fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <[email protected]>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "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                 spin_lock(&inode->i_lock);
 598                 list_del_init(&lo->plh_bulk_destroy);
 599                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 600                 if (is_bulk_recall)
 601                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 602                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 603                         ret = -EAGAIN;
 604                 spin_unlock(&inode->i_lock);
 605                 pnfs_free_lseg_list(&lseg_list);
 606                 pnfs_put_layout_hdr(lo);
 607                 iput(inode);
 608         }
 609         return ret;
 610 }
 611
 612 int
 613 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 614                 struct nfs_fsid *fsid,
 615                 bool is_recall)
 616 {
 617         struct nfs_server *server;
 618         LIST_HEAD(layout_list);
 619
 620         spin_lock(&clp->cl_lock);
 621         rcu_read_lock();
 622 restart:
 623         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 624                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 625                         continue;
 626                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 627                                 server,
 628                                 &layout_list) != 0)
 629                         goto restart;
 630         }
 631         rcu_read_unlock();
 632         spin_unlock(&clp->cl_lock);
 633
 634         if (list_empty(&layout_list))
 635                 return 0;
 636         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 637 }
 638
 639 int
 640 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 641                 bool is_recall)
 642 {
 643         struct nfs_server *server;
 644         LIST_HEAD(layout_list);
 645
 646         spin_lock(&clp->cl_lock);
 647         rcu_read_lock();
 648 restart:
 649         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 650                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 651                                         server,
 652                                         &layout_list) != 0)
 653                         goto restart;
 654         }
 655         rcu_read_unlock();
 656         spin_unlock(&clp->cl_lock);
 657
 658         if (list_empty(&layout_list))
 659                 return 0;
 660         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 661 }
 662
 663 /*
 664  * Called by the state manger to remove all layouts established under an
 665  * expired lease.
 666  */
 667 void
 668 pnfs_destroy_all_layouts(struct nfs_client *clp)
 669 {
 670         nfs4_deviceid_mark_client_invalid(clp);
 671         nfs4_deviceid_purge_client(clp);
 672
 673         pnfs_destroy_layouts_byclid(clp, false);
 674 }
 675
 676 /*
 677  * Compare 2 layout stateid sequence ids, to see which is newer,
 678  * taking into account wraparound issues.
 679  */
 680 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 681 {
 682         return (s32)(s1 - s2) > 0;
 683 }
 684
 685 static void
 686 pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
 687                 const nfs4_stateid *new,
 688                 struct list_head *free_me_list)
 689 {
 690         if (nfs4_stateid_match_other(&lo->plh_stateid, new))
 691                 return;
 692         /* Layout is new! Kill existing layout segments */
 693         pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
 694 }
 695
 696 /* update lo->plh_stateid with new if is more recent */
 697 void
 698 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 699                         bool update_barrier)
 700 {
 701         u32 oldseq, newseq, new_barrier;
 702         int empty = list_empty(&lo->plh_segs);
 703
 704         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 705         newseq = be32_to_cpu(new->seqid);
 706         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 707                 nfs4_stateid_copy(&lo->plh_stateid, new);
 708                 if (update_barrier) {
 709                         new_barrier = be32_to_cpu(new->seqid);
 710                 } else {
 711                         /* Because of wraparound, we want to keep the barrier
 712                          * "close" to the current seqids.
 713                          */
 714                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 715                 }
 716                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 717                         lo->plh_barrier = new_barrier;
 718         }
 719 }
 720
 721 static bool
 722 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 723                 const nfs4_stateid *stateid)
 724 {
 725         u32 seqid = be32_to_cpu(stateid->seqid);
 726
 727         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 728 }
 729
 730 /* lget is set to 1 if called from inside send_layoutget call chain */
 731 static bool
 732 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 733 {
 734         return lo->plh_block_lgets ||
 735                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 736                 (list_empty(&lo->plh_segs) &&
 737                  (atomic_read(&lo->plh_outstanding) > lget));
 738 }
 739
 740 int
 741 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 742                               struct nfs4_state *open_state)
 743 {
 744         int status = 0;
 745
 746         dprintk("--> %s\n", __func__);
 747         spin_lock(&lo->plh_inode->i_lock);
 748         if (pnfs_layoutgets_blocked(lo, 1)) {
 749                 status = -EAGAIN;
 750         } else if (!nfs4_valid_open_stateid(open_state)) {
 751                 status = -EBADF;
 752         } else if (list_empty(&lo->plh_segs)) {
 753                 int seq;
 754
 755                 do {
 756                         seq = read_seqbegin(&open_state->seqlock);
 757                         nfs4_stateid_copy(dst, &open_state->stateid);
 758                 } while (read_seqretry(&open_state->seqlock, seq));
 759         } else
 760                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 761         spin_unlock(&lo->plh_inode->i_lock);
 762         dprintk("<-- %s\n", __func__);
 763         return status;
 764 }
 765
 766 /*
 767 * Get layout from server.
