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 #include "delegation.h"
  38 #include "nfs42.h"
  39
  40 #define NFSDBG_FACILITY         NFSDBG_PNFS
  41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  42
  43 /* Locking:
  44  *
  45  * pnfs_spinlock:
  46  *      protects pnfs_modules_tbl.
  47  */
  48 static DEFINE_SPINLOCK(pnfs_spinlock);
  49
  50 /*
  51  * pnfs_modules_tbl holds all pnfs modules
  52  */
  53 static LIST_HEAD(pnfs_modules_tbl);
  54
  55 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  56
  57 /* Return the registered pnfs layout driver module matching given id */
  58 static struct pnfs_layoutdriver_type *
  59 find_pnfs_driver_locked(u32 id)
  60 {
  61         struct pnfs_layoutdriver_type *local;
  62
  63         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  64                 if (local->id == id)
  65                         goto out;
  66         local = NULL;
  67 out:
  68         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  69         return local;
  70 }
  71
  72 static struct pnfs_layoutdriver_type *
  73 find_pnfs_driver(u32 id)
  74 {
  75         struct pnfs_layoutdriver_type *local;
  76
  77         spin_lock(&pnfs_spinlock);
  78         local = find_pnfs_driver_locked(id);
  79         if (local != NULL && !try_module_get(local->owner)) {
  80                 dprintk("%s: Could not grab reference on module\n", __func__);
  81                 local = NULL;
  82         }
  83         spin_unlock(&pnfs_spinlock);
  84         return local;
  85 }
  86
  87 void
  88 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  89 {
  90         if (nfss->pnfs_curr_ld) {
  91                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  92                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  93                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  94                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  95                         nfs4_deviceid_purge_client(nfss->nfs_client);
  96                 module_put(nfss->pnfs_curr_ld->owner);
  97         }
  98         nfss->pnfs_curr_ld = NULL;
  99 }
 100
 101 /*
 102  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 103  * Currently only one pNFS layout driver per filesystem is supported.
 104  *
 105  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 106  */
 107 void
 108 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 109                       u32 id)
 110 {
 111         struct pnfs_layoutdriver_type *ld_type = NULL;
 112
 113         if (id == 0)
 114                 goto out_no_driver;
 115         if (!(server->nfs_client->cl_exchange_flags &
 116                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 117                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 118                         __func__, id, server->nfs_client->cl_exchange_flags);
 119                 goto out_no_driver;
 120         }
 121         ld_type = find_pnfs_driver(id);
 122         if (!ld_type) {
 123                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 124                 ld_type = find_pnfs_driver(id);
 125                 if (!ld_type) {
 126                         dprintk("%s: No pNFS module found for %u.\n",
 127                                 __func__, id);
 128                         goto out_no_driver;
 129                 }
 130         }
 131         server->pnfs_curr_ld = ld_type;
 132         if (ld_type->set_layoutdriver
 133             && ld_type->set_layoutdriver(server, mntfh)) {
 134                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 135                         "driver %u.\n", __func__, id);
 136                 module_put(ld_type->owner);
 137                 goto out_no_driver;
 138         }
 139         /* Bump the MDS count */
 140         atomic_inc(&server->nfs_client->cl_mds_count);
 141
 142         dprintk("%s: pNFS module for %u set\n", __func__, id);
 143         return;
 144
 145 out_no_driver:
 146         dprintk("%s: Using NFSv4 I/O\n", __func__);
 147         server->pnfs_curr_ld = NULL;
 148 }
 149
 150 int
 151 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 152 {
 153         int status = -EINVAL;
 154         struct pnfs_layoutdriver_type *tmp;
 155
 156         if (ld_type->id == 0) {
 157                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 158                 return status;
 159         }
 160         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 161                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 162                        "alloc_lseg and free_lseg.\n", __func__);
 163                 return status;
 164         }
 165
 166         spin_lock(&pnfs_spinlock);
 167         tmp = find_pnfs_driver_locked(ld_type->id);
 168         if (!tmp) {
 169                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 170                 status = 0;
 171                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 172                         ld_type->name);
 173         } else {
 174                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 175                         __func__, ld_type->id);
 176         }
 177         spin_unlock(&pnfs_spinlock);
 178
 179         return status;
 180 }
 181 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 182
 183 void
 184 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 185 {
 186         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 187         spin_lock(&pnfs_spinlock);
 188         list_del(&ld_type->pnfs_tblid);
 189         spin_unlock(&pnfs_spinlock);
 190 }
 191 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 192
 193 /*
 194  * pNFS client layout cache
 195  */
 196
 197 /* Need to hold i_lock if caller does not already hold reference */
 198 void
 199 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 200 {
 201         atomic_inc(&lo->plh_refcount);
 202 }
 203
 204 static struct pnfs_layout_hdr *
 205 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 206 {
 207         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 208         return ld->alloc_layout_hdr(ino, gfp_flags);
 209 }
 210
 211 static void
 212 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 213 {
 214         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 215         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 216
 217         if (!list_empty(&lo->plh_layouts)) {
 218                 struct nfs_client *clp = server->nfs_client;
 219
 220                 spin_lock(&clp->cl_lock);
 221                 list_del_init(&lo->plh_layouts);
 222                 spin_unlock(&clp->cl_lock);
 223         }
 224         put_rpccred(lo->plh_lc_cred);
 225         return ld->free_layout_hdr(lo);
 226 }
 227
 228 static void
 229 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 230 {
 231         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 232         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 233         nfsi->layout = NULL;
 234         /* Reset MDS Threshold I/O counters */
 235         nfsi->write_io = 0;
 236         nfsi->read_io = 0;
 237 }
 238
 239 void
 240 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 241 {
 242         struct inode *inode = lo->plh_inode;
 243
 244         pnfs_layoutreturn_before_put_layout_hdr(lo);
 245
 246         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 247                 if (!list_empty(&lo->plh_segs))
 248                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 249                 pnfs_detach_layout_hdr(lo);
 250                 spin_unlock(&inode->i_lock);
 251                 pnfs_free_layout_hdr(lo);
 252         }
 253 }
 254
 255 /*
 256  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 257  *
 258  * In order to continue using the pnfs_layout_hdr, a full recovery
 259  * is required.
 260  * Note that caller must hold inode->i_lock.
 261  */
 262 static void
 263 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 264                 struct list_head *lseg_list)
 265 {
 266         struct pnfs_layout_range range = {
 267                 .iomode = IOMODE_ANY,
 268                 .offset = 0,
 269                 .length = NFS4_MAX_UINT64,
 270         };
 271
 272         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 273         pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range);
 274 }
 275
 276 static int
 277 pnfs_iomode_to_fail_bit(u32 iomode)
 278 {
 279         return iomode == IOMODE_RW ?
 280                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 281 }
 282
 283 static void
 284 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 285 {
 286         lo->plh_retry_timestamp = jiffies;
 287         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 288                 atomic_inc(&lo->plh_refcount);
 289 }
 290
 291 static void
 292 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 293 {
 294         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 295                 atomic_dec(&lo->plh_refcount);
 296 }
 297
 298 static void
 299 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 300 {
 301         struct inode *inode = lo->plh_inode;
 302         struct pnfs_layout_range range = {
 303                 .iomode = iomode,
 304                 .offset = 0,
 305                 .length = NFS4_MAX_UINT64,
 306         };
 307         LIST_HEAD(head);
 308
 309         spin_lock(&inode->i_lock);
 310         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 311         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 312         spin_unlock(&inode->i_lock);
 313         pnfs_free_lseg_list(&head);
 314         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 315                         iomode == IOMODE_RW ?  "RW" : "READ");
 316 }
 317
 318 static bool
 319 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 320 {
 321         unsigned long start, end;
 322         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 323
 324         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 325                 return false;
 326         end = jiffies;
 327         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 328         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 329                 /* It is time to retry the failed layoutgets */
 330                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 331                 return false;
 332         }
 333         return true;
 334 }
 335
 336 static void
 337 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 338 {
 339         INIT_LIST_HEAD(&lseg->pls_list);
 340         INIT_LIST_HEAD(&lseg->pls_lc_list);
 341         atomic_set(&lseg->pls_refcount, 1);
 342         smp_mb();
 343         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 344         lseg->pls_layout = lo;
 345 }
 346
 347 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 348 {
 349         struct inode *ino = lseg->pls_layout->plh_inode;
 350
 351         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 352 }
 353
 354 static void
 355 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 356                 struct pnfs_layout_segment *lseg)
 357 {
 358         struct inode *inode = lo->plh_inode;
 359
 360         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 361         list_del_init(&lseg->pls_list);
 362         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 363         atomic_dec(&lo->plh_refcount);
 364         if (list_empty(&lo->plh_segs))
 365                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 366         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 367 }
 368
 369 void
 370 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 371 {
 372         struct pnfs_layout_hdr *lo;
 373         struct inode *inode;
 374
 375         if (!lseg)
 376                 return;
 377
 378         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 379                 atomic_read(&lseg->pls_refcount),
 380                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 381
 382         lo = lseg->pls_layout;
 383         inode = lo->plh_inode;
 384
 385         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 386                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 387                         spin_unlock(&inode->i_lock);
 388                         return;
 389                 }
 390                 pnfs_get_layout_hdr(lo);
 391                 pnfs_layout_remove_lseg(lo, lseg);
 392                 spin_unlock(&inode->i_lock);
 393                 pnfs_free_lseg(lseg);
 394                 pnfs_put_layout_hdr(lo);
 395         }
 396 }
 397 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 398
 399 static void pnfs_free_lseg_async_work(struct work_struct *work)
 400 {
 401         struct pnfs_layout_segment *lseg;
 402         struct pnfs_layout_hdr *lo;
 403
 404         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 405         lo = lseg->pls_layout;
 406
 407         pnfs_free_lseg(lseg);
 408         pnfs_put_layout_hdr(lo);
 409 }
 410
 411 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 412 {
 413         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 414         schedule_work(&lseg->pls_work);
 415 }
 416
 417 void
 418 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 419 {
 420         if (!lseg)
 421                 return;
 422
 423         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 424
 425         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 426                 atomic_read(&lseg->pls_refcount),
 427                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 428         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 429                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 430                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 431                         return;
 432                 pnfs_get_layout_hdr(lo);
 433                 pnfs_layout_remove_lseg(lo, lseg);
 434                 pnfs_free_lseg_async(lseg);
 435         }
 436 }
 437 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 438
 439 static u64
 440 end_offset(u64 start, u64 len)
 441 {
 442         u64 end;
 443
 444         end = start + len;
 445         return end >= start ? end : NFS4_MAX_UINT64;
 446 }
 447
 448 /*
 449  * is l2 fully contained in l1?
