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 void pnfs_put_lseg_async_work(struct work_struct *work)
 365 {
 366         struct pnfs_layout_segment *lseg;
 367
 368         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 369
 370         pnfs_put_lseg(lseg);
 371 }
 372
 373 void
 374 pnfs_put_lseg_async(struct pnfs_layout_segment *lseg)
 375 {
 376         INIT_WORK(&lseg->pls_work, pnfs_put_lseg_async_work);
 377         schedule_work(&lseg->pls_work);
 378 }
 379 EXPORT_SYMBOL_GPL(pnfs_put_lseg_async);
 380
 381 static u64
 382 end_offset(u64 start, u64 len)
 383 {
 384         u64 end;
 385
 386         end = start + len;
 387         return end >= start ? end : NFS4_MAX_UINT64;
 388 }
 389
 390 /*
 391  * is l2 fully contained in l1?
 392  *   start1                             end1
 393  *   [----------------------------------)
 394  *           start2           end2
 395  *           [----------------)
 396  */
 397 static bool
 398 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 399                  const struct pnfs_layout_range *l2)
 400 {
 401         u64 start1 = l1->offset;
 402         u64 end1 = end_offset(start1, l1->length);
 403         u64 start2 = l2->offset;
 404         u64 end2 = end_offset(start2, l2->length);
 405
 406         return (start1 <= start2) && (end1 >= end2);
 407 }
 408
 409 /*
 410  * is l1 and l2 intersecting?
 411  *   start1                             end1
 412  *   [----------------------------------)
 413  *                              start2           end2
 414  *                              [----------------)
 415  */
 416 static bool
 417 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 418                     const struct pnfs_layout_range *l2)
 419 {
 420         u64 start1 = l1->offset;
 421         u64 end1 = end_offset(start1, l1->length);
 422         u64 start2 = l2->offset;
 423         u64 end2 = end_offset(start2, l2->length);
 424
 425         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 426                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 427 }
 428
 429 static bool
 430 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 431                  const struct pnfs_layout_range *recall_range)
 432 {
 433         return (recall_range->iomode == IOMODE_ANY ||
 434                 lseg_range->iomode == recall_range->iomode) &&
 435                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 436 }
 437
 438 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 439                 struct list_head *tmp_list)
 440 {
 441         if (!atomic_dec_and_test(&lseg->pls_refcount))
 442                 return false;
 443         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 444         list_add(&lseg->pls_list, tmp_list);
 445         return true;
 446 }
 447
 448 /* Returns 1 if lseg is removed from list, 0 otherwise */
 449 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 450                              struct list_head *tmp_list)
 451 {
 452         int rv = 0;
 453
 454         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 455                 /* Remove the reference keeping the lseg in the
 456                  * list.  It will now be removed when all
 457                  * outstanding io is finished.
 458                  */
 459                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 460                         atomic_read(&lseg->pls_refcount));
 461                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 462                         rv = 1;
 463         }
 464         return rv;
 465 }
 466
 467 /* Returns count of number of matching invalid lsegs remaining in list
 468  * after call.
 469  */
 470 int
 471 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 472                             struct list_head *tmp_list,
 473                             struct pnfs_layout_range *recall_range)
 474 {
 475         struct pnfs_layout_segment *lseg, *next;
 476         int invalid = 0, removed = 0;
 477
 478         dprintk("%s:Begin lo %p\n", __func__, lo);
 479
 480         if (list_empty(&lo->plh_segs))
 481                 return 0;
 482         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 483                 if (!recall_range ||
 484                     should_free_lseg(&lseg->pls_range, recall_range)) {
 485                         dprintk("%s: freeing lseg %p iomode %d "
 486                                 "offset %llu length %llu\n", __func__,
 487                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 488                                 lseg->pls_range.length);
 489                         invalid++;
 490                         removed += mark_lseg_invalid(lseg, tmp_list);
 491                 }
 492         dprintk("%s:Return %i\n", __func__, invalid - removed);
 493         return invalid - removed;
 494 }
 495
 496 /* note free_me must contain lsegs from a single layout_hdr */
 497 void
 498 pnfs_free_lseg_list(struct list_head *free_me)
 499 {
 500         struct pnfs_layout_segment *lseg, *tmp;
 501
 502         if (list_empty(free_me))
 503                 return;
 504
 505         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 506                 list_del(&lseg->pls_list);
 507                 pnfs_free_lseg(lseg);
 508         }
 509 }
 510
 511 void
 512 pnfs_destroy_layout(struct nfs_inode *nfsi)
 513 {
 514         struct pnfs_layout_hdr *lo;
 515         LIST_HEAD(tmp_list);
 516
 517         spin_lock(&nfsi->vfs_inode.i_lock);
 518         lo = nfsi->layout;
 519         if (lo) {
 520                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 521                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 522                 pnfs_get_layout_hdr(lo);
 523                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 524                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 525                 spin_unlock(&nfsi->vfs_inode.i_lock);
 526                 pnfs_free_lseg_list(&tmp_list);
 527                 pnfs_put_layout_hdr(lo);
 528         } else
 529                 spin_unlock(&nfsi->vfs_inode.i_lock);
 530 }
 531 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 532
 533 static bool
 534 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 535                 struct list_head *layout_list)
 536 {
 537         struct pnfs_layout_hdr *lo;
 538         bool ret = false;
 539
 540         spin_lock(&inode->i_lock);
 541         lo = NFS_I(inode)->layout;
 542         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 543                 pnfs_get_layout_hdr(lo);
 544                 list_add(&lo->plh_bulk_destroy, layout_list);
 545                 ret = true;
 546         }
 547         spin_unlock(&inode->i_lock);
 548         return ret;
 549 }
 550
 551 /* Caller must hold rcu_read_lock and clp->cl_lock */
 552 static int
 553 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 554                 struct nfs_server *server,
 555                 struct list_head *layout_list)
 556 {
 557         struct pnfs_layout_hdr *lo, *next;
 558         struct inode *inode;
 559
 560         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 561                 inode = igrab(lo->plh_inode);
 562                 if (inode == NULL)
 563                         continue;