 768 *    for now, assume that whole file layouts are requested.
 769 *    arg->offset: 0
 770 *    arg->length: all ones
 771 */
 772 static struct pnfs_layout_segment *
 773 send_layoutget(struct pnfs_layout_hdr *lo,
 774            struct nfs_open_context *ctx,
 775            struct pnfs_layout_range *range,
 776            gfp_t gfp_flags)
 777 {
 778         struct inode *ino = lo->plh_inode;
 779         struct nfs_server *server = NFS_SERVER(ino);
 780         struct nfs4_layoutget *lgp;
 781         struct pnfs_layout_segment *lseg;
 782
 783         dprintk("--> %s\n", __func__);
 784
 785         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 786         if (lgp == NULL)
 787                 return NULL;
 788
 789         lgp->args.minlength = PAGE_CACHE_SIZE;
 790         if (lgp->args.minlength > range->length)
 791                 lgp->args.minlength = range->length;
 792         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 793         lgp->args.range = *range;
 794         lgp->args.type = server->pnfs_curr_ld->id;
 795         lgp->args.inode = ino;
 796         lgp->args.ctx = get_nfs_open_context(ctx);
 797         lgp->gfp_flags = gfp_flags;
 798         lgp->cred = lo->plh_lc_cred;
 799
 800         /* Synchronously retrieve layout information from server and
 801          * store in lseg.
 802          */
 803         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 804         if (IS_ERR(lseg)) {
 805                 switch (PTR_ERR(lseg)) {
 806                 case -ENOMEM:
 807                 case -ERESTARTSYS:
 808                         break;
 809                 default:
 810                         /* remember that LAYOUTGET failed and suspend trying */
 811                         pnfs_layout_io_set_failed(lo, range->iomode);
 812                 }
 813                 return NULL;
 814         }
 815
 816         return lseg;
 817 }
 818
 819 static void pnfs_clear_layoutcommit(struct inode *inode,
 820                 struct list_head *head)
 821 {
 822         struct nfs_inode *nfsi = NFS_I(inode);
 823         struct pnfs_layout_segment *lseg, *tmp;
 824
 825         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 826                 return;
 827         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 828                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 829                         continue;
 830                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 831         }
 832 }
 833
 834 /*
 835  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 836  * when the layout segment list is empty.
 837  *
 838  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 839  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 840  * deviceid is marked invalid.
 841  */
 842 int
 843 _pnfs_return_layout(struct inode *ino)
 844 {
 845         struct pnfs_layout_hdr *lo = NULL;
 846         struct nfs_inode *nfsi = NFS_I(ino);
 847         LIST_HEAD(tmp_list);
 848         struct nfs4_layoutreturn *lrp;
 849         nfs4_stateid stateid;
 850         int status = 0, empty;
 851
 852         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 853
 854         spin_lock(&ino->i_lock);
 855         lo = nfsi->layout;
 856         if (!lo) {
 857                 spin_unlock(&ino->i_lock);
 858                 dprintk("NFS: %s no layout to return\n", __func__);
 859                 goto out;
 860         }
 861         stateid = nfsi->layout->plh_stateid;
 862         /* Reference matched in nfs4_layoutreturn_release */
 863         pnfs_get_layout_hdr(lo);
 864         empty = list_empty(&lo->plh_segs);
 865         pnfs_clear_layoutcommit(ino, &tmp_list);
 866         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 867         /* Don't send a LAYOUTRETURN if list was initially empty */
 868         if (empty) {
 869                 spin_unlock(&ino->i_lock);
 870                 pnfs_put_layout_hdr(lo);
 871                 dprintk("NFS: %s no layout segments to return\n", __func__);
 872                 goto out;
 873         }
 874         lo->plh_block_lgets++;
 875         spin_unlock(&ino->i_lock);
 876         pnfs_free_lseg_list(&tmp_list);
 877
 878         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 879         if (unlikely(lrp == NULL)) {
 880                 status = -ENOMEM;
 881                 spin_lock(&ino->i_lock);
 882                 lo->plh_block_lgets--;
 883                 spin_unlock(&ino->i_lock);
 884                 pnfs_put_layout_hdr(lo);
 885                 goto out;
 886         }
