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
 861         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 862                 struct pnfs_layout_range range = {
 863                         .iomode         = IOMODE_ANY,
 864                         .offset         = 0,
 865                         .length         = NFS4_MAX_UINT64,
 866                 };
 867                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
 868         }
 869
 870         /* Don't send a LAYOUTRETURN if list was initially empty */
 871         if (empty) {
 872                 spin_unlock(&ino->i_lock);
 873                 pnfs_put_layout_hdr(lo);
 874                 dprintk("NFS: %s no layout segments to return\n", __func__);
 875                 goto out;
 876         }
 877
 878         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 879         lo->plh_block_lgets++;
 880         spin_unlock(&ino->i_lock);
 881         pnfs_free_lseg_list(&tmp_list);
 882
 883         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 884         if (unlikely(lrp == NULL)) {
 885                 status = -ENOMEM;
 886                 spin_lock(&ino->i_lock);
 887                 lo->plh_block_lgets--;
 888                 spin_unlock(&ino->i_lock);
 889                 pnfs_put_layout_hdr(lo);
 890                 goto out;
 891         }
 892
 893         lrp->args.stateid = stateid;
 894         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 895         lrp->args.inode = ino;
 896         lrp->args.layout = lo;
 897         lrp->clp = NFS_SERVER(ino)->nfs_client;
 898         lrp->cred = lo->plh_lc_cred;
 899
 900         status = nfs4_proc_layoutreturn(lrp);
 901 out:
 902         dprintk("<-- %s status: %d\n", __func__, status);
 903         return status;
 904 }
 905 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 906
 907 int
 908 pnfs_commit_and_return_layout(struct inode *inode)
 909 {
 910         struct pnfs_layout_hdr *lo;
 911         int ret;
 912
 913         spin_lock(&inode->i_lock);
 914         lo = NFS_I(inode)->layout;
 915         if (lo == NULL) {
 916                 spin_unlock(&inode->i_lock);
 917                 return 0;
 918         }
 919         pnfs_get_layout_hdr(lo);
 920         /* Block new layoutgets and read/write to ds */
 921         lo->plh_block_lgets++;
 922         spin_unlock(&inode->i_lock);
 923         filemap_fdatawait(inode->i_mapping);
 924         ret = pnfs_layoutcommit_inode(inode, true);
 925         if (ret == 0)
 926                 ret = _pnfs_return_layout(inode);
 927         spin_lock(&inode->i_lock);
 928         lo->plh_block_lgets--;
 929         spin_unlock(&inode->i_lock);
 930         pnfs_put_layout_hdr(lo);
 931         return ret;
 932 }
 933
 934 bool pnfs_roc(struct inode *ino)
 935 {
 936         struct pnfs_layout_hdr *lo;
 937         struct pnfs_layout_segment *lseg, *tmp;
 938         LIST_HEAD(tmp_list);
 939         bool found = false;
 940
 941         spin_lock(&ino->i_lock);
 942         lo = NFS_I(ino)->layout;
 943         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 944             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 945                 goto out_nolayout;
 946         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 947                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 948                         mark_lseg_invalid(lseg, &tmp_list);
 949                         found = true;
 950                 }
 951         if (!found)
 952                 goto out_nolayout;
 953         lo->plh_block_lgets++;
 954         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 955         spin_unlock(&ino->i_lock);
 956         pnfs_free_lseg_list(&tmp_list);
 957         return true;
 958
 959 out_nolayout:
 960         spin_unlock(&ino->i_lock);
 961         return false;
 962 }
 963
 964 void pnfs_roc_release(struct inode *ino)
 965 {
 966         struct pnfs_layout_hdr *lo;
 967
 968         spin_lock(&ino->i_lock);
 969         lo = NFS_I(ino)->layout;
 970         lo->plh_block_lgets--;
 971         if (atomic_dec_and_test(&lo->plh_refcount)) {
 972                 pnfs_detach_layout_hdr(lo);
 973                 spin_unlock(&ino->i_lock);
 974                 pnfs_free_layout_hdr(lo);
 975         } else
 976                 spin_unlock(&ino->i_lock);
 977 }
 978
 979 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 980 {
 981         struct pnfs_layout_hdr *lo;
 982
 983         spin_lock(&ino->i_lock);
 984         lo = NFS_I(ino)->layout;
 985         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 986                 lo->plh_barrier = barrier;
 987         spin_unlock(&ino->i_lock);
 988 }
 989
 990 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 991 {
 992         struct nfs_inode *nfsi = NFS_I(ino);
 993         struct pnfs_layout_hdr *lo;
 994         struct pnfs_layout_segment *lseg;
 995         u32 current_seqid;
 996         bool found = false;
 997
 998         spin_lock(&ino->i_lock);
 999         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
1000                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1001                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1002                         found = true;
1003                         goto out;
1004                 }
1005         lo = nfsi->layout;
1006         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1007
1008         /* Since close does not return a layout stateid for use as
1009          * a barrier, we choose the worst-case barrier.