 450  *   start1                             end1
 451  *   [----------------------------------)
 452  *           start2           end2
 453  *           [----------------)
 454  */
 455 static bool
 456 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 457                  const struct pnfs_layout_range *l2)
 458 {
 459         u64 start1 = l1->offset;
 460         u64 end1 = end_offset(start1, l1->length);
 461         u64 start2 = l2->offset;
 462         u64 end2 = end_offset(start2, l2->length);
 463
 464         return (start1 <= start2) && (end1 >= end2);
 465 }
 466
 467 /*
 468  * is l1 and l2 intersecting?
 469  *   start1                             end1
 470  *   [----------------------------------)
 471  *                              start2           end2
 472  *                              [----------------)
 473  */
 474 static bool
 475 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 476                     const struct pnfs_layout_range *l2)
 477 {
 478         u64 start1 = l1->offset;
 479         u64 end1 = end_offset(start1, l1->length);
 480         u64 start2 = l2->offset;
 481         u64 end2 = end_offset(start2, l2->length);
 482
 483         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 484                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 485 }
 486
 487 static bool
 488 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 489                  const struct pnfs_layout_range *recall_range)
 490 {
 491         return (recall_range->iomode == IOMODE_ANY ||
 492                 lseg_range->iomode == recall_range->iomode) &&
 493                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 494 }
 495
 496 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 497                 struct list_head *tmp_list)
 498 {
 499         if (!atomic_dec_and_test(&lseg->pls_refcount))
 500                 return false;
 501         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 502         list_add(&lseg->pls_list, tmp_list);
 503         return true;
 504 }
 505
 506 /* Returns 1 if lseg is removed from list, 0 otherwise */
 507 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 508                              struct list_head *tmp_list)
 509 {
 510         int rv = 0;
 511
 512         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 513                 /* Remove the reference keeping the lseg in the
 514                  * list.  It will now be removed when all
 515                  * outstanding io is finished.
 516                  */
 517                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 518                         atomic_read(&lseg->pls_refcount));
 519                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 520                         rv = 1;
 521         }
 522         return rv;
 523 }
 524
 525 /* Returns count of number of matching invalid lsegs remaining in list
 526  * after call.
 527  */
 528 int
 529 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 530                             struct list_head *tmp_list,
 531                             const struct pnfs_layout_range *recall_range)
 532 {
 533         struct pnfs_layout_segment *lseg, *next;
 534         int remaining = 0;
 535
 536         dprintk("%s:Begin lo %p\n", __func__, lo);
 537
 538         if (list_empty(&lo->plh_segs))
 539                 return 0;
 540         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 541                 if (!recall_range ||
 542                     should_free_lseg(&lseg->pls_range, recall_range)) {
 543                         dprintk("%s: freeing lseg %p iomode %d "
 544                                 "offset %llu length %llu\n", __func__,
 545                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 546                                 lseg->pls_range.length);
 547                         if (!mark_lseg_invalid(lseg, tmp_list))
 548                                 remaining++;
 549                 }
 550         dprintk("%s:Return %i\n", __func__, remaining);
 551         return remaining;
 552 }
 553
 554 /* note free_me must contain lsegs from a single layout_hdr */
 555 void
 556 pnfs_free_lseg_list(struct list_head *free_me)
 557 {
 558         struct pnfs_layout_segment *lseg, *tmp;
 559
 560         if (list_empty(free_me))
 561                 return;
 562
 563         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 564                 list_del(&lseg->pls_list);
 565                 pnfs_free_lseg(lseg);
 566         }
 567 }
 568
 569 void
 570 pnfs_destroy_layout(struct nfs_inode *nfsi)
 571 {
 572         struct pnfs_layout_hdr *lo;
 573         LIST_HEAD(tmp_list);
 574
 575         spin_lock(&nfsi->vfs_inode.i_lock);
 576         lo = nfsi->layout;
 577         if (lo) {
 578                 pnfs_get_layout_hdr(lo);
 579                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 580                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 581                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 582                 spin_unlock(&nfsi->vfs_inode.i_lock);
 583                 pnfs_free_lseg_list(&tmp_list);
 584                 pnfs_put_layout_hdr(lo);
 585         } else
 586                 spin_unlock(&nfsi->vfs_inode.i_lock);
 587 }
 588 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 589
 590 static bool
 591 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 592                 struct list_head *layout_list)
 593 {
 594         struct pnfs_layout_hdr *lo;
 595         bool ret = false;
 596
 597         spin_lock(&inode->i_lock);
 598         lo = NFS_I(inode)->layout;
 599         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 600                 pnfs_get_layout_hdr(lo);
 601                 list_add(&lo->plh_bulk_destroy, layout_list);
 602                 ret = true;
 603         }
 604         spin_unlock(&inode->i_lock);
 605         return ret;
 606 }
 607
 608 /* Caller must hold rcu_read_lock and clp->cl_lock */
 609 static int
 610 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 611                 struct nfs_server *server,
 612                 struct list_head *layout_list)
 613 {
 614         struct pnfs_layout_hdr *lo, *next;
 615         struct inode *inode;
 616
 617         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 618                 inode = igrab(lo->plh_inode);
 619                 if (inode == NULL)
 620                         continue;
 621                 list_del_init(&lo->plh_layouts);
 622                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 623                         continue;
 624                 rcu_read_unlock();
 625                 spin_unlock(&clp->cl_lock);
 626                 iput(inode);
 627                 spin_lock(&clp->cl_lock);
 628                 rcu_read_lock();
 629                 return -EAGAIN;
 630         }
 631         return 0;
 632 }
 633
 634 static int
 635 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 636                 bool is_bulk_recall)
 637 {
 638         struct pnfs_layout_hdr *lo;
 639         struct inode *inode;
 640         struct pnfs_layout_range range = {
 641                 .iomode = IOMODE_ANY,
 642                 .offset = 0,
 643                 .length = NFS4_MAX_UINT64,
 644         };
 645         LIST_HEAD(lseg_list);
 646         int ret = 0;
 647
 648         while (!list_empty(layout_list)) {
 649                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 650                                 plh_bulk_destroy);
 651                 dprintk("%s freeing layout for inode %lu\n", __func__,
 652                         lo->plh_inode->i_ino);
 653                 inode = lo->plh_inode;
 654
 655                 pnfs_layoutcommit_inode(inode, false);
 656
 657                 spin_lock(&inode->i_lock);
 658                 list_del_init(&lo->plh_bulk_destroy);
 659                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 660                 if (is_bulk_recall)
 661                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 662                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 663                         ret = -EAGAIN;
 664                 spin_unlock(&inode->i_lock);
 665                 pnfs_free_lseg_list(&lseg_list);
 666                 /* Free all lsegs that are attached to commit buckets */
 667                 nfs_commit_inode(inode, 0);
 668                 pnfs_put_layout_hdr(lo);
 669                 iput(inode);
 670         }
 671         return ret;
 672 }
 673
 674 int
 675 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 676                 struct nfs_fsid *fsid,
 677                 bool is_recall)
 678 {
 679         struct nfs_server *server;
 680         LIST_HEAD(layout_list);
 681
 682         spin_lock(&clp->cl_lock);
 683         rcu_read_lock();
 684 restart:
 685         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 686                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 687                         continue;
 688                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 689                                 server,
 690                                 &layout_list) != 0)
 691                         goto restart;
 692         }
 693         rcu_read_unlock();
 694         spin_unlock(&clp->cl_lock);
 695
 696         if (list_empty(&layout_list))
 697                 return 0;
 698         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 699 }
 700
 701 int
 702 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 703                 bool is_recall)
 704 {
 705         struct nfs_server *server;
 706         LIST_HEAD(layout_list);
 707
 708         spin_lock(&clp->cl_lock);
 709         rcu_read_lock();
 710 restart:
 711         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 712                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 713                                         server,
 714                                         &layout_list) != 0)
 715                         goto restart;
 716         }
 717         rcu_read_unlock();
 718         spin_unlock(&clp->cl_lock);
 719
 720         if (list_empty(&layout_list))
 721                 return 0;
 722         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 723 }
 724
 725 /*
 726  * Called by the state manger to remove all layouts established under an
 727  * expired lease.
