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 int
  56 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  57                        enum pnfs_iomode iomode, bool sync);
  58
  59 /* Return the registered pnfs layout driver module matching given id */
  60 static struct pnfs_layoutdriver_type *
  61 find_pnfs_driver_locked(u32 id)
  62 {
  63         struct pnfs_layoutdriver_type *local;
  64
  65         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  66                 if (local->id == id)
  67                         goto out;
  68         local = NULL;
  69 out:
  70         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  71         return local;
  72 }
  73
  74 static struct pnfs_layoutdriver_type *
  75 find_pnfs_driver(u32 id)
  76 {
  77         struct pnfs_layoutdriver_type *local;
  78
  79         spin_lock(&pnfs_spinlock);
  80         local = find_pnfs_driver_locked(id);
  81         if (local != NULL && !try_module_get(local->owner)) {
  82                 dprintk("%s: Could not grab reference on module\n", __func__);
  83                 local = NULL;
  84         }
  85         spin_unlock(&pnfs_spinlock);
  86         return local;
  87 }
  88
  89 void
  90 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  91 {
  92         if (nfss->pnfs_curr_ld) {
  93                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  94                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  95                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  96                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  97                         nfs4_deviceid_purge_client(nfss->nfs_client);
  98                 module_put(nfss->pnfs_curr_ld->owner);
  99         }
 100         nfss->pnfs_curr_ld = NULL;
 101 }
 102
 103 /*
 104  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 105  * Currently only one pNFS layout driver per filesystem is supported.
 106  *
 107  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 108  */
 109 void
 110 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 111                       u32 id)
 112 {
 113         struct pnfs_layoutdriver_type *ld_type = NULL;
 114
 115         if (id == 0)
 116                 goto out_no_driver;
 117         if (!(server->nfs_client->cl_exchange_flags &
 118                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 119                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 120                         __func__, id, server->nfs_client->cl_exchange_flags);
 121                 goto out_no_driver;
 122         }
 123         ld_type = find_pnfs_driver(id);
 124         if (!ld_type) {
 125                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 126                 ld_type = find_pnfs_driver(id);
 127                 if (!ld_type) {
 128                         dprintk("%s: No pNFS module found for %u.\n",
 129                                 __func__, id);
 130                         goto out_no_driver;
 131                 }
 132         }
 133         server->pnfs_curr_ld = ld_type;
 134         if (ld_type->set_layoutdriver
 135             && ld_type->set_layoutdriver(server, mntfh)) {
 136                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 137                         "driver %u.\n", __func__, id);
 138                 module_put(ld_type->owner);
 139                 goto out_no_driver;
 140         }
 141         /* Bump the MDS count */
 142         atomic_inc(&server->nfs_client->cl_mds_count);
 143
 144         dprintk("%s: pNFS module for %u set\n", __func__, id);
 145         return;
 146
 147 out_no_driver:
 148         dprintk("%s: Using NFSv4 I/O\n", __func__);
 149         server->pnfs_curr_ld = NULL;
 150 }
 151
 152 int
 153 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 154 {
 155         int status = -EINVAL;
 156         struct pnfs_layoutdriver_type *tmp;
 157
 158         if (ld_type->id == 0) {
 159                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 160                 return status;
 161         }
 162         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 163                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 164                        "alloc_lseg and free_lseg.\n", __func__);
 165                 return status;
 166         }
 167
 168         spin_lock(&pnfs_spinlock);
 169         tmp = find_pnfs_driver_locked(ld_type->id);
 170         if (!tmp) {
 171                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 172                 status = 0;
 173                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 174                         ld_type->name);
 175         } else {
 176                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 177                         __func__, ld_type->id);
 178         }
 179         spin_unlock(&pnfs_spinlock);
 180
 181         return status;
 182 }
 183 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 184
 185 void
 186 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 187 {
 188         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 189         spin_lock(&pnfs_spinlock);
 190         list_del(&ld_type->pnfs_tblid);
 191         spin_unlock(&pnfs_spinlock);
 192 }
 193 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 194
 195 /*
 196  * pNFS client layout cache
 197  */
 198
 199 /* Need to hold i_lock if caller does not already hold reference */
 200 void
 201 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 202 {
 203         atomic_inc(&lo->plh_refcount);
 204 }
 205
 206 static struct pnfs_layout_hdr *
 207 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 208 {
 209         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 210         return ld->alloc_layout_hdr(ino, gfp_flags);
 211 }
 212
 213 static void
 214 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 215 {
 216         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 217         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 218
 219         if (!list_empty(&lo->plh_layouts)) {
 220                 struct nfs_client *clp = server->nfs_client;
 221
 222                 spin_lock(&clp->cl_lock);
 223                 list_del_init(&lo->plh_layouts);
 224                 spin_unlock(&clp->cl_lock);
 225         }
 226         put_rpccred(lo->plh_lc_cred);
 227         return ld->free_layout_hdr(lo);
 228 }
 229
 230 static void
 231 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 232 {
 233         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 234         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 235         nfsi->layout = NULL;
 236         /* Reset MDS Threshold I/O counters */
 237         nfsi->write_io = 0;
 238         nfsi->read_io = 0;
 239 }
 240
 241 void
 242 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 243 {
 244         struct inode *inode = lo->plh_inode;
 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 static int
 256 pnfs_iomode_to_fail_bit(u32 iomode)
 257 {
 258         return iomode == IOMODE_RW ?
 259                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 260 }
 261
 262 static void
 263 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 264 {
 265         lo->plh_retry_timestamp = jiffies;
 266         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 267                 atomic_inc(&lo->plh_refcount);
 268 }
 269
 270 static void
 271 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 272 {
 273         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 274                 atomic_dec(&lo->plh_refcount);
 275 }
 276
 277 static void
 278 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 279 {
 280         struct inode *inode = lo->plh_inode;
 281         struct pnfs_layout_range range = {
 282                 .iomode = iomode,
 283                 .offset = 0,
 284                 .length = NFS4_MAX_UINT64,
 285         };
 286         LIST_HEAD(head);
 287
 288         spin_lock(&inode->i_lock);
 289         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 290         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 291         spin_unlock(&inode->i_lock);
 292         pnfs_free_lseg_list(&head);
 293         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 294                         iomode == IOMODE_RW ?  "RW" : "READ");
 295 }
 296
 297 static bool
 298 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 299 {
 300         unsigned long start, end;
 301         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 302
 303         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 304                 return false;
 305         end = jiffies;
 306         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 307         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 308                 /* It is time to retry the failed layoutgets */
 309                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 310                 return false;
 311         }
 312         return true;
 313 }
 314
 315 static void
 316 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 317 {
 318         INIT_LIST_HEAD(&lseg->pls_list);
 319         INIT_LIST_HEAD(&lseg->pls_lc_list);
 320         atomic_set(&lseg->pls_refcount, 1);
 321         smp_mb();
 322         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 323         lseg->pls_layout = lo;
 324 }
 325
 326 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *ino = lseg->pls_layout->plh_inode;
 329
 330         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 331 }
 332
 333 static void
 334 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 335                 struct pnfs_layout_segment *lseg)
 336 {
 337         struct inode *inode = lo->plh_inode;
 338
 339         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 340         list_del_init(&lseg->pls_list);
 341         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 342         atomic_dec(&lo->plh_refcount);
 343         if (list_empty(&lo->plh_segs))
 344                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 345         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 346 }
 347
 348 /* Return true if layoutreturn is needed */
 349 static bool
 350 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 351                         struct pnfs_layout_segment *lseg)
 352 {
 353         struct pnfs_layout_segment *s;
 354
 355         if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 356                 return false;
 357
 358         list_for_each_entry(s, &lo->plh_segs, pls_list)
 359                 if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 360                         return false;
 361
 362         return true;
 363 }
 364
 365 static bool
 366 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 367 {
 368         /* Serialise LAYOUTGET/LAYOUTRETURN */
 369         if (atomic_read(&lo->plh_outstanding) != 0)
 370                 return false;
 371         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 372                 return false;
 373         lo->plh_return_iomode = 0;
 374         pnfs_get_layout_hdr(lo);
 375         clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
 376         return true;
 377 }
 378
 379 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
 380                 struct pnfs_layout_hdr *lo, struct inode *inode)
 381 {
 382         lo = lseg->pls_layout;
 383         inode = lo->plh_inode;
 384
 385         spin_lock(&inode->i_lock);
 386         if (pnfs_layout_need_return(lo, lseg)) {
 387                 nfs4_stateid stateid;
 388                 enum pnfs_iomode iomode;
 389                 bool send;
 390
 391                 stateid = lo->plh_stateid;
 392                 iomode = lo->plh_return_iomode;
 393                 send = pnfs_prepare_layoutreturn(lo);
 394                 spin_unlock(&inode->i_lock);
 395                 if (send) {
 396                         /* Send an async layoutreturn so we dont deadlock */
 397                         pnfs_send_layoutreturn(lo, stateid, iomode, false);
 398                 }
 399         } else
 400                 spin_unlock(&inode->i_lock);
 401 }
 402
 403 void
 404 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 405 {
 406         struct pnfs_layout_hdr *lo;
 407         struct inode *inode;
 408
 409         if (!lseg)
 410                 return;
 411
 412         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 413                 atomic_read(&lseg->pls_refcount),
 414                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 415
 416         /* Handle the case where refcount != 1 */
 417         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
 418                 return;
 419
 420         lo = lseg->pls_layout;
 421         inode = lo->plh_inode;