1010          */
1011         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1012 out:
1013         spin_unlock(&ino->i_lock);
1014         return found;
1015 }
1016
1017 /*
1018  * Compare two layout segments for sorting into layout cache.
1019  * We want to preferentially return RW over RO layouts, so ensure those
1020  * are seen first.
1021  */
1022 static s64
1023 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1024            const struct pnfs_layout_range *l2)
1025 {
1026         s64 d;
1027
1028         /* high offset > low offset */
1029         d = l1->offset - l2->offset;
1030         if (d)
1031                 return d;
1032
1033         /* short length > long length */
1034         d = l2->length - l1->length;
1035         if (d)
1036                 return d;
1037
1038         /* read > read/write */
1039         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1040 }
1041
1042 static void
1043 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1044                    struct pnfs_layout_segment *lseg)
1045 {
1046         struct pnfs_layout_segment *lp;
1047
1048         dprintk("%s:Begin\n", __func__);
1049
1050         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1051                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1052                         continue;
1053                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1054                 dprintk("%s: inserted lseg %p "
1055                         "iomode %d offset %llu length %llu before "
1056                         "lp %p iomode %d offset %llu length %llu\n",
1057                         __func__, lseg, lseg->pls_range.iomode,
1058                         lseg->pls_range.offset, lseg->pls_range.length,
1059                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1060                         lp->pls_range.length);
1061                 goto out;
1062         }
1063         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1064         dprintk("%s: inserted lseg %p "
1065                 "iomode %d offset %llu length %llu at tail\n",
1066                 __func__, lseg, lseg->pls_range.iomode,
1067                 lseg->pls_range.offset, lseg->pls_range.length);
1068 out:
1069         pnfs_get_layout_hdr(lo);
1070
1071         dprintk("%s:Return\n", __func__);
1072 }
1073
1074 static struct pnfs_layout_hdr *
1075 alloc_init_layout_hdr(struct inode *ino,
1076                       struct nfs_open_context *ctx,
1077                       gfp_t gfp_flags)
1078 {
1079         struct pnfs_layout_hdr *lo;
1080
1081         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1082         if (!lo)
1083                 return NULL;
1084         atomic_set(&lo->plh_refcount, 1);
1085         INIT_LIST_HEAD(&lo->plh_layouts);
1086         INIT_LIST_HEAD(&lo->plh_segs);
1087         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1088         lo->plh_inode = ino;
1089         lo->plh_lc_cred = get_rpccred(ctx->cred);
1090         return lo;
1091 }
1092
1093 static struct pnfs_layout_hdr *
1094 pnfs_find_alloc_layout(struct inode *ino,
1095                        struct nfs_open_context *ctx,
1096                        gfp_t gfp_flags)
1097 {
1098         struct nfs_inode *nfsi = NFS_I(ino);
1099         struct pnfs_layout_hdr *new = NULL;
1100
1101         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1102
1103         if (nfsi->layout != NULL)
1104                 goto out_existing;
1105         spin_unlock(&ino->i_lock);
1106         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1107         spin_lock(&ino->i_lock);
1108
1109         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1110                 nfsi->layout = new;
1111                 return new;
1112         } else if (new != NULL)
1113                 pnfs_free_layout_hdr(new);
1114 out_existing:
1115         pnfs_get_layout_hdr(nfsi->layout);
1116         return nfsi->layout;
1117 }
1118
1119 /*
1120  * iomode matching rules:
1121  * iomode       lseg    match
1122  * -----        -----   -----
1123  * ANY          READ    true
1124  * ANY          RW      true
1125  * RW           READ    false
1126  * RW           RW      true
1127  * READ         READ    true
1128  * READ         RW      true
1129  */
1130 static bool
1131 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1132                  const struct pnfs_layout_range *range)
1133 {
1134         struct pnfs_layout_range range1;
1135
1136         if ((range->iomode == IOMODE_RW &&
