fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36
  37 #define NFSDBG_FACILITY         NFSDBG_PNFS
  38 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  39
  40 /* Locking:
  41  *
  42  * pnfs_spinlock:
  43  *      protects pnfs_modules_tbl.
  44  */
  45 static DEFINE_SPINLOCK(pnfs_spinlock);
  46
  47 /*
  48  * pnfs_modules_tbl holds all pnfs modules
  49  */
  50 static LIST_HEAD(pnfs_modules_tbl);
  51
  52 /* Return the registered pnfs layout driver module matching given id */
  53 static struct pnfs_layoutdriver_type *
  54 find_pnfs_driver_locked(u32 id)
  55 {
  56         struct pnfs_layoutdriver_type *local;
  57
  58         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  59                 if (local->id == id)
  60                         goto out;
  61         local = NULL;
  62 out:
  63         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  64         return local;
  65 }
  66
  67 static struct pnfs_layoutdriver_type *
  68 find_pnfs_driver(u32 id)
  69 {
  70         struct pnfs_layoutdriver_type *local;
  71
  72         spin_lock(&pnfs_spinlock);
  73         local = find_pnfs_driver_locked(id);
  74         if (local != NULL && !try_module_get(local->owner)) {
  75                 dprintk("%s: Could not grab reference on module\n", __func__);
  76                 local = NULL;
  77         }
  78         spin_unlock(&pnfs_spinlock);
  79         return local;
  80 }
  81
  82 void
  83 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  84 {
  85         if (nfss->pnfs_curr_ld) {
  86                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  87                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  88                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  89                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  90                         nfs4_deviceid_purge_client(nfss->nfs_client);
  91                 module_put(nfss->pnfs_curr_ld->owner);
  92         }
  93         nfss->pnfs_curr_ld = NULL;
  94 }
  95
  96 /*
  97  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  98  * Currently only one pNFS layout driver per filesystem is supported.
  99  *
 100  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 101  */
 102 void
 103 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 104                       u32 id)
 105 {
 106         struct pnfs_layoutdriver_type *ld_type = NULL;
 107
 108         if (id == 0)
 109                 goto out_no_driver;
 110         if (!(server->nfs_client->cl_exchange_flags &
 111                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 112                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 113                         __func__, id, server->nfs_client->cl_exchange_flags);
 114                 goto out_no_driver;
 115         }
 116         ld_type = find_pnfs_driver(id);
 117         if (!ld_type) {
 118                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 119                 ld_type = find_pnfs_driver(id);
 120                 if (!ld_type) {
 121                         dprintk("%s: No pNFS module found for %u.\n",
 122                                 __func__, id);
 123                         goto out_no_driver;
 124                 }
 125         }
 126         server->pnfs_curr_ld = ld_type;
 127         if (ld_type->set_layoutdriver
 128             && ld_type->set_layoutdriver(server, mntfh)) {
 129                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 130                         "driver %u.\n", __func__, id);
 131                 module_put(ld_type->owner);
 132                 goto out_no_driver;
 133         }
 134         /* Bump the MDS count */
 135         atomic_inc(&server->nfs_client->cl_mds_count);
 136
 137         dprintk("%s: pNFS module for %u set\n", __func__, id);
 138         return;
 139
 140 out_no_driver:
 141         dprintk("%s: Using NFSv4 I/O\n", __func__);
 142         server->pnfs_curr_ld = NULL;
 143 }
 144
 145 int
 146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 147 {
 148         int status = -EINVAL;
 149         struct pnfs_layoutdriver_type *tmp;
 150
 151         if (ld_type->id == 0) {
 152                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 153                 return status;
 154         }
 155         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 156                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 157                        "alloc_lseg and free_lseg.\n", __func__);
 158                 return status;
 159         }
 160
 161         spin_lock(&pnfs_spinlock);
 162         tmp = find_pnfs_driver_locked(ld_type->id);
 163         if (!tmp) {
 164                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 165                 status = 0;
 166                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 167                         ld_type->name);
 168         } else {
 169                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 170                         __func__, ld_type->id);
 171         }
 172         spin_unlock(&pnfs_spinlock);
 173
 174         return status;
 175 }
 176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 177
 178 void
 179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 180 {
 181         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 182         spin_lock(&pnfs_spinlock);
 183         list_del(&ld_type->pnfs_tblid);
 184         spin_unlock(&pnfs_spinlock);
 185 }
 186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 187
 188 /*
 189  * pNFS client layout cache
 190  */
 191
 192 /* Need to hold i_lock if caller does not already hold reference */
 193 void
 194 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 195 {
 196         atomic_inc(&lo->plh_refcount);
 197 }
 198
 199 static struct pnfs_layout_hdr *
 200 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 201 {
 202         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 203         return ld->alloc_layout_hdr(ino, gfp_flags);
 204 }
 205
 206 static void
 207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 208 {
 209         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 210         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 211
 212         if (!list_empty(&lo->plh_layouts)) {
 213                 struct nfs_client *clp = server->nfs_client;
 214
 215                 spin_lock(&clp->cl_lock);
 216                 list_del_init(&lo->plh_layouts);
 217                 spin_unlock(&clp->cl_lock);
 218         }
 219         put_rpccred(lo->plh_lc_cred);
 220         return ld->free_layout_hdr(lo);
 221 }
 222
 223 static void
 224 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 225 {
 226         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 227         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 228         nfsi->layout = NULL;
 229         /* Reset MDS Threshold I/O counters */
 230         nfsi->write_io = 0;
 231         nfsi->read_io = 0;
 232 }
 233
 234 void
 235 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 236 {
 237         struct inode *inode = lo->plh_inode;
 238
 239         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 240                 pnfs_detach_layout_hdr(lo);
 241                 spin_unlock(&inode->i_lock);
 242                 pnfs_free_layout_hdr(lo);
 243         }
 244 }
 245
 246 static int
 247 pnfs_iomode_to_fail_bit(u32 iomode)
 248 {
 249         return iomode == IOMODE_RW ?