 422         /* Do we need a layoutreturn? */
 423         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 424                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 425
 426         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 427                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 428                         spin_unlock(&inode->i_lock);
 429                         return;
 430                 }
 431                 pnfs_get_layout_hdr(lo);
 432                 pnfs_layout_remove_lseg(lo, lseg);
 433                 spin_unlock(&inode->i_lock);
 434                 pnfs_free_lseg(lseg);
 435                 pnfs_put_layout_hdr(lo);
 436         }
 437 }
 438 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 439
 440 static void pnfs_free_lseg_async_work(struct work_struct *work)
 441 {
 442         struct pnfs_layout_segment *lseg;
 443         struct pnfs_layout_hdr *lo;
 444
 445         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 446         lo = lseg->pls_layout;
 447
 448         pnfs_free_lseg(lseg);
 449         pnfs_put_layout_hdr(lo);
 450 }
 451
 452 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 453 {
 454         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 455         schedule_work(&lseg->pls_work);
 456 }
 457
 458 void
 459 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 460 {
 461         if (!lseg)
 462                 return;
 463
 464         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 465
 466         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 467                 atomic_read(&lseg->pls_refcount),
 468                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 469         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 470                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 471                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 472                         return;
 473                 pnfs_get_layout_hdr(lo);
 474                 pnfs_layout_remove_lseg(lo, lseg);
 475                 pnfs_free_lseg_async(lseg);
 476         }
 477 }
 478 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 479
 480 static u64
 481 end_offset(u64 start, u64 len)
 482 {
 483         u64 end;
 484
 485         end = start + len;
 486         return end >= start ? end : NFS4_MAX_UINT64;
 487 }
 488
 489 /*
 490  * is l2 fully contained in l1?
 491  *   start1                             end1
 492  *   [----------------------------------)
 493  *           start2           end2
 494  *           [----------------)
 495  */
 496 static bool
 497 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 498                  const struct pnfs_layout_range *l2)
 499 {
 500         u64 start1 = l1->offset;
 501         u64 end1 = end_offset(start1, l1->length);
 502         u64 start2 = l2->offset;
 503         u64 end2 = end_offset(start2, l2->length);
 504
 505         return (start1 <= start2) && (end1 >= end2);
 506 }
 507
 508 /*
 509  * is l1 and l2 intersecting?
 510  *   start1                             end1
 511  *   [----------------------------------)
 512  *                              start2           end2
 513  *                              [----------------)
 514  */
 515 static bool
 516 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 517                     const struct pnfs_layout_range *l2)
 518 {
 519         u64 start1 = l1->offset;
 520         u64 end1 = end_offset(start1, l1->length);
 521         u64 start2 = l2->offset;
 522         u64 end2 = end_offset(start2, l2->length);
 523
 524         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 525                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 526 }
 527
 528 static bool
 529 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 530                  const struct pnfs_layout_range *recall_range)
 531 {
 532         return (recall_range->iomode == IOMODE_ANY ||
 533                 lseg_range->iomode == recall_range->iomode) &&
 534                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 535 }
 536
 537 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 538                 struct list_head *tmp_list)
 539 {
 540         if (!atomic_dec_and_test(&lseg->pls_refcount))
 541                 return false;
 542         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 543         list_add(&lseg->pls_list, tmp_list);
 544         return true;
 545 }
 546
 547 /* Returns 1 if lseg is removed from list, 0 otherwise */
 548 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 549                              struct list_head *tmp_list)
 550 {
 551         int rv = 0;
 552
 553         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 554                 /* Remove the reference keeping the lseg in the
 555                  * list.  It will now be removed when all
 556                  * outstanding io is finished.
 557                  */
 558                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 559                         atomic_read(&lseg->pls_refcount));
 560                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 561                         rv = 1;
 562         }
 563         return rv;
 564 }
 565
 566 /* Returns count of number of matching invalid lsegs remaining in list
 567  * after call.
 568  */
 569 int
 570 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 571                             struct list_head *tmp_list,
 572                             struct pnfs_layout_range *recall_range)
 573 {
 574         struct pnfs_layout_segment *lseg, *next;
 575         int invalid = 0, removed = 0;
 576
 577         dprintk("%s:Begin lo %p\n", __func__, lo);
 578
 579         if (list_empty(&lo->plh_segs))
 580                 return 0;
 581         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 582                 if (!recall_range ||
 583                     should_free_lseg(&lseg->pls_range, recall_range)) {
 584                         dprintk("%s: freeing lseg %p iomode %d "
 585                                 "offset %llu length %llu\n", __func__,
 586                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 587                                 lseg->pls_range.length);
 588                         invalid++;
 589                         removed += mark_lseg_invalid(lseg, tmp_list);
 590                 }
 591         dprintk("%s:Return %i\n", __func__, invalid - removed);
 592         return invalid - removed;
 593 }
 594
 595 /* note free_me must contain lsegs from a single layout_hdr */
 596 void
 597 pnfs_free_lseg_list(struct list_head *free_me)
 598 {
 599         struct pnfs_layout_segment *lseg, *tmp;
 600
 601         if (list_empty(free_me))
 602                 return;
 603
 604         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 605                 list_del(&lseg->pls_list);
 606                 pnfs_free_lseg(lseg);
 607         }
 608 }
 609
 610 void
 611 pnfs_destroy_layout(struct nfs_inode *nfsi)
 612 {
 613         struct pnfs_layout_hdr *lo;
 614         LIST_HEAD(tmp_list);
 615
 616         spin_lock(&nfsi->vfs_inode.i_lock);
 617         lo = nfsi->layout;
 618         if (lo) {
 619                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 620                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 621                 pnfs_get_layout_hdr(lo);
 622                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 623                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 624                 pnfs_clear_retry_layoutget(lo);
 625                 spin_unlock(&nfsi->vfs_inode.i_lock);
 626                 pnfs_free_lseg_list(&tmp_list);
 627                 pnfs_put_layout_hdr(lo);
 628         } else
 629                 spin_unlock(&nfsi->vfs_inode.i_lock);
 630 }
 631 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 632
 633 static bool
 634 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 635                 struct list_head *layout_list)
 636 {
 637         struct pnfs_layout_hdr *lo;
 638         bool ret = false;
 639
 640         spin_lock(&inode->i_lock);
 641         lo = NFS_I(inode)->layout;
 642         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 643                 pnfs_get_layout_hdr(lo);
 644                 list_add(&lo->plh_bulk_destroy, layout_list);
 645                 ret = true;
 646         }
 647         spin_unlock(&inode->i_lock);
 648         return ret;
 649 }
 650
 651 /* Caller must hold rcu_read_lock and clp->cl_lock */
 652 static int
 653 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 654                 struct nfs_server *server,
 655                 struct list_head *layout_list)
 656 {
 657         struct pnfs_layout_hdr *lo, *next;
 658         struct inode *inode;
 659
 660         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 661                 inode = igrab(lo->plh_inode);
 662                 if (inode == NULL)
 663                         continue;
 664                 list_del_init(&lo->plh_layouts);
 665                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 666                         continue;
 667                 rcu_read_unlock();
 668                 spin_unlock(&clp->cl_lock);
 669                 iput(inode);
 670                 spin_lock(&clp->cl_lock);
 671                 rcu_read_lock();
 672                 return -EAGAIN;
 673         }
 674         return 0;
 675 }
 676
 677 static int
 678 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 679                 bool is_bulk_recall)
 680 {
 681         struct pnfs_layout_hdr *lo;
 682         struct inode *inode;
 683         struct pnfs_layout_range range = {
 684                 .iomode = IOMODE_ANY,
 685                 .offset = 0,
 686                 .length = NFS4_MAX_UINT64,
 687         };
 688         LIST_HEAD(lseg_list);
 689         int ret = 0;
 690
 691         while (!list_empty(layout_list)) {
 692                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 693                                 plh_bulk_destroy);
 694                 dprintk("%s freeing layout for inode %lu\n", __func__,
 695                         lo->plh_inode->i_ino);
 696                 inode = lo->plh_inode;
 697
 698                 pnfs_layoutcommit_inode(inode, false);
 699
 700                 spin_lock(&inode->i_lock);
 701                 list_del_init(&lo->plh_bulk_destroy);
 702                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 703                 if (is_bulk_recall)
 704                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 705                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 706                         ret = -EAGAIN;
 707                 spin_unlock(&inode->i_lock);
 708                 pnfs_free_lseg_list(&lseg_list);
 709                 pnfs_put_layout_hdr(lo);
 710                 iput(inode);
 711         }
 712         return ret;
 713 }
 714
 715 int
 716 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 717                 struct nfs_fsid *fsid,
 718                 bool is_recall)
 719 {
 720         struct nfs_server *server;
 721         LIST_HEAD(layout_list);
 722
 723         spin_lock(&clp->cl_lock);
 724         rcu_read_lock();
 725 restart:
 726         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 727                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 728                         continue;
 729                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 730                                 server,
 731                                 &layout_list) != 0)
 732                         goto restart;
 733         }
 734         rcu_read_unlock();
 735         spin_unlock(&clp->cl_lock);
 736
 737         if (list_empty(&layout_list))
 738                 return 0;
 739         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 740 }
 741
 742 int
 743 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 744                 bool is_recall)
 745 {
 746         struct nfs_server *server;
 747         LIST_HEAD(layout_list);
 748
 749         spin_lock(&clp->cl_lock);
 750         rcu_read_lock();
 751 restart:
 752         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 753                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 754                                         server,
 755                                         &layout_list) != 0)
 756                         goto restart;
 757         }
 758         rcu_read_unlock();
 759         spin_unlock(&clp->cl_lock);
 760
 761         if (list_empty(&layout_list))
 762                 return 0;
 763         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 764 }
 765
 766 /*
 767  * Called by the state manger to remove all layouts established under an
 768  * expired lease.