1137              ls_range->iomode != IOMODE_RW) ||
1138             !pnfs_lseg_range_intersecting(ls_range, range))
1139                 return 0;
1140
1141         /* range1 covers only the first byte in the range */
1142         range1 = *range;
1143         range1.length = 1;
1144         return pnfs_lseg_range_contained(ls_range, &range1);
1145 }
1146
1147 /*
1148  * lookup range in layout
1149  */
1150 static struct pnfs_layout_segment *
1151 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1152                 struct pnfs_layout_range *range)
1153 {
1154         struct pnfs_layout_segment *lseg, *ret = NULL;
1155
1156         dprintk("%s:Begin\n", __func__);
1157
1158         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1159                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1160                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1161                         ret = pnfs_get_lseg(lseg);
1162                         break;
1163                 }
1164                 if (lseg->pls_range.offset > range->offset)
1165                         break;
1166         }
1167
1168         dprintk("%s:Return lseg %p ref %d\n",
1169                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1170         return ret;
1171 }
1172
1173 /*
1174  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1175  * to the MDS or over pNFS
1176  *
1177  * The nfs_inode read_io and write_io fields are cumulative counters reset
1178  * when there are no layout segments. Note that in pnfs_update_layout iomode
1179  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1180  * WRITE request.
1181  *
1182  * A return of true means use MDS I/O.
1183  *
1184  * From rfc 5661:
1185  * If a file's size is smaller than the file size threshold, data accesses
1186  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1187  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1188  * server.  If both file size and I/O size are provided, the client SHOULD
1189  * reach or exceed  both thresholds before sending its read or write
1190  * requests to the data server.
1191  */
1192 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1193                                      struct inode *ino, int iomode)
1194 {
1195         struct nfs4_threshold *t = ctx->mdsthreshold;
1196         struct nfs_inode *nfsi = NFS_I(ino);
1197         loff_t fsize = i_size_read(ino);
1198         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1199
1200         if (t == NULL)
1201                 return ret;
1202
1203         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1204                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1205
1206         switch (iomode) {
1207         case IOMODE_READ:
1208                 if (t->bm & THRESHOLD_RD) {
1209                         dprintk("%s fsize %llu\n", __func__, fsize);
1210                         size_set = true;
1211                         if (fsize < t->rd_sz)
1212                                 size = true;
1213                 }
1214                 if (t->bm & THRESHOLD_RD_IO) {
1215                         dprintk("%s nfsi->read_io %llu\n", __func__,
1216                                 nfsi->read_io);
1217                         io_set = true;
1218                         if (nfsi->read_io < t->rd_io_sz)
1219                                 io = true;
1220                 }
1221                 break;
1222         case IOMODE_RW:
1223                 if (t->bm & THRESHOLD_WR) {
1224                         dprintk("%s fsize %llu\n", __func__, fsize);
1225                         size_set = true;
1226                         if (fsize < t->wr_sz)
1227                                 size = true;
1228                 }
1229                 if (t->bm & THRESHOLD_WR_IO) {
1230                         dprintk("%s nfsi->write_io %llu\n", __func__,
1231                                 nfsi->write_io);
1232                         io_set = true;
1233                         if (nfsi->write_io < t->wr_io_sz)
1234                                 io = true;
1235                 }
1236                 break;
1237         }
1238         if (size_set && io_set) {
1239                 if (size && io)
1240                         ret = true;
1241         } else if (size || io)
1242                 ret = true;
1243
1244         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1245         return ret;
1246 }
1247
1248 /*
1249  * Layout segment is retreived from the server if not cached.