 250                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 251 }
 252
 253 static void
 254 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 255 {
 256         lo->plh_retry_timestamp = jiffies;
 257         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 258                 atomic_inc(&lo->plh_refcount);
 259 }
 260
 261 static void
 262 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 263 {
 264         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 265                 atomic_dec(&lo->plh_refcount);
 266 }
 267
 268 static void
 269 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 270 {
 271         struct inode *inode = lo->plh_inode;
 272         struct pnfs_layout_range range = {
 273                 .iomode = iomode,
 274                 .offset = 0,
 275                 .length = NFS4_MAX_UINT64,
 276         };
 277         LIST_HEAD(head);
 278
 279         spin_lock(&inode->i_lock);
 280         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 281         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 282         spin_unlock(&inode->i_lock);
 283         pnfs_free_lseg_list(&head);
 284         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 285                         iomode == IOMODE_RW ?  "RW" : "READ");
 286 }
 287
 288 static bool
 289 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 290 {
 291         unsigned long start, end;
 292         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 293
 294         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 295                 return false;
 296         end = jiffies;
 297         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 298         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 299                 /* It is time to retry the failed layoutgets */
 300                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 301                 return false;
 302         }
 303         return true;
 304 }
 305
 306 static void
 307 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 308 {
 309         INIT_LIST_HEAD(&lseg->pls_list);
 310         INIT_LIST_HEAD(&lseg->pls_lc_list);
 311         atomic_set(&lseg->pls_refcount, 1);
 312         smp_mb();
 313         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 314         lseg->pls_layout = lo;
 315 }
 316
 317 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 318 {
 319         struct inode *ino = lseg->pls_layout->plh_inode;
 320
 321         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 322 }
 323
 324 static void
 325 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 326                 struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *inode = lo->plh_inode;
 329
 330         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 331         list_del_init(&lseg->pls_list);
 332         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 333         atomic_dec(&lo->plh_refcount);
 334         if (list_empty(&lo->plh_segs))
 335                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 336         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 337 }
 338
 339 void
 340 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 341 {
 342         struct pnfs_layout_hdr *lo;
 343         struct inode *inode;
 344
 345         if (!lseg)
 346                 return;
 347
 348         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 349                 atomic_read(&lseg->pls_refcount),
 350                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 351         lo = lseg->pls_layout;
 352         inode = lo->plh_inode;
 353         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 354                 pnfs_get_layout_hdr(lo);
 355                 pnfs_layout_remove_lseg(lo, lseg);
 356                 spin_unlock(&inode->i_lock);
 357                 pnfs_free_lseg(lseg);
 358                 pnfs_put_layout_hdr(lo);
 359         }
 360 }
 361 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 362
 363 static inline u64
 364 end_offset(u64 start, u64 len)
 365 {
 366         u64 end;
 367
 368         end = start + len;
 369         return end >= start ? end : NFS4_MAX_UINT64;
 370 }
 371
 372 /*
 373  * is l2 fully contained in l1?
 374  *   start1                             end1
 375  *   [----------------------------------)
 376  *           start2           end2
 377  *           [----------------)
 378  */
 379 static inline int
 380 lo_seg_contained(struct pnfs_layout_range *l1,
 381                  struct pnfs_layout_range *l2)
 382 {
 383         u64 start1 = l1->offset;
 384         u64 end1 = end_offset(start1, l1->length);
 385         u64 start2 = l2->offset;
 386         u64 end2 = end_offset(start2, l2->length);
 387
 388         return (start1 <= start2) && (end1 >= end2);
 389 }
 390
 391 /*
 392  * is l1 and l2 intersecting?
 393  *   start1                             end1
 394  *   [----------------------------------)
 395  *                              start2           end2
 396  *                              [----------------)
 397  */
 398 static inline int
 399 lo_seg_intersecting(struct pnfs_layout_range *l1,
 400                     struct pnfs_layout_range *l2)
 401 {
 402         u64 start1 = l1->offset;
 403         u64 end1 = end_offset(start1, l1->length);
 404         u64 start2 = l2->offset;
 405         u64 end2 = end_offset(start2, l2->length);
 406
 407         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 408                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 409 }
 410
 411 static bool
 412 should_free_lseg(struct pnfs_layout_range *lseg_range,
 413                  struct pnfs_layout_range *recall_range)
 414 {
 415         return (recall_range->iomode == IOMODE_ANY ||
 416                 lseg_range->iomode == recall_range->iomode) &&
 417                lo_seg_intersecting(lseg_range, recall_range);
 418 }
 419
 420 /* Returns 1 if lseg is removed from list, 0 otherwise */
 421 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 422                              struct list_head *tmp_list)
 423 {
 424         int rv = 0;
 425
 426         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 427                 /* Remove the reference keeping the lseg in the
 428                  * list.  It will now be removed when all
 429                  * outstanding io is finished.
 430                  */
 431                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 432                         atomic_read(&lseg->pls_refcount));
 433                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 434                         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 435                         list_add(&lseg->pls_list, tmp_list);
 436                         rv = 1;
 437                 }
 438         }
 439         return rv;
 440 }
 441
 442 /* Returns count of number of matching invalid lsegs remaining in list
 443  * after call.
 444  */
 445 int
 446 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 447                             struct list_head *tmp_list,
 448                             struct pnfs_layout_range *recall_range)
 449 {
 450         struct pnfs_layout_segment *lseg, *next;
 451         int invalid = 0, removed = 0;
 452
 453         dprintk("%s:Begin lo %p\n", __func__, lo);
 454
 455         if (list_empty(&lo->plh_segs))
 456                 return 0;
 457         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 458                 if (!recall_range ||
 459                     should_free_lseg(&lseg->pls_range, recall_range)) {
 460                         dprintk("%s: freeing lseg %p iomode %d "
 461                                 "offset %llu length %llu\n", __func__,
 462                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 463                                 lseg->pls_range.length);
 464                         invalid++;
 465                         removed += mark_lseg_invalid(lseg, tmp_list);
 466                 }
 467         dprintk("%s:Return %i\n", __func__, invalid - removed);
 468         return invalid - removed;
 469 }
 470
 471 /* note free_me must contain lsegs from a single layout_hdr */
 472 void
 473 pnfs_free_lseg_list(struct list_head *free_me)
 474 {
 475         struct pnfs_layout_segment *lseg, *tmp;
 476
 477         if (list_empty(free_me))
 478                 return;
 479
 480         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 481                 list_del(&lseg->pls_list);
 482                 pnfs_free_lseg(lseg);
 483         }
 484 }
 485
 486 void
 487 pnfs_destroy_layout(struct nfs_inode *nfsi)
 488 {
 489         struct pnfs_layout_hdr *lo;
 490         LIST_HEAD(tmp_list);
 491
 492         spin_lock(&nfsi->vfs_inode.i_lock);
 493         lo = nfsi->layout;
 494         if (lo) {
 495                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 496                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 497                 pnfs_get_layout_hdr(lo);
 498                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 499                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 500                 spin_unlock(&nfsi->vfs_inode.i_lock);
 501                 pnfs_free_lseg_list(&tmp_list);
 502                 pnfs_put_layout_hdr(lo);
 503         } else
 504                 spin_unlock(&nfsi->vfs_inode.i_lock);
 505 }