 887
 888         lrp->args.stateid = stateid;
 889         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 890         lrp->args.inode = ino;
 891         lrp->args.layout = lo;
 892         lrp->clp = NFS_SERVER(ino)->nfs_client;
 893         lrp->cred = lo->plh_lc_cred;
 894
 895         status = nfs4_proc_layoutreturn(lrp);
 896 out:
 897         dprintk("<-- %s status: %d\n", __func__, status);
 898         return status;
 899 }
 900 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 901
 902 int
 903 pnfs_commit_and_return_layout(struct inode *inode)
 904 {
 905         struct pnfs_layout_hdr *lo;
 906         int ret;
 907
 908         spin_lock(&inode->i_lock);
 909         lo = NFS_I(inode)->layout;
 910         if (lo == NULL) {
 911                 spin_unlock(&inode->i_lock);
 912                 return 0;
 913         }
 914         pnfs_get_layout_hdr(lo);
 915         /* Block new layoutgets and read/write to ds */
 916         lo->plh_block_lgets++;
 917         spin_unlock(&inode->i_lock);
 918         filemap_fdatawait(inode->i_mapping);
 919         ret = pnfs_layoutcommit_inode(inode, true);
 920         if (ret == 0)
 921                 ret = _pnfs_return_layout(inode);
 922         spin_lock(&inode->i_lock);
 923         lo->plh_block_lgets--;
 924         spin_unlock(&inode->i_lock);
 925         pnfs_put_layout_hdr(lo);
 926         return ret;
 927 }
 928
 929 bool pnfs_roc(struct inode *ino)
 930 {
 931         struct pnfs_layout_hdr *lo;
 932         struct pnfs_layout_segment *lseg, *tmp;
 933         LIST_HEAD(tmp_list);
 934         bool found = false;
 935
 936         spin_lock(&ino->i_lock);
 937         lo = NFS_I(ino)->layout;
 938         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 939             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 940                 goto out_nolayout;
 941         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 942                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 943                         mark_lseg_invalid(lseg, &tmp_list);
 944                         found = true;
 945                 }
 946         if (!found)
 947                 goto out_nolayout;
 948         lo->plh_block_lgets++;
 949         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 950         spin_unlock(&ino->i_lock);
 951         pnfs_free_lseg_list(&tmp_list);
 952         return true;
 953
 954 out_nolayout:
 955         spin_unlock(&ino->i_lock);
 956         return false;
 957 }
 958
 959 void pnfs_roc_release(struct inode *ino)
 960 {
 961         struct pnfs_layout_hdr *lo;
 962
 963         spin_lock(&ino->i_lock);
 964         lo = NFS_I(ino)->layout;
 965         lo->plh_block_lgets--;
 966         if (atomic_dec_and_test(&lo->plh_refcount)) {
 967                 pnfs_detach_layout_hdr(lo);
 968                 spin_unlock(&ino->i_lock);
 969                 pnfs_free_layout_hdr(lo);
 970         } else
 971                 spin_unlock(&ino->i_lock);
 972 }
 973
 974 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 975 {
 976         struct pnfs_layout_hdr *lo;
 977
 978         spin_lock(&ino->i_lock);
 979         lo = NFS_I(ino)->layout;
 980         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 981                 lo->plh_barrier = barrier;
 982         spin_unlock(&ino->i_lock);
 983 }
 984
 985 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 986 {
 987         struct nfs_inode *nfsi = NFS_I(ino);
 988         struct pnfs_layout_hdr *lo;
 989         struct pnfs_layout_segment *lseg;
 990         u32 current_seqid;
 991         bool found = false;
 992
 993         spin_lock(&ino->i_lock);
 994         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 995                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 996                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 997                         found = true;
 998                         goto out;
 999                 }
1000         lo = nfsi->layout;
1001         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1002
1003         /* Since close does not return a layout stateid for use as
1004          * a barrier, we choose the worst-case barrier.
1005          */
1006         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1007 out:
1008         spin_unlock(&ino->i_lock);
1009         return found;
1010 }
1011
1012 /*
1013  * Compare two layout segments for sorting into layout cache.
1014  * We want to preferentially return RW over RO layouts, so ensure those
1015  * are seen first.