1250  * The appropriate layout segment is referenced and returned to the caller.
1251  */
1252 struct pnfs_layout_segment *
1253 pnfs_update_layout(struct inode *ino,
1254                    struct nfs_open_context *ctx,
1255                    loff_t pos,
1256                    u64 count,
1257                    enum pnfs_iomode iomode,
1258                    gfp_t gfp_flags)
1259 {
1260         struct pnfs_layout_range arg = {
1261                 .iomode = iomode,
1262                 .offset = pos,
1263                 .length = count,
1264         };
1265         unsigned pg_offset;
1266         struct nfs_server *server = NFS_SERVER(ino);
1267         struct nfs_client *clp = server->nfs_client;
1268         struct pnfs_layout_hdr *lo;
1269         struct pnfs_layout_segment *lseg = NULL;
1270         bool first;
1271
1272         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1273                 goto out;
1274
1275         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1276                 goto out;
1277
1278         spin_lock(&ino->i_lock);
1279         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1280         if (lo == NULL) {
1281                 spin_unlock(&ino->i_lock);
1282                 goto out;
1283         }
1284
1285         /* Do we even need to bother with this? */
1286         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1287                 dprintk("%s matches recall, use MDS\n", __func__);
1288                 goto out_unlock;
1289         }
1290
1291         /* if LAYOUTGET already failed once we don't try again */
1292         if (pnfs_layout_io_test_failed(lo, iomode))
1293                 goto out_unlock;
1294
1295         /* Check to see if the layout for the given range already exists */
1296         lseg = pnfs_find_lseg(lo, &arg);
1297         if (lseg)
1298                 goto out_unlock;
1299
1300         if (pnfs_layoutgets_blocked(lo, 0))
1301                 goto out_unlock;
1302         atomic_inc(&lo->plh_outstanding);
1303
1304         first = list_empty(&lo->plh_layouts) ? true : false;
1305         spin_unlock(&ino->i_lock);
1306
1307         if (first) {
1308                 /* The lo must be on the clp list if there is any
1309                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1310                  */
1311                 spin_lock(&clp->cl_lock);
1312                 list_add_tail(&lo->plh_layouts, &server->layouts);
1313                 spin_unlock(&clp->cl_lock);
1314         }
1315
1316         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1317         if (pg_offset) {
1318                 arg.offset -= pg_offset;
1319                 arg.length += pg_offset;
1320         }
1321         if (arg.length != NFS4_MAX_UINT64)
1322                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1323
1324         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1325         atomic_dec(&lo->plh_outstanding);
1326 out_put_layout_hdr:
1327         pnfs_put_layout_hdr(lo);
1328 out:
1329         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1330                         "(%s, offset: %llu, length: %llu)\n",
1331                         __func__, ino->i_sb->s_id,
1332                         (unsigned long long)NFS_FILEID(ino),
1333                         lseg == NULL ? "not found" : "found",
1334                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1335                         (unsigned long long)pos,
1336                         (unsigned long long)count);
1337         return lseg;
1338 out_unlock:
1339         spin_unlock(&ino->i_lock);
1340         goto out_put_layout_hdr;
1341 }
1342 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1343
1344 struct pnfs_layout_segment *
1345 pnfs_layout_process(struct nfs4_layoutget *lgp)
1346 {
1347         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1348         struct nfs4_layoutget_res *res = &lgp->res;
1349         struct pnfs_layout_segment *lseg;
1350         struct inode *ino = lo->plh_inode;
1351         LIST_HEAD(free_me);
1352         int status = 0;
1353
1354         /* Inject layout blob into I/O device driver */
1355         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1356         if (!lseg || IS_ERR(lseg)) {
1357                 if (!lseg)
1358                         status = -ENOMEM;
1359                 else
1360                         status = PTR_ERR(lseg);
1361                 dprintk("%s: Could not allocate layout: error %d\n",
1362                        __func__, status);
1363                 goto out;
1364         }
1365
1366         init_lseg(lo, lseg);
1367         lseg->pls_range = res->range;
1368
1369         spin_lock(&ino->i_lock);
1370         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1371                 dprintk("%s forget reply due to recall\n", __func__);
1372                 goto out_forget_reply;
1373         }
1374
1375         if (pnfs_layoutgets_blocked(lo, 1)) {
1376                 dprintk("%s forget reply due to state\n", __func__);
1377                 goto out_forget_reply;
1378         }
1379
1380         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1381                 /* existing state ID, make sure the sequence number matches. */
1382                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1383                         dprintk("%s forget reply due to sequence\n", __func__);
1384                         goto out_forget_reply;
1385                 }
1386                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1387         } else {
1388                 /*
1389                  * We got an entirely new state ID.  Mark all segments for the
1390                  * inode invalid, and don't bother validating the stateid
1391                  * sequence number.