 769  */
 770 void
 771 pnfs_destroy_all_layouts(struct nfs_client *clp)
 772 {
 773         nfs4_deviceid_mark_client_invalid(clp);
 774         nfs4_deviceid_purge_client(clp);
 775
 776         pnfs_destroy_layouts_byclid(clp, false);
 777 }
 778
 779 /*
 780  * Compare 2 layout stateid sequence ids, to see which is newer,
 781  * taking into account wraparound issues.
 782  */
 783 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 784 {
 785         return (s32)(s1 - s2) > 0;
 786 }
 787
 788 /* update lo->plh_stateid with new if is more recent */
 789 void
 790 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 791                         bool update_barrier)
 792 {
 793         u32 oldseq, newseq, new_barrier;
 794         int empty = list_empty(&lo->plh_segs);
 795
 796         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 797         newseq = be32_to_cpu(new->seqid);
 798         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 799                 nfs4_stateid_copy(&lo->plh_stateid, new);
 800                 if (update_barrier) {
 801                         new_barrier = be32_to_cpu(new->seqid);
 802                 } else {
 803                         /* Because of wraparound, we want to keep the barrier
 804                          * "close" to the current seqids.
 805                          */
 806                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 807                 }
 808                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 809                         lo->plh_barrier = new_barrier;
 810         }
 811 }
 812
 813 static bool
 814 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 815                 const nfs4_stateid *stateid)
 816 {
 817         u32 seqid = be32_to_cpu(stateid->seqid);
 818
 819         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 820 }
 821
 822 /* lget is set to 1 if called from inside send_layoutget call chain */
 823 static bool
 824 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 825 {
 826         return lo->plh_block_lgets ||
 827                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 828 }
 829
 830 int
 831 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 832                               struct pnfs_layout_range *range,
 833                               struct nfs4_state *open_state)
 834 {
 835         int status = 0;
 836
 837         dprintk("--> %s\n", __func__);
 838         spin_lock(&lo->plh_inode->i_lock);
 839         if (pnfs_layoutgets_blocked(lo)) {
 840                 status = -EAGAIN;
 841         } else if (!nfs4_valid_open_stateid(open_state)) {
 842                 status = -EBADF;
 843         } else if (list_empty(&lo->plh_segs) ||
 844                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 845                 int seq;
 846
 847                 do {
 848                         seq = read_seqbegin(&open_state->seqlock);
 849                         nfs4_stateid_copy(dst, &open_state->stateid);
 850                 } while (read_seqretry(&open_state->seqlock, seq));
 851         } else
 852                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 853         spin_unlock(&lo->plh_inode->i_lock);
 854         dprintk("<-- %s\n", __func__);
 855         return status;
 856 }
 857
 858 /*
 859 * Get layout from server.
 860 *    for now, assume that whole file layouts are requested.
 861 *    arg->offset: 0
 862 *    arg->length: all ones
 863 */
 864 static struct pnfs_layout_segment *
 865 send_layoutget(struct pnfs_layout_hdr *lo,
 866            struct nfs_open_context *ctx,
 867            struct pnfs_layout_range *range,
 868            gfp_t gfp_flags)
 869 {
 870         struct inode *ino = lo->plh_inode;
 871         struct nfs_server *server = NFS_SERVER(ino);
 872         struct nfs4_layoutget *lgp;
 873         struct pnfs_layout_segment *lseg;
 874         loff_t i_size;
 875
 876         dprintk("--> %s\n", __func__);
 877
 878         /*
 879          * Synchronously retrieve layout information from server and
 880          * store in lseg. If we race with a concurrent seqid morphing
 881          * op, then re-send the LAYOUTGET.
 882          */
 883         do {
 884                 lgp = kzalloc(sizeof(*lgp), gfp_flags);
 885                 if (lgp == NULL)
 886                         return NULL;
 887
 888                 i_size = i_size_read(ino);
 889
 890                 lgp->args.minlength = PAGE_CACHE_SIZE;
 891                 if (lgp->args.minlength > range->length)
 892                         lgp->args.minlength = range->length;
 893                 if (range->iomode == IOMODE_READ) {
 894                         if (range->offset >= i_size)
 895                                 lgp->args.minlength = 0;
 896                         else if (i_size - range->offset < lgp->args.minlength)
 897                                 lgp->args.minlength = i_size - range->offset;
 898                 }
 899                 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 900                 lgp->args.range = *range;
 901                 lgp->args.type = server->pnfs_curr_ld->id;
 902                 lgp->args.inode = ino;
 903                 lgp->args.ctx = get_nfs_open_context(ctx);
 904                 lgp->gfp_flags = gfp_flags;
 905                 lgp->cred = lo->plh_lc_cred;
 906
 907                 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 908         } while (lseg == ERR_PTR(-EAGAIN));
 909
 910         if (IS_ERR(lseg)) {
 911                 switch (PTR_ERR(lseg)) {
 912                 case -ERESTARTSYS:
 913                 case -EIO:
 914                 case -ENOSPC:
 915                 case -EROFS:
 916                 case -E2BIG:
 917                         break;
 918                 default:
 919                         return NULL;
 920                 }
 921         } else
 922                 pnfs_layout_clear_fail_bit(lo,
 923                                 pnfs_iomode_to_fail_bit(range->iomode));
 924
 925         return lseg;
 926 }
 927
 928 static void pnfs_clear_layoutcommit(struct inode *inode,
 929                 struct list_head *head)
 930 {
 931         struct nfs_inode *nfsi = NFS_I(inode);
 932         struct pnfs_layout_segment *lseg, *tmp;
 933
 934         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 935                 return;
 936         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 937                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 938                         continue;
 939                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 940         }
 941 }
 942
 943 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 944 {
 945         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 946         smp_mb__after_atomic();
 947         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 948         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 949 }
 950
 951 static int
 952 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 953                        enum pnfs_iomode iomode, bool sync)
 954 {
 955         struct inode *ino = lo->plh_inode;
 956         struct nfs4_layoutreturn *lrp;
 957         int status = 0;
 958
 959         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 960         if (unlikely(lrp == NULL)) {
 961                 status = -ENOMEM;
 962                 spin_lock(&ino->i_lock);
 963                 pnfs_clear_layoutreturn_waitbit(lo);
 964                 spin_unlock(&ino->i_lock);
 965                 pnfs_put_layout_hdr(lo);
 966                 goto out;
 967         }
 968
 969         lrp->args.stateid = stateid;
 970         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 971         lrp->args.inode = ino;
 972         lrp->args.range.iomode = iomode;
 973         lrp->args.range.offset = 0;
 974         lrp->args.range.length = NFS4_MAX_UINT64;
 975         lrp->args.layout = lo;
 976         lrp->clp = NFS_SERVER(ino)->nfs_client;
 977         lrp->cred = lo->plh_lc_cred;
 978
 979         status = nfs4_proc_layoutreturn(lrp, sync);
 980 out:
 981         dprintk("<-- %s status: %d\n", __func__, status);
 982         return status;
 983 }
 984
 985 /*
 986  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 987  * when the layout segment list is empty.