 506 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 507
 508 static bool
 509 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 510                 struct list_head *layout_list)
 511 {
 512         struct pnfs_layout_hdr *lo;
 513         bool ret = false;
 514
 515         spin_lock(&inode->i_lock);
 516         lo = NFS_I(inode)->layout;
 517         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 518                 pnfs_get_layout_hdr(lo);
 519                 list_add(&lo->plh_bulk_destroy, layout_list);
 520                 ret = true;
 521         }
 522         spin_unlock(&inode->i_lock);
 523         return ret;
 524 }
 525
 526 /* Caller must hold rcu_read_lock and clp->cl_lock */
 527 static int
 528 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 529                 struct nfs_server *server,
 530                 struct list_head *layout_list)
 531 {
 532         struct pnfs_layout_hdr *lo, *next;
 533         struct inode *inode;
 534
 535         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 536                 inode = igrab(lo->plh_inode);
 537                 if (inode == NULL)
 538                         continue;
 539                 list_del_init(&lo->plh_layouts);
 540                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 541                         continue;
 542                 rcu_read_unlock();
 543                 spin_unlock(&clp->cl_lock);
 544                 iput(inode);
 545                 spin_lock(&clp->cl_lock);
 546                 rcu_read_lock();
 547                 return -EAGAIN;
 548         }
 549         return 0;
 550 }
 551
 552 static int
 553 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 554                 bool is_bulk_recall)
 555 {
 556         struct pnfs_layout_hdr *lo;
 557         struct inode *inode;
 558         struct pnfs_layout_range range = {
 559                 .iomode = IOMODE_ANY,
 560                 .offset = 0,
 561                 .length = NFS4_MAX_UINT64,
 562         };
 563         LIST_HEAD(lseg_list);
 564         int ret = 0;
 565
 566         while (!list_empty(layout_list)) {
 567                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 568                                 plh_bulk_destroy);
 569                 dprintk("%s freeing layout for inode %lu\n", __func__,
 570                         lo->plh_inode->i_ino);
 571                 inode = lo->plh_inode;
 572                 spin_lock(&inode->i_lock);
 573                 list_del_init(&lo->plh_bulk_destroy);
 574                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 575                 if (is_bulk_recall)
 576                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 577                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 578                         ret = -EAGAIN;
 579                 spin_unlock(&inode->i_lock);
 580                 pnfs_free_lseg_list(&lseg_list);
 581                 pnfs_put_layout_hdr(lo);
 582                 iput(inode);
 583         }
 584         return ret;
 585 }
 586
 587 int
 588 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 589                 struct nfs_fsid *fsid,
 590                 bool is_recall)
 591 {
 592         struct nfs_server *server;
 593         LIST_HEAD(layout_list);
 594
 595         spin_lock(&clp->cl_lock);
 596         rcu_read_lock();
 597 restart:
 598         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 599                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 600                         continue;
 601                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 602                                 server,
 603                                 &layout_list) != 0)
 604                         goto restart;
 605         }
 606         rcu_read_unlock();
 607         spin_unlock(&clp->cl_lock);
 608
 609         if (list_empty(&layout_list))
 610                 return 0;
 611         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 612 }
 613
 614 int
 615 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 616                 bool is_recall)
 617 {
 618         struct nfs_server *server;
 619         LIST_HEAD(layout_list);
 620
 621         spin_lock(&clp->cl_lock);
 622         rcu_read_lock();
 623 restart:
 624         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 625                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 626                                         server,
 627                                         &layout_list) != 0)
 628                         goto restart;
 629         }
 630         rcu_read_unlock();
 631         spin_unlock(&clp->cl_lock);
 632
 633         if (list_empty(&layout_list))
 634                 return 0;
 635         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 636 }
 637
 638 /*
 639  * Called by the state manger to remove all layouts established under an
 640  * expired lease.
 641  */
 642 void
 643 pnfs_destroy_all_layouts(struct nfs_client *clp)
 644 {
 645         nfs4_deviceid_mark_client_invalid(clp);
 646         nfs4_deviceid_purge_client(clp);
 647
 648         pnfs_destroy_layouts_byclid(clp, false);
 649 }
 650
 651 /*
 652  * Compare 2 layout stateid sequence ids, to see which is newer,
 653  * taking into account wraparound issues.
 654  */
 655 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 656 {
 657         return (s32)s1 - (s32)s2 > 0;
 658 }
 659
 660 /* update lo->plh_stateid with new if is more recent */
 661 void
 662 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 663                         bool update_barrier)
 664 {
 665         u32 oldseq, newseq, new_barrier;
 666         int empty = list_empty(&lo->plh_segs);
 667
 668         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 669         newseq = be32_to_cpu(new->seqid);
 670         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 671                 nfs4_stateid_copy(&lo->plh_stateid, new);
 672                 if (update_barrier) {
 673                         new_barrier = be32_to_cpu(new->seqid);
 674                 } else {
 675                         /* Because of wraparound, we want to keep the barrier
 676                          * "close" to the current seqids.
 677                          */
 678                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 679                 }
 680                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 681                         lo->plh_barrier = new_barrier;
 682         }
 683 }
 684
 685 static bool
 686 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 687                 const nfs4_stateid *stateid)
 688 {
 689         u32 seqid = be32_to_cpu(stateid->seqid);
 690
 691         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 692 }
 693
 694 /* lget is set to 1 if called from inside send_layoutget call chain */
 695 static bool
 696 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 697 {
 698         return lo->plh_block_lgets ||
 699                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 700                 (list_empty(&lo->plh_segs) &&
 701                  (atomic_read(&lo->plh_outstanding) > lget));
 702 }
 703
 704 int
 705 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 706                               struct nfs4_state *open_state)
 707 {
 708         int status = 0;
 709
 710         dprintk("--> %s\n", __func__);
 711         spin_lock(&lo->plh_inode->i_lock);
 712         if (pnfs_layoutgets_blocked(lo, 1)) {
 713                 status = -EAGAIN;
 714         } else if (list_empty(&lo->plh_segs)) {
 715                 int seq;
 716
 717                 do {
 718                         seq = read_seqbegin(&open_state->seqlock);
 719                         nfs4_stateid_copy(dst, &open_state->stateid);
 720                 } while (read_seqretry(&open_state->seqlock, seq));
 721         } else
 722                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 723         spin_unlock(&lo->plh_inode->i_lock);
 724         dprintk("<-- %s\n", __func__);
 725         return status;
 726 }
 727
 728 /*
 729 * Get layout from server.
 730 *    for now, assume that whole file layouts are requested.
 731 *    arg->offset: 0
 732 *    arg->length: all ones
 733 */
 734 static struct pnfs_layout_segment *
 735 send_layoutget(struct pnfs_layout_hdr *lo,
 736            struct nfs_open_context *ctx,
 737            struct pnfs_layout_range *range,
 738            gfp_t gfp_flags)
 739 {
 740         struct inode *ino = lo->plh_inode;
 741         struct nfs_server *server = NFS_SERVER(ino);
 742         struct nfs4_layoutget *lgp;
 743         struct pnfs_layout_segment *lseg;
 744
 745         dprintk("--> %s\n", __func__);
 746
 747         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 748         if (lgp == NULL)
 749                 return NULL;
 750
 751         lgp->args.minlength = PAGE_CACHE_SIZE;
 752         if (lgp->args.minlength > range->length)
 753                 lgp->args.minlength = range->length;
 754         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 755         lgp->args.range = *range;
 756         lgp->args.type = server->pnfs_curr_ld->id;
 757         lgp->args.inode = ino;
 758         lgp->args.ctx = get_nfs_open_context(ctx);
 759         lgp->gfp_flags = gfp_flags;
 760
 761         /* Synchronously retrieve layout information from server and
 762          * store in lseg.
