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 #include "nfs4trace.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PNFS
  39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  40
  41 /* Locking:
  42  *
  43  * pnfs_spinlock:
  44  *      protects pnfs_modules_tbl.
  45  */
  46 static DEFINE_SPINLOCK(pnfs_spinlock);
  47
  48 /*
  49  * pnfs_modules_tbl holds all pnfs modules
  50  */
  51 static LIST_HEAD(pnfs_modules_tbl);
  52
  53 /* Return the registered pnfs layout driver module matching given id */
  54 static struct pnfs_layoutdriver_type *
  55 find_pnfs_driver_locked(u32 id)
  56 {
  57         struct pnfs_layoutdriver_type *local;
  58
  59         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  60                 if (local->id == id)
  61                         goto out;
  62         local = NULL;
  63 out:
  64         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  65         return local;
  66 }
  67
  68 static struct pnfs_layoutdriver_type *
  69 find_pnfs_driver(u32 id)
  70 {
  71         struct pnfs_layoutdriver_type *local;
  72
  73         spin_lock(&pnfs_spinlock);
  74         local = find_pnfs_driver_locked(id);
  75         if (local != NULL && !try_module_get(local->owner)) {
  76                 dprintk("%s: Could not grab reference on module\n", __func__);
  77                 local = NULL;
  78         }
  79         spin_unlock(&pnfs_spinlock);
  80         return local;
  81 }
  82
  83 void
  84 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  85 {
  86         if (nfss->pnfs_curr_ld) {
  87                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  88                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  89                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  90                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  91                         nfs4_deviceid_purge_client(nfss->nfs_client);
  92                 module_put(nfss->pnfs_curr_ld->owner);
  93         }
  94         nfss->pnfs_curr_ld = NULL;
  95 }
  96
  97 /*
  98  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  99  * Currently only one pNFS layout driver per filesystem is supported.
 100  *
 101  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 102  */
 103 void
 104 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 105                       u32 id)
 106 {
 107         struct pnfs_layoutdriver_type *ld_type = NULL;
 108
 109         if (id == 0)
 110                 goto out_no_driver;
 111         if (!(server->nfs_client->cl_exchange_flags &
 112                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 113                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 114                         __func__, id, server->nfs_client->cl_exchange_flags);
 115                 goto out_no_driver;
 116         }
 117         ld_type = find_pnfs_driver(id);
 118         if (!ld_type) {
 119                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 120                 ld_type = find_pnfs_driver(id);
 121                 if (!ld_type) {
 122                         dprintk("%s: No pNFS module found for %u.\n",
 123                                 __func__, id);
 124                         goto out_no_driver;
 125                 }
 126         }
 127         server->pnfs_curr_ld = ld_type;
 128         if (ld_type->set_layoutdriver
 129             && ld_type->set_layoutdriver(server, mntfh)) {
 130                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 131                         "driver %u.\n", __func__, id);
 132                 module_put(ld_type->owner);
 133                 goto out_no_driver;
 134         }
 135         /* Bump the MDS count */
 136         atomic_inc(&server->nfs_client->cl_mds_count);
 137
 138         dprintk("%s: pNFS module for %u set\n", __func__, id);
 139         return;
 140
 141 out_no_driver:
 142         dprintk("%s: Using NFSv4 I/O\n", __func__);
 143         server->pnfs_curr_ld = NULL;
 144 }
 145
 146 int
 147 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 148 {
 149         int status = -EINVAL;
 150         struct pnfs_layoutdriver_type *tmp;
 151
 152         if (ld_type->id == 0) {
 153                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 154                 return status;
 155         }
 156         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 157                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 158                        "alloc_lseg and free_lseg.\n", __func__);
 159                 return status;
 160         }
 161
 162         spin_lock(&pnfs_spinlock);
 163         tmp = find_pnfs_driver_locked(ld_type->id);
 164         if (!tmp) {
 165                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 166                 status = 0;
 167                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 168                         ld_type->name);
 169         } else {
 170                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 171                         __func__, ld_type->id);
 172         }
 173         spin_unlock(&pnfs_spinlock);
 174
 175         return status;
 176 }
 177 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 178
 179 void
 180 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 181 {
 182         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 183         spin_lock(&pnfs_spinlock);
 184         list_del(&ld_type->pnfs_tblid);
 185         spin_unlock(&pnfs_spinlock);
 186 }
 187 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 188
 189 /*
 190  * pNFS client layout cache
 191  */
 192
 193 /* Need to hold i_lock if caller does not already hold reference */
 194 void
 195 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 196 {
 197         atomic_inc(&lo->plh_refcount);
 198 }
 199
 200 static struct pnfs_layout_hdr *
 201 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 202 {
 203         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 204         return ld->alloc_layout_hdr(ino, gfp_flags);
 205 }
 206
 207 static void
 208 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 209 {
 210         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 211         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 212
 213         if (!list_empty(&lo->plh_layouts)) {
 214                 struct nfs_client *clp = server->nfs_client;
 215
 216                 spin_lock(&clp->cl_lock);
 217                 list_del_init(&lo->plh_layouts);
 218                 spin_unlock(&clp->cl_lock);
 219         }
 220         put_rpccred(lo->plh_lc_cred);
 221         return ld->free_layout_hdr(lo);
 222 }
 223
 224 static void
 225 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 226 {
 227         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 228         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 229         nfsi->layout = NULL;
 230         /* Reset MDS Threshold I/O counters */
 231         nfsi->write_io = 0;
 232         nfsi->read_io = 0;
 233 }
 234
 235 void
 236 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 237 {
 238         struct inode *inode = lo->plh_inode;
 239
 240         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 241                 pnfs_detach_layout_hdr(lo);
 242                 spin_unlock(&inode->i_lock);
 243                 pnfs_free_layout_hdr(lo);
 244         }
 245 }
 246
 247 static int
 248 pnfs_iomode_to_fail_bit(u32 iomode)
 249 {
 250         return iomode == IOMODE_RW ?