1016  */
1017 static s64
1018 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1019            const struct pnfs_layout_range *l2)
1020 {
1021         s64 d;
1022
1023         /* high offset > low offset */
1024         d = l1->offset - l2->offset;
1025         if (d)
1026                 return d;
1027
1028         /* short length > long length */
1029         d = l2->length - l1->length;
1030         if (d)
1031                 return d;
1032
1033         /* read > read/write */
1034         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1035 }
1036
1037 static void
1038 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1039                    struct pnfs_layout_segment *lseg)
1040 {
1041         struct pnfs_layout_segment *lp;
1042
1043         dprintk("%s:Begin\n", __func__);
1044
1045         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1046                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1047                         continue;
1048                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1049                 dprintk("%s: inserted lseg %p "
1050                         "iomode %d offset %llu length %llu before "
1051                         "lp %p iomode %d offset %llu length %llu\n",
1052                         __func__, lseg, lseg->pls_range.iomode,
1053                         lseg->pls_range.offset, lseg->pls_range.length,
1054                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1055                         lp->pls_range.length);
1056                 goto out;
1057         }
1058         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1059         dprintk("%s: inserted lseg %p "
1060                 "iomode %d offset %llu length %llu at tail\n",
1061                 __func__, lseg, lseg->pls_range.iomode,
1062                 lseg->pls_range.offset, lseg->pls_range.length);
1063 out:
1064         pnfs_get_layout_hdr(lo);
1065
1066         dprintk("%s:Return\n", __func__);
1067 }
1068
1069 static struct pnfs_layout_hdr *
1070 alloc_init_layout_hdr(struct inode *ino,
1071                       struct nfs_open_context *ctx,
1072                       gfp_t gfp_flags)
1073 {
1074         struct pnfs_layout_hdr *lo;
1075
1076         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1077         if (!lo)
1078                 return NULL;
1079         atomic_set(&lo->plh_refcount, 1);
1080         INIT_LIST_HEAD(&lo->plh_layouts);
1081         INIT_LIST_HEAD(&lo->plh_segs);
1082         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1083         lo->plh_inode = ino;
1084         lo->plh_lc_cred = get_rpccred(ctx->cred);
1085         return lo;
1086 }
1087
1088 static struct pnfs_layout_hdr *
1089 pnfs_find_alloc_layout(struct inode *ino,
1090                        struct nfs_open_context *ctx,
1091                        gfp_t gfp_flags)
1092 {
1093         struct nfs_inode *nfsi = NFS_I(ino);
1094         struct pnfs_layout_hdr *new = NULL;
1095
1096         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1097
1098         if (nfsi->layout != NULL)
1099                 goto out_existing;
1100         spin_unlock(&ino->i_lock);
1101         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1102         spin_lock(&ino->i_lock);
1103
1104         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1105                 nfsi->layout = new;
1106                 return new;
1107         } else if (new != NULL)
1108                 pnfs_free_layout_hdr(new);
1109 out_existing:
1110         pnfs_get_layout_hdr(nfsi->layout);
1111         return nfsi->layout;
1112 }
1113
1114 /*
1115  * iomode matching rules:
1116  * iomode       lseg    match
1117  * -----        -----   -----
1118  * ANY          READ    true
1119  * ANY          RW      true
1120  * RW           READ    false
1121  * RW           RW      true
1122  * READ         READ    true
1123  * READ         RW      true
1124  */
1125 static bool
1126 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1127                  const struct pnfs_layout_range *range)
1128 {
1129         struct pnfs_layout_range range1;
1130
1131         if ((range->iomode == IOMODE_RW &&
1132              ls_range->iomode != IOMODE_RW) ||
1133             !pnfs_lseg_range_intersecting(ls_range, range))
1134                 return 0;
1135
1136         /* range1 covers only the first byte in the range */
1137         range1 = *range;
1138         range1.length = 1;
1139         return pnfs_lseg_range_contained(ls_range, &range1);
1140 }
1141
1142 /*
1143  * lookup range in layout
1144  */
1145 static struct pnfs_layout_segment *
1146 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1147                 struct pnfs_layout_range *range)
1148 {
1149         struct pnfs_layout_segment *lseg, *ret = NULL;
1150
1151         dprintk("%s:Begin\n", __func__);
1152
1153         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1154                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1155                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1156                         ret = pnfs_get_lseg(lseg);
1157                         break;
1158                 }
1159                 if (lseg->pls_range.offset > range->offset)
1160                         break;
1161         }
1162
1163         dprintk("%s:Return lseg %p ref %d\n",
1164                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1165         return ret;
1166 }
1167
1168 /*
1169  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1170  * to the MDS or over pNFS
1171  *
1172  * The nfs_inode read_io and write_io fields are cumulative counters reset
1173  * when there are no layout segments. Note that in pnfs_update_layout iomode
1174  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1175  * WRITE request.
1176  *
1177  * A return of true means use MDS I/O.
1178  *
1179  * From rfc 5661:
1180  * If a file's size is smaller than the file size threshold, data accesses
1181  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1182  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1183  * server.  If both file size and I/O size are provided, the client SHOULD
1184  * reach or exceed  both thresholds before sending its read or write
1185  * requests to the data server.