 763          */
 764         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 765         if (IS_ERR(lseg)) {
 766                 switch (PTR_ERR(lseg)) {
 767                 case -ENOMEM:
 768                 case -ERESTARTSYS:
 769                         break;
 770                 default:
 771                         /* remember that LAYOUTGET failed and suspend trying */
 772                         pnfs_layout_io_set_failed(lo, range->iomode);
 773                 }
 774                 return NULL;
 775         }
 776
 777         return lseg;
 778 }
 779
 780 /*
 781  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 782  * when the layout segment list is empty.
 783  *
 784  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 785  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 786  * deviceid is marked invalid.
 787  */
 788 int
 789 _pnfs_return_layout(struct inode *ino)
 790 {
 791         struct pnfs_layout_hdr *lo = NULL;
 792         struct nfs_inode *nfsi = NFS_I(ino);
 793         LIST_HEAD(tmp_list);
 794         struct nfs4_layoutreturn *lrp;
 795         nfs4_stateid stateid;
 796         int status = 0, empty;
 797
 798         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 799
 800         spin_lock(&ino->i_lock);
 801         lo = nfsi->layout;
 802         if (!lo) {
 803                 spin_unlock(&ino->i_lock);
 804                 dprintk("NFS: %s no layout to return\n", __func__);
 805                 goto out;
 806         }
 807         stateid = nfsi->layout->plh_stateid;
 808         /* Reference matched in nfs4_layoutreturn_release */
 809         pnfs_get_layout_hdr(lo);
 810         empty = list_empty(&lo->plh_segs);
 811         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 812         /* Don't send a LAYOUTRETURN if list was initially empty */
 813         if (empty) {
 814                 spin_unlock(&ino->i_lock);
 815                 pnfs_put_layout_hdr(lo);
 816                 dprintk("NFS: %s no layout segments to return\n", __func__);
 817                 goto out;
 818         }
 819         lo->plh_block_lgets++;
 820         spin_unlock(&ino->i_lock);
 821         pnfs_free_lseg_list(&tmp_list);
 822
 823         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 824
 825         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 826         if (unlikely(lrp == NULL)) {
 827                 status = -ENOMEM;
 828                 spin_lock(&ino->i_lock);
 829                 lo->plh_block_lgets--;
 830                 spin_unlock(&ino->i_lock);
 831                 pnfs_put_layout_hdr(lo);
 832                 goto out;
 833         }
 834
 835         lrp->args.stateid = stateid;
 836         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 837         lrp->args.inode = ino;
 838         lrp->args.layout = lo;
 839         lrp->clp = NFS_SERVER(ino)->nfs_client;
 840
 841         status = nfs4_proc_layoutreturn(lrp);
 842 out:
 843         dprintk("<-- %s status: %d\n", __func__, status);
 844         return status;
 845 }
 846 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 847
 848 bool pnfs_roc(struct inode *ino)
 849 {
 850         struct pnfs_layout_hdr *lo;
 851         struct pnfs_layout_segment *lseg, *tmp;
 852         LIST_HEAD(tmp_list);
 853         bool found = false;
 854
 855         spin_lock(&ino->i_lock);
 856         lo = NFS_I(ino)->layout;
 857         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 858             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 859                 goto out_nolayout;
 860         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 861                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 862                         mark_lseg_invalid(lseg, &tmp_list);
 863                         found = true;
 864                 }
 865         if (!found)
 866                 goto out_nolayout;
 867         lo->plh_block_lgets++;
 868         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 869         spin_unlock(&ino->i_lock);
 870         pnfs_free_lseg_list(&tmp_list);
 871         return true;
 872
 873 out_nolayout:
 874         spin_unlock(&ino->i_lock);
 875         return false;
 876 }
 877
 878 void pnfs_roc_release(struct inode *ino)
 879 {
 880         struct pnfs_layout_hdr *lo;
 881
 882         spin_lock(&ino->i_lock);
 883         lo = NFS_I(ino)->layout;
 884         lo->plh_block_lgets--;
 885         if (atomic_dec_and_test(&lo->plh_refcount)) {
 886                 pnfs_detach_layout_hdr(lo);
 887                 spin_unlock(&ino->i_lock);
 888                 pnfs_free_layout_hdr(lo);
 889         } else
 890                 spin_unlock(&ino->i_lock);
 891 }
 892
 893 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 894 {
 895         struct pnfs_layout_hdr *lo;
 896
 897         spin_lock(&ino->i_lock);
 898         lo = NFS_I(ino)->layout;
 899         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 900                 lo->plh_barrier = barrier;
 901         spin_unlock(&ino->i_lock);
 902 }
 903
 904 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 905 {
 906         struct nfs_inode *nfsi = NFS_I(ino);
 907         struct pnfs_layout_hdr *lo;
 908         struct pnfs_layout_segment *lseg;
 909         u32 current_seqid;
 910         bool found = false;
 911
 912         spin_lock(&ino->i_lock);
 913         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 914                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 915                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 916                         found = true;
 917                         goto out;
 918                 }
 919         lo = nfsi->layout;
 920         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 921
 922         /* Since close does not return a layout stateid for use as
 923          * a barrier, we choose the worst-case barrier.
 924          */
 925         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 926 out:
 927         spin_unlock(&ino->i_lock);
 928         return found;
 929 }
 930
 931 /*
 932  * Compare two layout segments for sorting into layout cache.
 933  * We want to preferentially return RW over RO layouts, so ensure those
 934  * are seen first.