 251                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 252 }
 253
 254 static void
 255 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 256 {
 257         lo->plh_retry_timestamp = jiffies;
 258         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 259                 atomic_inc(&lo->plh_refcount);
 260 }
 261
 262 static void
 263 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 264 {
 265         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 266                 atomic_dec(&lo->plh_refcount);
 267 }
 268
 269 static void
 270 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 271 {
 272         struct inode *inode = lo->plh_inode;
 273         struct pnfs_layout_range range = {
 274                 .iomode = iomode,
 275                 .offset = 0,
 276                 .length = NFS4_MAX_UINT64,
 277         };
 278         LIST_HEAD(head);
 279
 280         spin_lock(&inode->i_lock);
 281         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 282         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 283         spin_unlock(&inode->i_lock);
 284         pnfs_free_lseg_list(&head);
 285         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 286                         iomode == IOMODE_RW ?  "RW" : "READ");
 287 }
 288
 289 static bool
 290 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 291 {
 292         unsigned long start, end;
 293         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 294
 295         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 296                 return false;
 297         end = jiffies;
 298         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 299         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 300                 /* It is time to retry the failed layoutgets */
 301                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 302                 return false;
 303         }
 304         return true;
 305 }
 306
 307 static void
 308 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 309 {
 310         INIT_LIST_HEAD(&lseg->pls_list);
 311         INIT_LIST_HEAD(&lseg->pls_lc_list);
 312         atomic_set(&lseg->pls_refcount, 1);
 313         smp_mb();
 314         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 315         lseg->pls_layout = lo;
 316 }
 317
 318 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 319 {
 320         struct inode *ino = lseg->pls_layout->plh_inode;
 321
 322         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 323 }
 324
 325 static void
 326 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 327                 struct pnfs_layout_segment *lseg)
 328 {
 329         struct inode *inode = lo->plh_inode;
 330
 331         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 332         list_del_init(&lseg->pls_list);
 333         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 334         atomic_dec(&lo->plh_refcount);
 335         if (list_empty(&lo->plh_segs))
 336                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 337         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 338 }
 339
 340 void
 341 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 342 {
 343         struct pnfs_layout_hdr *lo;
 344         struct inode *inode;
 345
 346         if (!lseg)
 347                 return;
 348
 349         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 350                 atomic_read(&lseg->pls_refcount),
 351                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 352         lo = lseg->pls_layout;
 353         inode = lo->plh_inode;
 354         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 355                 pnfs_get_layout_hdr(lo);
 356                 pnfs_layout_remove_lseg(lo, lseg);
 357                 spin_unlock(&inode->i_lock);
 358                 pnfs_free_lseg(lseg);
 359                 pnfs_put_layout_hdr(lo);
 360         }
 361 }
 362 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 363
 364 static u64
 365 end_offset(u64 start, u64 len)
 366 {
 367         u64 end;
 368
 369         end = start + len;
 370         return end >= start ? end : NFS4_MAX_UINT64;
 371 }
 372
 373 /*
 374  * is l2 fully contained in l1?
 375  *   start1                             end1
 376  *   [----------------------------------)
 377  *           start2           end2
 378  *           [----------------)
 379  */
 380 static bool
 381 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 382                  const struct pnfs_layout_range *l2)
 383 {
 384         u64 start1 = l1->offset;
 385         u64 end1 = end_offset(start1, l1->length);
 386         u64 start2 = l2->offset;
 387         u64 end2 = end_offset(start2, l2->length);
 388
 389         return (start1 <= start2) && (end1 >= end2);
 390 }
 391
 392 /*
 393  * is l1 and l2 intersecting?
 394  *   start1                             end1
 395  *   [----------------------------------)
 396  *                              start2           end2
 397  *                              [----------------)
 398  */
 399 static bool
 400 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 401                     const struct pnfs_layout_range *l2)
 402 {
 403         u64 start1 = l1->offset;
 404         u64 end1 = end_offset(start1, l1->length);
 405         u64 start2 = l2->offset;
 406         u64 end2 = end_offset(start2, l2->length);
 407
 408         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 409                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 410 }
 411
 412 static bool
 413 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 414                  const struct pnfs_layout_range *recall_range)
 415 {
 416         return (recall_range->iomode == IOMODE_ANY ||
 417                 lseg_range->iomode == recall_range->iomode) &&
 418                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 419 }
 420
 421 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 422                 struct list_head *tmp_list)
 423 {
 424         if (!atomic_dec_and_test(&lseg->pls_refcount))
 425                 return false;
 426         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 427         list_add(&lseg->pls_list, tmp_list);
 428         return true;
 429 }
 430
 431 /* Returns 1 if lseg is removed from list, 0 otherwise */
 432 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 433                              struct list_head *tmp_list)
 434 {
 435         int rv = 0;
 436
 437         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 438                 /* Remove the reference keeping the lseg in the
 439                  * list.  It will now be removed when all
 440                  * outstanding io is finished.
 441                  */
 442                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 443                         atomic_read(&lseg->pls_refcount));
 444                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 445                         rv = 1;
 446         }
 447         return rv;
 448 }
 449
 450 /* Returns count of number of matching invalid lsegs remaining in list
 451  * after call.
 452  */
 453 int
 454 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 455                             struct list_head *tmp_list,
 456                             struct pnfs_layout_range *recall_range)
 457 {
 458         struct pnfs_layout_segment *lseg, *next;
 459         int invalid = 0, removed = 0;
 460
 461         dprintk("%s:Begin lo %p\n", __func__, lo);
 462
 463         if (list_empty(&lo->plh_segs))
 464                 return 0;
 465         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 466                 if (!recall_range ||
 467                     should_free_lseg(&lseg->pls_range, recall_range)) {
 468                         dprintk("%s: freeing lseg %p iomode %d "
 469                                 "offset %llu length %llu\n", __func__,
 470                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 471                                 lseg->pls_range.length);
 472                         invalid++;
 473                         removed += mark_lseg_invalid(lseg, tmp_list);
 474                 }
 475         dprintk("%s:Return %i\n", __func__, invalid - removed);
 476         return invalid - removed;
 477 }
 478
 479 /* note free_me must contain lsegs from a single layout_hdr */
 480 void
 481 pnfs_free_lseg_list(struct list_head *free_me)
 482 {
 483         struct pnfs_layout_segment *lseg, *tmp;
 484
 485         if (list_empty(free_me))
 486                 return;
 487
 488         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 489                 list_del(&lseg->pls_list);
 490                 pnfs_free_lseg(lseg);
 491         }
 492 }
 493
 494 void
 495 pnfs_destroy_layout(struct nfs_inode *nfsi)
 496 {
 497         struct pnfs_layout_hdr *lo;
 498         LIST_HEAD(tmp_list);
 499
 500         spin_lock(&nfsi->vfs_inode.i_lock);
 501         lo = nfsi->layout;
 502         if (lo) {
 503                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 504                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 505                 pnfs_get_layout_hdr(lo);
 506                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 507                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 508                 spin_unlock(&nfsi->vfs_inode.i_lock);
 509                 pnfs_free_lseg_list(&tmp_list);
 510                 pnfs_put_layout_hdr(lo);
 511         } else
 512                 spin_unlock(&nfsi->vfs_inode.i_lock);