 564                 list_del_init(&lo->plh_layouts);
 565                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 566                         continue;
 567                 rcu_read_unlock();
 568                 spin_unlock(&clp->cl_lock);
 569                 iput(inode);
 570                 spin_lock(&clp->cl_lock);
 571                 rcu_read_lock();
 572                 return -EAGAIN;
 573         }
 574         return 0;
 575 }
 576
 577 static int
 578 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 579                 bool is_bulk_recall)
 580 {
 581         struct pnfs_layout_hdr *lo;
 582         struct inode *inode;
 583         struct pnfs_layout_range range = {
 584                 .iomode = IOMODE_ANY,
 585                 .offset = 0,
 586                 .length = NFS4_MAX_UINT64,
 587         };
 588         LIST_HEAD(lseg_list);
 589         int ret = 0;
 590
 591         while (!list_empty(layout_list)) {
 592                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 593                                 plh_bulk_destroy);
 594                 dprintk("%s freeing layout for inode %lu\n", __func__,
 595                         lo->plh_inode->i_ino);
 596                 inode = lo->plh_inode;
 597
 598                 pnfs_layoutcommit_inode(inode, false);
 599
 600                 spin_lock(&inode->i_lock);
 601                 list_del_init(&lo->plh_bulk_destroy);
 602                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 603                 if (is_bulk_recall)
 604                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 605                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 606                         ret = -EAGAIN;
 607                 spin_unlock(&inode->i_lock);
 608                 pnfs_free_lseg_list(&lseg_list);
 609                 pnfs_put_layout_hdr(lo);
 610                 iput(inode);
 611         }
 612         return ret;
 613 }
 614
 615 int
 616 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 617                 struct nfs_fsid *fsid,
 618                 bool is_recall)
 619 {
 620         struct nfs_server *server;
 621         LIST_HEAD(layout_list);
 622
 623         spin_lock(&clp->cl_lock);
 624         rcu_read_lock();
 625 restart:
 626         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 627                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 628                         continue;
 629                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 630                                 server,
 631                                 &layout_list) != 0)
 632                         goto restart;
 633         }
 634         rcu_read_unlock();
 635         spin_unlock(&clp->cl_lock);
 636
 637         if (list_empty(&layout_list))
 638                 return 0;
 639         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 640 }
 641
 642 int
 643 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 644                 bool is_recall)
 645 {
 646         struct nfs_server *server;
 647         LIST_HEAD(layout_list);
 648
 649         spin_lock(&clp->cl_lock);
 650         rcu_read_lock();
 651 restart:
 652         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 653                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 654                                         server,
 655                                         &layout_list) != 0)
 656                         goto restart;
 657         }
 658         rcu_read_unlock();
 659         spin_unlock(&clp->cl_lock);
 660
 661         if (list_empty(&layout_list))
 662                 return 0;
 663         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 664 }
 665
 666 /*
 667  * Called by the state manger to remove all layouts established under an
 668  * expired lease.
 669  */
 670 void
 671 pnfs_destroy_all_layouts(struct nfs_client *clp)
 672 {
 673         nfs4_deviceid_mark_client_invalid(clp);
 674         nfs4_deviceid_purge_client(clp);
 675
 676         pnfs_destroy_layouts_byclid(clp, false);
 677 }
 678
 679 /*
 680  * Compare 2 layout stateid sequence ids, to see which is newer,
 681  * taking into account wraparound issues.
 682  */
 683 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 684 {
 685         return (s32)(s1 - s2) > 0;
 686 }
 687
 688 /* update lo->plh_stateid with new if is more recent */
 689 void
 690 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 691                         bool update_barrier)
 692 {
 693         u32 oldseq, newseq, new_barrier;
 694         int empty = list_empty(&lo->plh_segs);
 695
 696         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 697         newseq = be32_to_cpu(new->seqid);
 698         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 699                 nfs4_stateid_copy(&lo->plh_stateid, new);
 700                 if (update_barrier) {
 701                         new_barrier = be32_to_cpu(new->seqid);
 702                 } else {
 703                         /* Because of wraparound, we want to keep the barrier
 704                          * "close" to the current seqids.
 705                          */
 706                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 707                 }
 708                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 709                         lo->plh_barrier = new_barrier;
 710         }
 711 }
 712
 713 static bool
 714 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 715                 const nfs4_stateid *stateid)
 716 {
 717         u32 seqid = be32_to_cpu(stateid->seqid);
 718
 719         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 720 }
 721
 722 /* lget is set to 1 if called from inside send_layoutget call chain */
 723 static bool
 724 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 725 {
 726         return lo->plh_block_lgets ||
 727                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 728                 (list_empty(&lo->plh_segs) &&
 729                  (atomic_read(&lo->plh_outstanding) > lget));
 730 }
 731
 732 int
 733 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 734                               struct nfs4_state *open_state)
 735 {
 736         int status = 0;
 737
 738         dprintk("--> %s\n", __func__);
 739         spin_lock(&lo->plh_inode->i_lock);
 740         if (pnfs_layoutgets_blocked(lo, 1)) {
 741                 status = -EAGAIN;
 742         } else if (!nfs4_valid_open_stateid(open_state)) {
 743                 status = -EBADF;
 744         } else if (list_empty(&lo->plh_segs) ||
 745                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 746                 int seq;
 747
 748                 do {
 749                         seq = read_seqbegin(&open_state->seqlock);
 750                         nfs4_stateid_copy(dst, &open_state->stateid);
 751                 } while (read_seqretry(&open_state->seqlock, seq));
 752         } else
 753                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 754         spin_unlock(&lo->plh_inode->i_lock);
 755         dprintk("<-- %s\n", __func__);
 756         return status;
 757 }
 758
 759 /*
 760 * Get layout from server.
 761 *    for now, assume that whole file layouts are requested.