 935  */
 936 static s64
 937 cmp_layout(struct pnfs_layout_range *l1,
 938            struct pnfs_layout_range *l2)
 939 {
 940         s64 d;
 941
 942         /* high offset > low offset */
 943         d = l1->offset - l2->offset;
 944         if (d)
 945                 return d;
 946
 947         /* short length > long length */
 948         d = l2->length - l1->length;
 949         if (d)
 950                 return d;
 951
 952         /* read > read/write */
 953         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 954 }
 955
 956 static void
 957 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
 958                    struct pnfs_layout_segment *lseg)
 959 {
 960         struct pnfs_layout_segment *lp;
 961
 962         dprintk("%s:Begin\n", __func__);
 963
 964         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 965                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 966                         continue;
 967                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 968                 dprintk("%s: inserted lseg %p "
 969                         "iomode %d offset %llu length %llu before "
 970                         "lp %p iomode %d offset %llu length %llu\n",
 971                         __func__, lseg, lseg->pls_range.iomode,
 972                         lseg->pls_range.offset, lseg->pls_range.length,
 973                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 974                         lp->pls_range.length);
 975                 goto out;
 976         }
 977         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 978         dprintk("%s: inserted lseg %p "
 979                 "iomode %d offset %llu length %llu at tail\n",
 980                 __func__, lseg, lseg->pls_range.iomode,
 981                 lseg->pls_range.offset, lseg->pls_range.length);
 982 out:
 983         pnfs_get_layout_hdr(lo);
 984
 985         dprintk("%s:Return\n", __func__);
 986 }
 987
 988 static struct pnfs_layout_hdr *
 989 alloc_init_layout_hdr(struct inode *ino,
 990                       struct nfs_open_context *ctx,
 991                       gfp_t gfp_flags)
 992 {
 993         struct pnfs_layout_hdr *lo;
 994
 995         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 996         if (!lo)
 997                 return NULL;
 998         atomic_set(&lo->plh_refcount, 1);
 999         INIT_LIST_HEAD(&lo->plh_layouts);
1000         INIT_LIST_HEAD(&lo->plh_segs);
1001         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1002         lo->plh_inode = ino;
1003         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
1004         return lo;
1005 }
1006
1007 static struct pnfs_layout_hdr *
1008 pnfs_find_alloc_layout(struct inode *ino,
1009                        struct nfs_open_context *ctx,
1010                        gfp_t gfp_flags)
1011 {
1012         struct nfs_inode *nfsi = NFS_I(ino);
1013         struct pnfs_layout_hdr *new = NULL;
1014
1015         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1016
1017         if (nfsi->layout != NULL)
1018                 goto out_existing;
1019         spin_unlock(&ino->i_lock);
1020         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1021         spin_lock(&ino->i_lock);
1022
1023         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1024                 nfsi->layout = new;
1025                 return new;
1026         } else if (new != NULL)
1027                 pnfs_free_layout_hdr(new);
1028 out_existing:
1029         pnfs_get_layout_hdr(nfsi->layout);
1030         return nfsi->layout;
1031 }
1032
1033 /*
1034  * iomode matching rules:
1035  * iomode       lseg    match
1036  * -----        -----   -----
1037  * ANY          READ    true
1038  * ANY          RW      true
1039  * RW           READ    false
1040  * RW           RW      true
1041  * READ         READ    true
1042  * READ         RW      true
1043  */
1044 static int
1045 is_matching_lseg(struct pnfs_layout_range *ls_range,
1046                  struct pnfs_layout_range *range)
1047 {
1048         struct pnfs_layout_range range1;
1049
1050         if ((range->iomode == IOMODE_RW &&
1051              ls_range->iomode != IOMODE_RW) ||
1052             !lo_seg_intersecting(ls_range, range))
1053                 return 0;
1054
1055         /* range1 covers only the first byte in the range */
1056         range1 = *range;
1057         range1.length = 1;
1058         return lo_seg_contained(ls_range, &range1);
1059 }
1060
1061 /*
1062  * lookup range in layout
1063  */
1064 static struct pnfs_layout_segment *
1065 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1066                 struct pnfs_layout_range *range)
1067 {
1068         struct pnfs_layout_segment *lseg, *ret = NULL;
1069
1070         dprintk("%s:Begin\n", __func__);
1071
1072         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1073                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1074                     is_matching_lseg(&lseg->pls_range, range)) {
1075                         ret = pnfs_get_lseg(lseg);
1076                         break;
1077                 }
1078                 if (lseg->pls_range.offset > range->offset)
1079                         break;
1080         }
1081
1082         dprintk("%s:Return lseg %p ref %d\n",
1083                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1084         return ret;
1085 }
1086
1087 /*
1088  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1089  * to the MDS or over pNFS
1090  *
1091  * The nfs_inode read_io and write_io fields are cumulative counters reset
1092  * when there are no layout segments. Note that in pnfs_update_layout iomode
1093  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1094  * WRITE request.
1095  *
1096  * A return of true means use MDS I/O.
1097  *
1098  * From rfc 5661:
1099  * If a file's size is smaller than the file size threshold, data accesses
1100  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1101  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1102  * server.  If both file size and I/O size are provided, the client SHOULD
1103  * reach or exceed  both thresholds before sending its read or write
1104  * requests to the data server.
1105  */
1106 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1107                                      struct inode *ino, int iomode)
1108 {
1109         struct nfs4_threshold *t = ctx->mdsthreshold;
1110         struct nfs_inode *nfsi = NFS_I(ino);
1111         loff_t fsize = i_size_read(ino);
1112         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1113
1114         if (t == NULL)
1115                 return ret;
1116
1117         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1118                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1119
1120         switch (iomode) {
1121         case IOMODE_READ:
1122                 if (t->bm & THRESHOLD_RD) {
1123                         dprintk("%s fsize %llu\n", __func__, fsize);
1124                         size_set = true;
1125                         if (fsize < t->rd_sz)
1126                                 size = true;
1127                 }
1128                 if (t->bm & THRESHOLD_RD_IO) {
1129                         dprintk("%s nfsi->read_io %llu\n", __func__,
1130                                 nfsi->read_io);
1131                         io_set = true;
1132                         if (nfsi->read_io < t->rd_io_sz)
1133                                 io = true;
1134                 }
1135                 break;
1136         case IOMODE_RW:
1137                 if (t->bm & THRESHOLD_WR) {
1138                         dprintk("%s fsize %llu\n", __func__, fsize);
1139                         size_set = true;
1140                         if (fsize < t->wr_sz)
1141                                 size = true;
1142                 }
1143                 if (t->bm & THRESHOLD_WR_IO) {
1144                         dprintk("%s nfsi->write_io %llu\n", __func__,
1145                                 nfsi->write_io);
1146                         io_set = true;
1147                         if (nfsi->write_io < t->wr_io_sz)
1148                                 io = true;
1149                 }
1150                 break;
1151         }
1152         if (size_set && io_set) {
1153                 if (size && io)
1154                         ret = true;
1155         } else if (size || io)
1156                 ret = true;
1157
1158         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1159         return ret;
1160 }
1161
1162 /*
1163  * Layout segment is retreived from the server if not cached.
1164  * The appropriate layout segment is referenced and returned to the caller.