 513 }
 514 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 515
 516 static bool
 517 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 518                 struct list_head *layout_list)
 519 {
 520         struct pnfs_layout_hdr *lo;
 521         bool ret = false;
 522
 523         spin_lock(&inode->i_lock);
 524         lo = NFS_I(inode)->layout;
 525         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 526                 pnfs_get_layout_hdr(lo);
 527                 list_add(&lo->plh_bulk_destroy, layout_list);
 528                 ret = true;
 529         }
 530         spin_unlock(&inode->i_lock);
 531         return ret;
 532 }
 533
 534 /* Caller must hold rcu_read_lock and clp->cl_lock */
 535 static int
 536 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 537                 struct nfs_server *server,
 538                 struct list_head *layout_list)
 539 {
 540         struct pnfs_layout_hdr *lo, *next;
 541         struct inode *inode;
 542
 543         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 544                 inode = igrab(lo->plh_inode);
 545                 if (inode == NULL)
 546                         continue;
 547                 list_del_init(&lo->plh_layouts);
 548                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 549                         continue;
 550                 rcu_read_unlock();
 551                 spin_unlock(&clp->cl_lock);
 552                 iput(inode);
 553                 spin_lock(&clp->cl_lock);
 554                 rcu_read_lock();
 555                 return -EAGAIN;
 556         }
 557         return 0;
 558 }
 559
 560 static int
 561 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 562                 bool is_bulk_recall)
 563 {
 564         struct pnfs_layout_hdr *lo;
 565         struct inode *inode;
 566         struct pnfs_layout_range range = {
 567                 .iomode = IOMODE_ANY,
 568                 .offset = 0,
 569                 .length = NFS4_MAX_UINT64,
 570         };
 571         LIST_HEAD(lseg_list);
 572         int ret = 0;
 573
 574         while (!list_empty(layout_list)) {
 575                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 576                                 plh_bulk_destroy);
 577                 dprintk("%s freeing layout for inode %lu\n", __func__,
 578                         lo->plh_inode->i_ino);
 579                 inode = lo->plh_inode;
 580                 spin_lock(&inode->i_lock);
 581                 list_del_init(&lo->plh_bulk_destroy);
 582                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 583                 if (is_bulk_recall)
 584                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 585                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 586                         ret = -EAGAIN;
 587                 spin_unlock(&inode->i_lock);
 588                 pnfs_free_lseg_list(&lseg_list);
 589                 pnfs_put_layout_hdr(lo);
 590                 iput(inode);
 591         }
 592         return ret;
 593 }
 594
 595 int
 596 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 597                 struct nfs_fsid *fsid,
 598                 bool is_recall)
 599 {
 600         struct nfs_server *server;
 601         LIST_HEAD(layout_list);
 602
 603         spin_lock(&clp->cl_lock);
 604         rcu_read_lock();
 605 restart:
 606         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 607                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 608                         continue;
 609                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 610                                 server,
 611                                 &layout_list) != 0)
 612                         goto restart;
 613         }
 614         rcu_read_unlock();
 615         spin_unlock(&clp->cl_lock);
 616
 617         if (list_empty(&layout_list))
 618                 return 0;
 619         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 620 }
 621
 622 int
 623 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 624                 bool is_recall)
 625 {
 626         struct nfs_server *server;
 627         LIST_HEAD(layout_list);
 628
 629         spin_lock(&clp->cl_lock);
 630         rcu_read_lock();
 631 restart:
 632         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 633                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 634                                         server,
 635                                         &layout_list) != 0)
 636                         goto restart;
 637         }
 638         rcu_read_unlock();
 639         spin_unlock(&clp->cl_lock);
 640
 641         if (list_empty(&layout_list))
 642                 return 0;
 643         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 644 }
 645
 646 /*
 647  * Called by the state manger to remove all layouts established under an
 648  * expired lease.
 649  */
 650 void
 651 pnfs_destroy_all_layouts(struct nfs_client *clp)
 652 {
 653         nfs4_deviceid_mark_client_invalid(clp);
 654         nfs4_deviceid_purge_client(clp);
 655
 656         pnfs_destroy_layouts_byclid(clp, false);
 657 }
 658
 659 /*
 660  * Compare 2 layout stateid sequence ids, to see which is newer,
 661  * taking into account wraparound issues.
 662  */
 663 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 664 {
 665         return (s32)s1 - (s32)s2 > 0;
 666 }
 667
 668 /* update lo->plh_stateid with new if is more recent */
 669 void
 670 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 671                         bool update_barrier)
 672 {
 673         u32 oldseq, newseq, new_barrier;
 674         int empty = list_empty(&lo->plh_segs);
 675
 676         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 677         newseq = be32_to_cpu(new->seqid);
 678         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 679                 nfs4_stateid_copy(&lo->plh_stateid, new);
 680                 if (update_barrier) {
 681                         new_barrier = be32_to_cpu(new->seqid);
 682                 } else {
 683                         /* Because of wraparound, we want to keep the barrier
 684                          * "close" to the current seqids.
 685                          */
 686                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 687                 }
 688                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 689                         lo->plh_barrier = new_barrier;
 690         }
 691 }
 692
 693 static bool
 694 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 695                 const nfs4_stateid *stateid)
 696 {
 697         u32 seqid = be32_to_cpu(stateid->seqid);
 698
 699         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 700 }
 701
 702 /* lget is set to 1 if called from inside send_layoutget call chain */
 703 static bool
 704 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 705 {
 706         return lo->plh_block_lgets ||
 707                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 708                 (list_empty(&lo->plh_segs) &&
 709                  (atomic_read(&lo->plh_outstanding) > lget));
 710 }
 711
 712 int
 713 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 714                               struct nfs4_state *open_state)
 715 {
 716         int status = 0;
 717
 718         dprintk("--> %s\n", __func__);
 719         spin_lock(&lo->plh_inode->i_lock);
 720         if (pnfs_layoutgets_blocked(lo, 1)) {
 721                 status = -EAGAIN;
 722         } else if (!nfs4_valid_open_stateid(open_state)) {
 723                 status = -EBADF;
 724         } else if (list_empty(&lo->plh_segs)) {
 725                 int seq;
 726
 727                 do {
 728                         seq = read_seqbegin(&open_state->seqlock);
 729                         nfs4_stateid_copy(dst, &open_state->stateid);
 730                 } while (read_seqretry(&open_state->seqlock, seq));
 731         } else
 732                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 733         spin_unlock(&lo->plh_inode->i_lock);
 734         dprintk("<-- %s\n", __func__);
 735         return status;
 736 }
 737
 738 /*
 739 * Get layout from server.
 740 *    for now, assume that whole file layouts are requested.
 741 *    arg->offset: 0
 742 *    arg->length: all ones
 743 */
 744 static struct pnfs_layout_segment *
 745 send_layoutget(struct pnfs_layout_hdr *lo,
 746            struct nfs_open_context *ctx,
 747            struct pnfs_layout_range *range,
 748            gfp_t gfp_flags)
 749 {
 750         struct inode *ino = lo->plh_inode;
 751         struct nfs_server *server = NFS_SERVER(ino);
 752         struct nfs4_layoutget *lgp;
 753         struct pnfs_layout_segment *lseg;
 754
 755         dprintk("--> %s\n", __func__);
 756
 757         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 758         if (lgp == NULL)
 759                 return NULL;
 760
 761         lgp->args.minlength = PAGE_CACHE_SIZE;
 762         if (lgp->args.minlength > range->length)
 763                 lgp->args.minlength = range->length;
 764         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 765         lgp->args.range = *range;
 766         lgp->args.type = server->pnfs_curr_ld->id;
 767         lgp->args.inode = ino;
 768         lgp->args.ctx = get_nfs_open_context(ctx);
 769         lgp->gfp_flags = gfp_flags;
 770         lgp->cred = lo->plh_lc_cred;
 771
 772         /* Synchronously retrieve layout information from server and
 773          * store in lseg.