 762 *    arg->offset: 0
 763 *    arg->length: all ones
 764 */
 765 static struct pnfs_layout_segment *
 766 send_layoutget(struct pnfs_layout_hdr *lo,
 767            struct nfs_open_context *ctx,
 768            struct pnfs_layout_range *range,
 769            gfp_t gfp_flags)
 770 {
 771         struct inode *ino = lo->plh_inode;
 772         struct nfs_server *server = NFS_SERVER(ino);
 773         struct nfs4_layoutget *lgp;
 774         struct pnfs_layout_segment *lseg;
 775
 776         dprintk("--> %s\n", __func__);
 777
 778         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 779         if (lgp == NULL)
 780                 return NULL;
 781
 782         lgp->args.minlength = PAGE_CACHE_SIZE;
 783         if (lgp->args.minlength > range->length)
 784                 lgp->args.minlength = range->length;
 785         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 786         lgp->args.range = *range;
 787         lgp->args.type = server->pnfs_curr_ld->id;
 788         lgp->args.inode = ino;
 789         lgp->args.ctx = get_nfs_open_context(ctx);
 790         lgp->gfp_flags = gfp_flags;
 791         lgp->cred = lo->plh_lc_cred;
 792
 793         /* Synchronously retrieve layout information from server and
 794          * store in lseg.
 795          */
 796         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 797         if (IS_ERR(lseg)) {
 798                 switch (PTR_ERR(lseg)) {
 799                 case -ENOMEM:
 800                 case -ERESTARTSYS:
 801                         break;
 802                 default:
 803                         /* remember that LAYOUTGET failed and suspend trying */
 804                         pnfs_layout_io_set_failed(lo, range->iomode);
 805                 }
 806                 return NULL;
 807         }
 808
 809         return lseg;
 810 }
 811
 812 static void pnfs_clear_layoutcommit(struct inode *inode,
 813                 struct list_head *head)
 814 {
 815         struct nfs_inode *nfsi = NFS_I(inode);
 816         struct pnfs_layout_segment *lseg, *tmp;
 817
 818         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 819                 return;
 820         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 821                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 822                         continue;
 823                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 824         }
 825 }
 826
 827 /*
 828  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 829  * when the layout segment list is empty.
 830  *
 831  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 832  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 833  * deviceid is marked invalid.
 834  */
 835 int
 836 _pnfs_return_layout(struct inode *ino)
 837 {
 838         struct pnfs_layout_hdr *lo = NULL;
 839         struct nfs_inode *nfsi = NFS_I(ino);
 840         LIST_HEAD(tmp_list);
 841         struct nfs4_layoutreturn *lrp;
 842         nfs4_stateid stateid;
 843         int status = 0, empty;
 844
 845         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 846
 847         spin_lock(&ino->i_lock);
 848         lo = nfsi->layout;
 849         if (!lo) {
 850                 spin_unlock(&ino->i_lock);
 851                 dprintk("NFS: %s no layout to return\n", __func__);
 852                 goto out;
 853         }
 854         stateid = nfsi->layout->plh_stateid;
 855         /* Reference matched in nfs4_layoutreturn_release */
 856         pnfs_get_layout_hdr(lo);
 857         empty = list_empty(&lo->plh_segs);
 858         pnfs_clear_layoutcommit(ino, &tmp_list);
 859         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 860         /* Don't send a LAYOUTRETURN if list was initially empty */
 861         if (empty) {
 862                 spin_unlock(&ino->i_lock);
 863                 pnfs_put_layout_hdr(lo);
 864                 dprintk("NFS: %s no layout segments to return\n", __func__);
 865                 goto out;
 866         }
 867
 868         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 869         lo->plh_block_lgets++;
 870         spin_unlock(&ino->i_lock);
 871         pnfs_free_lseg_list(&tmp_list);
 872
 873         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 874         if (unlikely(lrp == NULL)) {
 875                 status = -ENOMEM;
 876                 spin_lock(&ino->i_lock);
 877                 lo->plh_block_lgets--;
 878                 spin_unlock(&ino->i_lock);
 879                 pnfs_put_layout_hdr(lo);
 880                 goto out;
 881         }
 882
 883         lrp->args.stateid = stateid;
 884         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 885         lrp->args.inode = ino;
 886         lrp->args.layout = lo;
 887         lrp->clp = NFS_SERVER(ino)->nfs_client;
 888         lrp->cred = lo->plh_lc_cred;
 889
 890         status = nfs4_proc_layoutreturn(lrp);
 891 out:
 892         dprintk("<-- %s status: %d\n", __func__, status);
 893         return status;
 894 }
 895 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 896
 897 int
 898 pnfs_commit_and_return_layout(struct inode *inode)
 899 {
 900         struct pnfs_layout_hdr *lo;
 901         int ret;
 902
 903         spin_lock(&inode->i_lock);
 904         lo = NFS_I(inode)->layout;
 905         if (lo == NULL) {
 906                 spin_unlock(&inode->i_lock);
 907                 return 0;
 908         }
 909         pnfs_get_layout_hdr(lo);
 910         /* Block new layoutgets and read/write to ds */
 911         lo->plh_block_lgets++;
 912         spin_unlock(&inode->i_lock);
 913         filemap_fdatawait(inode->i_mapping);
 914         ret = pnfs_layoutcommit_inode(inode, true);
 915         if (ret == 0)
 916                 ret = _pnfs_return_layout(inode);
 917         spin_lock(&inode->i_lock);
 918         lo->plh_block_lgets--;
 919         spin_unlock(&inode->i_lock);
 920         pnfs_put_layout_hdr(lo);
 921         return ret;
 922 }
 923
 924 bool pnfs_roc(struct inode *ino)
 925 {
 926         struct pnfs_layout_hdr *lo;
 927         struct pnfs_layout_segment *lseg, *tmp;
 928         LIST_HEAD(tmp_list);
 929         bool found = false;
 930
 931         spin_lock(&ino->i_lock);
 932         lo = NFS_I(ino)->layout;
 933         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 934             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 935                 goto out_nolayout;
 936         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 937                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 938                         mark_lseg_invalid(lseg, &tmp_list);
 939                         found = true;
 940                 }
 941         if (!found)
 942                 goto out_nolayout;
 943         lo->plh_block_lgets++;
 944         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 945         spin_unlock(&ino->i_lock);
 946         pnfs_free_lseg_list(&tmp_list);
 947         return true;
 948
 949 out_nolayout:
 950         spin_unlock(&ino->i_lock);
 951         return false;
 952 }
 953
 954 void pnfs_roc_release(struct inode *ino)
 955 {
 956         struct pnfs_layout_hdr *lo;
 957
 958         spin_lock(&ino->i_lock);
 959         lo = NFS_I(ino)->layout;
 960         lo->plh_block_lgets--;
 961         if (atomic_dec_and_test(&lo->plh_refcount)) {
 962                 pnfs_detach_layout_hdr(lo);
 963                 spin_unlock(&ino->i_lock);
