size = nfs_generic_pg_test(pgio, prev, req);
2477         if (!size)
2478                 return 0;
2479
2480         /*
2481          * 'size' contains the number of bytes left in the current page (up
2482          * to the original size asked for in @req->wb_bytes).
2483          *
2484          * Calculate how many bytes are left in the layout segment
2485          * and if there are less bytes than 'size', return that instead.
2486          *
2487          * Please also note that 'end_offset' is actually the offset of the
2488          * first byte that lies outside the pnfs_layout_range. FIXME?
2489          *
2490          */
2491         if (pgio->pg_lseg) {
2492                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2493                                      pgio->pg_lseg->pls_range.length);
2494                 req_start = req_offset(req);
2495
2496                 /* start of request is past the last byte of this segment */
2497                 if (req_start >= seg_end)
2498                         return 0;
2499
2500                 /* adjust 'size' iff there are fewer bytes left in the
2501                  * segment than what nfs_generic_pg_test returned */
2502                 seg_left = seg_end - req_start;
2503                 if (seg_left < size)
2504                         size = (unsigned int)seg_left;
2505         }
2506
2507         return size;
2508 }
2509 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2510
2511 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2512 {
2513         struct nfs_pageio_descriptor pgio;
2514
2515         /* Resend all requests through the MDS */
2516         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2517                               hdr->completion_ops);
2518         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2519         return nfs_pageio_resend(&pgio, hdr);
2520 }
2521 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2522
2523 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2524 {
2525
2526         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2527         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2528             PNFS_LAYOUTRET_ON_ERROR) {
2529                 pnfs_return_layout(hdr->inode);
2530         }
2531         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2532                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2533 }
2534
2535 /*
2536  * Called by non rpc-based layout drivers
2537  */
2538 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2539 {
2540         if (likely(!hdr->pnfs_error)) {
2541                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2542                                 hdr->mds_offset + hdr->res.count);
2543                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2544         }
2545         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2546         if (unlikely(hdr->pnfs_error))
2547                 pnfs_ld_handle_write_error(hdr);
2548         hdr->mds_ops->rpc_release(hdr);
2549 }
2550 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2551
2552 static void
2553 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2554                 struct nfs_pgio_header *hdr)
2555 {
2556         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2557
2558         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2559                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2560                 nfs_pageio_reset_write_mds(desc);
2561                 mirror->pg_recoalesce = 1;
2562         }
2563         hdr->completion_ops->completion(hdr);
2564 }
2565
2566 static enum pnfs_try_status
2567 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2568                         const struct rpc_call_ops *call_ops,
2569                         struct pnfs_layout_segment *lseg,
2570                         int how)
2571 {
2572         struct inode *inode = hdr->inode;
2573         enum pnfs_try_status trypnfs;
2574         struct nfs_server *nfss = NFS_SERVER(inode);
2575
2576         hdr->mds_ops = call_ops;
2577
2578         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2579                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2580         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2581         if (trypnfs != PNFS_NOT_ATTEMPTED)
2582                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2583         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2584         return trypnfs;
2585 }
2586
2587 static void
2588 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2589               struct nfs_pgio_header *hdr, int how)
2590 {
2591         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2592         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2593         enum pnfs_try_status trypnfs;
2594
2595         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2596         switch (trypnfs) {
2597         case PNFS_NOT_ATTEMPTED:
2598                 pnfs_write_through_mds(desc, hdr);
2599         case PNFS_ATTEMPTED:
2600                 break;
2601         case PNFS_TRY_AGAIN:
2602                 /* cleanup hdr and prepare to redo pnfs */
2603                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2604                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2605                         list_splice_init(&hdr->pages, &mirror->pg_list);
2606                         mirror->pg_recoalesce = 1;
2607                 }
2608                 hdr->mds_ops->rpc_release(hdr);
2609         }
2610 }
2611
2612 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2613 {
2614         pnfs_put_lseg(hdr->lseg);
2615         nfs_pgio_header_free(hdr);
2616 }
2617
2618 int
2619 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2620 {
2621         struct nfs_pgio_header *hdr;
2622         int ret;
2623
2624         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2625         if (!hdr) {
2626                 desc->pg_error = -ENOMEM;
2627                 return desc->pg_error;
2628         }