1392                  */
1393                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1394
1395                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1396                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1397         }
1398
1399         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1400
1401         pnfs_get_lseg(lseg);
1402         pnfs_layout_insert_lseg(lo, lseg);
1403
1404         if (res->return_on_close) {
1405                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1406                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1407         }
1408
1409         spin_unlock(&ino->i_lock);
1410         pnfs_free_lseg_list(&free_me);
1411         return lseg;
1412 out:
1413         return ERR_PTR(status);
1414
1415 out_forget_reply:
1416         spin_unlock(&ino->i_lock);
1417         lseg->pls_layout = lo;
1418         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1419         goto out;
1420 }
1421
1422 void
1423 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1424 {
1425         u64 rd_size = req->wb_bytes;
1426
1427         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1428
1429         if (pgio->pg_dreq == NULL)
1430                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1431         else
1432                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1433
1434         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1435                                            req->wb_context,
1436                                            req_offset(req),
1437                                            rd_size,
1438                                            IOMODE_READ,
1439                                            GFP_KERNEL);
1440         /* If no lseg, fall back to read through mds */
1441         if (pgio->pg_lseg == NULL)
1442                 nfs_pageio_reset_read_mds(pgio);
1443
1444 }
1445 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1446
1447 void
1448 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1449                            struct nfs_page *req, u64 wb_size)
1450 {
1451         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1452
1453         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1454                                            req->wb_context,
1455                                            req_offset(req),
1456                                            wb_size,
1457                                            IOMODE_RW,
1458                                            GFP_NOFS);
1459         /* If no lseg, fall back to write through mds */
1460         if (pgio->pg_lseg == NULL)
1461                 nfs_pageio_reset_write_mds(pgio);
1462 }
1463 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1464
1465 /*
1466  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1467  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1468  */
1469 size_t
1470 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1471                      struct nfs_page *req)
1472 {
1473         unsigned int size;
1474         u64 seg_end, req_start, seg_left;
1475
1476         size = nfs_generic_pg_test(pgio, prev, req);
1477         if (!size)
1478                 return 0;
1479
1480         /*
1481          * 'size' contains the number of bytes left in the current page (up
1482          * to the original size asked for in @req->wb_bytes).
1483          *
1484          * Calculate how many bytes are left in the layout segment
1485          * and if there are less bytes than 'size', return that instead.
1486          *
1487          * Please also note that 'end_offset' is actually the offset of the
1488          * first byte that lies outside the pnfs_layout_range. FIXME?