 964                 pnfs_free_layout_hdr(lo);
 965         } else
 966                 spin_unlock(&ino->i_lock);
 967 }
 968
 969 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 970 {
 971         struct pnfs_layout_hdr *lo;
 972
 973         spin_lock(&ino->i_lock);
 974         lo = NFS_I(ino)->layout;
 975         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 976                 lo->plh_barrier = barrier;
 977         spin_unlock(&ino->i_lock);
 978 }
 979
 980 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 981 {
 982         struct nfs_inode *nfsi = NFS_I(ino);
 983         struct pnfs_layout_hdr *lo;
 984         struct pnfs_layout_segment *lseg;
 985         u32 current_seqid;
 986         bool found = false;
 987
 988         spin_lock(&ino->i_lock);
 989         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 990                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 991                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 992                         found = true;
 993                         goto out;
 994                 }
 995         lo = nfsi->layout;
 996         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 997
 998         /* Since close does not return a layout stateid for use as
 999          * a barrier, we choose the worst-case barrier.
1000          */
1001         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1002 out:
1003         spin_unlock(&ino->i_lock);
1004         return found;
1005 }
1006
1007 /*
1008  * Compare two layout segments for sorting into layout cache.
1009  * We want to preferentially return RW over RO layouts, so ensure those
1010  * are seen first.
1011  */
1012 static s64
1013 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1014            const struct pnfs_layout_range *l2)
1015 {
1016         s64 d;
1017
1018         /* high offset > low offset */
1019         d = l1->offset - l2->offset;
1020         if (d)
1021                 return d;
1022
1023         /* short length > long length */
1024         d = l2->length - l1->length;
1025         if (d)
1026                 return d;
1027
1028         /* read > read/write */
1029         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1030 }
1031
1032 static void
1033 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1034                    struct pnfs_layout_segment *lseg)
1035 {
1036         struct pnfs_layout_segment *lp;
1037
1038         dprintk("%s:Begin\n", __func__);
1039
1040         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1041                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1042                         continue;
1043                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1044                 dprintk("%s: inserted lseg %p "
1045                         "iomode %d offset %llu length %llu before "
1046                         "lp %p iomode %d offset %llu length %llu\n",
1047                         __func__, lseg, lseg->pls_range.iomode,
1048                         lseg->pls_range.offset, lseg->pls_range.length,
1049                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1050                         lp->pls_range.length);
1051                 goto out;
1052         }
1053         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1054         dprintk("%s: inserted lseg %p "
1055                 "iomode %d offset %llu length %llu at tail\n",
1056                 __func__, lseg, lseg->pls_range.iomode,
1057                 lseg->pls_range.offset, lseg->pls_range.length);
1058 out:
1059         pnfs_get_layout_hdr(lo);
1060
1061         dprintk("%s:Return\n", __func__);
1062 }
1063
1064 static struct pnfs_layout_hdr *
1065 alloc_init_layout_hdr(struct inode *ino,
1066                       struct nfs_open_context *ctx,
1067                       gfp_t gfp_flags)
1068 {
1069         struct pnfs_layout_hdr *lo;
1070
1071         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1072         if (!lo)
1073                 return NULL;
1074         atomic_set(&lo->plh_refcount, 1);
1075         INIT_LIST_HEAD(&lo->plh_layouts);
1076         INIT_LIST_HEAD(&lo->plh_segs);
1077         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1078         lo->plh_inode = ino;
1079         lo->plh_lc_cred = get_rpccred(ctx->cred);
1080         return lo;
1081 }
1082
1083 static struct pnfs_layout_hdr *
1084 pnfs_find_alloc_layout(struct inode *ino,
1085                        struct nfs_open_context *ctx,
1086                        gfp_t gfp_flags)
1087 {
1088         struct nfs_inode *nfsi = NFS_I(ino);
1089         struct pnfs_layout_hdr *new = NULL;
1090
1091         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1092
1093         if (nfsi->layout != NULL)
1094                 goto out_existing;
1095         spin_unlock(&ino->i_lock);
1096         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1097         spin_lock(&ino->i_lock);
1098
1099         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1100                 nfsi->layout = new;
1101                 return new;
1102         } else if (new != NULL)
1103                 pnfs_free_layout_hdr(new);
1104 out_existing:
1105         pnfs_get_layout_hdr(nfsi->layout);
1106         return nfsi->layout;
1107 }
1108
1109 /*
1110  * iomode matching rules:
1111  * iomode       lseg    match
1112  * -----        -----   -----
1113  * ANY          READ    true
1114  * ANY          RW      true
1115  * RW           READ    false
1116  * RW           RW      true
1117  * READ         READ    true
1118  * READ         RW      true
1119  */
1120 static bool
1121 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1122                  const struct pnfs_layout_range *range)
1123 {
1124         struct pnfs_layout_range range1;
1125
1126         if ((range->iomode == IOMODE_RW &&
1127              ls_range->iomode != IOMODE_RW) ||
1128             !pnfs_lseg_range_intersecting(ls_range, range))
1129                 return 0;
1130
1131         /* range1 covers only the first byte in the range */
1132         range1 = *range;
1133         range1.length = 1;
1134         return pnfs_lseg_range_contained(ls_range, &range1);
1135 }
1136
1137 /*
1138  * lookup range in layout
1139  */
1140 static struct pnfs_layout_segment *
1141 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1142                 struct pnfs_layout_range *range)
1143 {
1144         struct pnfs_layout_segment *lseg, *ret = NULL;
1145
1146         dprintk("%s:Begin\n", __func__);
1147
1148         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1149                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1150                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1151                         ret = pnfs_get_lseg(lseg);
1152                         break;
1153                 }
1154                 if (lseg->pls_range.offset > range->offset)
1155                         break;
1156         }
1157
1158         dprintk("%s:Return lseg %p ref %d\n",
1159                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1160         return ret;
1161 }
1162
1163 /*
1164  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1165  * to the MDS or over pNFS
1166  *
1167  * The nfs_inode read_io and write_io fields are cumulative counters reset
1168  * when there are no layout segments. Note that in pnfs_update_layout iomode
1169  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1170  * WRITE request.