 728  */
 729 void
 730 pnfs_destroy_all_layouts(struct nfs_client *clp)
 731 {
 732         nfs4_deviceid_mark_client_invalid(clp);
 733         nfs4_deviceid_purge_client(clp);
 734
 735         pnfs_destroy_layouts_byclid(clp, false);
 736 }
 737
 738 /*
 739  * Compare 2 layout stateid sequence ids, to see which is newer,
 740  * taking into account wraparound issues.
 741  */
 742 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 743 {
 744         return (s32)(s1 - s2) > 0;
 745 }
 746
 747 /* update lo->plh_stateid with new if is more recent */
 748 void
 749 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 750                         bool update_barrier)
 751 {
 752         u32 oldseq, newseq, new_barrier;
 753         int empty = list_empty(&lo->plh_segs);
 754
 755         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 756         newseq = be32_to_cpu(new->seqid);
 757         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 758                 nfs4_stateid_copy(&lo->plh_stateid, new);
 759                 if (update_barrier) {
 760                         new_barrier = be32_to_cpu(new->seqid);
 761                 } else {
 762                         /* Because of wraparound, we want to keep the barrier
 763                          * "close" to the current seqids.
 764                          */
 765                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 766                 }
 767                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 768                         lo->plh_barrier = new_barrier;
 769         }
 770 }
 771
 772 static bool
 773 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 774                 const nfs4_stateid *stateid)
 775 {
 776         u32 seqid = be32_to_cpu(stateid->seqid);
 777
 778         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 779 }
 780
 781 /* lget is set to 1 if called from inside send_layoutget call chain */
 782 static bool
 783 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 784 {
 785         return lo->plh_block_lgets ||
 786                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 787 }
 788
 789 int
 790 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 791                               const struct pnfs_layout_range *range,
 792                               struct nfs4_state *open_state)
 793 {
 794         int status = 0;
 795
 796         dprintk("--> %s\n", __func__);
 797         spin_lock(&lo->plh_inode->i_lock);
 798         if (pnfs_layoutgets_blocked(lo)) {
 799                 status = -EAGAIN;
 800         } else if (!nfs4_valid_open_stateid(open_state)) {
 801                 status = -EBADF;
 802         } else if (list_empty(&lo->plh_segs) ||
 803                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 804                 int seq;
 805
 806                 do {
 807                         seq = read_seqbegin(&open_state->seqlock);
 808                         nfs4_stateid_copy(dst, &open_state->stateid);
 809                 } while (read_seqretry(&open_state->seqlock, seq));
 810         } else
 811                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 812         spin_unlock(&lo->plh_inode->i_lock);
 813         dprintk("<-- %s\n", __func__);
 814         return status;
 815 }
 816
 817 /*
 818 * Get layout from server.
 819 *    for now, assume that whole file layouts are requested.
 820 *    arg->offset: 0
 821 *    arg->length: all ones
 822 */
 823 static struct pnfs_layout_segment *
 824 send_layoutget(struct pnfs_layout_hdr *lo,
 825            struct nfs_open_context *ctx,
 826            const struct pnfs_layout_range *range,
 827            gfp_t gfp_flags)
 828 {
 829         struct inode *ino = lo->plh_inode;
 830         struct nfs_server *server = NFS_SERVER(ino);
 831         struct nfs4_layoutget *lgp;
 832         struct pnfs_layout_segment *lseg;
 833         loff_t i_size;
 834
 835         dprintk("--> %s\n", __func__);
 836
 837         /*
 838          * Synchronously retrieve layout information from server and
 839          * store in lseg. If we race with a concurrent seqid morphing
 840          * op, then re-send the LAYOUTGET.
 841          */
 842         do {
 843                 lgp = kzalloc(sizeof(*lgp), gfp_flags);
 844                 if (lgp == NULL)
 845                         return NULL;
 846
 847                 i_size = i_size_read(ino);
 848
 849                 lgp->args.minlength = PAGE_CACHE_SIZE;
 850                 if (lgp->args.minlength > range->length)
 851                         lgp->args.minlength = range->length;
 852                 if (range->iomode == IOMODE_READ) {
 853                         if (range->offset >= i_size)
 854                                 lgp->args.minlength = 0;
 855                         else if (i_size - range->offset < lgp->args.minlength)
 856                                 lgp->args.minlength = i_size - range->offset;
 857                 }
 858                 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 859                 pnfs_copy_range(&lgp->args.range, range);
 860                 lgp->args.type = server->pnfs_curr_ld->id;
 861                 lgp->args.inode = ino;
 862                 lgp->args.ctx = get_nfs_open_context(ctx);
 863                 lgp->gfp_flags = gfp_flags;
 864                 lgp->cred = lo->plh_lc_cred;
 865
 866                 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 867         } while (lseg == ERR_PTR(-EAGAIN));
 868
 869         if (IS_ERR(lseg) && !nfs_error_is_fatal(PTR_ERR(lseg)))
 870                 lseg = NULL;
 871         else
 872                 pnfs_layout_clear_fail_bit(lo,
 873                                 pnfs_iomode_to_fail_bit(range->iomode));
 874
 875         return lseg;
 876 }
 877
 878 static void pnfs_clear_layoutcommit(struct inode *inode,
 879                 struct list_head *head)
 880 {
 881         struct nfs_inode *nfsi = NFS_I(inode);
 882         struct pnfs_layout_segment *lseg, *tmp;
 883
 884         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 885                 return;
 886         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 887                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 888                         continue;
 889                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 890         }
 891 }
 892
 893 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 894 {
 895         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 896         smp_mb__after_atomic();
 897         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 898         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 899 }
 900
 901 static bool
 902 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 903 {
 904         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 905                 return false;
 906         lo->plh_return_iomode = 0;
 907         pnfs_get_layout_hdr(lo);
 908         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 909         return true;
 910 }
 911
 912 static int
 913 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
 914                        enum pnfs_iomode iomode, bool sync)
 915 {
 916         struct inode *ino = lo->plh_inode;
 917         struct nfs4_layoutreturn *lrp;
 918         int status = 0;
 919
 920         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 921         if (unlikely(lrp == NULL)) {
 922                 status = -ENOMEM;
 923                 spin_lock(&ino->i_lock);
 924                 pnfs_clear_layoutreturn_waitbit(lo);
 925                 spin_unlock(&ino->i_lock);
 926                 pnfs_put_layout_hdr(lo);
 927                 goto out;
 928         }
 929
 930         nfs4_stateid_copy(&lrp->args.stateid, stateid);
 931         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 932         lrp->args.inode = ino;
 933         lrp->args.range.iomode = iomode;
 934         lrp->args.range.offset = 0;
 935         lrp->args.range.length = NFS4_MAX_UINT64;
 936         lrp->args.layout = lo;
 937         lrp->clp = NFS_SERVER(ino)->nfs_client;
 938         lrp->cred = lo->plh_lc_cred;
 939
 940         status = nfs4_proc_layoutreturn(lrp, sync);
 941 out:
 942         dprintk("<-- %s status: %d\n", __func__, status);
 943         return status;
 944 }
 945
 946 /* Return true if layoutreturn is needed */
 947 static bool
 948 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
 949 {
 950         struct pnfs_layout_segment *s;
 951
 952         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 953                 return false;
 954
 955         /* Defer layoutreturn until all lsegs are done */
 956         list_for_each_entry(s, &lo->plh_segs, pls_list) {
 957                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 958                         return false;
 959         }
 960
 961         return true;
 962 }
 963
 964 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
 965 {
 966         struct inode *inode= lo->plh_inode;
 967
 968         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 969                 return;
 970         spin_lock(&inode->i_lock);
 971         if (pnfs_layout_need_return(lo)) {
 972                 nfs4_stateid stateid;
 973                 enum pnfs_iomode iomode;
 974                 bool send;
 975
 976                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
 977                 iomode = lo->plh_return_iomode;
 978                 send = pnfs_prepare_layoutreturn(lo);
 979                 spin_unlock(&inode->i_lock);
 980                 if (send) {
 981                         /* Send an async layoutreturn so we dont deadlock */
 982                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
 983                 }
 984         } else
 985                 spin_unlock(&inode->i_lock);
 986 }
 987
 988 /*
 989  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 990  * when the layout segment list is empty.