 988  *
 989  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 990  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 991  * deviceid is marked invalid.
 992  */
 993 int
 994 _pnfs_return_layout(struct inode *ino)
 995 {
 996         struct pnfs_layout_hdr *lo = NULL;
 997         struct nfs_inode *nfsi = NFS_I(ino);
 998         LIST_HEAD(tmp_list);
 999         nfs4_stateid stateid;
1000         int status = 0, empty;
1001         bool send;
1002
1003         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1004
1005         spin_lock(&ino->i_lock);
1006         lo = nfsi->layout;
1007         if (!lo) {
1008                 spin_unlock(&ino->i_lock);
1009                 dprintk("NFS: %s no layout to return\n", __func__);
1010                 goto out;
1011         }
1012         stateid = nfsi->layout->plh_stateid;
1013         /* Reference matched in nfs4_layoutreturn_release */
1014         pnfs_get_layout_hdr(lo);
1015         empty = list_empty(&lo->plh_segs);
1016         pnfs_clear_layoutcommit(ino, &tmp_list);
1017         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1018
1019         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1020                 struct pnfs_layout_range range = {
1021                         .iomode         = IOMODE_ANY,
1022                         .offset         = 0,
1023                         .length         = NFS4_MAX_UINT64,
1024                 };
1025                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1026         }
1027
1028         /* Don't send a LAYOUTRETURN if list was initially empty */
1029         if (empty) {
1030                 spin_unlock(&ino->i_lock);
1031                 dprintk("NFS: %s no layout segments to return\n", __func__);
1032                 goto out_put_layout_hdr;
1033         }
1034
1035         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1036         send = pnfs_prepare_layoutreturn(lo);
1037         spin_unlock(&ino->i_lock);
1038         pnfs_free_lseg_list(&tmp_list);
1039         if (send)
1040                 status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1041 out_put_layout_hdr:
1042         pnfs_put_layout_hdr(lo);
1043 out:
1044         dprintk("<-- %s status: %d\n", __func__, status);
1045         return status;
1046 }
1047 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1048
1049 int
1050 pnfs_commit_and_return_layout(struct inode *inode)
1051 {
1052         struct pnfs_layout_hdr *lo;
1053         int ret;
1054
1055         spin_lock(&inode->i_lock);
1056         lo = NFS_I(inode)->layout;
1057         if (lo == NULL) {
1058                 spin_unlock(&inode->i_lock);
1059                 return 0;
1060         }
1061         pnfs_get_layout_hdr(lo);
1062         /* Block new layoutgets and read/write to ds */
1063         lo->plh_block_lgets++;
1064         spin_unlock(&inode->i_lock);
1065         filemap_fdatawait(inode->i_mapping);
1066         ret = pnfs_layoutcommit_inode(inode, true);
1067         if (ret == 0)
1068                 ret = _pnfs_return_layout(inode);
1069         spin_lock(&inode->i_lock);
1070         lo->plh_block_lgets--;
1071         spin_unlock(&inode->i_lock);
1072         pnfs_put_layout_hdr(lo);
1073         return ret;
1074 }
1075
1076 bool pnfs_roc(struct inode *ino)
1077 {
1078         struct nfs_inode *nfsi = NFS_I(ino);
1079         struct nfs_open_context *ctx;
1080         struct nfs4_state *state;
1081         struct pnfs_layout_hdr *lo;
1082         struct pnfs_layout_segment *lseg, *tmp;
1083         nfs4_stateid stateid;
1084         LIST_HEAD(tmp_list);
1085         bool found = false, layoutreturn = false, roc = false;
1086
1087         spin_lock(&ino->i_lock);
1088         lo = nfsi->layout;
1089         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1090                 goto out_noroc;
1091
1092         /* no roc if we hold a delegation */
1093         if (nfs4_check_delegation(ino, FMODE_READ))
1094                 goto out_noroc;
1095
1096         list_for_each_entry(ctx, &nfsi->open_files, list) {
1097                 state = ctx->state;
1098                 /* Don't return layout if there is open file state */
1099                 if (state != NULL && state->state != 0)
1100                         goto out_noroc;
1101         }
1102
1103         stateid = lo->plh_stateid;
1104         /* always send layoutreturn if being marked so */
1105         if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1106                                    &lo->plh_flags))
1107                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1108
1109         pnfs_clear_retry_layoutget(lo);
1110         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1111                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1112                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1113                         mark_lseg_invalid(lseg, &tmp_list);
1114                         found = true;
1115                 }
1116         /* ROC in two conditions:
1117          * 1. there are ROC lsegs
1118          * 2. we don't send layoutreturn
1119          */
1120         if (found && !layoutreturn) {
1121                 /* lo ref dropped in pnfs_roc_release() */
1122                 pnfs_get_layout_hdr(lo);
1123                 roc = true;
1124         }
1125
1126 out_noroc:
1127         spin_unlock(&ino->i_lock);
1128         pnfs_free_lseg_list(&tmp_list);
1129         pnfs_layoutcommit_inode(ino, true);
1130         if (layoutreturn)
1131                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1132         return roc;
1133 }
1134
1135 void pnfs_roc_release(struct inode *ino)
1136 {
1137         struct pnfs_layout_hdr *lo;
1138
1139         spin_lock(&ino->i_lock);
1140         lo = NFS_I(ino)->layout;
1141         pnfs_clear_layoutreturn_waitbit(lo);
1142         if (atomic_dec_and_test(&lo->plh_refcount)) {
1143                 pnfs_detach_layout_hdr(lo);
1144                 spin_unlock(&ino->i_lock);
1145                 pnfs_free_layout_hdr(lo);
1146         } else
1147                 spin_unlock(&ino->i_lock);
1148 }
1149
1150 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1151 {
1152         struct pnfs_layout_hdr *lo;
1153
1154         spin_lock(&ino->i_lock);
1155         lo = NFS_I(ino)->layout;
1156         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1157                 lo->plh_barrier = barrier;
1158         spin_unlock(&ino->i_lock);
1159         trace_nfs4_layoutreturn_on_close(ino, 0);
1160 }
1161
1162 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1163 {
1164         struct nfs_inode *nfsi = NFS_I(ino);
1165         struct pnfs_layout_hdr *lo;
1166         u32 current_seqid;
1167
1168         spin_lock(&ino->i_lock);
1169         lo = nfsi->layout;
1170         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1171
1172         /* Since close does not return a layout stateid for use as
1173          * a barrier, we choose the worst-case barrier.
1174          */
1175         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1176         spin_unlock(&ino->i_lock);
1177 }
1178
1179 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1180 {
1181         struct nfs_inode *nfsi = NFS_I(ino);
1182         struct pnfs_layout_hdr *lo;
1183         bool sleep = false;
1184
1185         /* we might not have grabbed lo reference. so need to check under
1186          * i_lock */
1187         spin_lock(&ino->i_lock);
1188         lo = nfsi->layout;
1189         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1190                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1191                 sleep = true;
1192         }
1193         spin_unlock(&ino->i_lock);
1194         return sleep;
1195 }
1196
1197 /*
1198  * Compare two layout segments for sorting into layout cache.
1199  * We want to preferentially return RW over RO layouts, so ensure those
1200  * are seen first.