1171  *
1172  * A return of true means use MDS I/O.
1173  *
1174  * From rfc 5661:
1175  * If a file's size is smaller than the file size threshold, data accesses
1176  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1177  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1178  * server.  If both file size and I/O size are provided, the client SHOULD
1179  * reach or exceed  both thresholds before sending its read or write
1180  * requests to the data server.
1181  */
1182 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1183                                      struct inode *ino, int iomode)
1184 {
1185         struct nfs4_threshold *t = ctx->mdsthreshold;
1186         struct nfs_inode *nfsi = NFS_I(ino);
1187         loff_t fsize = i_size_read(ino);
1188         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1189
1190         if (t == NULL)
1191                 return ret;
1192
1193         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1194                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1195
1196         switch (iomode) {
1197         case IOMODE_READ:
1198                 if (t->bm & THRESHOLD_RD) {
1199                         dprintk("%s fsize %llu\n", __func__, fsize);
1200                         size_set = true;
1201                         if (fsize < t->rd_sz)
1202                                 size = true;
1203                 }
1204                 if (t->bm & THRESHOLD_RD_IO) {
1205                         dprintk("%s nfsi->read_io %llu\n", __func__,
1206                                 nfsi->read_io);
1207                         io_set = true;
1208                         if (nfsi->read_io < t->rd_io_sz)
1209                                 io = true;
1210                 }
1211                 break;
1212         case IOMODE_RW:
1213                 if (t->bm & THRESHOLD_WR) {
1214                         dprintk("%s fsize %llu\n", __func__, fsize);
1215                         size_set = true;
1216                         if (fsize < t->wr_sz)
1217                                 size = true;
1218                 }
1219                 if (t->bm & THRESHOLD_WR_IO) {
1220                         dprintk("%s nfsi->write_io %llu\n", __func__,
1221                                 nfsi->write_io);
1222                         io_set = true;
1223                         if (nfsi->write_io < t->wr_io_sz)
1224                                 io = true;
1225                 }
1226                 break;
1227         }
1228         if (size_set && io_set) {
1229                 if (size && io)
1230                         ret = true;
1231         } else if (size || io)
1232                 ret = true;
1233
1234         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1235         return ret;
1236 }
1237
1238 /*
1239  * Layout segment is retreived from the server if not cached.
1240  * The appropriate layout segment is referenced and returned to the caller.
1241  */
1242 struct pnfs_layout_segment *
1243 pnfs_update_layout(struct inode *ino,
1244                    struct nfs_open_context *ctx,
1245                    loff_t pos,
1246                    u64 count,
1247                    enum pnfs_iomode iomode,
1248                    gfp_t gfp_flags)
1249 {
1250         struct pnfs_layout_range arg = {
1251                 .iomode = iomode,
1252                 .offset = pos,
1253                 .length = count,
1254         };
1255         unsigned pg_offset;
1256         struct nfs_server *server = NFS_SERVER(ino);
1257         struct nfs_client *clp = server->nfs_client;
1258         struct pnfs_layout_hdr *lo;
1259         struct pnfs_layout_segment *lseg = NULL;
1260         bool first;
1261
1262         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1263                 goto out;
1264
1265         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1266                 goto out;
1267
1268         spin_lock(&ino->i_lock);
1269         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1270         if (lo == NULL) {
1271                 spin_unlock(&ino->i_lock);
1272                 goto out;
1273         }
1274
1275         /* Do we even need to bother with this? */
1276         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1277                 dprintk("%s matches recall, use MDS\n", __func__);
1278                 goto out_unlock;
1279         }
1280
1281         /* if LAYOUTGET already failed once we don't try again */
1282         if (pnfs_layout_io_test_failed(lo, iomode))
1283                 goto out_unlock;
1284
1285         /* Check to see if the layout for the given range already exists */
1286         lseg = pnfs_find_lseg(lo, &arg);
1287         if (lseg)
1288                 goto out_unlock;
1289
1290         if (pnfs_layoutgets_blocked(lo, 0))
1291                 goto out_unlock;
1292         atomic_inc(&lo->plh_outstanding);
1293
1294         first = list_empty(&lo->plh_layouts) ? true : false;
1295         spin_unlock(&ino->i_lock);
1296
1297         if (first) {
1298                 /* The lo must be on the clp list if there is any
1299                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1300                  */
1301                 spin_lock(&clp->cl_lock);
1302                 list_add_tail(&lo->plh_layouts, &server->layouts);
1303                 spin_unlock(&clp->cl_lock);
1304         }
1305
1306         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1307         if (pg_offset) {
1308                 arg.offset -= pg_offset;
1309                 arg.length += pg_offset;
1310         }
1311         if (arg.length != NFS4_MAX_UINT64)
1312                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1313
1314         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1315         atomic_dec(&lo->plh_outstanding);
1316 out_put_layout_hdr:
1317         pnfs_put_layout_hdr(lo);
1318 out:
1319         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1320                         "(%s, offset: %llu, length: %llu)\n",
1321                         __func__, ino->i_sb->s_id,
1322                         (unsigned long long)NFS_FILEID(ino),
1323                         lseg == NULL ? "not found" : "found",