1165  */
1166 struct pnfs_layout_segment *
1167 pnfs_update_layout(struct inode *ino,
1168                    struct nfs_open_context *ctx,
1169                    loff_t pos,
1170                    u64 count,
1171                    enum pnfs_iomode iomode,
1172                    gfp_t gfp_flags)
1173 {
1174         struct pnfs_layout_range arg = {
1175                 .iomode = iomode,
1176                 .offset = pos,
1177                 .length = count,
1178         };
1179         unsigned pg_offset;
1180         struct nfs_server *server = NFS_SERVER(ino);
1181         struct nfs_client *clp = server->nfs_client;
1182         struct pnfs_layout_hdr *lo;
1183         struct pnfs_layout_segment *lseg = NULL;
1184         bool first = false;
1185
1186         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1187                 goto out;
1188
1189         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1190                 goto out;
1191
1192         spin_lock(&ino->i_lock);
1193         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1194         if (lo == NULL) {
1195                 spin_unlock(&ino->i_lock);
1196                 goto out;
1197         }
1198
1199         /* Do we even need to bother with this? */
1200         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1201                 dprintk("%s matches recall, use MDS\n", __func__);
1202                 goto out_unlock;
1203         }
1204
1205         /* if LAYOUTGET already failed once we don't try again */
1206         if (pnfs_layout_io_test_failed(lo, iomode))
1207                 goto out_unlock;
1208
1209         /* Check to see if the layout for the given range already exists */
1210         lseg = pnfs_find_lseg(lo, &arg);
1211         if (lseg)
1212                 goto out_unlock;
1213
1214         if (pnfs_layoutgets_blocked(lo, 0))
1215                 goto out_unlock;
1216         atomic_inc(&lo->plh_outstanding);
1217
1218         if (list_empty(&lo->plh_segs))
1219                 first = true;
1220
1221         spin_unlock(&ino->i_lock);
1222         if (first) {
1223                 /* The lo must be on the clp list if there is any
1224                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1225                  */
1226                 spin_lock(&clp->cl_lock);
1227                 list_add_tail(&lo->plh_layouts, &server->layouts);
1228                 spin_unlock(&clp->cl_lock);
1229         }
1230
1231         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1232         if (pg_offset) {
1233                 arg.offset -= pg_offset;
1234                 arg.length += pg_offset;
1235         }
1236         if (arg.length != NFS4_MAX_UINT64)
1237                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1238
1239         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1240         atomic_dec(&lo->plh_outstanding);
1241 out_put_layout_hdr:
1242         pnfs_put_layout_hdr(lo);
1243 out:
1244         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1245                         "(%s, offset: %llu, length: %llu)\n",
1246                         __func__, ino->i_sb->s_id,
1247                         (unsigned long long)NFS_FILEID(ino),
1248                         lseg == NULL ? "not found" : "found",
1249                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1250                         (unsigned long long)pos,
1251                         (unsigned long long)count);
1252         return lseg;
1253 out_unlock:
1254         spin_unlock(&ino->i_lock);
1255         goto out_put_layout_hdr;
1256 }
1257 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1258
1259 struct pnfs_layout_segment *
1260 pnfs_layout_process(struct nfs4_layoutget *lgp)
1261 {
1262         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1263         struct nfs4_layoutget_res *res = &lgp->res;
1264         struct pnfs_layout_segment *lseg;
1265         struct inode *ino = lo->plh_inode;
1266         int status = 0;
1267
1268         /* Inject layout blob into I/O device driver */
1269         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1270         if (!lseg || IS_ERR(lseg)) {
1271                 if (!lseg)
1272                         status = -ENOMEM;
1273                 else
1274                         status = PTR_ERR(lseg);
1275                 dprintk("%s: Could not allocate layout: error %d\n",
1276                        __func__, status);
1277                 goto out;
1278         }
1279
1280         spin_lock(&ino->i_lock);
1281         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1282                 dprintk("%s forget reply due to recall\n", __func__);
1283                 goto out_forget_reply;
1284         }
1285
1286         if (pnfs_layoutgets_blocked(lo, 1) ||
1287             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1288                 dprintk("%s forget reply due to state\n", __func__);
1289                 goto out_forget_reply;
1290         }
1291
1292         /* Done processing layoutget. Set the layout stateid */
1293         pnfs_set_layout_stateid(lo, &res->stateid, false);
1294
1295         init_lseg(lo, lseg);
1296         lseg->pls_range = res->range;
1297         pnfs_get_lseg(lseg);
1298         pnfs_layout_insert_lseg(lo, lseg);
1299
1300         if (res->return_on_close) {
1301                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1302                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1303         }
1304
1305         spin_unlock(&ino->i_lock);
1306         return lseg;
1307 out:
1308         return ERR_PTR(status);
1309
1310 out_forget_reply:
1311         spin_unlock(&ino->i_lock);
1312         lseg->pls_layout = lo;
1313         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1314         goto out;
1315 }
1316
1317 void
1318 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1319 {
1320         u64 rd_size = req->wb_bytes;
1321
1322         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1323
1324         if (req->wb_offset != req->wb_pgbase) {
1325                 nfs_pageio_reset_read_mds(pgio);
1326                 return;
1327         }
1328
1329         if (pgio->pg_dreq == NULL)
1330                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1331         else
1332                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1333
1334         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1335                                            req->wb_context,
1336                                            req_offset(req),
1337                                            rd_size,
1338                                            IOMODE_READ,
1339                                            GFP_KERNEL);
1340         /* If no lseg, fall back to read through mds */
1341         if (pgio->pg_lseg == NULL)
1342                 nfs_pageio_reset_read_mds(pgio);
1343
1344 }
1345 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1346
1347 void
1348 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1349                            struct nfs_page *req, u64 wb_size)
1350 {
1351         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1352
1353         if (req->wb_offset != req->wb_pgbase) {
1354                 nfs_pageio_reset_write_mds(pgio);
1355                 return;
1356         }
1357
1358         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1359                                            req->wb_context,
1360                                            req_offset(req),
1361                                            wb_size,
1362                                            IOMODE_RW,
1363                                            GFP_NOFS);
1364         /* If no lseg, fall back to write through mds */
1365         if (pgio->pg_lseg == NULL)
1366                 nfs_pageio_reset_write_mds(pgio);
1367 }
1368 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1369
1370 void
1371 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1372                       const struct nfs_pgio_completion_ops *compl_ops)
1373 {
1374         struct nfs_server *server = NFS_SERVER(inode);
1375         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1376
1377         if (ld == NULL)
1378                 nfs_pageio_init_read(pgio, inode, compl_ops);
1379         else
1380                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1381 }
1382
1383 void
1384 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1385                        int ioflags,
1386                        const struct nfs_pgio_completion_ops *compl_ops)
1387 {
1388         struct nfs_server *server = NFS_SERVER(inode);
1389         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1390
1391         if (ld == NULL)
1392                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1393         else
1394                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1395 }
1396
1397 bool
1398 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1399                      struct nfs_page *req)
1400 {
1401         if (pgio->pg_lseg == NULL)
1402                 return nfs_generic_pg_test(pgio, prev, req);
1403
1404         /*
1405          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1406          * Note that this test makes several assumptions:
1407          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1408          *   is known to lie within the range.