1186  */
1187 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1188                                      struct inode *ino, int iomode)
1189 {
1190         struct nfs4_threshold *t = ctx->mdsthreshold;
1191         struct nfs_inode *nfsi = NFS_I(ino);
1192         loff_t fsize = i_size_read(ino);
1193         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1194
1195         if (t == NULL)
1196                 return ret;
1197
1198         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1199                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1200
1201         switch (iomode) {
1202         case IOMODE_READ:
1203                 if (t->bm & THRESHOLD_RD) {
1204                         dprintk("%s fsize %llu\n", __func__, fsize);
1205                         size_set = true;
1206                         if (fsize < t->rd_sz)
1207                                 size = true;
1208                 }
1209                 if (t->bm & THRESHOLD_RD_IO) {
1210                         dprintk("%s nfsi->read_io %llu\n", __func__,
1211                                 nfsi->read_io);
1212                         io_set = true;
1213                         if (nfsi->read_io < t->rd_io_sz)
1214                                 io = true;
1215                 }
1216                 break;
1217         case IOMODE_RW:
1218                 if (t->bm & THRESHOLD_WR) {
1219                         dprintk("%s fsize %llu\n", __func__, fsize);
1220                         size_set = true;
1221                         if (fsize < t->wr_sz)
1222                                 size = true;
1223                 }
1224                 if (t->bm & THRESHOLD_WR_IO) {
1225                         dprintk("%s nfsi->write_io %llu\n", __func__,
1226                                 nfsi->write_io);
1227                         io_set = true;
1228                         if (nfsi->write_io < t->wr_io_sz)
1229                                 io = true;
1230                 }
1231                 break;
1232         }
1233         if (size_set && io_set) {
1234                 if (size && io)
1235                         ret = true;
1236         } else if (size || io)
1237                 ret = true;
1238
1239         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1240         return ret;
1241 }
1242
1243 /*
1244  * Layout segment is retreived from the server if not cached.
1245  * The appropriate layout segment is referenced and returned to the caller.
1246  */
1247 struct pnfs_layout_segment *
1248 pnfs_update_layout(struct inode *ino,
1249                    struct nfs_open_context *ctx,
1250                    loff_t pos,
1251                    u64 count,
1252                    enum pnfs_iomode iomode,
1253                    gfp_t gfp_flags)
1254 {
1255         struct pnfs_layout_range arg = {
1256                 .iomode = iomode,
1257                 .offset = pos,
1258                 .length = count,
1259         };
1260         unsigned pg_offset;
1261         struct nfs_server *server = NFS_SERVER(ino);
1262         struct nfs_client *clp = server->nfs_client;
1263         struct pnfs_layout_hdr *lo;
1264         struct pnfs_layout_segment *lseg = NULL;
1265         bool first;
1266
1267         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1268                 goto out;
1269
1270         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1271                 goto out;
1272
1273         spin_lock(&ino->i_lock);
1274         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1275         if (lo == NULL) {
1276                 spin_unlock(&ino->i_lock);
1277                 goto out;
1278         }
1279
1280         /* Do we even need to bother with this? */
1281         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1282                 dprintk("%s matches recall, use MDS\n", __func__);
1283                 goto out_unlock;
1284         }
1285
1286         /* if LAYOUTGET already failed once we don't try again */
1287         if (pnfs_layout_io_test_failed(lo, iomode))
1288                 goto out_unlock;
1289
1290         /* Check to see if the layout for the given range already exists */
1291         lseg = pnfs_find_lseg(lo, &arg);
1292         if (lseg)
1293                 goto out_unlock;
1294
1295         if (pnfs_layoutgets_blocked(lo, 0))
1296                 goto out_unlock;
1297         atomic_inc(&lo->plh_outstanding);
1298
1299         first = list_empty(&lo->plh_layouts) ? true : false;
1300         spin_unlock(&ino->i_lock);
1301
1302         if (first) {
1303                 /* The lo must be on the clp list if there is any
1304                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1305                  */
1306                 spin_lock(&clp->cl_lock);
1307                 list_add_tail(&lo->plh_layouts, &server->layouts);
1308                 spin_unlock(&clp->cl_lock);
1309         }
1310
1311         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1312         if (pg_offset) {
1313                 arg.offset -= pg_offset;
1314                 arg.length += pg_offset;
1315         }
1316         if (arg.length != NFS4_MAX_UINT64)
1317                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1318
1319         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1320         atomic_dec(&lo->plh_outstanding);
1321 out_put_layout_hdr:
1322         pnfs_put_layout_hdr(lo);
1323 out:
1324         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1325                         "(%s, offset: %llu, length: %llu)\n",
1326                         __func__, ino->i_sb->s_id,
1327                         (unsigned long long)NFS_FILEID(ino),
1328                         lseg == NULL ? "not found" : "found",
1329                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1330                         (unsigned long long)pos,
1331                         (unsigned long long)count);
1332         return lseg;
1333 out_unlock:
1334         spin_unlock(&ino->i_lock);
1335         goto out_put_layout_hdr;
1336 }
1337 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1338
1339 struct pnfs_layout_segment *
1340 pnfs_layout_process(struct nfs4_layoutget *lgp)
1341 {
1342         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1343         struct nfs4_layoutget_res *res = &lgp->res;
1344         struct pnfs_layout_segment *lseg;
1345         struct inode *ino = lo->plh_inode;
1346         LIST_HEAD(free_me);
1347         int status = 0;
1348
1349         /* Inject layout blob into I/O device driver */
1350         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1351         if (!lseg || IS_ERR(lseg)) {
1352                 if (!lseg)
1353                         status = -ENOMEM;
1354                 else
1355                         status = PTR_ERR(lseg);
1356                 dprintk("%s: Could not allocate layout: error %d\n",
1357                        __func__, status);
1358                 goto out;
1359         }
1360
1361         spin_lock(&ino->i_lock);