 991  *
 992  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 993  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 994  * deviceid is marked invalid.
 995  */
 996 int
 997 _pnfs_return_layout(struct inode *ino)
 998 {
 999         struct pnfs_layout_hdr *lo = NULL;
1000         struct nfs_inode *nfsi = NFS_I(ino);
1001         LIST_HEAD(tmp_list);
1002         nfs4_stateid stateid;
1003         int status = 0, empty;
1004         bool send;
1005
1006         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1007
1008         spin_lock(&ino->i_lock);
1009         lo = nfsi->layout;
1010         if (!lo) {
1011                 spin_unlock(&ino->i_lock);
1012                 dprintk("NFS: %s no layout to return\n", __func__);
1013                 goto out;
1014         }
1015         nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
1016         /* Reference matched in nfs4_layoutreturn_release */
1017         pnfs_get_layout_hdr(lo);
1018         empty = list_empty(&lo->plh_segs);
1019         pnfs_clear_layoutcommit(ino, &tmp_list);
1020         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1021
1022         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1023                 struct pnfs_layout_range range = {
1024                         .iomode         = IOMODE_ANY,
1025                         .offset         = 0,
1026                         .length         = NFS4_MAX_UINT64,
1027                 };
1028                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1029         }
1030
1031         /* Don't send a LAYOUTRETURN if list was initially empty */
1032         if (empty) {
1033                 spin_unlock(&ino->i_lock);
1034                 dprintk("NFS: %s no layout segments to return\n", __func__);
1035                 goto out_put_layout_hdr;
1036         }
1037
1038         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1039         send = pnfs_prepare_layoutreturn(lo);
1040         spin_unlock(&ino->i_lock);
1041         pnfs_free_lseg_list(&tmp_list);
1042         if (send)
1043                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1044 out_put_layout_hdr:
1045         pnfs_put_layout_hdr(lo);
1046 out:
1047         dprintk("<-- %s status: %d\n", __func__, status);
1048         return status;
1049 }
1050 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1051
1052 int
1053 pnfs_commit_and_return_layout(struct inode *inode)
1054 {
1055         struct pnfs_layout_hdr *lo;
1056         int ret;
1057
1058         spin_lock(&inode->i_lock);
1059         lo = NFS_I(inode)->layout;
1060         if (lo == NULL) {
1061                 spin_unlock(&inode->i_lock);
1062                 return 0;
1063         }
1064         pnfs_get_layout_hdr(lo);
1065         /* Block new layoutgets and read/write to ds */
1066         lo->plh_block_lgets++;
1067         spin_unlock(&inode->i_lock);
1068         filemap_fdatawait(inode->i_mapping);
1069         ret = pnfs_layoutcommit_inode(inode, true);
1070         if (ret == 0)
1071                 ret = _pnfs_return_layout(inode);
1072         spin_lock(&inode->i_lock);
1073         lo->plh_block_lgets--;
1074         spin_unlock(&inode->i_lock);
1075         pnfs_put_layout_hdr(lo);
1076         return ret;
1077 }
1078
1079 bool pnfs_roc(struct inode *ino)
1080 {
1081         struct nfs_inode *nfsi = NFS_I(ino);
1082         struct nfs_open_context *ctx;
1083         struct nfs4_state *state;
1084         struct pnfs_layout_hdr *lo;
1085         struct pnfs_layout_segment *lseg, *tmp;
1086         nfs4_stateid stateid;
1087         LIST_HEAD(tmp_list);
1088         bool found = false, layoutreturn = false, roc = false;
1089
1090         spin_lock(&ino->i_lock);
1091         lo = nfsi->layout;
1092         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1093                 goto out_noroc;
1094
1095         /* no roc if we hold a delegation */
1096         if (nfs4_check_delegation(ino, FMODE_READ))
1097                 goto out_noroc;
1098
1099         list_for_each_entry(ctx, &nfsi->open_files, list) {
1100                 state = ctx->state;
1101                 /* Don't return layout if there is open file state */
1102                 if (state != NULL && state->state != 0)
1103                         goto out_noroc;
1104         }
1105
1106         nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1107         /* always send layoutreturn if being marked so */
1108         if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
1109                                    &lo->plh_flags))
1110                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1111
1112         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1113                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1114                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1115                         mark_lseg_invalid(lseg, &tmp_list);
1116                         found = true;
1117                 }
1118         /* ROC in two conditions:
1119          * 1. there are ROC lsegs
1120          * 2. we don't send layoutreturn
1121          */
1122         if (found && !layoutreturn) {
1123                 /* lo ref dropped in pnfs_roc_release() */
1124                 pnfs_get_layout_hdr(lo);
1125                 roc = true;
1126         }
1127
1128 out_noroc:
1129         spin_unlock(&ino->i_lock);
1130         pnfs_free_lseg_list(&tmp_list);
1131         pnfs_layoutcommit_inode(ino, true);
1132         if (layoutreturn)
1133                 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1134         return roc;
1135 }
1136
1137 void pnfs_roc_release(struct inode *ino)
1138 {
1139         struct pnfs_layout_hdr *lo;
1140
1141         spin_lock(&ino->i_lock);
1142         lo = NFS_I(ino)->layout;
1143         pnfs_clear_layoutreturn_waitbit(lo);
1144         if (atomic_dec_and_test(&lo->plh_refcount)) {
1145                 pnfs_detach_layout_hdr(lo);
1146                 spin_unlock(&ino->i_lock);
1147                 pnfs_free_layout_hdr(lo);
1148         } else
1149                 spin_unlock(&ino->i_lock);
1150 }
1151
1152 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1153 {
1154         struct pnfs_layout_hdr *lo;
1155
1156         spin_lock(&ino->i_lock);
1157         lo = NFS_I(ino)->layout;
1158         pnfs_mark_layout_returned_if_empty(lo);
1159         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1160                 lo->plh_barrier = barrier;
1161         spin_unlock(&ino->i_lock);
1162         trace_nfs4_layoutreturn_on_close(ino, 0);
1163 }
1164
1165 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1166 {
1167         struct nfs_inode *nfsi = NFS_I(ino);
1168         struct pnfs_layout_hdr *lo;
1169         u32 current_seqid;
1170
1171         spin_lock(&ino->i_lock);
1172         lo = nfsi->layout;
1173         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1174
1175         /* Since close does not return a layout stateid for use as
1176          * a barrier, we choose the worst-case barrier.
1177          */
1178         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1179         spin_unlock(&ino->i_lock);
1180 }
1181
1182 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1183 {
1184         struct nfs_inode *nfsi = NFS_I(ino);
1185         struct pnfs_layout_hdr *lo;
1186         bool sleep = false;
1187
1188         /* we might not have grabbed lo reference. so need to check under
1189          * i_lock */
1190         spin_lock(&ino->i_lock);
1191         lo = nfsi->layout;
1192         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1193                 sleep = true;
1194         spin_unlock(&ino->i_lock);
1195
1196         if (sleep)
1197                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1198
1199         return sleep;
1200 }
1201
1202 /*
1203  * Compare two layout segments for sorting into layout cache.
1204  * We want to preferentially return RW over RO layouts, so ensure those
1205  * are seen first.