2629         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2630
2631         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2632         ret = nfs_generic_pgio(desc, hdr);
2633         if (!ret)
2634                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2635
2636         return ret;
2637 }
2638 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2639
2640 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2641 {
2642         struct nfs_pageio_descriptor pgio;
2643
2644         /* Resend all requests through the MDS */
2645         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2646         return nfs_pageio_resend(&pgio, hdr);
2647 }
2648 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2649
2650 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2651 {
2652         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2653         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2654             PNFS_LAYOUTRET_ON_ERROR) {
2655                 pnfs_return_layout(hdr->inode);
2656         }
2657         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2658                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2659 }
2660
2661 /*
2662  * Called by non rpc-based layout drivers
2663  */
2664 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2665 {
2666         if (likely(!hdr->pnfs_error))
2667                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2668         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2669         if (unlikely(hdr->pnfs_error))
2670                 pnfs_ld_handle_read_error(hdr);
2671         hdr->mds_ops->rpc_release(hdr);
2672 }
2673 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2674
2675 static void
2676 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2677                 struct nfs_pgio_header *hdr)
2678 {
2679         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2680
2681         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2682                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2683                 nfs_pageio_reset_read_mds(desc);
2684                 mirror->pg_recoalesce = 1;
2685         }
2686         hdr->completion_ops->completion(hdr);
2687 }
2688
2689 /*
2690  * Call the appropriate parallel I/O subsystem read function.
2691  */
2692 static enum pnfs_try_status
2693 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2694                        const struct rpc_call_ops *call_ops,
2695                        struct pnfs_layout_segment *lseg)
2696 {
2697         struct inode *inode = hdr->inode;
2698         struct nfs_server *nfss = NFS_SERVER(inode);
2699         enum pnfs_try_status trypnfs;
2700
2701         hdr->mds_ops = call_ops;
2702
2703         dprintk("%s: Reading ino:%lu %u@%llu\n",
2704                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2705
2706         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2707         if (trypnfs != PNFS_NOT_ATTEMPTED)
2708                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2709         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2710         return trypnfs;
2711 }
2712
2713 /* Resend all requests through pnfs. */
2714 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2715 {
2716         struct nfs_pageio_descriptor pgio;
2717
2718         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2719                 /* Prevent deadlocks with layoutreturn! */
2720                 pnfs_put_lseg(hdr->lseg);
2721                 hdr->lseg = NULL;
2722
2723                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2724                                         hdr->completion_ops);
2725                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2726         }
2727 }
2728 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2729
2730 static void
2731 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2732 {
2733         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2734         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2735         enum pnfs_try_status trypnfs;
2736
2737         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2738         switch (trypnfs) {
2739         case PNFS_NOT_ATTEMPTED:
2740                 pnfs_read_through_mds(desc, hdr);
2741         case PNFS_ATTEMPTED:
2742                 break;
2743         case PNFS_TRY_AGAIN:
2744                 /* cleanup hdr and prepare to redo pnfs */
2745                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2746                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2747                         list_splice_init(&hdr->pages, &mirror->pg_list);
2748                         mirror->pg_recoalesce = 1;
2749                 }
2750                 hdr->mds_ops->rpc_release(hdr);
2751         }
2752 }
2753
2754 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2755 {
2756         pnfs_put_lseg(hdr->lseg);
2757         nfs_pgio_header_free(hdr);
2758 }
2759
2760 int
2761 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2762 {
2763         struct nfs_pgio_header *hdr;
2764         int ret;
2765
2766         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2767         if (!hdr) {
2768                 desc->pg_error = -ENOMEM;
2769                 return desc->pg_error;
2770         }
2771         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2772         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2773         ret = nfs_generic_pgio(desc, hdr);
2774         if (!ret)
2775                 pnfs_do_read(desc, hdr);
2776         return ret;
2777 }
2778 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2779
2780 static void pnfs_clear_layoutcommitting(struct inode *inode)
2781 {
2782         unsigned long *bitlock = &NFS_I(inode)->flags;
2783
2784         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2785         smp_mb__after_atomic();
2786         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2787 }
2788
2789 /*
2790  * There can be multiple RW segments.