1201  */
1202 static s64
1203 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1204            const struct pnfs_layout_range *l2)
1205 {
1206         s64 d;
1207
1208         /* high offset > low offset */
1209         d = l1->offset - l2->offset;
1210         if (d)
1211                 return d;
1212
1213         /* short length > long length */
1214         d = l2->length - l1->length;
1215         if (d)
1216                 return d;
1217
1218         /* read > read/write */
1219         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1220 }
1221
1222 static bool
1223 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1224                 const struct pnfs_layout_range *l2)
1225 {
1226         return pnfs_lseg_range_cmp(l1, l2) > 0;
1227 }
1228
1229 static bool
1230 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1231                 struct pnfs_layout_segment *old)
1232 {
1233         return false;
1234 }
1235
1236 void
1237 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1238                    struct pnfs_layout_segment *lseg,
1239                    bool (*is_after)(const struct pnfs_layout_range *,
1240                            const struct pnfs_layout_range *),
1241                    bool (*do_merge)(struct pnfs_layout_segment *,
1242                            struct pnfs_layout_segment *),
1243                    struct list_head *free_me)
1244 {
1245         struct pnfs_layout_segment *lp, *tmp;
1246
1247         dprintk("%s:Begin\n", __func__);
1248
1249         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1250                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1251                         continue;
1252                 if (do_merge(lseg, lp)) {
1253                         mark_lseg_invalid(lp, free_me);
1254                         continue;
1255                 }
1256                 if (is_after(&lseg->pls_range, &lp->pls_range))
1257                         continue;
1258                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1259                 dprintk("%s: inserted lseg %p "
1260                         "iomode %d offset %llu length %llu before "
1261                         "lp %p iomode %d offset %llu length %llu\n",
1262                         __func__, lseg, lseg->pls_range.iomode,
1263                         lseg->pls_range.offset, lseg->pls_range.length,
1264                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1265                         lp->pls_range.length);
1266                 goto out;
1267         }
1268         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1269         dprintk("%s: inserted lseg %p "
1270                 "iomode %d offset %llu length %llu at tail\n",
1271                 __func__, lseg, lseg->pls_range.iomode,
1272                 lseg->pls_range.offset, lseg->pls_range.length);
1273 out:
1274         pnfs_get_layout_hdr(lo);
1275
1276         dprintk("%s:Return\n", __func__);
1277 }
1278 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1279
1280 static void
1281 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1282                    struct pnfs_layout_segment *lseg,
1283                    struct list_head *free_me)
1284 {
1285         struct inode *inode = lo->plh_inode;
1286         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1287
1288         if (ld->add_lseg != NULL)
1289                 ld->add_lseg(lo, lseg, free_me);
1290         else
1291                 pnfs_generic_layout_insert_lseg(lo, lseg,
1292                                 pnfs_lseg_range_is_after,
1293                                 pnfs_lseg_no_merge,
1294                                 free_me);
1295 }
1296
1297 static struct pnfs_layout_hdr *
1298 alloc_init_layout_hdr(struct inode *ino,
1299                       struct nfs_open_context *ctx,
1300                       gfp_t gfp_flags)
1301 {
1302         struct pnfs_layout_hdr *lo;
1303
1304         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1305         if (!lo)
1306                 return NULL;
1307         atomic_set(&lo->plh_refcount, 1);
1308         INIT_LIST_HEAD(&lo->plh_layouts);
1309         INIT_LIST_HEAD(&lo->plh_segs);
1310         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1311         lo->plh_inode = ino;
1312         lo->plh_lc_cred = get_rpccred(ctx->cred);
1313         return lo;
1314 }
1315
1316 static struct pnfs_layout_hdr *
1317 pnfs_find_alloc_layout(struct inode *ino,
1318                        struct nfs_open_context *ctx,
1319                        gfp_t gfp_flags)
1320 {
1321         struct nfs_inode *nfsi = NFS_I(ino);
1322         struct pnfs_layout_hdr *new = NULL;
1323
1324         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1325
1326         if (nfsi->layout != NULL)
1327                 goto out_existing;
1328         spin_unlock(&ino->i_lock);
1329         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1330         spin_lock(&ino->i_lock);
1331
1332         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1333                 nfsi->layout = new;
1334                 return new;
1335         } else if (new != NULL)
1336                 pnfs_free_layout_hdr(new);
1337 out_existing:
1338         pnfs_get_layout_hdr(nfsi->layout);
1339         return nfsi->layout;
1340 }
1341
1342 /*
1343  * iomode matching rules:
1344  * iomode       lseg    match
1345  * -----        -----   -----
1346  * ANY          READ    true
1347  * ANY          RW      true
1348  * RW           READ    false
1349  * RW           RW      true
1350  * READ         READ    true
1351  * READ         RW      true
1352  */
1353 static bool
1354 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1355                  const struct pnfs_layout_range *range)
1356 {
1357         struct pnfs_layout_range range1;
1358
1359         if ((range->iomode == IOMODE_RW &&
1360              ls_range->iomode != IOMODE_RW) ||
1361             !pnfs_lseg_range_intersecting(ls_range, range))
1362                 return 0;
1363
1364         /* range1 covers only the first byte in the range */
1365         range1 = *range;
1366         range1.length = 1;
1367         return pnfs_lseg_range_contained(ls_range, &range1);
1368 }
1369
1370 /*
1371  * lookup range in layout
1372  */
1373 static struct pnfs_layout_segment *
1374 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1375                 struct pnfs_layout_range *range)
1376 {
1377         struct pnfs_layout_segment *lseg, *ret = NULL;
1378
1379         dprintk("%s:Begin\n", __func__);
1380
1381         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1382                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1383                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1384                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1385                         ret = pnfs_get_lseg(lseg);
1386                         break;
1387                 }
1388         }
1389
1390         dprintk("%s:Return lseg %p ref %d\n",
1391                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1392         return ret;
1393 }
1394
1395 /*
1396  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1397  * to the MDS or over pNFS
1398  *
1399  * The nfs_inode read_io and write_io fields are cumulative counters reset
1400  * when there are no layout segments. Note that in pnfs_update_layout iomode
1401  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1402  * WRITE request.
1403  *
1404  * A return of true means use MDS I/O.
1405  *
1406  * From rfc 5661:
1407  * If a file's size is smaller than the file size threshold, data accesses
1408  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1409  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1410  * server.  If both file size and I/O size are provided, the client SHOULD
1411  * reach or exceed  both thresholds before sending its read or write
1412  * requests to the data server.
1413  */
1414 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1415                                      struct inode *ino, int iomode)
1416 {
1417         struct nfs4_threshold *t = ctx->mdsthreshold;
1418         struct nfs_inode *nfsi = NFS_I(ino);
1419         loff_t fsize = i_size_read(ino);
1420         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1421
1422         if (t == NULL)
1423                 return ret;
1424
1425         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1426                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1427
1428         switch (iomode) {
1429         case IOMODE_READ:
1430                 if (t->bm & THRESHOLD_RD) {
1431                         dprintk("%s fsize %llu\n", __func__, fsize);
1432                         size_set = true;
1433                         if (fsize < t->rd_sz)
1434                                 size = true;
1435                 }
1436                 if (t->bm & THRESHOLD_RD_IO) {
1437                         dprintk("%s nfsi->read_io %llu\n", __func__,
1438                                 nfsi->read_io);
1439                         io_set = true;
1440                         if (nfsi->read_io < t->rd_io_sz)
1441                                 io = true;
1442                 }
1443                 break;
1444         case IOMODE_RW:
1445                 if (t->bm & THRESHOLD_WR) {
1446                         dprintk("%s fsize %llu\n", __func__, fsize);
1447                         size_set = true;
1448                         if (fsize < t->wr_sz)
1449                                 size = true;
1450                 }
1451                 if (t->bm & THRESHOLD_WR_IO) {
1452                         dprintk("%s nfsi->write_io %llu\n", __func__,
1453                                 nfsi->write_io);
1454                         io_set = true;
1455                         if (nfsi->write_io < t->wr_io_sz)
1456                                 io = true;
1457                 }
1458                 break;
1459         }
1460         if (size_set && io_set) {
1461                 if (size && io)
1462                         ret = true;
1463         } else if (size || io)
1464                 ret = true;
1465
1466         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1467         return ret;
1468 }
1469
1470 /* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1471 static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key, int mode)
1472 {
1473         if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1474                 return 1;
1475         return nfs_wait_bit_killable(key, mode);
1476 }
1477
1478 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1479 {
1480         if (!pnfs_should_retry_layoutget(lo))
1481                 return false;
1482         /*
1483          * send layoutcommit as it can hold up layoutreturn due to lseg
1484          * reference
1485          */
1486         pnfs_layoutcommit_inode(lo->plh_inode, false);
1487         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1488                                    pnfs_layoutget_retry_bit_wait,
1489                                    TASK_UNINTERRUPTIBLE);
1490 }
1491
1492 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1493 {
1494         unsigned long *bitlock = &lo->plh_flags;
1495
1496         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1497         smp_mb__after_atomic();
1498         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1499 }
1500
1501 /*
1502  * Layout segment is retreived from the server if not cached.