1362         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1363                 dprintk("%s forget reply due to recall\n", __func__);
1364                 goto out_forget_reply;
1365         }
1366
1367         if (pnfs_layoutgets_blocked(lo, 1) ||
1368             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1369                 dprintk("%s forget reply due to state\n", __func__);
1370                 goto out_forget_reply;
1371         }
1372
1373         /* Check that the new stateid matches the old stateid */
1374         pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
1375         /* Done processing layoutget. Set the layout stateid */
1376         pnfs_set_layout_stateid(lo, &res->stateid, false);
1377
1378         init_lseg(lo, lseg);
1379         lseg->pls_range = res->range;
1380         pnfs_get_lseg(lseg);
1381         pnfs_layout_insert_lseg(lo, lseg);
1382
1383         if (res->return_on_close) {
1384                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1385                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1386         }
1387
1388         spin_unlock(&ino->i_lock);
1389         pnfs_free_lseg_list(&free_me);
1390         return lseg;
1391 out:
1392         return ERR_PTR(status);
1393
1394 out_forget_reply:
1395         spin_unlock(&ino->i_lock);
1396         lseg->pls_layout = lo;
1397         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1398         goto out;
1399 }
1400
1401 void
1402 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1403 {
1404         u64 rd_size = req->wb_bytes;
1405
1406         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1407
1408         if (pgio->pg_dreq == NULL)
1409                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1410         else
1411                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1412
1413         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1414                                            req->wb_context,
1415                                            req_offset(req),
1416                                            rd_size,
1417                                            IOMODE_READ,
1418                                            GFP_KERNEL);
1419         /* If no lseg, fall back to read through mds */
1420         if (pgio->pg_lseg == NULL)
1421                 nfs_pageio_reset_read_mds(pgio);
1422
1423 }
1424 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1425
1426 void
1427 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1428                            struct nfs_page *req, u64 wb_size)
1429 {
1430         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1431
1432         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1433                                            req->wb_context,
1434                                            req_offset(req),
1435                                            wb_size,
1436                                            IOMODE_RW,
1437                                            GFP_NOFS);
1438         /* If no lseg, fall back to write through mds */
1439         if (pgio->pg_lseg == NULL)
1440                 nfs_pageio_reset_write_mds(pgio);
1441 }
1442 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1443
1444 /*
1445  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1446  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1447  */
1448 size_t
1449 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1450                      struct nfs_page *req)
1451 {
1452         unsigned int size;
1453         u64 seg_end, req_start, seg_left;
1454
1455         size = nfs_generic_pg_test(pgio, prev, req);
1456         if (!size)
1457                 return 0;
1458
1459         /*
1460          * 'size' contains the number of bytes left in the current page (up
1461          * to the original size asked for in @req->wb_bytes).
1462          *
1463          * Calculate how many bytes are left in the layout segment
1464          * and if there are less bytes than 'size', return that instead.
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         if (pgio->pg_lseg) {
1471                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1472                                      pgio->pg_lseg->pls_range.length);
1473                 req_start = req_offset(req);
1474                 WARN_ON_ONCE(req_start > seg_end);
1475                 /* start of request is past the last byte of this segment */
1476                 if (req_start >= seg_end)
1477                         return 0;
1478
1479                 /* adjust 'size' iff there are fewer bytes left in the
1480                  * segment than what nfs_generic_pg_test returned */
1481                 seg_left = seg_end - req_start;
1482                 if (seg_left < size)
1483                         size = (unsigned int)seg_left;
1484         }
1485
1486         return size;
1487 }
1488 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1489
1490 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1491 {
1492         struct nfs_pageio_descriptor pgio;
1493
1494         /* Resend all requests through the MDS */
1495         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1496                               hdr->completion_ops);
1497         return nfs_pageio_resend(&pgio, hdr);
1498 }
1499 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1500
1501 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1502 {
1503
1504         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1505         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1506             PNFS_LAYOUTRET_ON_ERROR) {
1507                 pnfs_return_layout(hdr->inode);
1508         }
1509         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1510                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1511 }
1512
1513 /*
1514  * Called by non rpc-based layout drivers
1515  */
1516 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1517 {
1518         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1519         if (!hdr->pnfs_error) {
1520                 pnfs_set_layoutcommit(hdr);
1521                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1522         } else
1523                 pnfs_ld_handle_write_error(hdr);
1524         hdr->mds_ops->rpc_release(hdr);
1525 }
1526 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1527
1528 static void
1529 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1530                 struct nfs_pgio_header *hdr)
1531 {
1532         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1533                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1534                 nfs_pageio_reset_write_mds(desc);
1535                 desc->pg_recoalesce = 1;
1536         }
1537         nfs_pgio_data_destroy(hdr);
1538 }
1539
1540 static enum pnfs_try_status
1541 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1542                         const struct rpc_call_ops *call_ops,
1543                         struct pnfs_layout_segment *lseg,
1544                         int how)
1545 {