 774          */
 775         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 776         if (IS_ERR(lseg)) {
 777                 switch (PTR_ERR(lseg)) {
 778                 case -ENOMEM:
 779                 case -ERESTARTSYS:
 780                         break;
 781                 default:
 782                         /* remember that LAYOUTGET failed and suspend trying */
 783                         pnfs_layout_io_set_failed(lo, range->iomode);
 784                 }
 785                 return NULL;
 786         }
 787
 788         return lseg;
 789 }
 790
 791 static void pnfs_clear_layoutcommit(struct inode *inode,
 792                 struct list_head *head)
 793 {
 794         struct nfs_inode *nfsi = NFS_I(inode);
 795         struct pnfs_layout_segment *lseg, *tmp;
 796
 797         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 798                 return;
 799         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 800                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 801                         continue;
 802                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 803         }
 804 }
 805
 806 /*
 807  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 808  * when the layout segment list is empty.
 809  *
 810  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 811  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 812  * deviceid is marked invalid.
 813  */
 814 int
 815 _pnfs_return_layout(struct inode *ino)
 816 {
 817         struct pnfs_layout_hdr *lo = NULL;
 818         struct nfs_inode *nfsi = NFS_I(ino);
 819         LIST_HEAD(tmp_list);
 820         struct nfs4_layoutreturn *lrp;
 821         nfs4_stateid stateid;
 822         int status = 0, empty;
 823
 824         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 825
 826         spin_lock(&ino->i_lock);
 827         lo = nfsi->layout;
 828         if (!lo) {
 829                 spin_unlock(&ino->i_lock);
 830                 dprintk("NFS: %s no layout to return\n", __func__);
 831                 goto out;
 832         }
 833         stateid = nfsi->layout->plh_stateid;
 834         /* Reference matched in nfs4_layoutreturn_release */
 835         pnfs_get_layout_hdr(lo);
 836         empty = list_empty(&lo->plh_segs);
 837         pnfs_clear_layoutcommit(ino, &tmp_list);
 838         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 839         /* Don't send a LAYOUTRETURN if list was initially empty */
 840         if (empty) {
 841                 spin_unlock(&ino->i_lock);
 842                 pnfs_put_layout_hdr(lo);
 843                 dprintk("NFS: %s no layout segments to return\n", __func__);
 844                 goto out;
 845         }
 846         lo->plh_block_lgets++;
 847         spin_unlock(&ino->i_lock);
 848         pnfs_free_lseg_list(&tmp_list);
 849
 850         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 851         if (unlikely(lrp == NULL)) {
 852                 status = -ENOMEM;
 853                 spin_lock(&ino->i_lock);
 854                 lo->plh_block_lgets--;
 855                 spin_unlock(&ino->i_lock);
 856                 pnfs_put_layout_hdr(lo);
 857                 goto out;
 858         }
 859
 860         lrp->args.stateid = stateid;
 861         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 862         lrp->args.inode = ino;
 863         lrp->args.layout = lo;
 864         lrp->clp = NFS_SERVER(ino)->nfs_client;
 865         lrp->cred = lo->plh_lc_cred;
 866
 867         status = nfs4_proc_layoutreturn(lrp);
 868 out:
 869         dprintk("<-- %s status: %d\n", __func__, status);
 870         return status;
 871 }
 872 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 873
 874 int
 875 pnfs_commit_and_return_layout(struct inode *inode)
 876 {
 877         struct pnfs_layout_hdr *lo;
 878         int ret;
 879
 880         spin_lock(&inode->i_lock);
 881         lo = NFS_I(inode)->layout;
 882         if (lo == NULL) {
 883                 spin_unlock(&inode->i_lock);
 884                 return 0;
 885         }
 886         pnfs_get_layout_hdr(lo);
 887         /* Block new layoutgets and read/write to ds */
 888         lo->plh_block_lgets++;
 889         spin_unlock(&inode->i_lock);
 890         filemap_fdatawait(inode->i_mapping);
 891         ret = pnfs_layoutcommit_inode(inode, true);
 892         if (ret == 0)
 893                 ret = _pnfs_return_layout(inode);
 894         spin_lock(&inode->i_lock);
 895         lo->plh_block_lgets--;
 896         spin_unlock(&inode->i_lock);
 897         pnfs_put_layout_hdr(lo);
 898         return ret;
 899 }
 900
 901 bool pnfs_roc(struct inode *ino)
 902 {
 903         struct pnfs_layout_hdr *lo;
 904         struct pnfs_layout_segment *lseg, *tmp;
 905         LIST_HEAD(tmp_list);
 906         bool found = false;
 907
 908         spin_lock(&ino->i_lock);
 909         lo = NFS_I(ino)->layout;
 910         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 911             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 912                 goto out_nolayout;
 913         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 914                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 915                         mark_lseg_invalid(lseg, &tmp_list);
 916                         found = true;
 917                 }
 918         if (!found)
 919                 goto out_nolayout;
 920         lo->plh_block_lgets++;
 921         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 922         spin_unlock(&ino->i_lock);
 923         pnfs_free_lseg_list(&tmp_list);
 924         return true;
 925
 926 out_nolayout:
 927         spin_unlock(&ino->i_lock);
 928         return false;
 929 }
 930
 931 void pnfs_roc_release(struct inode *ino)
 932 {
 933         struct pnfs_layout_hdr *lo;
 934
 935         spin_lock(&ino->i_lock);
 936         lo = NFS_I(ino)->layout;
 937         lo->plh_block_lgets--;
 938         if (atomic_dec_and_test(&lo->plh_refcount)) {
 939                 pnfs_detach_layout_hdr(lo);
 940                 spin_unlock(&ino->i_lock);
 941                 pnfs_free_layout_hdr(lo);
 942         } else
 943                 spin_unlock(&ino->i_lock);
 944 }
 945
 946 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 947 {
 948         struct pnfs_layout_hdr *lo;
 949
 950         spin_lock(&ino->i_lock);
 951         lo = NFS_I(ino)->layout;
 952         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 953                 lo->plh_barrier = barrier;
 954         spin_unlock(&ino->i_lock);
 955 }
 956
 957 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 958 {
 959         struct nfs_inode *nfsi = NFS_I(ino);
 960         struct pnfs_layout_hdr *lo;
 961         struct pnfs_layout_segment *lseg;
 962         u32 current_seqid;
 963         bool found = false;
 964
 965         spin_lock(&ino->i_lock);
 966         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 967                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 968                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 969                         found = true;
 970                         goto out;
 971                 }
 972         lo = nfsi->layout;
 973         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 974
 975         /* Since close does not return a layout stateid for use as
 976          * a barrier, we choose the worst-case barrier.
 977          */
 978         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 979 out:
 980         spin_unlock(&ino->i_lock);
 981         return found;
 982 }
 983
 984 /*
 985  * Compare two layout segments for sorting into layout cache.
 986  * We want to preferentially return RW over RO layouts, so ensure those
 987  * are seen first.