1324                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1325                         (unsigned long long)pos,
1326                         (unsigned long long)count);
1327         return lseg;
1328 out_unlock:
1329         spin_unlock(&ino->i_lock);
1330         goto out_put_layout_hdr;
1331 }
1332 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1333
1334 struct pnfs_layout_segment *
1335 pnfs_layout_process(struct nfs4_layoutget *lgp)
1336 {
1337         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1338         struct nfs4_layoutget_res *res = &lgp->res;
1339         struct pnfs_layout_segment *lseg;
1340         struct inode *ino = lo->plh_inode;
1341         LIST_HEAD(free_me);
1342         int status = 0;
1343
1344         /* Inject layout blob into I/O device driver */
1345         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1346         if (!lseg || IS_ERR(lseg)) {
1347                 if (!lseg)
1348                         status = -ENOMEM;
1349                 else
1350                         status = PTR_ERR(lseg);
1351                 dprintk("%s: Could not allocate layout: error %d\n",
1352                        __func__, status);
1353                 goto out;
1354         }
1355
1356         init_lseg(lo, lseg);
1357         lseg->pls_range = res->range;
1358
1359         spin_lock(&ino->i_lock);
1360         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1361                 dprintk("%s forget reply due to recall\n", __func__);
1362                 goto out_forget_reply;
1363         }
1364
1365         if (pnfs_layoutgets_blocked(lo, 1)) {
1366                 dprintk("%s forget reply due to state\n", __func__);
1367                 goto out_forget_reply;
1368         }
1369
1370         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1371                 /* existing state ID, make sure the sequence number matches. */
1372                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1373                         dprintk("%s forget reply due to sequence\n", __func__);
1374                         goto out_forget_reply;
1375                 }
1376                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1377         } else {
1378                 /*
1379                  * We got an entirely new state ID.  Mark all segments for the
1380                  * inode invalid, and don't bother validating the stateid
1381                  * sequence number.
1382                  */
1383                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1384
1385                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1386                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1387         }
1388
1389         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1390
1391         pnfs_get_lseg(lseg);
1392         pnfs_layout_insert_lseg(lo, lseg);
1393
1394         if (res->return_on_close) {
1395                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1396                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1397         }
1398
1399         spin_unlock(&ino->i_lock);
1400         pnfs_free_lseg_list(&free_me);
1401         return lseg;
1402 out:
1403         return ERR_PTR(status);
1404
1405 out_forget_reply:
1406         spin_unlock(&ino->i_lock);
1407         lseg->pls_layout = lo;
1408         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1409         goto out;
1410 }
1411
1412 void
1413 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1414 {
1415         u64 rd_size = req->wb_bytes;
1416
1417         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1418
1419         if (pgio->pg_dreq == NULL)
1420                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1421         else
1422                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1423
1424         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1425                                            req->wb_context,
1426                                            req_offset(req),
1427                                            rd_size,
1428                                            IOMODE_READ,
1429                                            GFP_KERNEL);
1430         /* If no lseg, fall back to read through mds */
1431         if (pgio->pg_lseg == NULL)
1432                 nfs_pageio_reset_read_mds(pgio);
1433
1434 }
1435 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1436
1437 void
1438 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1439                            struct nfs_page *req, u64 wb_size)
1440 {
1441         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1442
1443         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1444                                            req->wb_context,
1445                                            req_offset(req),
1446                                            wb_size,
1447                                            IOMODE_RW,
1448                                            GFP_NOFS);
1449         /* If no lseg, fall back to write through mds */
1450         if (pgio->pg_lseg == NULL)
1451                 nfs_pageio_reset_write_mds(pgio);
1452 }
1453 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1454
1455 /*
1456  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1457  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1458  */
1459 size_t
1460 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1461                      struct nfs_page *req)
1462 {
1463         unsigned int size;
1464         u64 seg_end, req_start, seg_left;
1465
1466         size = nfs_generic_pg_test(pgio, prev, req);
1467         if (!size)
1468                 return 0;
1469
1470         /*
1471          * 'size' contains the number of bytes left in the current page (up
1472          * to the original size asked for in @req->wb_bytes).
1473          *
1474          * Calculate how many bytes are left in the layout segment
1475          * and if there are less bytes than 'size', return that instead.
1476          *
1477          * Please also note that 'end_offset' is actually the offset of the
1478          * first byte that lies outside the pnfs_layout_range. FIXME?