1546         struct inode *inode = hdr->inode;
1547         enum pnfs_try_status trypnfs;
1548         struct nfs_server *nfss = NFS_SERVER(inode);
1549
1550         hdr->mds_ops = call_ops;
1551
1552         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1553                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1554         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1555         if (trypnfs != PNFS_NOT_ATTEMPTED)
1556                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1557         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1558         return trypnfs;
1559 }
1560
1561 static void
1562 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1563               struct nfs_pgio_header *hdr, int how)
1564 {
1565         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1566         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1567         enum pnfs_try_status trypnfs;
1568
1569         desc->pg_lseg = NULL;
1570         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1571         if (trypnfs == PNFS_NOT_ATTEMPTED)
1572                 pnfs_write_through_mds(desc, hdr);
1573         pnfs_put_lseg(lseg);
1574 }
1575
1576 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1577 {
1578         pnfs_put_lseg(hdr->lseg);
1579         nfs_pgio_header_free(hdr);
1580 }
1581 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1582
1583 int
1584 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1585 {
1586         struct nfs_pgio_header *hdr;
1587         int ret;
1588
1589         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1590         if (!hdr) {
1591                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1592                 pnfs_put_lseg(desc->pg_lseg);
1593                 desc->pg_lseg = NULL;
1594                 return -ENOMEM;
1595         }
1596         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1597         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1598         ret = nfs_generic_pgio(desc, hdr);
1599         if (ret != 0) {
1600                 pnfs_put_lseg(desc->pg_lseg);
1601                 desc->pg_lseg = NULL;
1602         } else
1603                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1604         return ret;
1605 }
1606 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1607
1608 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1609 {
1610         struct nfs_pageio_descriptor pgio;
1611
1612         /* Resend all requests through the MDS */
1613         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1614         return nfs_pageio_resend(&pgio, hdr);
1615 }
1616 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1617
1618 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1619 {
1620         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1621         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1622             PNFS_LAYOUTRET_ON_ERROR) {
1623                 pnfs_return_layout(hdr->inode);
1624         }
1625         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1626                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1627 }
1628
1629 /*
1630  * Called by non rpc-based layout drivers
1631  */
1632 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1633 {
1634         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1635         if (likely(!hdr->pnfs_error)) {
1636                 __nfs4_read_done_cb(hdr);
1637                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1638         } else
1639                 pnfs_ld_handle_read_error(hdr);
1640         hdr->mds_ops->rpc_release(hdr);
1641 }
1642 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1643
1644 static void
1645 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1646                 struct nfs_pgio_header *hdr)
1647 {
1648         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1649                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1650                 nfs_pageio_reset_read_mds(desc);
1651                 desc->pg_recoalesce = 1;
1652         }
1653         nfs_pgio_data_destroy(hdr);
1654 }
1655
1656 /*
1657  * Call the appropriate parallel I/O subsystem read function.
1658  */
1659 static enum pnfs_try_status
1660 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1661                        const struct rpc_call_ops *call_ops,
1662                        struct pnfs_layout_segment *lseg)
1663 {
1664         struct inode *inode = hdr->inode;
1665         struct nfs_server *nfss = NFS_SERVER(inode);
1666         enum pnfs_try_status trypnfs;
1667
1668         hdr->mds_ops = call_ops;
1669
1670         dprintk("%s: Reading ino:%lu %u@%llu\n",
1671                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1672
1673         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1674         if (trypnfs != PNFS_NOT_ATTEMPTED)
1675                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1676         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1677         return trypnfs;
1678 }
1679
1680 static void
1681 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1682 {
1683         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1684         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1685         enum pnfs_try_status trypnfs;
1686
1687         desc->pg_lseg = NULL;
1688         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1689         if (trypnfs == PNFS_NOT_ATTEMPTED)
1690                 pnfs_read_through_mds(desc, hdr);
1691         pnfs_put_lseg(lseg);
1692 }
1693
1694 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1695 {
1696         pnfs_put_lseg(hdr->lseg);
1697         nfs_pgio_header_free(hdr);
1698 }
1699 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1700
1701 int
1702 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1703 {
1704         struct nfs_pgio_header *hdr;
1705         int ret;
1706
1707         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1708         if (!hdr) {
1709                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1710                 ret = -ENOMEM;
1711                 pnfs_put_lseg(desc->pg_lseg);
1712                 desc->pg_lseg = NULL;
1713                 return ret;
1714         }
1715         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1716         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1717         ret = nfs_generic_pgio(desc, hdr);
1718         if (ret != 0) {
1719                 pnfs_put_lseg(desc->pg_lseg);
1720                 desc->pg_lseg = NULL;
1721         } else
1722                 pnfs_do_read(desc, hdr);
1723         return ret;
1724 }
1725 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1726
1727 static void pnfs_clear_layoutcommitting(struct inode *inode)
1728 {
1729         unsigned long *bitlock = &NFS_I(inode)->flags;
1730
1731         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1732         smp_mb__after_atomic();
1733         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1734 }
1735
1736 /*
1737  * There can be multiple RW segments.