1206  */
1207 static s64
1208 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1209            const struct pnfs_layout_range *l2)
1210 {
1211         s64 d;
1212
1213         /* high offset > low offset */
1214         d = l1->offset - l2->offset;
1215         if (d)
1216                 return d;
1217
1218         /* short length > long length */
1219         d = l2->length - l1->length;
1220         if (d)
1221                 return d;
1222
1223         /* read > read/write */
1224         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1225 }
1226
1227 static bool
1228 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1229                 const struct pnfs_layout_range *l2)
1230 {
1231         return pnfs_lseg_range_cmp(l1, l2) > 0;
1232 }
1233
1234 static bool
1235 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1236                 struct pnfs_layout_segment *old)
1237 {
1238         return false;
1239 }
1240
1241 void
1242 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1243                    struct pnfs_layout_segment *lseg,
1244                    bool (*is_after)(const struct pnfs_layout_range *,
1245                            const struct pnfs_layout_range *),
1246                    bool (*do_merge)(struct pnfs_layout_segment *,
1247                            struct pnfs_layout_segment *),
1248                    struct list_head *free_me)
1249 {
1250         struct pnfs_layout_segment *lp, *tmp;
1251
1252         dprintk("%s:Begin\n", __func__);
1253
1254         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1255                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1256                         continue;
1257                 if (do_merge(lseg, lp)) {
1258                         mark_lseg_invalid(lp, free_me);
1259                         continue;
1260                 }
1261                 if (is_after(&lseg->pls_range, &lp->pls_range))
1262                         continue;
1263                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1264                 dprintk("%s: inserted lseg %p "
1265                         "iomode %d offset %llu length %llu before "
1266                         "lp %p iomode %d offset %llu length %llu\n",
1267                         __func__, lseg, lseg->pls_range.iomode,
1268                         lseg->pls_range.offset, lseg->pls_range.length,
1269                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1270                         lp->pls_range.length);
1271                 goto out;
1272         }
1273         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1274         dprintk("%s: inserted lseg %p "
1275                 "iomode %d offset %llu length %llu at tail\n",
1276                 __func__, lseg, lseg->pls_range.iomode,
1277                 lseg->pls_range.offset, lseg->pls_range.length);
1278 out:
1279         pnfs_get_layout_hdr(lo);
1280
1281         dprintk("%s:Return\n", __func__);
1282 }
1283 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1284
1285 static void
1286 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1287                    struct pnfs_layout_segment *lseg,
1288                    struct list_head *free_me)
1289 {
1290         struct inode *inode = lo->plh_inode;
1291         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1292
1293         if (ld->add_lseg != NULL)
1294                 ld->add_lseg(lo, lseg, free_me);
1295         else
1296                 pnfs_generic_layout_insert_lseg(lo, lseg,
1297                                 pnfs_lseg_range_is_after,
1298                                 pnfs_lseg_no_merge,
1299                                 free_me);
1300 }
1301
1302 static struct pnfs_layout_hdr *
1303 alloc_init_layout_hdr(struct inode *ino,
1304                       struct nfs_open_context *ctx,
1305                       gfp_t gfp_flags)
1306 {
1307         struct pnfs_layout_hdr *lo;
1308
1309         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1310         if (!lo)
1311                 return NULL;
1312         atomic_set(&lo->plh_refcount, 1);
1313         INIT_LIST_HEAD(&lo->plh_layouts);
1314         INIT_LIST_HEAD(&lo->plh_segs);
1315         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1316         lo->plh_inode = ino;
1317         lo->plh_lc_cred = get_rpccred(ctx->cred);
1318         return lo;
1319 }
1320
1321 static struct pnfs_layout_hdr *
1322 pnfs_find_alloc_layout(struct inode *ino,
1323                        struct nfs_open_context *ctx,
1324                        gfp_t gfp_flags)
1325 {
1326         struct nfs_inode *nfsi = NFS_I(ino);
1327         struct pnfs_layout_hdr *new = NULL;
1328
1329         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1330
1331         if (nfsi->layout != NULL)
1332                 goto out_existing;
1333         spin_unlock(&ino->i_lock);
1334         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1335         spin_lock(&ino->i_lock);
1336
1337         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1338                 nfsi->layout = new;
1339                 return new;
1340         } else if (new != NULL)
1341                 pnfs_free_layout_hdr(new);
1342 out_existing:
1343         pnfs_get_layout_hdr(nfsi->layout);
1344         return nfsi->layout;
1345 }
1346
1347 /*
1348  * iomode matching rules:
1349  * iomode       lseg    match
1350  * -----        -----   -----
1351  * ANY          READ    true
1352  * ANY          RW      true
1353  * RW           READ    false
1354  * RW           RW      true
1355  * READ         READ    true
1356  * READ         RW      true
1357  */
1358 static bool
1359 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1360                  const struct pnfs_layout_range *range)
1361 {
1362         struct pnfs_layout_range range1;
1363
1364         if ((range->iomode == IOMODE_RW &&
1365              ls_range->iomode != IOMODE_RW) ||
1366             !pnfs_lseg_range_intersecting(ls_range, range))
1367                 return 0;
1368
1369         /* range1 covers only the first byte in the range */
1370         range1 = *range;
1371         range1.length = 1;
1372         return pnfs_lseg_range_contained(ls_range, &range1);
1373 }
1374
1375 /*
1376  * lookup range in layout
1377  */
1378 static struct pnfs_layout_segment *
1379 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1380                 struct pnfs_layout_range *range)
1381 {
1382         struct pnfs_layout_segment *lseg, *ret = NULL;
1383
1384         dprintk("%s:Begin\n", __func__);
1385
1386         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1387                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1388                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1389                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1390                         ret = pnfs_get_lseg(lseg);
1391                         break;
1392                 }
1393         }
1394
1395         dprintk("%s:Return lseg %p ref %d\n",
1396                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1397         return ret;
1398 }
1399
1400 /*
1401  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1402  * to the MDS or over pNFS
1403  *
1404  * The nfs_inode read_io and write_io fields are cumulative counters reset
1405  * when there are no layout segments. Note that in pnfs_update_layout iomode
1406  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1407  * WRITE request.
1408  *
1409  * A return of true means use MDS I/O.
1410  *
1411  * From rfc 5661:
1412  * If a file's size is smaller than the file size threshold, data accesses
1413  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1414  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1415  * server.  If both file size and I/O size are provided, the client SHOULD
1416  * reach or exceed  both thresholds before sending its read or write
1417  * requests to the data server.
1418  */
1419 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1420                                      struct inode *ino, int iomode)
1421 {
1422         struct nfs4_threshold *t = ctx->mdsthreshold;
1423         struct nfs_inode *nfsi = NFS_I(ino);
1424         loff_t fsize = i_size_read(ino);
1425         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1426
1427         if (t == NULL)
1428                 return ret;
1429
1430         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1431                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1432
1433         switch (iomode) {
1434         case IOMODE_READ:
1435                 if (t->bm & THRESHOLD_RD) {
1436                         dprintk("%s fsize %llu\n", __func__, fsize);
1437                         size_set = true;
1438                         if (fsize < t->rd_sz)
1439                                 size = true;
1440                 }
1441                 if (t->bm & THRESHOLD_RD_IO) {
1442                         dprintk("%s nfsi->read_io %llu\n", __func__,
1443                                 nfsi->read_io);
1444                         io_set = true;
1445                         if (nfsi->read_io < t->rd_io_sz)
1446                                 io = true;
1447                 }
1448                 break;
1449         case IOMODE_RW:
1450                 if (t->bm & THRESHOLD_WR) {
1451                         dprintk("%s fsize %llu\n", __func__, fsize);
1452                         size_set = true;
1453                         if (fsize < t->wr_sz)
1454                                 size = true;
1455                 }
1456                 if (t->bm & THRESHOLD_WR_IO) {
1457                         dprintk("%s nfsi->write_io %llu\n", __func__,
1458                                 nfsi->write_io);
1459                         io_set = true;
1460                         if (nfsi->write_io < t->wr_io_sz)
1461                                 io = true;
1462                 }
1463                 break;
1464         }
1465         if (size_set && io_set) {
1466                 if (size && io)
1467                         ret = true;
1468         } else if (size || io)
1469                 ret = true;
1470
1471         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1472         return ret;
1473 }
1474
1475 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1476 {
1477         /*
1478          * send layoutcommit as it can hold up layoutreturn due to lseg
1479          * reference
1480          */
1481         pnfs_layoutcommit_inode(lo->plh_inode, false);
1482         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1483                                    nfs_wait_bit_killable,
1484                                    TASK_UNINTERRUPTIBLE);
1485 }
1486
1487 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1488 {
1489         unsigned long *bitlock = &lo->plh_flags;
1490
1491         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1492         smp_mb__after_atomic();
1493         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1494 }
1495
1496 /*
1497  * Layout segment is retreived from the server if not cached.
1498  * The appropriate layout segment is referenced and returned to the caller.