2791  */
2792 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2793 {
2794         struct pnfs_layout_segment *lseg;
2795
2796         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2797                 if (lseg->pls_range.iomode == IOMODE_RW &&
2798                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2799                         list_add(&lseg->pls_lc_list, listp);
2800         }
2801 }
2802
2803 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2804 {
2805         struct pnfs_layout_segment *lseg, *tmp;
2806
2807         /* Matched by references in pnfs_set_layoutcommit */
2808         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2809                 list_del_init(&lseg->pls_lc_list);
2810                 pnfs_put_lseg(lseg);
2811         }
2812
2813         pnfs_clear_layoutcommitting(inode);
2814 }
2815
2816 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2817 {
2818         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2819 }
2820 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2821
2822 void
2823 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2824                 loff_t end_pos)
2825 {
2826         struct nfs_inode *nfsi = NFS_I(inode);
2827         bool mark_as_dirty = false;
2828
2829         spin_lock(&inode->i_lock);
2830         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2831                 nfsi->layout->plh_lwb = end_pos;
2832                 mark_as_dirty = true;
2833                 dprintk("%s: Set layoutcommit for inode %lu ",
2834                         __func__, inode->i_ino);
2835         } else if (end_pos > nfsi->layout->plh_lwb)
2836                 nfsi->layout->plh_lwb = end_pos;
2837         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2838                 /* references matched in nfs4_layoutcommit_release */
2839                 pnfs_get_lseg(lseg);
2840         }
2841         spin_unlock(&inode->i_lock);
2842         dprintk("%s: lseg %p end_pos %llu\n",
2843                 __func__, lseg, nfsi->layout->plh_lwb);
2844
2845         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2846          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2847         if (mark_as_dirty)
2848                 mark_inode_dirty_sync(inode);
2849 }
2850 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2851
2852 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2853 {
2854         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2855
2856         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2857                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2858         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2859 }
2860
2861 /*
2862  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2863  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2864  * data to disk to allow the server to recover the data if it crashes.
2865  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2866  * is off, and a COMMIT is sent to a data server, or
2867  * if WRITEs to a data server return NFS_DATA_SYNC.