1479          *
1480          */
1481         if (pgio->pg_lseg) {
1482                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1483                                      pgio->pg_lseg->pls_range.length);
1484                 req_start = req_offset(req);
1485                 WARN_ON_ONCE(req_start > seg_end);
1486                 /* start of request is past the last byte of this segment */
1487                 if (req_start >= seg_end)
1488                         return 0;
1489
1490                 /* adjust 'size' iff there are fewer bytes left in the
1491                  * segment than what nfs_generic_pg_test returned */
1492                 seg_left = seg_end - req_start;
1493                 if (seg_left < size)
1494                         size = (unsigned int)seg_left;
1495         }
1496
1497         return size;
1498 }
1499 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1500
1501 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1502 {
1503         struct nfs_pageio_descriptor pgio;
1504
1505         /* Resend all requests through the MDS */
1506         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1507                               hdr->completion_ops);
1508         return nfs_pageio_resend(&pgio, hdr);
1509 }
1510 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1511
1512 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1513 {
1514
1515         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1516         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1517             PNFS_LAYOUTRET_ON_ERROR) {
1518                 pnfs_return_layout(hdr->inode);
1519         }
1520         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1521                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1522 }
1523
1524 /*
1525  * Called by non rpc-based layout drivers
1526  */
1527 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1528 {
1529         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1530         if (!hdr->pnfs_error) {
1531                 pnfs_set_layoutcommit(hdr);
1532                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1533         } else
1534                 pnfs_ld_handle_write_error(hdr);
1535         hdr->mds_ops->rpc_release(hdr);
1536 }
1537 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1538
1539 static void
1540 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1541                 struct nfs_pgio_header *hdr)
1542 {
1543         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1544                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1545                 nfs_pageio_reset_write_mds(desc);
1546                 desc->pg_recoalesce = 1;
1547         }
1548         nfs_pgio_data_destroy(hdr);
1549 }
1550
1551 static enum pnfs_try_status
1552 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1553                         const struct rpc_call_ops *call_ops,
1554                         struct pnfs_layout_segment *lseg,
1555                         int how)
1556 {
1557         struct inode *inode = hdr->inode;
1558         enum pnfs_try_status trypnfs;
1559         struct nfs_server *nfss = NFS_SERVER(inode);
1560
1561         hdr->mds_ops = call_ops;
1562
1563         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1564                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1565         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1566         if (trypnfs != PNFS_NOT_ATTEMPTED)
1567                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1568         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1569         return trypnfs;
1570 }
1571
1572 static void
1573 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1574               struct nfs_pgio_header *hdr, int how)
1575 {
1576         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1577         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1578         enum pnfs_try_status trypnfs;
1579
1580         desc->pg_lseg = NULL;
1581         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1582         if (trypnfs == PNFS_NOT_ATTEMPTED)
1583                 pnfs_write_through_mds(desc, hdr);
1584         pnfs_put_lseg(lseg);
1585 }
1586
1587 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1588 {
1589         pnfs_put_lseg(hdr->lseg);
1590         nfs_pgio_header_free(hdr);
1591 }
1592 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1593
1594 int
1595 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1596 {
1597         struct nfs_pgio_header *hdr;
1598         int ret;
1599
1600         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1601         if (!hdr) {
1602                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1603                 pnfs_put_lseg(desc->pg_lseg);
1604                 desc->pg_lseg = NULL;
1605                 return -ENOMEM;
1606         }
1607         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1608         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1609         ret = nfs_generic_pgio(desc, hdr);
1610         if (ret != 0) {
1611                 pnfs_put_lseg(desc->pg_lseg);
1612                 desc->pg_lseg = NULL;
1613         } else
1614                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1615         return ret;
1616 }
1617 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1618
1619 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1620 {
1621         struct nfs_pageio_descriptor pgio;
1622
1623         /* Resend all requests through the MDS */
1624         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1625         return nfs_pageio_resend(&pgio, hdr);
1626 }
1627 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1628
1629 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1630 {
1631         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1632         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1633             PNFS_LAYOUTRET_ON_ERROR) {
1634                 pnfs_return_layout(hdr->inode);
1635         }
1636         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1637                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1638 }
1639
1640 /*
1641  * Called by non rpc-based layout drivers
1642  */
1643 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1644 {
1645         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1646         if (likely(!hdr->pnfs_error)) {
1647                 __nfs4_read_done_cb(hdr);
1648                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1649         } else
1650                 pnfs_ld_handle_read_error(hdr);
1651         hdr->mds_ops->rpc_release(hdr);
1652 }
1653 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1654
1655 static void
1656 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1657                 struct nfs_pgio_header *hdr)
1658 {
1659         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1660                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1661                 nfs_pageio_reset_read_mds(desc);
1662                 desc->pg_recoalesce = 1;
1663         }
1664         nfs_pgio_data_destroy(hdr);
1665 }
1666
1667 /*
1668  * Call the appropriate parallel I/O subsystem read function.
1669  */
1670 static enum pnfs_try_status
1671 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1672                        const struct rpc_call_ops *call_ops,
1673                        struct pnfs_layout_segment *lseg)
1674 {
1675         struct inode *inode = hdr->inode;
1676         struct nfs_server *nfss = NFS_SERVER(inode);
1677         enum pnfs_try_status trypnfs;
1678
1679         hdr->mds_ops = call_ops;
1680
1681         dprintk("%s: Reading ino:%lu %u@%llu\n",
1682                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1683
1684         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1685         if (trypnfs != PNFS_NOT_ATTEMPTED)
1686                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1687         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1688         return trypnfs;
1689 }
1690
1691 static void
1692 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1693 {
1694         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1695         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1696         enum pnfs_try_status trypnfs;
1697
1698         desc->pg_lseg = NULL;
1699         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1700         if (trypnfs == PNFS_NOT_ATTEMPTED)
1701                 pnfs_read_through_mds(desc, hdr);
1702         pnfs_put_lseg(lseg);
1703 }
1704
1705 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1706 {
1707         pnfs_put_lseg(hdr->lseg);
1708         nfs_pgio_header_free(hdr);
1709 }
1710 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1711
1712 int
1713 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1714 {
1715         struct nfs_pgio_header *hdr;
1716         int ret;
1717
1718         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1719         if (!hdr) {
1720                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1721                 ret = -ENOMEM;
1722                 pnfs_put_lseg(desc->pg_lseg);
1723                 desc->pg_lseg = NULL;
1724                 return ret;
1725         }
1726         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1727         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1728         ret = nfs_generic_pgio(desc, hdr);
1729         if (ret != 0) {
1730                 pnfs_put_lseg(desc->pg_lseg);
1731                 desc->pg_lseg = NULL;
1732         } else
1733                 pnfs_do_read(desc, hdr);
1734         return ret;
1735 }
1736 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1737
1738 static void pnfs_clear_layoutcommitting(struct inode *inode)
1739 {
1740         unsigned long *bitlock = &NFS_I(inode)->flags;
1741
1742         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1743         smp_mb__after_atomic();
1744         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1745 }
1746
1747 /*
1748  * There can be multiple RW segments.