1499  */
1500 struct pnfs_layout_segment *
1501 pnfs_update_layout(struct inode *ino,
1502                    struct nfs_open_context *ctx,
1503                    loff_t pos,
1504                    u64 count,
1505                    enum pnfs_iomode iomode,
1506                    gfp_t gfp_flags)
1507 {
1508         struct pnfs_layout_range arg = {
1509                 .iomode = iomode,
1510                 .offset = pos,
1511                 .length = count,
1512         };
1513         unsigned pg_offset;
1514         struct nfs_server *server = NFS_SERVER(ino);
1515         struct nfs_client *clp = server->nfs_client;
1516         struct pnfs_layout_hdr *lo;
1517         struct pnfs_layout_segment *lseg = NULL;
1518         bool first;
1519
1520         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1521                 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1522                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1523                 goto out;
1524         }
1525
1526         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1527                 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1528                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1529                 goto out;
1530         }
1531
1532         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1533                 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1534                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1535                 goto out;
1536         }
1537
1538 lookup_again:
1539         first = false;
1540         spin_lock(&ino->i_lock);
1541         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1542         if (lo == NULL) {
1543                 spin_unlock(&ino->i_lock);
1544                 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1545                                  PNFS_UPDATE_LAYOUT_NOMEM);
1546                 goto out;
1547         }
1548
1549         /* Do we even need to bother with this? */
1550         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1551                 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1552                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1553                 dprintk("%s matches recall, use MDS\n", __func__);
1554                 goto out_unlock;
1555         }
1556
1557         /* if LAYOUTGET already failed once we don't try again */
1558         if (pnfs_layout_io_test_failed(lo, iomode)) {
1559                 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1560                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1561                 goto out_unlock;
1562         }
1563
1564         first = list_empty(&lo->plh_segs);
1565         if (first) {
1566                 /* The first layoutget for the file. Need to serialize per
1567                  * RFC 5661 Errata 3208.
1568                  */
1569                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1570                                      &lo->plh_flags)) {
1571                         spin_unlock(&ino->i_lock);
1572                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1573                                     TASK_UNINTERRUPTIBLE);
1574                         pnfs_put_layout_hdr(lo);
1575                         goto lookup_again;
1576                 }
1577         } else {
1578                 /* Check to see if the layout for the given range
1579                  * already exists
1580                  */
1581                 lseg = pnfs_find_lseg(lo, &arg);
1582                 if (lseg) {
1583                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1584                                         PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1585                         goto out_unlock;
1586                 }
1587         }
1588
1589         /*
1590          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1591          * for LAYOUTRETURN even if first is true.
1592          */
1593         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1594                 spin_unlock(&ino->i_lock);
1595                 dprintk("%s wait for layoutreturn\n", __func__);
1596                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1597                         if (first)
1598                                 pnfs_clear_first_layoutget(lo);
1599                         pnfs_put_layout_hdr(lo);
1600                         dprintk("%s retrying\n", __func__);
1601                         goto lookup_again;
1602                 }
1603                 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1604                                 PNFS_UPDATE_LAYOUT_RETURN);
1605                 goto out_put_layout_hdr;
1606         }
1607
1608         if (pnfs_layoutgets_blocked(lo)) {
1609                 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1610                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1611                 goto out_unlock;
1612         }
1613         atomic_inc(&lo->plh_outstanding);
1614         spin_unlock(&ino->i_lock);
1615
1616         if (list_empty(&lo->plh_layouts)) {
1617                 /* The lo must be on the clp list if there is any
1618                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1619                  */
1620                 spin_lock(&clp->cl_lock);
1621                 if (list_empty(&lo->plh_layouts))
1622                         list_add_tail(&lo->plh_layouts, &server->layouts);
1623                 spin_unlock(&clp->cl_lock);
1624         }
1625
1626         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1627         if (pg_offset) {
1628                 arg.offset -= pg_offset;
1629                 arg.length += pg_offset;
1630         }
1631         if (arg.length != NFS4_MAX_UINT64)
1632                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1633
1634         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1635         atomic_dec(&lo->plh_outstanding);
1636         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1637                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1638 out_put_layout_hdr:
1639         if (first)
1640                 pnfs_clear_first_layoutget(lo);
1641         pnfs_put_layout_hdr(lo);
1642 out:
1643         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1644                         "(%s, offset: %llu, length: %llu)\n",
1645                         __func__, ino->i_sb->s_id,
1646                         (unsigned long long)NFS_FILEID(ino),
1647                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1648                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1649                         (unsigned long long)pos,
1650                         (unsigned long long)count);
1651         return lseg;
1652 out_unlock:
1653         spin_unlock(&ino->i_lock);
1654         goto out_put_layout_hdr;
1655 }
1656 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1657
1658 static bool
1659 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1660 {
1661         switch (range->iomode) {
1662         case IOMODE_READ:
1663         case IOMODE_RW:
1664                 break;
1665         default:
1666                 return false;
1667         }
1668         if (range->offset == NFS4_MAX_UINT64)
1669                 return false;
1670         if (range->length == 0)
1671                 return false;
1672         if (range->length != NFS4_MAX_UINT64 &&
1673             range->length > NFS4_MAX_UINT64 - range->offset)
1674                 return false;
1675         return true;
1676 }
1677
1678 struct pnfs_layout_segment *
1679 pnfs_layout_process(struct nfs4_layoutget *lgp)
1680 {
1681         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1682         struct nfs4_layoutget_res *res = &lgp->res;
1683         struct pnfs_layout_segment *lseg;
1684         struct inode *ino = lo->plh_inode;
1685         LIST_HEAD(free_me);
1686         int status = -EINVAL;
1687
1688         if (!pnfs_sanity_check_layout_range(&res->range))
1689                 goto out;
1690
1691         /* Inject layout blob into I/O device driver */
1692         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1693         if (!lseg || IS_ERR(lseg)) {
1694                 if (!lseg)
1695                         status = -ENOMEM;
1696                 else
1697                         status = PTR_ERR(lseg);
1698                 dprintk("%s: Could not allocate layout: error %d\n",
1699                        __func__, status);
1700                 goto out;
1701         }
1702
1703         init_lseg(lo, lseg);
1704         lseg->pls_range = res->range;
1705
1706         spin_lock(&ino->i_lock);
1707         if (pnfs_layoutgets_blocked(lo)) {
1708                 dprintk("%s forget reply due to state\n", __func__);
1709                 goto out_forget_reply;
1710         }
1711
1712         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1713                 /* existing state ID, make sure the sequence number matches. */
1714                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1715                         dprintk("%s forget reply due to sequence\n", __func__);
1716                         status = -EAGAIN;
1717                         goto out_forget_reply;
1718                 }
1719                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1720         } else {
1721                 /*
1722                  * We got an entirely new state ID.  Mark all segments for the
1723                  * inode invalid, and don't bother validating the stateid
1724                  * sequence number.
1725                  */
1726                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1727
1728                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1729                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1730         }
1731
1732         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1733
1734         pnfs_get_lseg(lseg);
1735         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1736
1737         if (res->return_on_close)
1738                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1739
1740         spin_unlock(&ino->i_lock);
1741         pnfs_free_lseg_list(&free_me);
1742         return lseg;
1743 out:
1744         return ERR_PTR(status);
1745
1746 out_forget_reply:
1747         spin_unlock(&ino->i_lock);
1748         lseg->pls_layout = lo;
1749         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1750         goto out;
1751 }
1752
1753 static void
1754 pnfs_set_plh_return_iomode(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode)
1755 {
1756         if (lo->plh_return_iomode == iomode)
1757                 return;
1758         if (lo->plh_return_iomode != 0)
1759                 iomode = IOMODE_ANY;
1760         lo->plh_return_iomode = iomode;
1761         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
1762 }
1763
1764 /**
1765  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1766  * @lo: pointer to layout header
1767  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1768  * @return_range: describe layout segment ranges to be returned
1769  *
1770  * This function is mainly intended for use by layoutrecall. It attempts
1771  * to free the layout segment immediately, or else to mark it for return
1772  * as soon as its reference count drops to zero.
1773  */
1774 int
1775 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1776                                 struct list_head *tmp_list,
1777                                 const struct pnfs_layout_range *return_range)
1778 {
1779         struct pnfs_layout_segment *lseg, *next;
1780         int remaining = 0;
1781
1782         dprintk("%s:Begin lo %p\n", __func__, lo);
1783
1784         if (list_empty(&lo->plh_segs))
1785                 return 0;
1786
1787         assert_spin_locked(&lo->plh_inode->i_lock);
1788
1789         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1790                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1791                         dprintk("%s: marking lseg %p iomode %d "
1792                                 "offset %llu length %llu\n", __func__,
1793                                 lseg, lseg->pls_range.iomode,
1794                                 lseg->pls_range.offset,
1795                                 lseg->pls_range.length);
1796                         if (mark_lseg_invalid(lseg, tmp_list))
1797                                 continue;
1798                         remaining++;
1799                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1800                         pnfs_set_plh_return_iomode(lo, return_range->iomode);
1801                 }
1802         return remaining;
1803 }
1804
1805 void pnfs_error_mark_layout_for_return(struct inode *inode,
1806                                        struct pnfs_layout_segment *lseg)
1807 {
1808         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1809         struct pnfs_layout_range range = {
1810                 .iomode = lseg->pls_range.iomode,
1811                 .offset = 0,
1812                 .length = NFS4_MAX_UINT64,
1813         };
1814         LIST_HEAD(free_me);
1815         bool return_now = false;
1816
1817         spin_lock(&inode->i_lock);
1818         pnfs_set_plh_return_iomode(lo, range.iomode);
1819         /*
1820          * mark all matching lsegs so that we are sure to have no live
1821          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1822          * for how it works.