2868  */
2869 int
2870 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2871 {
2872         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2873         struct nfs4_layoutcommit_data *data;
2874         struct nfs_inode *nfsi = NFS_I(inode);
2875         loff_t end_pos;
2876         int status;
2877
2878         if (!pnfs_layoutcommit_outstanding(inode))
2879                 return 0;
2880
2881         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2882
2883         status = -EAGAIN;
2884         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2885                 if (!sync)
2886                         goto out;
2887                 status = wait_on_bit_lock_action(&nfsi->flags,
2888                                 NFS_INO_LAYOUTCOMMITTING,
2889                                 nfs_wait_bit_killable,
2890                                 TASK_KILLABLE);
2891                 if (status)
2892                         goto out;
2893         }
2894
2895         status = -ENOMEM;
2896         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2897         data = kzalloc(sizeof(*data), GFP_NOFS);
2898         if (!data)
2899                 goto clear_layoutcommitting;
2900
2901         status = 0;
2902         spin_lock(&inode->i_lock);
2903         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2904                 goto out_unlock;
2905
2906         INIT_LIST_HEAD(&data->lseg_list);
2907         pnfs_list_write_lseg(inode, &data->lseg_list);
2908
2909         end_pos = nfsi->layout->plh_lwb;
2910
2911         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2912         spin_unlock(&inode->i_lock);
2913
2914         data->args.inode = inode;
2915         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2916         nfs_fattr_init(&data->fattr);
2917         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2918         data->res.fattr = &data->fattr;
2919         if (end_pos != 0)
2920                 data->args.lastbytewritten = end_pos - 1;
2921         else
2922                 data->args.lastbytewritten = U64_MAX;
2923         data->res.server = NFS_SERVER(inode);
2924
2925         if (ld->prepare_layoutcommit) {
2926                 status = ld->prepare_layoutcommit(&data->args);
2927                 if (status) {
2928                         put_rpccred(data->cred);
2929                         spin_lock(&inode->i_lock);
2930                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2931                         if (end_pos > nfsi->layout->plh_lwb)
2932                                 nfsi->layout->plh_lwb = end_pos;
2933                         goto out_unlock;
2934                 }
2935         }
2936
2937
2938         status = nfs4_proc_layoutcommit(data, sync);
2939 out:
2940         if (status)
2941                 mark_inode_dirty_sync(inode);
2942         dprintk("<-- %s status %d\n", __func__, status);
2943         return status;
2944 out_unlock:
2945         spin_unlock(&inode->i_lock);
2946         kfree(data);
2947 clear_layoutcommitting:
2948         pnfs_clear_layoutcommitting(inode);
2949         goto out;
2950 }
2951 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2952
2953 int
2954 pnfs_generic_sync(struct inode *inode, bool datasync)
2955 {
2956         return pnfs_layoutcommit_inode(inode, true);
2957 }
2958 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2959
2960 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2961 {
2962         struct nfs4_threshold *thp;
2963
2964         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2965         if (!thp) {
2966                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2967                 return NULL;
2968         }
2969         return thp;
2970 }
2971
2972 #if IS_ENABLED(CONFIG_NFS_V4_2)
2973 int
2974 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2975 {
2976         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2977         struct nfs_server *server = NFS_SERVER(inode);
2978         struct nfs_inode *nfsi = NFS_I(inode);
2979         struct nfs42_layoutstat_data *data;
2980         struct pnfs_layout_hdr *hdr;
2981         int status = 0;
2982
2983         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2984                 goto out;
2985
2986         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2987                 goto out;
2988
2989         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2990                 goto out;
2991
2992         spin_lock(&inode->i_lock);
2993         if (!NFS_I(inode)->layout) {
2994                 spin_unlock(&inode->i_lock);
2995                 goto out_clear_layoutstats;
2996         }
2997         hdr = NFS_I(inode)->layout;
2998         pnfs_get_layout_hdr(hdr);
2999         spin_unlock(&inode->i_lock);
3000
3001         data = kzalloc(sizeof(*data), gfp_flags);
3002         if (!data) {
3003                 status = -ENOMEM;
3004                 goto out_put;
3005         }
3006
3007         data->args.fh = NFS_FH(inode);
3008         data->args.inode = inode;
3009         status = ld->prepare_layoutstats(&data->args);
3010         if (status)
3011                 goto out_free;
3012
3013         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3014
3015 out:
3016         dprintk("%s returns %d\n", __func__, status);
3017         return status;
3018
3019 out_free:
3020         kfree(data);
3021 out_put:
3022         pnfs_put_layout_hdr(hdr);
3023 out_clear_layoutstats:
3024         smp_mb__before_atomic();
3025         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3026         smp_mb__after_atomic();
3027         goto out;
3028 }
3029 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3030 #endif
3031
3032 unsigned int layoutstats_timer;
3033 module_param(layoutstats_timer, uint, 0644);
3034 EXPORT_SYMBOL_GPL(layoutstats_timer);