1489          *
1490          */
1491         if (pgio->pg_lseg) {
1492                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1493                                      pgio->pg_lseg->pls_range.length);
1494                 req_start = req_offset(req);
1495                 WARN_ON_ONCE(req_start > seg_end);
1496                 /* start of request is past the last byte of this segment */
1497                 if (req_start >= seg_end)
1498                         return 0;
1499
1500                 /* adjust 'size' iff there are fewer bytes left in the
1501                  * segment than what nfs_generic_pg_test returned */
1502                 seg_left = seg_end - req_start;
1503                 if (seg_left < size)
1504                         size = (unsigned int)seg_left;
1505         }
1506
1507         return size;
1508 }
1509 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1510
1511 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1512 {
1513         struct nfs_pageio_descriptor pgio;
1514
1515         /* Resend all requests through the MDS */
1516         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1517                               hdr->completion_ops);
1518         return nfs_pageio_resend(&pgio, hdr);
1519 }
1520 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1521
1522 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1523 {
1524
1525         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1526         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1527             PNFS_LAYOUTRET_ON_ERROR) {
1528                 pnfs_return_layout(hdr->inode);
1529         }
1530         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1531                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1532 }
1533
1534 /*
1535  * Called by non rpc-based layout drivers
1536  */
1537 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1538 {
1539         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1540         if (!hdr->pnfs_error) {
1541                 pnfs_set_layoutcommit(hdr);
1542                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1543         } else
1544                 pnfs_ld_handle_write_error(hdr);
1545         hdr->mds_ops->rpc_release(hdr);
1546 }
1547 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1548
1549 static void
1550 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1551                 struct nfs_pgio_header *hdr)
1552 {
1553         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1554                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1555                 nfs_pageio_reset_write_mds(desc);
1556                 desc->pg_recoalesce = 1;
1557         }
1558         nfs_pgio_data_destroy(hdr);
1559 }
1560
1561 static enum pnfs_try_status
1562 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1563                         const struct rpc_call_ops *call_ops,
1564                         struct pnfs_layout_segment *lseg,
1565                         int how)
1566 {
1567         struct inode *inode = hdr->inode;
1568         enum pnfs_try_status trypnfs;
1569         struct nfs_server *nfss = NFS_SERVER(inode);
1570
1571         hdr->mds_ops = call_ops;
1572
1573         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1574                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1575         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1576         if (trypnfs != PNFS_NOT_ATTEMPTED)
1577                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1578         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1579         return trypnfs;
1580 }
1581
1582 static void
1583 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1584               struct nfs_pgio_header *hdr, int how)
1585 {
1586         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1587         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1588         enum pnfs_try_status trypnfs;
1589
1590         desc->pg_lseg = NULL;
1591         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1592         if (trypnfs == PNFS_NOT_ATTEMPTED)
1593                 pnfs_write_through_mds(desc, hdr);
1594         pnfs_put_lseg(lseg);
1595 }
1596
1597 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1598 {
1599         pnfs_put_lseg(hdr->lseg);
1600         nfs_pgio_header_free(hdr);
1601 }
1602 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1603
1604 int
1605 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1606 {
1607         struct nfs_pgio_header *hdr;
1608         int ret;
1609
1610         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1611         if (!hdr) {
1612                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1613                 pnfs_put_lseg(desc->pg_lseg);
1614                 desc->pg_lseg = NULL;
1615                 return -ENOMEM;
1616         }
1617         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1618         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1619         ret = nfs_generic_pgio(desc, hdr);
1620         if (ret != 0) {
1621                 pnfs_put_lseg(desc->pg_lseg);
1622                 desc->pg_lseg = NULL;
1623         } else
1624                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1625         return ret;
1626 }
1627 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1628
1629 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1630 {
1631         struct nfs_pageio_descriptor pgio;
1632
1633         /* Resend all requests through the MDS */
1634         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1635         return nfs_pageio_resend(&pgio, hdr);
1636 }
1637 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1638
1639 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1640 {
1641         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1642         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1643             PNFS_LAYOUTRET_ON_ERROR) {
1644                 pnfs_return_layout(hdr->inode);
1645         }
1646         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1647                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1648 }
1649
1650 /*
1651  * Called by non rpc-based layout drivers
1652  */
1653 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1654 {
1655         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1656         if (likely(!hdr->pnfs_error)) {
1657                 __nfs4_read_done_cb(hdr);
1658                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1659         } else
1660                 pnfs_ld_handle_read_error(hdr);
1661         hdr->mds_ops->rpc_release(hdr);
1662 }
1663 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1664
1665 static void
1666 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1667                 struct nfs_pgio_header *hdr)
1668 {
1669         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1670                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1671                 nfs_pageio_reset_read_mds(desc);
1672                 desc->pg_recoalesce = 1;
1673         }
1674         nfs_pgio_data_destroy(hdr);
1675 }
1676
1677 /*
1678  * Call the appropriate parallel I/O subsystem read function.