 988  */
 989 static s64
 990 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
 991            const struct pnfs_layout_range *l2)
 992 {
 993         s64 d;
 994
 995         /* high offset > low offset */
 996         d = l1->offset - l2->offset;
 997         if (d)
 998                 return d;
 999
1000         /* short length > long length */
1001         d = l2->length - l1->length;
1002         if (d)
1003                 return d;
1004
1005         /* read > read/write */
1006         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1007 }
1008
1009 static void
1010 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1011                    struct pnfs_layout_segment *lseg)
1012 {
1013         struct pnfs_layout_segment *lp;
1014
1015         dprintk("%s:Begin\n", __func__);
1016
1017         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1018                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1019                         continue;
1020                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1021                 dprintk("%s: inserted lseg %p "
1022                         "iomode %d offset %llu length %llu before "
1023                         "lp %p iomode %d offset %llu length %llu\n",
1024                         __func__, lseg, lseg->pls_range.iomode,
1025                         lseg->pls_range.offset, lseg->pls_range.length,
1026                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1027                         lp->pls_range.length);
1028                 goto out;
1029         }
1030         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1031         dprintk("%s: inserted lseg %p "
1032                 "iomode %d offset %llu length %llu at tail\n",
1033                 __func__, lseg, lseg->pls_range.iomode,
1034                 lseg->pls_range.offset, lseg->pls_range.length);
1035 out:
1036         pnfs_get_layout_hdr(lo);
1037
1038         dprintk("%s:Return\n", __func__);
1039 }
1040
1041 static struct pnfs_layout_hdr *
1042 alloc_init_layout_hdr(struct inode *ino,
1043                       struct nfs_open_context *ctx,
1044                       gfp_t gfp_flags)
1045 {
1046         struct pnfs_layout_hdr *lo;
1047
1048         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1049         if (!lo)
1050                 return NULL;
1051         atomic_set(&lo->plh_refcount, 1);
1052         INIT_LIST_HEAD(&lo->plh_layouts);
1053         INIT_LIST_HEAD(&lo->plh_segs);
1054         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1055         lo->plh_inode = ino;
1056         lo->plh_lc_cred = get_rpccred(ctx->cred);
1057         return lo;
1058 }
1059
1060 static struct pnfs_layout_hdr *
1061 pnfs_find_alloc_layout(struct inode *ino,
1062                        struct nfs_open_context *ctx,
1063                        gfp_t gfp_flags)
1064 {
1065         struct nfs_inode *nfsi = NFS_I(ino);
1066         struct pnfs_layout_hdr *new = NULL;
1067
1068         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1069
1070         if (nfsi->layout != NULL)
1071                 goto out_existing;
1072         spin_unlock(&ino->i_lock);
1073         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1074         spin_lock(&ino->i_lock);
1075
1076         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1077                 nfsi->layout = new;
1078                 return new;
1079         } else if (new != NULL)
1080                 pnfs_free_layout_hdr(new);
1081 out_existing:
1082         pnfs_get_layout_hdr(nfsi->layout);
1083         return nfsi->layout;
1084 }
1085
1086 /*
1087  * iomode matching rules:
1088  * iomode       lseg    match
1089  * -----        -----   -----
1090  * ANY          READ    true
1091  * ANY          RW      true
1092  * RW           READ    false
1093  * RW           RW      true
1094  * READ         READ    true
1095  * READ         RW      true
1096  */
1097 static bool
1098 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1099                  const struct pnfs_layout_range *range)
1100 {
1101         struct pnfs_layout_range range1;
1102
1103         if ((range->iomode == IOMODE_RW &&
1104              ls_range->iomode != IOMODE_RW) ||
1105             !pnfs_lseg_range_intersecting(ls_range, range))
1106                 return 0;
1107
1108         /* range1 covers only the first byte in the range */
1109         range1 = *range;
1110         range1.length = 1;
1111         return pnfs_lseg_range_contained(ls_range, &range1);
1112 }
1113
1114 /*
1115  * lookup range in layout
1116  */
1117 static struct pnfs_layout_segment *
1118 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1119                 struct pnfs_layout_range *range)
1120 {
1121         struct pnfs_layout_segment *lseg, *ret = NULL;
1122
1123         dprintk("%s:Begin\n", __func__);
1124
1125         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1126                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1127                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1128                         ret = pnfs_get_lseg(lseg);
1129                         break;
1130                 }
1131                 if (lseg->pls_range.offset > range->offset)
1132                         break;
1133         }
1134
1135         dprintk("%s:Return lseg %p ref %d\n",
1136                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1137         return ret;
1138 }
1139
1140 /*
1141  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1142  * to the MDS or over pNFS
1143  *
1144  * The nfs_inode read_io and write_io fields are cumulative counters reset
1145  * when there are no layout segments. Note that in pnfs_update_layout iomode
1146  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1147  * WRITE request.
1148  *
1149  * A return of true means use MDS I/O.
1150  *
1151  * From rfc 5661:
1152  * If a file's size is smaller than the file size threshold, data accesses
1153  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1154  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1155  * server.  If both file size and I/O size are provided, the client SHOULD
1156  * reach or exceed  both thresholds before sending its read or write
1157  * requests to the data server.
1158  */
1159 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1160                                      struct inode *ino, int iomode)
1161 {
1162         struct nfs4_threshold *t = ctx->mdsthreshold;
1163         struct nfs_inode *nfsi = NFS_I(ino);
1164         loff_t fsize = i_size_read(ino);
1165         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1166
1167         if (t == NULL)
1168                 return ret;
1169
1170         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1171                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1172
1173         switch (iomode) {
1174         case IOMODE_READ:
1175                 if (t->bm & THRESHOLD_RD) {
1176                         dprintk("%s fsize %llu\n", __func__, fsize);
1177                         size_set = true;
1178                         if (fsize < t->rd_sz)
1179                                 size = true;
1180                 }
1181                 if (t->bm & THRESHOLD_RD_IO) {
1182                         dprintk("%s nfsi->read_io %llu\n", __func__,
1183                                 nfsi->read_io);
1184                         io_set = true;
1185                         if (nfsi->read_io < t->rd_io_sz)
1186                                 io = true;
1187                 }
1188                 break;
1189         case IOMODE_RW:
1190                 if (t->bm & THRESHOLD_WR) {
1191                         dprintk("%s fsize %llu\n", __func__, fsize);
1192                         size_set = true;
1193                         if (fsize < t->wr_sz)
1194                                 size = true;
1195                 }
1196                 if (t->bm & THRESHOLD_WR_IO) {
1197                         dprintk("%s nfsi->write_io %llu\n", __func__,
1198                                 nfsi->write_io);
1199                         io_set = true;
1200                         if (nfsi->write_io < t->wr_io_sz)
1201                                 io = true;
1202                 }
1203                 break;
1204         }
1205         if (size_set && io_set) {
1206                 if (size && io)
1207                         ret = true;
1208         } else if (size || io)
1209                 ret = true;
1210
1211         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1212         return ret;
1213 }
1214
1215 /*
1216  * Layout segment is retreived from the server if not cached.
1217  * The appropriate layout segment is referenced and returned to the caller.