1749  */
1750 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1751 {
1752         struct pnfs_layout_segment *lseg;
1753
1754         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1755                 if (lseg->pls_range.iomode == IOMODE_RW &&
1756                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1757                         list_add(&lseg->pls_lc_list, listp);
1758         }
1759 }
1760
1761 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1762 {
1763         struct pnfs_layout_segment *lseg, *tmp;
1764
1765         /* Matched by references in pnfs_set_layoutcommit */
1766         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1767                 list_del_init(&lseg->pls_lc_list);
1768                 pnfs_put_lseg(lseg);
1769         }
1770
1771         pnfs_clear_layoutcommitting(inode);
1772 }
1773
1774 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1775 {
1776         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1777 }
1778 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1779
1780 void
1781 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1782 {
1783         struct inode *inode = hdr->inode;
1784         struct nfs_inode *nfsi = NFS_I(inode);
1785         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1786         bool mark_as_dirty = false;
1787
1788         spin_lock(&inode->i_lock);
1789         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1790                 mark_as_dirty = true;
1791                 dprintk("%s: Set layoutcommit for inode %lu ",
1792                         __func__, inode->i_ino);
1793         }
1794         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1795                 /* references matched in nfs4_layoutcommit_release */
1796                 pnfs_get_lseg(hdr->lseg);
1797         }
1798         if (end_pos > nfsi->layout->plh_lwb)
1799                 nfsi->layout->plh_lwb = end_pos;
1800         spin_unlock(&inode->i_lock);
1801         dprintk("%s: lseg %p end_pos %llu\n",
1802                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1803
1804         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1805          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1806         if (mark_as_dirty)
1807                 mark_inode_dirty_sync(inode);
1808 }
1809 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1810
1811 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
1812 {
1813         struct inode *inode = data->inode;
1814         struct nfs_inode *nfsi = NFS_I(inode);
1815         bool mark_as_dirty = false;
1816
1817         spin_lock(&inode->i_lock);
1818         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1819                 mark_as_dirty = true;
1820                 dprintk("%s: Set layoutcommit for inode %lu ",
1821                         __func__, inode->i_ino);
1822         }
1823         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
1824                 /* references matched in nfs4_layoutcommit_release */
1825                 pnfs_get_lseg(data->lseg);
1826         }
1827         if (data->lwb > nfsi->layout->plh_lwb)
1828                 nfsi->layout->plh_lwb = data->lwb;
1829         spin_unlock(&inode->i_lock);
1830         dprintk("%s: lseg %p end_pos %llu\n",
1831                 __func__, data->lseg, nfsi->layout->plh_lwb);
1832
1833         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1834          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1835         if (mark_as_dirty)
1836                 mark_inode_dirty_sync(inode);
1837 }
1838 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
1839
1840 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1841 {
1842         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1843
1844         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1845                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1846         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1847 }
1848
1849 /*
1850  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1851  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1852  * data to disk to allow the server to recover the data if it crashes.
1853  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1854  * is off, and a COMMIT is sent to a data server, or
1855  * if WRITEs to a data server return NFS_DATA_SYNC.
1856  */
1857 int
1858 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1859 {
1860         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1861         struct nfs4_layoutcommit_data *data;
1862         struct nfs_inode *nfsi = NFS_I(inode);
1863         loff_t end_pos;
1864         int status;
1865
1866         if (!pnfs_layoutcommit_outstanding(inode))
1867                 return 0;
1868
1869         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1870
1871         status = -EAGAIN;
1872         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1873                 if (!sync)
1874                         goto out;
1875                 status = wait_on_bit_lock_action(&nfsi->flags,
1876                                 NFS_INO_LAYOUTCOMMITTING,
1877                                 nfs_wait_bit_killable,
1878                                 TASK_KILLABLE);
1879                 if (status)
1880                         goto out;
1881         }
1882
1883         status = -ENOMEM;
1884         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1885         data = kzalloc(sizeof(*data), GFP_NOFS);
1886         if (!data)
1887                 goto clear_layoutcommitting;
1888
1889         status = 0;
1890         spin_lock(&inode->i_lock);
1891         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1892                 goto out_unlock;
1893
1894         INIT_LIST_HEAD(&data->lseg_list);
1895         pnfs_list_write_lseg(inode, &data->lseg_list);
1896
1897         end_pos = nfsi->layout->plh_lwb;
1898         nfsi->layout->plh_lwb = 0;
1899
1900         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1901         spin_unlock(&inode->i_lock);
1902
1903         data->args.inode = inode;
1904         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1905         nfs_fattr_init(&data->fattr);
1906         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1907         data->res.fattr = &data->fattr;
1908         data->args.lastbytewritten = end_pos - 1;
1909         data->res.server = NFS_SERVER(inode);
1910
1911         if (ld->prepare_layoutcommit) {
1912                 status = ld->prepare_layoutcommit(&data->args);
1913                 if (status) {
1914                         spin_lock(&inode->i_lock);
1915                         if (end_pos < nfsi->layout->plh_lwb)
1916                                 nfsi->layout->plh_lwb = end_pos;
1917                         spin_unlock(&inode->i_lock);
1918                         put_rpccred(data->cred);
1919                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
1920                         goto clear_layoutcommitting;
1921                 }
1922         }
1923
1924
1925         status = nfs4_proc_layoutcommit(data, sync);
1926 out:
1927         if (status)
1928                 mark_inode_dirty_sync(inode);
1929         dprintk("<-- %s status %d\n", __func__, status);
1930         return status;
1931 out_unlock:
1932         spin_unlock(&inode->i_lock);
1933         kfree(data);
1934 clear_layoutcommitting:
1935         pnfs_clear_layoutcommitting(inode);
1936         goto out;
1937 }
1938
1939 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1940 {
1941         struct nfs4_threshold *thp;
1942
1943         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1944         if (!thp) {
1945                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1946                 return NULL;
1947         }
1948         return thp;
1949 }