1679  */
1680 static enum pnfs_try_status
1681 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1682                        const struct rpc_call_ops *call_ops,
1683                        struct pnfs_layout_segment *lseg)
1684 {
1685         struct inode *inode = hdr->inode;
1686         struct nfs_server *nfss = NFS_SERVER(inode);
1687         enum pnfs_try_status trypnfs;
1688
1689         hdr->mds_ops = call_ops;
1690
1691         dprintk("%s: Reading ino:%lu %u@%llu\n",
1692                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1693
1694         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1695         if (trypnfs != PNFS_NOT_ATTEMPTED)
1696                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1697         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1698         return trypnfs;
1699 }
1700
1701 static void
1702 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1703 {
1704         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1705         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1706         enum pnfs_try_status trypnfs;
1707
1708         desc->pg_lseg = NULL;
1709         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1710         if (trypnfs == PNFS_NOT_ATTEMPTED)
1711                 pnfs_read_through_mds(desc, hdr);
1712         pnfs_put_lseg(lseg);
1713 }
1714
1715 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1716 {
1717         pnfs_put_lseg(hdr->lseg);
1718         nfs_pgio_header_free(hdr);
1719 }
1720 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1721
1722 int
1723 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1724 {
1725         struct nfs_pgio_header *hdr;
1726         int ret;
1727
1728         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1729         if (!hdr) {
1730                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1731                 ret = -ENOMEM;
1732                 pnfs_put_lseg(desc->pg_lseg);
1733                 desc->pg_lseg = NULL;
1734                 return ret;
1735         }
1736         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1737         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1738         ret = nfs_generic_pgio(desc, hdr);
1739         if (ret != 0) {
1740                 pnfs_put_lseg(desc->pg_lseg);
1741                 desc->pg_lseg = NULL;
1742         } else
1743                 pnfs_do_read(desc, hdr);
1744         return ret;
1745 }
1746 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1747
1748 static void pnfs_clear_layoutcommitting(struct inode *inode)
1749 {
1750         unsigned long *bitlock = &NFS_I(inode)->flags;
1751
1752         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1753         smp_mb__after_atomic();
1754         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1755 }
1756
1757 /*
1758  * There can be multiple RW segments.
1759  */
1760 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1761 {
1762         struct pnfs_layout_segment *lseg;
1763
1764         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1765                 if (lseg->pls_range.iomode == IOMODE_RW &&
1766                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1767                         list_add(&lseg->pls_lc_list, listp);
1768         }
1769 }
1770
1771 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1772 {
1773         struct pnfs_layout_segment *lseg, *tmp;
1774
1775         /* Matched by references in pnfs_set_layoutcommit */
1776         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1777                 list_del_init(&lseg->pls_lc_list);
1778                 pnfs_put_lseg(lseg);
1779         }
1780
1781         pnfs_clear_layoutcommitting(inode);
1782 }
1783
1784 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1785 {
1786         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1787 }
1788 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1789
1790 void
1791 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1792 {
1793         struct inode *inode = hdr->inode;
1794         struct nfs_inode *nfsi = NFS_I(inode);
1795         loff_t end_pos = hdr->mds_offset + hdr->res.count;
1796         bool mark_as_dirty = false;
1797
1798         spin_lock(&inode->i_lock);
1799         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1800                 mark_as_dirty = true;
1801                 dprintk("%s: Set layoutcommit for inode %lu ",
1802                         __func__, inode->i_ino);
1803         }
1804         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1805                 /* references matched in nfs4_layoutcommit_release */
1806                 pnfs_get_lseg(hdr->lseg);
1807         }
1808         if (end_pos > nfsi->layout->plh_lwb)
1809                 nfsi->layout->plh_lwb = end_pos;
1810         spin_unlock(&inode->i_lock);
1811         dprintk("%s: lseg %p end_pos %llu\n",
1812                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1813
1814         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1815          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1816         if (mark_as_dirty)
1817                 mark_inode_dirty_sync(inode);
1818 }
1819 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1820
1821 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
1822 {
1823         struct inode *inode = data->inode;
1824         struct nfs_inode *nfsi = NFS_I(inode);
1825         bool mark_as_dirty = false;
1826
1827         spin_lock(&inode->i_lock);
1828         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1829                 mark_as_dirty = true;
1830                 dprintk("%s: Set layoutcommit for inode %lu ",
1831                         __func__, inode->i_ino);
1832         }
1833         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
1834                 /* references matched in nfs4_layoutcommit_release */
1835                 pnfs_get_lseg(data->lseg);
1836         }
1837         if (data->lwb > nfsi->layout->plh_lwb)
1838                 nfsi->layout->plh_lwb = data->lwb;
1839         spin_unlock(&inode->i_lock);
1840         dprintk("%s: lseg %p end_pos %llu\n",
1841                 __func__, data->lseg, nfsi->layout->plh_lwb);
1842
1843         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1844          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1845         if (mark_as_dirty)
1846                 mark_inode_dirty_sync(inode);
1847 }
1848 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
1849
1850 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1851 {
1852         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1853
1854         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1855                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1856         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1857 }
1858
1859 /*
1860  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1861  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1862  * data to disk to allow the server to recover the data if it crashes.