1823          */
1824         if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range)) {
1825                 nfs4_stateid stateid;
1826                 enum pnfs_iomode iomode = lo->plh_return_iomode;
1827
1828                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1829                 return_now = pnfs_prepare_layoutreturn(lo);
1830                 spin_unlock(&inode->i_lock);
1831                 if (return_now)
1832                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1833         } else {
1834                 spin_unlock(&inode->i_lock);
1835                 nfs_commit_inode(inode, 0);
1836         }
1837         pnfs_free_lseg_list(&free_me);
1838 }
1839 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1840
1841 void
1842 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1843 {
1844         u64 rd_size = req->wb_bytes;
1845
1846         if (pgio->pg_lseg == NULL) {
1847                 if (pgio->pg_dreq == NULL)
1848                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1849                 else
1850                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1851
1852                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1853                                                    req->wb_context,
1854                                                    req_offset(req),
1855                                                    rd_size,
1856                                                    IOMODE_READ,
1857                                                    GFP_KERNEL);
1858                 if (IS_ERR(pgio->pg_lseg)) {
1859                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1860                         pgio->pg_lseg = NULL;
1861                         return;
1862                 }
1863         }
1864         /* If no lseg, fall back to read through mds */
1865         if (pgio->pg_lseg == NULL)
1866                 nfs_pageio_reset_read_mds(pgio);
1867
1868 }
1869 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1870
1871 void
1872 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1873                            struct nfs_page *req, u64 wb_size)
1874 {
1875         if (pgio->pg_lseg == NULL) {
1876                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1877                                                    req->wb_context,
1878                                                    req_offset(req),
1879                                                    wb_size,
1880                                                    IOMODE_RW,
1881                                                    GFP_NOFS);
1882                 if (IS_ERR(pgio->pg_lseg)) {
1883                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1884                         pgio->pg_lseg = NULL;
1885                         return;
1886                 }
1887         }
1888         /* If no lseg, fall back to write through mds */
1889         if (pgio->pg_lseg == NULL)
1890                 nfs_pageio_reset_write_mds(pgio);
1891 }
1892 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1893
1894 void
1895 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1896 {
1897         if (desc->pg_lseg) {
1898                 pnfs_put_lseg(desc->pg_lseg);
1899                 desc->pg_lseg = NULL;
1900         }
1901 }
1902 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1903
1904 /*
1905  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1906  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1907  */
1908 size_t
1909 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1910                      struct nfs_page *prev, struct nfs_page *req)
1911 {
1912         unsigned int size;
1913         u64 seg_end, req_start, seg_left;
1914
1915         size = nfs_generic_pg_test(pgio, prev, req);
1916         if (!size)
1917                 return 0;
1918
1919         /*
1920          * 'size' contains the number of bytes left in the current page (up
1921          * to the original size asked for in @req->wb_bytes).
1922          *
1923          * Calculate how many bytes are left in the layout segment
1924          * and if there are less bytes than 'size', return that instead.
1925          *
1926          * Please also note that 'end_offset' is actually the offset of the
1927          * first byte that lies outside the pnfs_layout_range. FIXME?
1928          *
1929          */
1930         if (pgio->pg_lseg) {
1931                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1932                                      pgio->pg_lseg->pls_range.length);
1933                 req_start = req_offset(req);
1934                 WARN_ON_ONCE(req_start >= seg_end);
1935                 /* start of request is past the last byte of this segment */
1936                 if (req_start >= seg_end) {
1937                         /* reference the new lseg */
1938                         if (pgio->pg_ops->pg_cleanup)
1939                                 pgio->pg_ops->pg_cleanup(pgio);
1940                         if (pgio->pg_ops->pg_init)
1941                                 pgio->pg_ops->pg_init(pgio, req);
1942                         return 0;
1943                 }
1944
1945                 /* adjust 'size' iff there are fewer bytes left in the
1946                  * segment than what nfs_generic_pg_test returned */
1947                 seg_left = seg_end - req_start;
1948                 if (seg_left < size)
1949                         size = (unsigned int)seg_left;
1950         }
1951
1952         return size;
1953 }
1954 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1955
1956 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1957 {
1958         struct nfs_pageio_descriptor pgio;
1959
1960         /* Resend all requests through the MDS */
1961         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1962                               hdr->completion_ops);
1963         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1964         return nfs_pageio_resend(&pgio, hdr);
1965 }
1966 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1967
1968 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1969 {
1970
1971         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1972         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1973             PNFS_LAYOUTRET_ON_ERROR) {
1974                 pnfs_return_layout(hdr->inode);
1975         }
1976         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1977                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1978 }
1979
1980 /*
1981  * Called by non rpc-based layout drivers
1982  */
1983 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1984 {
1985         if (likely(!hdr->pnfs_error)) {
1986                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1987                                 hdr->mds_offset + hdr->res.count);
1988                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1989         }
1990         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1991         if (unlikely(hdr->pnfs_error))
1992                 pnfs_ld_handle_write_error(hdr);
1993         hdr->mds_ops->rpc_release(hdr);
1994 }
1995 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1996
1997 static void
1998 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1999                 struct nfs_pgio_header *hdr)
2000 {
2001         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2002
2003         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2004                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2005                 nfs_pageio_reset_write_mds(desc);
2006                 mirror->pg_recoalesce = 1;
2007         }
2008         nfs_pgio_data_destroy(hdr);
2009         hdr->release(hdr);
2010 }
2011
2012 static enum pnfs_try_status
2013 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2014                         const struct rpc_call_ops *call_ops,
2015                         struct pnfs_layout_segment *lseg,
2016                         int how)
2017 {
2018         struct inode *inode = hdr->inode;
2019         enum pnfs_try_status trypnfs;
2020         struct nfs_server *nfss = NFS_SERVER(inode);
2021
2022         hdr->mds_ops = call_ops;
2023
2024         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2025                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2026         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2027         if (trypnfs != PNFS_NOT_ATTEMPTED)
2028                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2029         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2030         return trypnfs;
2031 }
2032
2033 static void
2034 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2035               struct nfs_pgio_header *hdr, int how)
2036 {
2037         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2038         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2039         enum pnfs_try_status trypnfs;
2040
2041         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2042         if (trypnfs == PNFS_NOT_ATTEMPTED)
2043                 pnfs_write_through_mds(desc, hdr);
2044 }
2045
2046 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2047 {
2048         pnfs_put_lseg(hdr->lseg);
2049         nfs_pgio_header_free(hdr);
2050 }
2051
2052 int
2053 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2054 {
2055         struct nfs_pgio_header *hdr;
2056         int ret;
2057
2058         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2059         if (!hdr) {
2060                 desc->pg_error = -ENOMEM;
2061                 return desc->pg_error;
2062         }
2063         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2064
2065         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2066         ret = nfs_generic_pgio(desc, hdr);
2067         if (!ret)
2068                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2069
2070         return ret;
2071 }
2072 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2073
2074 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2075 {
2076         struct nfs_pageio_descriptor pgio;
2077
2078         /* Resend all requests through the MDS */
2079         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2080         return nfs_pageio_resend(&pgio, hdr);
2081 }
2082 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2083
2084 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2085 {
2086         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2087         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2088             PNFS_LAYOUTRET_ON_ERROR) {
2089                 pnfs_return_layout(hdr->inode);
2090         }
2091         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2092                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2093 }
2094
2095 /*
2096  * Called by non rpc-based layout drivers
2097  */
2098 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2099 {
2100         if (likely(!hdr->pnfs_error)) {
2101                 __nfs4_read_done_cb(hdr);
2102                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2103         }
2104         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2105         if (unlikely(hdr->pnfs_error))
2106                 pnfs_ld_handle_read_error(hdr);
2107         hdr->mds_ops->rpc_release(hdr);
2108 }
2109 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2110
2111 static void
2112 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2113                 struct nfs_pgio_header *hdr)
2114 {
2115         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2116
2117         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2118                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2119                 nfs_pageio_reset_read_mds(desc);
2120                 mirror->pg_recoalesce = 1;
2121         }
2122         nfs_pgio_data_destroy(hdr);
2123         hdr->release(hdr);
2124 }
2125
2126 /*
2127  * Call the appropriate parallel I/O subsystem read function.