1738  */
1739 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1740 {
1741         struct pnfs_layout_segment *lseg;
1742
1743         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1744                 if (lseg->pls_range.iomode == IOMODE_RW &&
1745                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1746                         list_add(&lseg->pls_lc_list, listp);
1747         }
1748 }
1749
1750 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1751 {
1752         struct pnfs_layout_segment *lseg, *tmp;
1753
1754         /* Matched by references in pnfs_set_layoutcommit */
1755         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1756                 list_del_init(&lseg->pls_lc_list);
1757                 pnfs_put_lseg(lseg);
1758         }
1759
1760         pnfs_clear_layoutcommitting(inode);
1761 }
1762
1763 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1764 {
1765         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1766 }
1767 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1768
1769 void
1770 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1771 {
1772         struct inode *inode = hdr->inode;
1773         struct nfs_inode *nfsi = NFS_I(inode);
1774         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1775         bool mark_as_dirty = false;
1776
1777         spin_lock(&inode->i_lock);
1778         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1779                 mark_as_dirty = true;
1780                 dprintk("%s: Set layoutcommit for inode %lu ",
1781                         __func__, inode->i_ino);
1782         }
1783         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1784                 /* references matched in nfs4_layoutcommit_release */
1785                 pnfs_get_lseg(hdr->lseg);
1786         }
1787         if (end_pos > nfsi->layout->plh_lwb)
1788                 nfsi->layout->plh_lwb = end_pos;
1789         spin_unlock(&inode->i_lock);
1790         dprintk("%s: lseg %p end_pos %llu\n",
1791                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1792
1793         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1794          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1795         if (mark_as_dirty)
1796                 mark_inode_dirty_sync(inode);
1797 }
1798 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1799
1800 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1801 {
1802         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1803
1804         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1805                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1806         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1807 }
1808
1809 /*
1810  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1811  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1812  * data to disk to allow the server to recover the data if it crashes.
1813  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1814  * is off, and a COMMIT is sent to a data server, or
1815  * if WRITEs to a data server return NFS_DATA_SYNC.
1816  */
1817 int
1818 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1819 {
1820         struct nfs4_layoutcommit_data *data;
1821         struct nfs_inode *nfsi = NFS_I(inode);
1822         loff_t end_pos;
1823         int status;
1824
1825         if (!pnfs_layoutcommit_outstanding(inode))
1826                 return 0;
1827
1828         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1829
1830         status = -EAGAIN;
1831         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1832                 if (!sync)
1833                         goto out;
1834                 status = wait_on_bit_lock_action(&nfsi->flags,
1835                                 NFS_INO_LAYOUTCOMMITTING,
1836                                 nfs_wait_bit_killable,
1837                                 TASK_KILLABLE);
1838                 if (status)
1839                         goto out;
1840         }
1841
1842         status = -ENOMEM;
1843         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1844         data = kzalloc(sizeof(*data), GFP_NOFS);
1845         if (!data)
1846                 goto clear_layoutcommitting;
1847
1848         status = 0;
1849         spin_lock(&inode->i_lock);
1850         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1851                 goto out_unlock;
1852
1853         INIT_LIST_HEAD(&data->lseg_list);
1854         pnfs_list_write_lseg(inode, &data->lseg_list);
1855
1856         end_pos = nfsi->layout->plh_lwb;
1857         nfsi->layout->plh_lwb = 0;
1858
1859         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1860         spin_unlock(&inode->i_lock);
1861
1862         data->args.inode = inode;
1863         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1864         nfs_fattr_init(&data->fattr);
1865         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1866         data->res.fattr = &data->fattr;
1867         data->args.lastbytewritten = end_pos - 1;
1868         data->res.server = NFS_SERVER(inode);
1869
1870         status = nfs4_proc_layoutcommit(data, sync);
1871 out:
1872         if (status)
1873                 mark_inode_dirty_sync(inode);
1874         dprintk("<-- %s status %d\n", __func__, status);
1875         return status;
1876 out_unlock:
1877         spin_unlock(&inode->i_lock);
1878         kfree(data);
1879 clear_layoutcommitting:
1880         pnfs_clear_layoutcommitting(inode);
1881         goto out;
1882 }
1883
1884 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1885 {
1886         struct nfs4_threshold *thp;
1887
1888         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1889         if (!thp) {
1890                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1891                 return NULL;
1892         }
1893         return thp;
1894 }