1863  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1864  * is off, and a COMMIT is sent to a data server, or
1865  * if WRITEs to a data server return NFS_DATA_SYNC.
1866  */
1867 int
1868 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1869 {
1870         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1871         struct nfs4_layoutcommit_data *data;
1872         struct nfs_inode *nfsi = NFS_I(inode);
1873         loff_t end_pos;
1874         int status;
1875
1876         if (!pnfs_layoutcommit_outstanding(inode))
1877                 return 0;
1878
1879         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1880
1881         status = -EAGAIN;
1882         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1883                 if (!sync)
1884                         goto out;
1885                 status = wait_on_bit_lock_action(&nfsi->flags,
1886                                 NFS_INO_LAYOUTCOMMITTING,
1887                                 nfs_wait_bit_killable,
1888                                 TASK_KILLABLE);
1889                 if (status)
1890                         goto out;
1891         }
1892
1893         status = -ENOMEM;
1894         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1895         data = kzalloc(sizeof(*data), GFP_NOFS);
1896         if (!data)
1897                 goto clear_layoutcommitting;
1898
1899         status = 0;
1900         spin_lock(&inode->i_lock);
1901         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1902                 goto out_unlock;
1903
1904         INIT_LIST_HEAD(&data->lseg_list);
1905         pnfs_list_write_lseg(inode, &data->lseg_list);
1906
1907         end_pos = nfsi->layout->plh_lwb;
1908         nfsi->layout->plh_lwb = 0;
1909
1910         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1911         spin_unlock(&inode->i_lock);
1912
1913         data->args.inode = inode;
1914         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1915         nfs_fattr_init(&data->fattr);
1916         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1917         data->res.fattr = &data->fattr;
1918         data->args.lastbytewritten = end_pos - 1;
1919         data->res.server = NFS_SERVER(inode);
1920
1921         if (ld->prepare_layoutcommit) {
1922                 status = ld->prepare_layoutcommit(&data->args);
1923                 if (status) {
1924                         spin_lock(&inode->i_lock);
1925                         if (end_pos < nfsi->layout->plh_lwb)
1926                                 nfsi->layout->plh_lwb = end_pos;
1927                         spin_unlock(&inode->i_lock);
1928                         put_rpccred(data->cred);
1929                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
1930                         goto clear_layoutcommitting;
1931                 }
1932         }
1933
1934
1935         status = nfs4_proc_layoutcommit(data, sync);
1936 out:
1937         if (status)
1938                 mark_inode_dirty_sync(inode);
1939         dprintk("<-- %s status %d\n", __func__, status);
1940         return status;
1941 out_unlock:
1942         spin_unlock(&inode->i_lock);
1943         kfree(data);
1944 clear_layoutcommitting:
1945         pnfs_clear_layoutcommitting(inode);
1946         goto out;
1947 }
1948
1949 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1950 {
1951         struct nfs4_threshold *thp;
1952
1953         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1954         if (!thp) {
1955                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1956                 return NULL;
1957         }
1958         return thp;
1959 }