1218  */
1219 struct pnfs_layout_segment *
1220 pnfs_update_layout(struct inode *ino,
1221                    struct nfs_open_context *ctx,
1222                    loff_t pos,
1223                    u64 count,
1224                    enum pnfs_iomode iomode,
1225                    gfp_t gfp_flags)
1226 {
1227         struct pnfs_layout_range arg = {
1228                 .iomode = iomode,
1229                 .offset = pos,
1230                 .length = count,
1231         };
1232         unsigned pg_offset;
1233         struct nfs_server *server = NFS_SERVER(ino);
1234         struct nfs_client *clp = server->nfs_client;
1235         struct pnfs_layout_hdr *lo;
1236         struct pnfs_layout_segment *lseg = NULL;
1237         bool first;
1238
1239         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1240                 goto out;
1241
1242         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1243                 goto out;
1244
1245         spin_lock(&ino->i_lock);
1246         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1247         if (lo == NULL) {
1248                 spin_unlock(&ino->i_lock);
1249                 goto out;
1250         }
1251
1252         /* Do we even need to bother with this? */
1253         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1254                 dprintk("%s matches recall, use MDS\n", __func__);
1255                 goto out_unlock;
1256         }
1257
1258         /* if LAYOUTGET already failed once we don't try again */
1259         if (pnfs_layout_io_test_failed(lo, iomode))
1260                 goto out_unlock;
1261
1262         /* Check to see if the layout for the given range already exists */
1263         lseg = pnfs_find_lseg(lo, &arg);
1264         if (lseg)
1265                 goto out_unlock;
1266
1267         if (pnfs_layoutgets_blocked(lo, 0))
1268                 goto out_unlock;
1269         atomic_inc(&lo->plh_outstanding);
1270
1271         first = list_empty(&lo->plh_layouts) ? true : false;
1272         spin_unlock(&ino->i_lock);
1273
1274         if (first) {
1275                 /* The lo must be on the clp list if there is any
1276                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1277                  */
1278                 spin_lock(&clp->cl_lock);
1279                 list_add_tail(&lo->plh_layouts, &server->layouts);
1280                 spin_unlock(&clp->cl_lock);
1281         }
1282
1283         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1284         if (pg_offset) {
1285                 arg.offset -= pg_offset;
1286                 arg.length += pg_offset;
1287         }
1288         if (arg.length != NFS4_MAX_UINT64)
1289                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1290
1291         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1292         atomic_dec(&lo->plh_outstanding);
1293 out_put_layout_hdr:
1294         pnfs_put_layout_hdr(lo);
1295 out:
1296         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1297                         "(%s, offset: %llu, length: %llu)\n",
1298                         __func__, ino->i_sb->s_id,
1299                         (unsigned long long)NFS_FILEID(ino),
1300                         lseg == NULL ? "not found" : "found",
1301                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1302                         (unsigned long long)pos,
1303                         (unsigned long long)count);
1304         return lseg;
1305 out_unlock:
1306         spin_unlock(&ino->i_lock);
1307         goto out_put_layout_hdr;
1308 }
1309 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1310
1311 struct pnfs_layout_segment *
1312 pnfs_layout_process(struct nfs4_layoutget *lgp)
1313 {
1314         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1315         struct nfs4_layoutget_res *res = &lgp->res;
1316         struct pnfs_layout_segment *lseg;
1317         struct inode *ino = lo->plh_inode;
1318         int status = 0;
1319
1320         /* Inject layout blob into I/O device driver */
1321         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1322         if (!lseg || IS_ERR(lseg)) {
1323                 if (!lseg)
1324                         status = -ENOMEM;
1325                 else
1326                         status = PTR_ERR(lseg);
1327                 dprintk("%s: Could not allocate layout: error %d\n",
1328                        __func__, status);
1329                 goto out;
1330         }
1331
1332         spin_lock(&ino->i_lock);
1333         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1334                 dprintk("%s forget reply due to recall\n", __func__);
1335                 goto out_forget_reply;
1336         }
1337
1338         if (pnfs_layoutgets_blocked(lo, 1) ||
1339             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1340                 dprintk("%s forget reply due to state\n", __func__);
1341                 goto out_forget_reply;
1342         }
1343
1344         /* Done processing layoutget. Set the layout stateid */
1345         pnfs_set_layout_stateid(lo, &res->stateid, false);
1346
1347         init_lseg(lo, lseg);
1348         lseg->pls_range = res->range;
1349         pnfs_get_lseg(lseg);
1350         pnfs_layout_insert_lseg(lo, lseg);
1351
1352         if (res->return_on_close) {
1353                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1354                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1355         }
1356
1357         spin_unlock(&ino->i_lock);
1358         return lseg;
1359 out:
1360         return ERR_PTR(status);
1361
1362 out_forget_reply:
1363         spin_unlock(&ino->i_lock);
1364         lseg->pls_layout = lo;
1365         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1366         goto out;
1367 }
1368
1369 void
1370 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1371 {
1372         u64 rd_size = req->wb_bytes;
1373
1374         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1375
1376         if (req->wb_offset != req->wb_pgbase) {
1377                 nfs_pageio_reset_read_mds(pgio);
1378                 return;
1379         }
1380
1381         if (pgio->pg_dreq == NULL)
1382                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1383         else
1384                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1385
1386         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1387                                            req->wb_context,
1388                                            req_offset(req),
1389                                            rd_size,
1390                                            IOMODE_READ,
1391                                            GFP_KERNEL);
1392         /* If no lseg, fall back to read through mds */
1393         if (pgio->pg_lseg == NULL)
1394                 nfs_pageio_reset_read_mds(pgio);
1395
1396 }
1397 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1398
1399 void
1400 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1401                            struct nfs_page *req, u64 wb_size)
1402 {
1403         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1404
1405         if (req->wb_offset != req->wb_pgbase) {
1406                 nfs_pageio_reset_write_mds(pgio);
1407                 return;
1408         }
1409
1410         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1411                                            req->wb_context,
1412                                            req_offset(req),
1413                                            wb_size,
1414                                            IOMODE_RW,
1415                                            GFP_NOFS);
1416         /* If no lseg, fall back to write through mds */
1417         if (pgio->pg_lseg == NULL)
1418                 nfs_pageio_reset_write_mds(pgio);
1419 }
1420 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1421
1422 void
1423 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1424                       const struct nfs_pgio_completion_ops *compl_ops)
1425 {
1426         struct nfs_server *server = NFS_SERVER(inode);
1427         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1428
1429         if (ld == NULL)
1430                 nfs_pageio_init_read(pgio, inode, compl_ops);
1431         else
1432                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1433 }
1434
1435 void
1436 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1437                        int ioflags,
1438                        const struct nfs_pgio_completion_ops *compl_ops)
1439 {
1440         struct nfs_server *server = NFS_SERVER(inode);
1441         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1442
1443         if (ld == NULL)
1444                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1445         else
1446                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1447 }
1448
1449 bool
1450 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1451                      struct nfs_page *req)
1452 {
1453         if (pgio->pg_lseg == NULL)
1454                 return nfs_generic_pg_test(pgio, prev, req);
1455
1456         /*
1457          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1458          * Note that this test makes several assumptions:
1459          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1460          *   is known to lie within the range.
1461          *   - that the nfs_page being tested is known to be contiguous with the
1462          *   previous nfs_page.
1463          *   - Layout ranges are page aligned, so we only have to test the
1464          *   start offset of the request.