1409          *   - that the nfs_page being tested is known to be contiguous with the
1410          *   previous nfs_page.
1411          *   - Layout ranges are page aligned, so we only have to test the
1412          *   start offset of the request.
1413          *
1414          * Please also note that 'end_offset' is actually the offset of the
1415          * first byte that lies outside the pnfs_layout_range. FIXME?
1416          *
1417          */
1418         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1419                                          pgio->pg_lseg->pls_range.length);
1420 }
1421 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1422
1423 int pnfs_write_done_resend_to_mds(struct inode *inode,
1424                                 struct list_head *head,
1425                                 const struct nfs_pgio_completion_ops *compl_ops)
1426 {
1427         struct nfs_pageio_descriptor pgio;
1428         LIST_HEAD(failed);
1429
1430         /* Resend all requests through the MDS */
1431         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1432         while (!list_empty(head)) {
1433                 struct nfs_page *req = nfs_list_entry(head->next);
1434
1435                 nfs_list_remove_request(req);
1436                 if (!nfs_pageio_add_request(&pgio, req))
1437                         nfs_list_add_request(req, &failed);
1438         }
1439         nfs_pageio_complete(&pgio);
1440
1441         if (!list_empty(&failed)) {
1442                 /* For some reason our attempt to resend pages. Mark the
1443                  * overall send request as having failed, and let
1444                  * nfs_writeback_release_full deal with the error.
1445                  */
1446                 list_move(&failed, head);
1447                 return -EIO;
1448         }
1449         return 0;
1450 }
1451 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1452
1453 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1454 {
1455         struct nfs_pgio_header *hdr = data->header;
1456
1457         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1458         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1459             PNFS_LAYOUTRET_ON_ERROR) {
1460                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1461                 pnfs_return_layout(hdr->inode);
1462         }
1463         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1464                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1465                                                         &hdr->pages,
1466                                                         hdr->completion_ops);
1467 }
1468
1469 /*
1470  * Called by non rpc-based layout drivers
1471  */
1472 void pnfs_ld_write_done(struct nfs_write_data *data)
1473 {
1474         struct nfs_pgio_header *hdr = data->header;
1475
1476         if (!hdr->pnfs_error) {
1477                 pnfs_set_layoutcommit(data);
1478                 hdr->mds_ops->rpc_call_done(&data->task, data);
1479         } else
1480                 pnfs_ld_handle_write_error(data);
1481         hdr->mds_ops->rpc_release(data);
1482 }
1483 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1484
1485 static void
1486 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1487                 struct nfs_write_data *data)
1488 {
1489         struct nfs_pgio_header *hdr = data->header;
1490
1491         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1492                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1493                 nfs_pageio_reset_write_mds(desc);
1494                 desc->pg_recoalesce = 1;
1495         }
1496         nfs_writedata_release(data);
1497 }
1498
1499 static enum pnfs_try_status
1500 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1501                         const struct rpc_call_ops *call_ops,
1502                         struct pnfs_layout_segment *lseg,
1503                         int how)
1504 {
1505         struct nfs_pgio_header *hdr = wdata->header;
1506         struct inode *inode = hdr->inode;
1507         enum pnfs_try_status trypnfs;
1508         struct nfs_server *nfss = NFS_SERVER(inode);
1509
1510         hdr->mds_ops = call_ops;
1511
1512         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1513                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1514         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1515         if (trypnfs != PNFS_NOT_ATTEMPTED)
1516                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1517         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1518         return trypnfs;
1519 }
1520
1521 static void
1522 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1523 {
1524         struct nfs_write_data *data;
1525         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1526         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1527
1528         desc->pg_lseg = NULL;
1529         while (!list_empty(head)) {
1530                 enum pnfs_try_status trypnfs;
1531
1532                 data = list_first_entry(head, struct nfs_write_data, list);
1533                 list_del_init(&data->list);
1534
1535                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1536                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1537                         pnfs_write_through_mds(desc, data);
1538         }
1539         pnfs_put_lseg(lseg);
1540 }
1541
1542 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1543 {
1544         pnfs_put_lseg(hdr->lseg);
1545         nfs_writehdr_free(hdr);
1546 }
1547 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1548
1549 int
1550 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1551 {
1552         struct nfs_write_header *whdr;
1553         struct nfs_pgio_header *hdr;
1554         int ret;
1555
1556         whdr = nfs_writehdr_alloc();
1557         if (!whdr) {
1558                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1559                 pnfs_put_lseg(desc->pg_lseg);
1560                 desc->pg_lseg = NULL;
1561                 return -ENOMEM;
1562         }
1563         hdr = &whdr->header;
1564         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1565         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1566         atomic_inc(&hdr->refcnt);
1567         ret = nfs_generic_flush(desc, hdr);
1568         if (ret != 0) {
1569                 pnfs_put_lseg(desc->pg_lseg);
1570                 desc->pg_lseg = NULL;
1571         } else
1572                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1573         if (atomic_dec_and_test(&hdr->refcnt))
1574                 hdr->completion_ops->completion(hdr);
1575         return ret;
1576 }
1577 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1578
1579 int pnfs_read_done_resend_to_mds(struct inode *inode,
1580                                 struct list_head *head,
1581                                 const struct nfs_pgio_completion_ops *compl_ops)
1582 {
1583         struct nfs_pageio_descriptor pgio;
1584         LIST_HEAD(failed);
1585
1586         /* Resend all requests through the MDS */
1587         nfs_pageio_init_read(&pgio, inode, compl_ops);
1588         while (!list_empty(head)) {
1589                 struct nfs_page *req = nfs_list_entry(head->next);
1590
1591                 nfs_list_remove_request(req);
1592                 if (!nfs_pageio_add_request(&pgio, req))
1593                         nfs_list_add_request(req, &failed);
1594         }
1595         nfs_pageio_complete(&pgio);
1596
1597         if (!list_empty(&failed)) {
1598                 list_move(&failed, head);
1599                 return -EIO;
1600         }
1601         return 0;
1602 }
1603 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1604
1605 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1606 {
1607         struct nfs_pgio_header *hdr = data->header;
1608
1609         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1610         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1611             PNFS_LAYOUTRET_ON_ERROR) {
1612                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1613                 pnfs_return_layout(hdr->inode);
1614         }
1615         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1616                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1617                                                         &hdr->pages,
1618                                                         hdr->completion_ops);
1619 }
1620
1621 /*
1622  * Called by non rpc-based layout drivers
1623  */
1624 void pnfs_ld_read_done(struct nfs_read_data *data)
1625 {
1626         struct nfs_pgio_header *hdr = data->header;
1627
1628         if (likely(!hdr->pnfs_error)) {
1629                 __nfs4_read_done_cb(data);
1630                 hdr->mds_ops->rpc_call_done(&data->task, data);
1631         } else
1632                 pnfs_ld_handle_read_error(data);
1633         hdr->mds_ops->rpc_release(data);
1634 }
1635 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1636
1637 static void
1638 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1639                 struct nfs_read_data *data)
1640 {
1641         struct nfs_pgio_header *hdr = data->header;
1642
1643         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1644                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1645                 nfs_pageio_reset_read_mds(desc);
1646                 desc->pg_recoalesce = 1;
1647         }
1648         nfs_readdata_release(data);
1649 }
1650
1651 /*
1652  * Call the appropriate parallel I/O subsystem read function.