1503  * The appropriate layout segment is referenced and returned to the caller.
1504  */
1505 struct pnfs_layout_segment *
1506 pnfs_update_layout(struct inode *ino,
1507                    struct nfs_open_context *ctx,
1508                    loff_t pos,
1509                    u64 count,
1510                    enum pnfs_iomode iomode,
1511                    gfp_t gfp_flags)
1512 {
1513         struct pnfs_layout_range arg = {
1514                 .iomode = iomode,
1515                 .offset = pos,
1516                 .length = count,
1517         };
1518         unsigned pg_offset;
1519         struct nfs_server *server = NFS_SERVER(ino);
1520         struct nfs_client *clp = server->nfs_client;
1521         struct pnfs_layout_hdr *lo;
1522         struct pnfs_layout_segment *lseg = NULL;
1523         bool first;
1524
1525         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1526                 goto out;
1527
1528         if (iomode == IOMODE_READ && i_size_read(ino) == 0)
1529                 goto out;
1530
1531         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1532                 goto out;
1533
1534 lookup_again:
1535         nfs4_client_recover_expired_lease(clp);
1536         first = false;
1537         spin_lock(&ino->i_lock);
1538         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1539         if (lo == NULL) {
1540                 spin_unlock(&ino->i_lock);
1541                 goto out;
1542         }
1543
1544         /* Do we even need to bother with this? */
1545         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1546                 dprintk("%s matches recall, use MDS\n", __func__);
1547                 goto out_unlock;
1548         }
1549
1550         /* if LAYOUTGET already failed once we don't try again */
1551         if (pnfs_layout_io_test_failed(lo, iomode) &&
1552             !pnfs_should_retry_layoutget(lo))
1553                 goto out_unlock;
1554
1555         first = list_empty(&lo->plh_segs);
1556         if (first) {
1557                 /* The first layoutget for the file. Need to serialize per
1558                  * RFC 5661 Errata 3208.
1559                  */
1560                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1561                                      &lo->plh_flags)) {
1562                         spin_unlock(&ino->i_lock);
1563                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1564                                     TASK_UNINTERRUPTIBLE);
1565                         pnfs_put_layout_hdr(lo);
1566                         goto lookup_again;
1567                 }
1568         } else {
1569                 /* Check to see if the layout for the given range
1570                  * already exists
1571                  */
1572                 lseg = pnfs_find_lseg(lo, &arg);
1573                 if (lseg)
1574                         goto out_unlock;
1575         }
1576
1577         /*
1578          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1579          * for LAYOUTRETURN even if first is true.
1580          */
1581         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1582                 spin_unlock(&ino->i_lock);
1583                 dprintk("%s wait for layoutreturn\n", __func__);
1584                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1585                         if (first)
1586                                 pnfs_clear_first_layoutget(lo);
1587                         pnfs_put_layout_hdr(lo);
1588                         dprintk("%s retrying\n", __func__);
1589                         goto lookup_again;
1590                 }
1591                 goto out_put_layout_hdr;
1592         }
1593
1594         if (pnfs_layoutgets_blocked(lo))
1595                 goto out_unlock;
1596         atomic_inc(&lo->plh_outstanding);
1597         spin_unlock(&ino->i_lock);
1598
1599         if (list_empty(&lo->plh_layouts)) {
1600                 /* The lo must be on the clp list if there is any
1601                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1602                  */
1603                 spin_lock(&clp->cl_lock);
1604                 if (list_empty(&lo->plh_layouts))
1605                         list_add_tail(&lo->plh_layouts, &server->layouts);
1606                 spin_unlock(&clp->cl_lock);
1607         }
1608
1609         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1610         if (pg_offset) {
1611                 arg.offset -= pg_offset;
1612                 arg.length += pg_offset;
1613         }
1614         if (arg.length != NFS4_MAX_UINT64)
1615                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1616
1617         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1618         pnfs_clear_retry_layoutget(lo);
1619         atomic_dec(&lo->plh_outstanding);
1620 out_put_layout_hdr:
1621         if (first)
1622                 pnfs_clear_first_layoutget(lo);
1623         pnfs_put_layout_hdr(lo);
1624 out:
1625         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1626                         "(%s, offset: %llu, length: %llu)\n",
1627                         __func__, ino->i_sb->s_id,
1628                         (unsigned long long)NFS_FILEID(ino),
1629                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1630                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1631                         (unsigned long long)pos,
1632                         (unsigned long long)count);
1633         return lseg;
1634 out_unlock:
1635         spin_unlock(&ino->i_lock);
1636         goto out_put_layout_hdr;
1637 }
1638 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1639
1640 static bool
1641 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1642 {
1643         switch (range->iomode) {
1644         case IOMODE_READ:
1645         case IOMODE_RW:
1646                 break;
1647         default:
1648                 return false;
1649         }
1650         if (range->offset == NFS4_MAX_UINT64)
1651                 return false;
1652         if (range->length == 0)
1653                 return false;
1654         if (range->length != NFS4_MAX_UINT64 &&
1655             range->length > NFS4_MAX_UINT64 - range->offset)
1656                 return false;
1657         return true;
1658 }
1659
1660 struct pnfs_layout_segment *
1661 pnfs_layout_process(struct nfs4_layoutget *lgp)
1662 {
1663         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1664         struct nfs4_layoutget_res *res = &lgp->res;
1665         struct pnfs_layout_segment *lseg;
1666         struct inode *ino = lo->plh_inode;
1667         LIST_HEAD(free_me);
1668         int status = -EINVAL;
1669
1670         if (!pnfs_sanity_check_layout_range(&res->range))
1671                 goto out;
1672
1673         /* Inject layout blob into I/O device driver */
1674         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1675         if (!lseg || IS_ERR(lseg)) {
1676                 if (!lseg)
1677                         status = -ENOMEM;
1678                 else
1679                         status = PTR_ERR(lseg);
1680                 dprintk("%s: Could not allocate layout: error %d\n",
1681                        __func__, status);
1682                 goto out;
1683         }
1684
1685         init_lseg(lo, lseg);
1686         lseg->pls_range = res->range;
1687
1688         spin_lock(&ino->i_lock);
1689         if (pnfs_layoutgets_blocked(lo)) {
1690                 dprintk("%s forget reply due to state\n", __func__);
1691                 goto out_forget_reply;
1692         }
1693
1694         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1695                 /* existing state ID, make sure the sequence number matches. */
1696                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1697                         dprintk("%s forget reply due to sequence\n", __func__);
1698                         status = -EAGAIN;
1699                         goto out_forget_reply;
1700                 }
1701                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1702         } else {
1703                 /*
1704                  * We got an entirely new state ID.  Mark all segments for the
1705                  * inode invalid, and don't bother validating the stateid
1706                  * sequence number.
1707                  */
1708                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1709
1710                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1711                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1712         }
1713
1714         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1715
1716         pnfs_get_lseg(lseg);
1717         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1718
1719         if (res->return_on_close)
1720                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1721
1722         spin_unlock(&ino->i_lock);
1723         pnfs_free_lseg_list(&free_me);
1724         return lseg;
1725 out:
1726         return ERR_PTR(status);
1727
1728 out_forget_reply:
1729         spin_unlock(&ino->i_lock);
1730         lseg->pls_layout = lo;
1731         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1732         goto out;
1733 }
1734
1735 static void
1736 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1737                                 struct list_head *tmp_list,
1738                                 struct pnfs_layout_range *return_range)
1739 {
1740         struct pnfs_layout_segment *lseg, *next;
1741
1742         dprintk("%s:Begin lo %p\n", __func__, lo);
1743
1744         if (list_empty(&lo->plh_segs))
1745                 return;
1746
1747         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1748                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1749                         dprintk("%s: marking lseg %p iomode %d "
1750                                 "offset %llu length %llu\n", __func__,
1751                                 lseg, lseg->pls_range.iomode,
1752                                 lseg->pls_range.offset,
1753                                 lseg->pls_range.length);
1754                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1755                         mark_lseg_invalid(lseg, tmp_list);
1756                         set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1757                                         &lo->plh_flags);
1758                 }
1759 }
1760
1761 void pnfs_error_mark_layout_for_return(struct inode *inode,
1762                                        struct pnfs_layout_segment *lseg)
1763 {
1764         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1765         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1766         struct pnfs_layout_range range = {
1767                 .iomode = lseg->pls_range.iomode,
1768                 .offset = 0,
1769                 .length = NFS4_MAX_UINT64,
1770         };
1771         LIST_HEAD(free_me);
1772
1773         spin_lock(&inode->i_lock);
1774         /* set failure bit so that pnfs path will be retried later */
1775         pnfs_layout_set_fail_bit(lo, iomode);
1776         if (lo->plh_return_iomode == 0)
1777                 lo->plh_return_iomode = range.iomode;
1778         else if (lo->plh_return_iomode != range.iomode)
1779                 lo->plh_return_iomode = IOMODE_ANY;
1780         /*
1781          * mark all matching lsegs so that we are sure to have no live
1782          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1783          * for how it works.