1465          *
1466          * Please also note that 'end_offset' is actually the offset of the
1467          * first byte that lies outside the pnfs_layout_range. FIXME?
1468          *
1469          */
1470         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1471                                          pgio->pg_lseg->pls_range.length);
1472 }
1473 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1474
1475 int pnfs_write_done_resend_to_mds(struct inode *inode,
1476                                 struct list_head *head,
1477                                 const struct nfs_pgio_completion_ops *compl_ops,
1478                                 struct nfs_direct_req *dreq)
1479 {
1480         struct nfs_pageio_descriptor pgio;
1481         LIST_HEAD(failed);
1482
1483         /* Resend all requests through the MDS */
1484         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1485         pgio.pg_dreq = dreq;
1486         while (!list_empty(head)) {
1487                 struct nfs_page *req = nfs_list_entry(head->next);
1488
1489                 nfs_list_remove_request(req);
1490                 if (!nfs_pageio_add_request(&pgio, req))
1491                         nfs_list_add_request(req, &failed);
1492         }
1493         nfs_pageio_complete(&pgio);
1494
1495         if (!list_empty(&failed)) {
1496                 /* For some reason our attempt to resend pages. Mark the
1497                  * overall send request as having failed, and let
1498                  * nfs_writeback_release_full deal with the error.
1499                  */
1500                 list_move(&failed, head);
1501                 return -EIO;
1502         }
1503         return 0;
1504 }
1505 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1506
1507 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1508 {
1509         struct nfs_pgio_header *hdr = data->header;
1510
1511         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1512         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1513             PNFS_LAYOUTRET_ON_ERROR) {
1514                 pnfs_return_layout(hdr->inode);
1515         }
1516         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1517                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1518                                                         &hdr->pages,
1519                                                         hdr->completion_ops,
1520                                                         hdr->dreq);
1521 }
1522
1523 /*
1524  * Called by non rpc-based layout drivers
1525  */
1526 void pnfs_ld_write_done(struct nfs_write_data *data)
1527 {
1528         struct nfs_pgio_header *hdr = data->header;
1529
1530         trace_nfs4_pnfs_write(data, hdr->pnfs_error);
1531         if (!hdr->pnfs_error) {
1532                 pnfs_set_layoutcommit(data);
1533                 hdr->mds_ops->rpc_call_done(&data->task, data);
1534         } else
1535                 pnfs_ld_handle_write_error(data);
1536         hdr->mds_ops->rpc_release(data);
1537 }
1538 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1539
1540 static void
1541 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1542                 struct nfs_write_data *data)
1543 {
1544         struct nfs_pgio_header *hdr = data->header;
1545
1546         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1547                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1548                 nfs_pageio_reset_write_mds(desc);
1549                 desc->pg_recoalesce = 1;
1550         }
1551         nfs_writedata_release(data);
1552 }
1553
1554 static enum pnfs_try_status
1555 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1556                         const struct rpc_call_ops *call_ops,
1557                         struct pnfs_layout_segment *lseg,
1558                         int how)
1559 {
1560         struct nfs_pgio_header *hdr = wdata->header;
1561         struct inode *inode = hdr->inode;
1562         enum pnfs_try_status trypnfs;
1563         struct nfs_server *nfss = NFS_SERVER(inode);
1564
1565         hdr->mds_ops = call_ops;
1566
1567         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1568                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1569         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1570         if (trypnfs != PNFS_NOT_ATTEMPTED)
1571                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1572         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1573         return trypnfs;
1574 }
1575
1576 static void
1577 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1578 {
1579         struct nfs_write_data *data;
1580         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1581         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1582
1583         desc->pg_lseg = NULL;
1584         while (!list_empty(head)) {
1585                 enum pnfs_try_status trypnfs;
1586
1587                 data = list_first_entry(head, struct nfs_write_data, list);
1588                 list_del_init(&data->list);
1589
1590                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1591                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1592                         pnfs_write_through_mds(desc, data);
1593         }
1594         pnfs_put_lseg(lseg);
1595 }
1596
1597 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1598 {
1599         pnfs_put_lseg(hdr->lseg);
1600         nfs_writehdr_free(hdr);
1601 }
1602 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1603
1604 int
1605 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1606 {
1607         struct nfs_write_header *whdr;
1608         struct nfs_pgio_header *hdr;
1609         int ret;
1610
1611         whdr = nfs_writehdr_alloc();
1612         if (!whdr) {
1613                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1614                 pnfs_put_lseg(desc->pg_lseg);
1615                 desc->pg_lseg = NULL;
1616                 return -ENOMEM;
1617         }
1618         hdr = &whdr->header;
1619         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1620         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1621         atomic_inc(&hdr->refcnt);
1622         ret = nfs_generic_flush(desc, hdr);
1623         if (ret != 0) {
1624                 pnfs_put_lseg(desc->pg_lseg);
1625                 desc->pg_lseg = NULL;
1626         } else
1627                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1628         if (atomic_dec_and_test(&hdr->refcnt))
1629                 hdr->completion_ops->completion(hdr);
1630         return ret;
1631 }
1632 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1633
1634 int pnfs_read_done_resend_to_mds(struct inode *inode,
1635                                 struct list_head *head,
1636                                 const struct nfs_pgio_completion_ops *compl_ops,
1637                                 struct nfs_direct_req *dreq)
1638 {
1639         struct nfs_pageio_descriptor pgio;
1640         LIST_HEAD(failed);
1641
1642         /* Resend all requests through the MDS */
1643         nfs_pageio_init_read(&pgio, inode, compl_ops);
1644         pgio.pg_dreq = dreq;
1645         while (!list_empty(head)) {
1646                 struct nfs_page *req = nfs_list_entry(head->next);
1647
1648                 nfs_list_remove_request(req);
1649                 if (!nfs_pageio_add_request(&pgio, req))
1650                         nfs_list_add_request(req, &failed);
1651         }
1652         nfs_pageio_complete(&pgio);
1653
1654         if (!list_empty(&failed)) {
1655                 list_move(&failed, head);
1656                 return -EIO;
1657         }
1658         return 0;
1659 }
1660 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1661
1662 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1663 {
1664         struct nfs_pgio_header *hdr = data->header;
1665
1666         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1667         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1668             PNFS_LAYOUTRET_ON_ERROR) {
1669                 pnfs_return_layout(hdr->inode);
1670         }
1671         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1672                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1673                                                         &hdr->pages,
1674                                                         hdr->completion_ops,
1675                                                         hdr->dreq);
1676 }
1677
1678 /*
1679  * Called by non rpc-based layout drivers
1680  */
1681 void pnfs_ld_read_done(struct nfs_read_data *data)
1682 {
1683         struct nfs_pgio_header *hdr = data->header;
1684
1685         trace_nfs4_pnfs_read(data, hdr->pnfs_error);
1686         if (likely(!hdr->pnfs_error)) {
1687                 __nfs4_read_done_cb(data);
1688                 hdr->mds_ops->rpc_call_done(&data->task, data);
1689         } else
1690                 pnfs_ld_handle_read_error(data);
1691         hdr->mds_ops->rpc_release(data);
1692 }
1693 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1694
1695 static void
1696 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1697                 struct nfs_read_data *data)
1698 {
1699         struct nfs_pgio_header *hdr = data->header;
1700
1701         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1702                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1703                 nfs_pageio_reset_read_mds(desc);
1704                 desc->pg_recoalesce = 1;
1705         }
1706         nfs_readdata_release(data);
1707 }
1708
1709 /*
1710  * Call the appropriate parallel I/O subsystem read function.