2128  */
2129 static enum pnfs_try_status
2130 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2131                        const struct rpc_call_ops *call_ops,
2132                        struct pnfs_layout_segment *lseg)
2133 {
2134         struct inode *inode = hdr->inode;
2135         struct nfs_server *nfss = NFS_SERVER(inode);
2136         enum pnfs_try_status trypnfs;
2137
2138         hdr->mds_ops = call_ops;
2139
2140         dprintk("%s: Reading ino:%lu %u@%llu\n",
2141                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2142
2143         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2144         if (trypnfs != PNFS_NOT_ATTEMPTED)
2145                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2146         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2147         return trypnfs;
2148 }
2149
2150 /* Resend all requests through pnfs. */
2151 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2152 {
2153         struct nfs_pageio_descriptor pgio;
2154
2155         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2156         return nfs_pageio_resend(&pgio, hdr);
2157 }
2158 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2159
2160 static void
2161 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2162 {
2163         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2164         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2165         enum pnfs_try_status trypnfs;
2166         int err = 0;
2167
2168         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2169         if (trypnfs == PNFS_TRY_AGAIN)
2170                 err = pnfs_read_resend_pnfs(hdr);
2171         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2172                 pnfs_read_through_mds(desc, hdr);
2173 }
2174
2175 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2176 {
2177         pnfs_put_lseg(hdr->lseg);
2178         nfs_pgio_header_free(hdr);
2179 }
2180
2181 int
2182 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2183 {
2184         struct nfs_pgio_header *hdr;
2185         int ret;
2186
2187         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2188         if (!hdr) {
2189                 desc->pg_error = -ENOMEM;
2190                 return desc->pg_error;
2191         }
2192         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2193         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2194         ret = nfs_generic_pgio(desc, hdr);
2195         if (!ret)
2196                 pnfs_do_read(desc, hdr);
2197         return ret;
2198 }
2199 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2200
2201 static void pnfs_clear_layoutcommitting(struct inode *inode)
2202 {
2203         unsigned long *bitlock = &NFS_I(inode)->flags;
2204
2205         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2206         smp_mb__after_atomic();
2207         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2208 }
2209
2210 /*
2211  * There can be multiple RW segments.
2212  */
2213 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2214 {
2215         struct pnfs_layout_segment *lseg;
2216
2217         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2218                 if (lseg->pls_range.iomode == IOMODE_RW &&
2219                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2220                         list_add(&lseg->pls_lc_list, listp);
2221         }
2222 }
2223
2224 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2225 {
2226         struct pnfs_layout_segment *lseg, *tmp;
2227
2228         /* Matched by references in pnfs_set_layoutcommit */
2229         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2230                 list_del_init(&lseg->pls_lc_list);
2231                 pnfs_put_lseg(lseg);
2232         }
2233
2234         pnfs_clear_layoutcommitting(inode);
2235 }
2236
2237 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2238 {
2239         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2240 }
2241 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2242
2243 void
2244 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2245                 loff_t end_pos)
2246 {
2247         struct nfs_inode *nfsi = NFS_I(inode);
2248         bool mark_as_dirty = false;
2249
2250         spin_lock(&inode->i_lock);
2251         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2252                 nfsi->layout->plh_lwb = end_pos;
2253                 mark_as_dirty = true;
2254                 dprintk("%s: Set layoutcommit for inode %lu ",
2255                         __func__, inode->i_ino);
2256         } else if (end_pos > nfsi->layout->plh_lwb)
2257                 nfsi->layout->plh_lwb = end_pos;
2258         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2259                 /* references matched in nfs4_layoutcommit_release */
2260                 pnfs_get_lseg(lseg);
2261         }
2262         spin_unlock(&inode->i_lock);
2263         dprintk("%s: lseg %p end_pos %llu\n",
2264                 __func__, lseg, nfsi->layout->plh_lwb);
2265
2266         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2267          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2268         if (mark_as_dirty)
2269                 mark_inode_dirty_sync(inode);
2270 }
2271 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2272
2273 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2274 {
2275         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2276
2277         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2278                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2279         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2280 }
2281
2282 /*
2283  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2284  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2285  * data to disk to allow the server to recover the data if it crashes.
2286  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2287  * is off, and a COMMIT is sent to a data server, or
2288  * if WRITEs to a data server return NFS_DATA_SYNC.
2289  */
2290 int
2291 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2292 {
2293         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2294         struct nfs4_layoutcommit_data *data;
2295         struct nfs_inode *nfsi = NFS_I(inode);
2296         loff_t end_pos;
2297         int status;
2298
2299         if (!pnfs_layoutcommit_outstanding(inode))
2300                 return 0;
2301
2302         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2303
2304         status = -EAGAIN;
2305         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2306                 if (!sync)
2307                         goto out;
2308                 status = wait_on_bit_lock_action(&nfsi->flags,
2309                                 NFS_INO_LAYOUTCOMMITTING,
2310                                 nfs_wait_bit_killable,
2311                                 TASK_KILLABLE);
2312                 if (status)
2313                         goto out;
2314         }
2315
2316         status = -ENOMEM;
2317         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2318         data = kzalloc(sizeof(*data), GFP_NOFS);
2319         if (!data)
2320                 goto clear_layoutcommitting;
2321
2322         status = 0;
2323         spin_lock(&inode->i_lock);
2324         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2325                 goto out_unlock;
2326
2327         INIT_LIST_HEAD(&data->lseg_list);
2328         pnfs_list_write_lseg(inode, &data->lseg_list);
2329
2330         end_pos = nfsi->layout->plh_lwb;
2331
2332         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2333         spin_unlock(&inode->i_lock);
2334
2335         data->args.inode = inode;
2336         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2337         nfs_fattr_init(&data->fattr);
2338         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2339         data->res.fattr = &data->fattr;
2340         data->args.lastbytewritten = end_pos - 1;
2341         data->res.server = NFS_SERVER(inode);
2342
2343         if (ld->prepare_layoutcommit) {
2344                 status = ld->prepare_layoutcommit(&data->args);
2345                 if (status) {
2346                         put_rpccred(data->cred);
2347                         spin_lock(&inode->i_lock);
2348                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2349                         if (end_pos > nfsi->layout->plh_lwb)
2350                                 nfsi->layout->plh_lwb = end_pos;
2351                         goto out_unlock;
2352                 }
2353         }
2354
2355
2356         status = nfs4_proc_layoutcommit(data, sync);
2357 out:
2358         if (status)
2359                 mark_inode_dirty_sync(inode);
2360         dprintk("<-- %s status %d\n", __func__, status);
2361         return status;
2362 out_unlock:
2363         spin_unlock(&inode->i_lock);
2364         kfree(data);
2365 clear_layoutcommitting:
2366         pnfs_clear_layoutcommitting(inode);
2367         goto out;
2368 }
2369 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2370
2371 int
2372 pnfs_generic_sync(struct inode *inode, bool datasync)
2373 {
2374         return pnfs_layoutcommit_inode(inode, true);
2375 }
2376 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2377
2378 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2379 {
2380         struct nfs4_threshold *thp;
2381
2382         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2383         if (!thp) {
2384                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2385                 return NULL;
2386         }
2387         return thp;
2388 }
2389
2390 #if IS_ENABLED(CONFIG_NFS_V4_2)
2391 int
2392 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2393 {
2394         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2395         struct nfs_server *server = NFS_SERVER(inode);
2396         struct nfs_inode *nfsi = NFS_I(inode);
2397         struct nfs42_layoutstat_data *data;
2398         struct pnfs_layout_hdr *hdr;
2399         int status = 0;
2400
2401         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2402                 goto out;
2403
2404         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2405                 goto out;
2406
2407         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2408                 goto out;
2409
2410         spin_lock(&inode->i_lock);
2411         if (!NFS_I(inode)->layout) {
2412                 spin_unlock(&inode->i_lock);
2413                 goto out;
2414         }
2415         hdr = NFS_I(inode)->layout;
2416         pnfs_get_layout_hdr(hdr);
2417         spin_unlock(&inode->i_lock);
2418
2419         data = kzalloc(sizeof(*data), gfp_flags);
2420         if (!data) {
2421                 status = -ENOMEM;
2422                 goto out_put;
2423         }
2424
2425         data->args.fh = NFS_FH(inode);
2426         data->args.inode = inode;
2427         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2428         status = ld->prepare_layoutstats(&data->args);
2429         if (status)
2430                 goto out_free;
2431
2432         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2433
2434 out:
2435         dprintk("%s returns %d\n", __func__, status);
2436         return status;
2437
2438 out_free:
2439         kfree(data);
2440 out_put:
2441         pnfs_put_layout_hdr(hdr);
2442         smp_mb__before_atomic();
2443         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2444         smp_mb__after_atomic();
2445         goto out;
2446 }
2447 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2448 #endif
2449
2450 unsigned int layoutstats_timer;
2451 module_param(layoutstats_timer, uint, 0644);
2452 EXPORT_SYMBOL_GPL(layoutstats_timer);