1784          */
1785         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1786         spin_unlock(&inode->i_lock);
1787         pnfs_free_lseg_list(&free_me);
1788 }
1789 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1790
1791 void
1792 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1793 {
1794         u64 rd_size = req->wb_bytes;
1795
1796         if (pgio->pg_lseg == NULL) {
1797                 if (pgio->pg_dreq == NULL)
1798                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1799                 else
1800                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1801
1802                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1803                                                    req->wb_context,
1804                                                    req_offset(req),
1805                                                    rd_size,
1806                                                    IOMODE_READ,
1807                                                    GFP_KERNEL);
1808                 if (IS_ERR(pgio->pg_lseg)) {
1809                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1810                         pgio->pg_lseg = NULL;
1811                         return;
1812                 }
1813         }
1814         /* If no lseg, fall back to read through mds */
1815         if (pgio->pg_lseg == NULL)
1816                 nfs_pageio_reset_read_mds(pgio);
1817
1818 }
1819 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1820
1821 void
1822 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1823                            struct nfs_page *req, u64 wb_size)
1824 {
1825         if (pgio->pg_lseg == NULL) {
1826                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1827                                                    req->wb_context,
1828                                                    req_offset(req),
1829                                                    wb_size,
1830                                                    IOMODE_RW,
1831                                                    GFP_NOFS);
1832                 if (IS_ERR(pgio->pg_lseg)) {
1833                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1834                         pgio->pg_lseg = NULL;
1835                         return;
1836                 }
1837         }
1838         /* If no lseg, fall back to write through mds */
1839         if (pgio->pg_lseg == NULL)
1840                 nfs_pageio_reset_write_mds(pgio);
1841 }
1842 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1843
1844 void
1845 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1846 {
1847         if (desc->pg_lseg) {
1848                 pnfs_put_lseg(desc->pg_lseg);
1849                 desc->pg_lseg = NULL;
1850         }
1851 }
1852 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1853
1854 /*
1855  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1856  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1857  */
1858 size_t
1859 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1860                      struct nfs_page *prev, struct nfs_page *req)
1861 {
1862         unsigned int size;
1863         u64 seg_end, req_start, seg_left;
1864
1865         size = nfs_generic_pg_test(pgio, prev, req);
1866         if (!size)
1867                 return 0;
1868
1869         /*
1870          * 'size' contains the number of bytes left in the current page (up
1871          * to the original size asked for in @req->wb_bytes).
1872          *
1873          * Calculate how many bytes are left in the layout segment
1874          * and if there are less bytes than 'size', return that instead.
1875          *
1876          * Please also note that 'end_offset' is actually the offset of the
1877          * first byte that lies outside the pnfs_layout_range. FIXME?
1878          *
1879          */
1880         if (pgio->pg_lseg) {
1881                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1882                                      pgio->pg_lseg->pls_range.length);
1883                 req_start = req_offset(req);
1884                 WARN_ON_ONCE(req_start >= seg_end);
1885                 /* start of request is past the last byte of this segment */
1886                 if (req_start >= seg_end) {
1887                         /* reference the new lseg */
1888                         if (pgio->pg_ops->pg_cleanup)
1889                                 pgio->pg_ops->pg_cleanup(pgio);
1890                         if (pgio->pg_ops->pg_init)
1891                                 pgio->pg_ops->pg_init(pgio, req);
1892                         return 0;
1893                 }
1894
1895                 /* adjust 'size' iff there are fewer bytes left in the
1896                  * segment than what nfs_generic_pg_test returned */
1897                 seg_left = seg_end - req_start;
1898                 if (seg_left < size)
1899                         size = (unsigned int)seg_left;
1900         }
1901
1902         return size;
1903 }
1904 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1905
1906 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1907 {
1908         struct nfs_pageio_descriptor pgio;
1909
1910         /* Resend all requests through the MDS */
1911         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1912                               hdr->completion_ops);
1913         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1914         return nfs_pageio_resend(&pgio, hdr);
1915 }
1916 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1917
1918 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1919 {
1920
1921         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1922         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1923             PNFS_LAYOUTRET_ON_ERROR) {
1924                 pnfs_return_layout(hdr->inode);
1925         }
1926         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1927                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1928 }
1929
1930 /*
1931  * Called by non rpc-based layout drivers
1932  */
1933 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1934 {
1935         if (likely(!hdr->pnfs_error)) {
1936                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1937                                 hdr->mds_offset + hdr->res.count);
1938                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1939         }
1940         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1941         if (unlikely(hdr->pnfs_error))
1942                 pnfs_ld_handle_write_error(hdr);
1943         hdr->mds_ops->rpc_release(hdr);
1944 }
1945 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1946
1947 static void
1948 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1949                 struct nfs_pgio_header *hdr)
1950 {
1951         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1952
1953         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1954                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1955                 nfs_pageio_reset_write_mds(desc);
1956                 mirror->pg_recoalesce = 1;
1957         }
1958         hdr->completion_ops->completion(hdr);
1959 }
1960
1961 static enum pnfs_try_status
1962 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1963                         const struct rpc_call_ops *call_ops,
1964                         struct pnfs_layout_segment *lseg,
1965                         int how)
1966 {
1967         struct inode *inode = hdr->inode;
1968         enum pnfs_try_status trypnfs;
1969         struct nfs_server *nfss = NFS_SERVER(inode);
1970
1971         hdr->mds_ops = call_ops;
1972
1973         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1974                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1975         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1976         if (trypnfs != PNFS_NOT_ATTEMPTED)
1977                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1978         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1979         return trypnfs;
1980 }
1981
1982 static void
1983 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1984               struct nfs_pgio_header *hdr, int how)
1985 {
1986         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1987         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1988         enum pnfs_try_status trypnfs;
1989
1990         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1991         if (trypnfs == PNFS_NOT_ATTEMPTED)
1992                 pnfs_write_through_mds(desc, hdr);
1993 }
1994
1995 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1996 {
1997         pnfs_put_lseg(hdr->lseg);
1998         nfs_pgio_header_free(hdr);
1999 }
2000
2001 int
2002 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2003 {
2004         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2005
2006         struct nfs_pgio_header *hdr;
2007         int ret;
2008
2009         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2010         if (!hdr) {
2011                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2012                 return -ENOMEM;
2013         }
2014         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2015
2016         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2017         ret = nfs_generic_pgio(desc, hdr);
2018         if (!ret)
2019                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2020
2021         return ret;
2022 }
2023 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2024
2025 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2026 {
2027         struct nfs_pageio_descriptor pgio;
2028
2029         /* Resend all requests through the MDS */
2030         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2031         return nfs_pageio_resend(&pgio, hdr);
2032 }
2033 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2034
2035 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2036 {
2037         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2038         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2039             PNFS_LAYOUTRET_ON_ERROR) {
2040                 pnfs_return_layout(hdr->inode);
2041         }
2042         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2043                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2044 }
2045
2046 /*
2047  * Called by non rpc-based layout drivers
2048  */
2049 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2050 {
2051         if (likely(!hdr->pnfs_error)) {
2052                 __nfs4_read_done_cb(hdr);
2053                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2054         }
2055         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2056         if (unlikely(hdr->pnfs_error))
2057                 pnfs_ld_handle_read_error(hdr);
2058         hdr->mds_ops->rpc_release(hdr);
2059 }
2060 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2061
2062 static void
2063 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2064                 struct nfs_pgio_header *hdr)
2065 {
2066         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2067
2068         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2069                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2070                 nfs_pageio_reset_read_mds(desc);
2071                 mirror->pg_recoalesce = 1;
2072         }
2073         hdr->completion_ops->completion(hdr);
2074 }
2075
2076 /*
2077  * Call the appropriate parallel I/O subsystem read function.