1711  */
1712 static enum pnfs_try_status
1713 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1714                        const struct rpc_call_ops *call_ops,
1715                        struct pnfs_layout_segment *lseg)
1716 {
1717         struct nfs_pgio_header *hdr = rdata->header;
1718         struct inode *inode = hdr->inode;
1719         struct nfs_server *nfss = NFS_SERVER(inode);
1720         enum pnfs_try_status trypnfs;
1721
1722         hdr->mds_ops = call_ops;
1723
1724         dprintk("%s: Reading ino:%lu %u@%llu\n",
1725                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1726
1727         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1728         if (trypnfs != PNFS_NOT_ATTEMPTED)
1729                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1730         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1731         return trypnfs;
1732 }
1733
1734 static void
1735 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1736 {
1737         struct nfs_read_data *data;
1738         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1739         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1740
1741         desc->pg_lseg = NULL;
1742         while (!list_empty(head)) {
1743                 enum pnfs_try_status trypnfs;
1744
1745                 data = list_first_entry(head, struct nfs_read_data, list);
1746                 list_del_init(&data->list);
1747
1748                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1749                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1750                         pnfs_read_through_mds(desc, data);
1751         }
1752         pnfs_put_lseg(lseg);
1753 }
1754
1755 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1756 {
1757         pnfs_put_lseg(hdr->lseg);
1758         nfs_readhdr_free(hdr);
1759 }
1760 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1761
1762 int
1763 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1764 {
1765         struct nfs_read_header *rhdr;
1766         struct nfs_pgio_header *hdr;
1767         int ret;
1768
1769         rhdr = nfs_readhdr_alloc();
1770         if (!rhdr) {
1771                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1772                 ret = -ENOMEM;
1773                 pnfs_put_lseg(desc->pg_lseg);
1774                 desc->pg_lseg = NULL;
1775                 return ret;
1776         }
1777         hdr = &rhdr->header;
1778         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1779         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1780         atomic_inc(&hdr->refcnt);
1781         ret = nfs_generic_pagein(desc, hdr);
1782         if (ret != 0) {
1783                 pnfs_put_lseg(desc->pg_lseg);
1784                 desc->pg_lseg = NULL;
1785         } else
1786                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1787         if (atomic_dec_and_test(&hdr->refcnt))
1788                 hdr->completion_ops->completion(hdr);
1789         return ret;
1790 }
1791 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1792
1793 /*
1794  * There can be multiple RW segments.
1795  */
1796 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1797 {
1798         struct pnfs_layout_segment *lseg;
1799
1800         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1801                 if (lseg->pls_range.iomode == IOMODE_RW &&
1802                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1803                         list_add(&lseg->pls_lc_list, listp);
1804         }
1805 }
1806
1807 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1808 {
1809         struct pnfs_layout_segment *lseg, *tmp;
1810         unsigned long *bitlock = &NFS_I(inode)->flags;
1811
1812         /* Matched by references in pnfs_set_layoutcommit */
1813         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1814                 list_del_init(&lseg->pls_lc_list);
1815                 pnfs_put_lseg(lseg);
1816         }
1817
1818         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1819         smp_mb__after_clear_bit();
1820         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1821 }
1822
1823 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1824 {
1825         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1826 }
1827 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1828
1829 void
1830 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1831 {
1832         struct nfs_pgio_header *hdr = wdata->header;
1833         struct inode *inode = hdr->inode;
1834         struct nfs_inode *nfsi = NFS_I(inode);
1835         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1836         bool mark_as_dirty = false;
1837
1838         spin_lock(&inode->i_lock);
1839         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1840                 mark_as_dirty = true;
1841                 dprintk("%s: Set layoutcommit for inode %lu ",
1842                         __func__, inode->i_ino);
1843         }
1844         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1845                 /* references matched in nfs4_layoutcommit_release */
1846                 pnfs_get_lseg(hdr->lseg);
1847         }
1848         if (end_pos > nfsi->layout->plh_lwb)
1849                 nfsi->layout->plh_lwb = end_pos;
1850         spin_unlock(&inode->i_lock);
1851         dprintk("%s: lseg %p end_pos %llu\n",
1852                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1853
1854         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1855          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1856         if (mark_as_dirty)
1857                 mark_inode_dirty_sync(inode);
1858 }
1859 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1860
1861 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1862 {
1863         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1864
1865         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1866                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1867         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1868 }
1869
1870 /*
1871  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1872  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1873  * data to disk to allow the server to recover the data if it crashes.
1874  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1875  * is off, and a COMMIT is sent to a data server, or
1876  * if WRITEs to a data server return NFS_DATA_SYNC.
1877  */
1878 int
1879 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1880 {
1881         struct nfs4_layoutcommit_data *data;
1882         struct nfs_inode *nfsi = NFS_I(inode);
1883         loff_t end_pos;
1884         int status = 0;
1885
1886         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1887
1888         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1889                 return 0;
1890
1891         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1892         data = kzalloc(sizeof(*data), GFP_NOFS);
1893         if (!data) {
1894                 status = -ENOMEM;
1895                 goto out;
1896         }
1897
1898         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1899                 goto out_free;
1900
1901         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1902                 if (!sync) {
1903                         status = -EAGAIN;
1904                         goto out_free;
1905                 }
1906                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1907                                         nfs_wait_bit_killable, TASK_KILLABLE);
1908                 if (status)
1909                         goto out_free;
1910         }
1911
1912         INIT_LIST_HEAD(&data->lseg_list);
1913         spin_lock(&inode->i_lock);
1914         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1915                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1916                 spin_unlock(&inode->i_lock);
1917                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1918                 goto out_free;
1919         }
1920
1921         pnfs_list_write_lseg(inode, &data->lseg_list);
1922
1923         end_pos = nfsi->layout->plh_lwb;
1924         nfsi->layout->plh_lwb = 0;
1925
1926         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1927         spin_unlock(&inode->i_lock);
1928
1929         data->args.inode = inode;
1930         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1931         nfs_fattr_init(&data->fattr);
1932         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1933         data->res.fattr = &data->fattr;
1934         data->args.lastbytewritten = end_pos - 1;
1935         data->res.server = NFS_SERVER(inode);
1936
1937         status = nfs4_proc_layoutcommit(data, sync);
1938 out:
1939         if (status)
1940                 mark_inode_dirty_sync(inode);
1941         dprintk("<-- %s status %d\n", __func__, status);
1942         return status;
1943 out_free:
1944         kfree(data);
1945         goto out;
1946 }
1947
1948 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1949 {
1950         struct nfs4_threshold *thp;
1951
1952         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1953         if (!thp) {
1954                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1955                 return NULL;
1956         }
1957         return thp;
1958 }