1653  */
1654 static enum pnfs_try_status
1655 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1656                        const struct rpc_call_ops *call_ops,
1657                        struct pnfs_layout_segment *lseg)
1658 {
1659         struct nfs_pgio_header *hdr = rdata->header;
1660         struct inode *inode = hdr->inode;
1661         struct nfs_server *nfss = NFS_SERVER(inode);
1662         enum pnfs_try_status trypnfs;
1663
1664         hdr->mds_ops = call_ops;
1665
1666         dprintk("%s: Reading ino:%lu %u@%llu\n",
1667                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1668
1669         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1670         if (trypnfs != PNFS_NOT_ATTEMPTED)
1671                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1672         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1673         return trypnfs;
1674 }
1675
1676 static void
1677 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1678 {
1679         struct nfs_read_data *data;
1680         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1681         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1682
1683         desc->pg_lseg = NULL;
1684         while (!list_empty(head)) {
1685                 enum pnfs_try_status trypnfs;
1686
1687                 data = list_first_entry(head, struct nfs_read_data, list);
1688                 list_del_init(&data->list);
1689
1690                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1691                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1692                         pnfs_read_through_mds(desc, data);
1693         }
1694         pnfs_put_lseg(lseg);
1695 }
1696
1697 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1698 {
1699         pnfs_put_lseg(hdr->lseg);
1700         nfs_readhdr_free(hdr);
1701 }
1702 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1703
1704 int
1705 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1706 {
1707         struct nfs_read_header *rhdr;
1708         struct nfs_pgio_header *hdr;
1709         int ret;
1710
1711         rhdr = nfs_readhdr_alloc();
1712         if (!rhdr) {
1713                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1714                 ret = -ENOMEM;
1715                 pnfs_put_lseg(desc->pg_lseg);
1716                 desc->pg_lseg = NULL;
1717                 return ret;
1718         }
1719         hdr = &rhdr->header;
1720         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1721         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1722         atomic_inc(&hdr->refcnt);
1723         ret = nfs_generic_pagein(desc, hdr);
1724         if (ret != 0) {
1725                 pnfs_put_lseg(desc->pg_lseg);
1726                 desc->pg_lseg = NULL;
1727         } else
1728                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1729         if (atomic_dec_and_test(&hdr->refcnt))
1730                 hdr->completion_ops->completion(hdr);
1731         return ret;
1732 }
1733 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1734
1735 /*
1736  * There can be multiple RW segments.
1737  */
1738 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1739 {
1740         struct pnfs_layout_segment *lseg;
1741
1742         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1743                 if (lseg->pls_range.iomode == IOMODE_RW &&
1744                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1745                         list_add(&lseg->pls_lc_list, listp);
1746         }
1747 }
1748
1749 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1750 {
1751         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1752 }
1753 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1754
1755 void
1756 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1757 {
1758         struct nfs_pgio_header *hdr = wdata->header;
1759         struct inode *inode = hdr->inode;
1760         struct nfs_inode *nfsi = NFS_I(inode);
1761         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1762         bool mark_as_dirty = false;
1763
1764         spin_lock(&inode->i_lock);
1765         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1766                 mark_as_dirty = true;
1767                 dprintk("%s: Set layoutcommit for inode %lu ",
1768                         __func__, inode->i_ino);
1769         }
1770         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1771                 /* references matched in nfs4_layoutcommit_release */
1772                 pnfs_get_lseg(hdr->lseg);
1773         }
1774         if (end_pos > nfsi->layout->plh_lwb)
1775                 nfsi->layout->plh_lwb = end_pos;
1776         spin_unlock(&inode->i_lock);
1777         dprintk("%s: lseg %p end_pos %llu\n",
1778                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1779
1780         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1781          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1782         if (mark_as_dirty)
1783                 mark_inode_dirty_sync(inode);
1784 }
1785 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1786
1787 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1788 {
1789         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1790
1791         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1792                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1793 }
1794
1795 /*
1796  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1797  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1798  * data to disk to allow the server to recover the data if it crashes.
1799  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1800  * is off, and a COMMIT is sent to a data server, or
1801  * if WRITEs to a data server return NFS_DATA_SYNC.
1802  */
1803 int
1804 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1805 {
1806         struct nfs4_layoutcommit_data *data;
1807         struct nfs_inode *nfsi = NFS_I(inode);
1808         loff_t end_pos;
1809         int status = 0;
1810
1811         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1812
1813         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1814                 return 0;
1815
1816         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1817         data = kzalloc(sizeof(*data), GFP_NOFS);
1818         if (!data) {
1819                 status = -ENOMEM;
1820                 goto out;
1821         }
1822
1823         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1824                 goto out_free;
1825
1826         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1827                 if (!sync) {
1828                         status = -EAGAIN;
1829                         goto out_free;
1830                 }
1831                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1832                                         nfs_wait_bit_killable, TASK_KILLABLE);
1833                 if (status)
1834                         goto out_free;
1835         }
1836
1837         INIT_LIST_HEAD(&data->lseg_list);
1838         spin_lock(&inode->i_lock);
1839         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1840                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1841                 spin_unlock(&inode->i_lock);
1842                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1843                 goto out_free;
1844         }
1845
1846         pnfs_list_write_lseg(inode, &data->lseg_list);
1847
1848         end_pos = nfsi->layout->plh_lwb;
1849         nfsi->layout->plh_lwb = 0;
1850
1851         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1852         spin_unlock(&inode->i_lock);
1853
1854         data->args.inode = inode;
1855         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1856         nfs_fattr_init(&data->fattr);
1857         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1858         data->res.fattr = &data->fattr;
1859         data->args.lastbytewritten = end_pos - 1;
1860         data->res.server = NFS_SERVER(inode);
1861
1862         status = nfs4_proc_layoutcommit(data, sync);
1863 out:
1864         if (status)
1865                 mark_inode_dirty_sync(inode);
1866         dprintk("<-- %s status %d\n", __func__, status);
1867         return status;
1868 out_free:
1869         kfree(data);
1870         goto out;
1871 }
1872
1873 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1874 {
1875         struct nfs4_threshold *thp;
1876
1877         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1878         if (!thp) {
1879                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1880                 return NULL;
1881         }
1882         return thp;
1883 }