2078  */
2079 static enum pnfs_try_status
2080 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2081                        const struct rpc_call_ops *call_ops,
2082                        struct pnfs_layout_segment *lseg)
2083 {
2084         struct inode *inode = hdr->inode;
2085         struct nfs_server *nfss = NFS_SERVER(inode);
2086         enum pnfs_try_status trypnfs;
2087
2088         hdr->mds_ops = call_ops;
2089
2090         dprintk("%s: Reading ino:%lu %u@%llu\n",
2091                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2092
2093         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2094         if (trypnfs != PNFS_NOT_ATTEMPTED)
2095                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2096         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2097         return trypnfs;
2098 }
2099
2100 /* Resend all requests through pnfs. */
2101 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2102 {
2103         struct nfs_pageio_descriptor pgio;
2104
2105         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2106         return nfs_pageio_resend(&pgio, hdr);
2107 }
2108 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2109
2110 static void
2111 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2112 {
2113         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2114         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2115         enum pnfs_try_status trypnfs;
2116         int err = 0;
2117
2118         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2119         if (trypnfs == PNFS_TRY_AGAIN)
2120                 err = pnfs_read_resend_pnfs(hdr);
2121         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2122                 pnfs_read_through_mds(desc, hdr);
2123 }
2124
2125 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2126 {
2127         pnfs_put_lseg(hdr->lseg);
2128         nfs_pgio_header_free(hdr);
2129 }
2130
2131 int
2132 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2133 {
2134         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2135
2136         struct nfs_pgio_header *hdr;
2137         int ret;
2138
2139         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2140         if (!hdr) {
2141                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2142                 return -ENOMEM;
2143         }
2144         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2145         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2146         ret = nfs_generic_pgio(desc, hdr);
2147         if (!ret)
2148                 pnfs_do_read(desc, hdr);
2149         return ret;
2150 }
2151 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2152
2153 static void pnfs_clear_layoutcommitting(struct inode *inode)
2154 {
2155         unsigned long *bitlock = &NFS_I(inode)->flags;
2156
2157         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2158         smp_mb__after_atomic();
2159         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2160 }
2161
2162 /*
2163  * There can be multiple RW segments.
2164  */
2165 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2166 {
2167         struct pnfs_layout_segment *lseg;
2168
2169         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2170                 if (lseg->pls_range.iomode == IOMODE_RW &&
2171                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2172                         list_add(&lseg->pls_lc_list, listp);
2173         }
2174 }
2175
2176 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2177 {
2178         struct pnfs_layout_segment *lseg, *tmp;
2179
2180         /* Matched by references in pnfs_set_layoutcommit */
2181         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2182                 list_del_init(&lseg->pls_lc_list);
2183                 pnfs_put_lseg(lseg);
2184         }
2185
2186         pnfs_clear_layoutcommitting(inode);
2187 }
2188
2189 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2190 {
2191         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2192 }
2193 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2194
2195 void
2196 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2197                 loff_t end_pos)
2198 {
2199         struct nfs_inode *nfsi = NFS_I(inode);
2200         bool mark_as_dirty = false;
2201
2202         spin_lock(&inode->i_lock);
2203         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2204                 nfsi->layout->plh_lwb = end_pos;
2205                 mark_as_dirty = true;
2206                 dprintk("%s: Set layoutcommit for inode %lu ",
2207                         __func__, inode->i_ino);
2208         } else if (end_pos > nfsi->layout->plh_lwb)
2209                 nfsi->layout->plh_lwb = end_pos;
2210         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2211                 /* references matched in nfs4_layoutcommit_release */
2212                 pnfs_get_lseg(lseg);
2213         }
2214         spin_unlock(&inode->i_lock);
2215         dprintk("%s: lseg %p end_pos %llu\n",
2216                 __func__, lseg, nfsi->layout->plh_lwb);
2217
2218         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2219          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2220         if (mark_as_dirty)
2221                 mark_inode_dirty_sync(inode);
2222 }
2223 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2224
2225 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2226 {
2227         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2228
2229         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2230                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2231         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2232 }
2233
2234 /*
2235  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2236  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2237  * data to disk to allow the server to recover the data if it crashes.
2238  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2239  * is off, and a COMMIT is sent to a data server, or
2240  * if WRITEs to a data server return NFS_DATA_SYNC.
2241  */
2242 int
2243 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2244 {
2245         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2246         struct nfs4_layoutcommit_data *data;
2247         struct nfs_inode *nfsi = NFS_I(inode);
2248         loff_t end_pos;
2249         int status;
2250
2251         if (!pnfs_layoutcommit_outstanding(inode))
2252                 return 0;
2253
2254         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2255
2256         status = -EAGAIN;
2257         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2258                 if (!sync)
2259                         goto out;
2260                 status = wait_on_bit_lock_action(&nfsi->flags,
2261                                 NFS_INO_LAYOUTCOMMITTING,
2262                                 nfs_wait_bit_killable,
2263                                 TASK_KILLABLE);
2264                 if (status)
2265                         goto out;
2266         }
2267
2268         status = -ENOMEM;
2269         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2270         data = kzalloc(sizeof(*data), GFP_NOFS);
2271         if (!data)
2272                 goto clear_layoutcommitting;
2273
2274         status = 0;
2275         spin_lock(&inode->i_lock);
2276         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2277                 goto out_unlock;
2278
2279         INIT_LIST_HEAD(&data->lseg_list);
2280         pnfs_list_write_lseg(inode, &data->lseg_list);
2281
2282         end_pos = nfsi->layout->plh_lwb;
2283
2284         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2285         spin_unlock(&inode->i_lock);
2286
2287         data->args.inode = inode;
2288         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2289         nfs_fattr_init(&data->fattr);
2290         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2291         data->res.fattr = &data->fattr;
2292         data->args.lastbytewritten = end_pos - 1;
2293         data->res.server = NFS_SERVER(inode);
2294
2295         if (ld->prepare_layoutcommit) {
2296                 status = ld->prepare_layoutcommit(&data->args);
2297                 if (status) {
2298                         put_rpccred(data->cred);
2299                         spin_lock(&inode->i_lock);
2300                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2301                         if (end_pos > nfsi->layout->plh_lwb)
2302                                 nfsi->layout->plh_lwb = end_pos;
2303                         goto out_unlock;
2304                 }
2305         }
2306
2307
2308         status = nfs4_proc_layoutcommit(data, sync);
2309 out:
2310         if (status)
2311                 mark_inode_dirty_sync(inode);
2312         dprintk("<-- %s status %d\n", __func__, status);
2313         return status;
2314 out_unlock:
2315         spin_unlock(&inode->i_lock);
2316         kfree(data);
2317 clear_layoutcommitting:
2318         pnfs_clear_layoutcommitting(inode);
2319         goto out;
2320 }
2321 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2322
2323 int
2324 pnfs_generic_sync(struct inode *inode, bool datasync)
2325 {
2326         return pnfs_layoutcommit_inode(inode, true);
2327 }
2328 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2329
2330 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2331 {
2332         struct nfs4_threshold *thp;
2333
2334         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2335         if (!thp) {
2336                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2337                 return NULL;
2338         }
2339         return thp;
2340 }
2341
2342 #if IS_ENABLED(CONFIG_NFS_V4_2)
2343 int
2344 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2345 {
2346         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2347         struct nfs_server *server = NFS_SERVER(inode);
2348         struct nfs_inode *nfsi = NFS_I(inode);
2349         struct nfs42_layoutstat_data *data;
2350         struct pnfs_layout_hdr *hdr;
2351         int status = 0;
2352
2353         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2354                 goto out;
2355
2356         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2357                 goto out;
2358
2359         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2360                 goto out;
2361
2362         spin_lock(&inode->i_lock);
2363         if (!NFS_I(inode)->layout) {
2364                 spin_unlock(&inode->i_lock);
2365                 goto out;
2366         }
2367         hdr = NFS_I(inode)->layout;
2368         pnfs_get_layout_hdr(hdr);
2369         spin_unlock(&inode->i_lock);
2370
2371         data = kzalloc(sizeof(*data), gfp_flags);
2372         if (!data) {
2373                 status = -ENOMEM;
2374                 goto out_put;
2375         }
2376
2377         data->args.fh = NFS_FH(inode);
2378         data->args.inode = inode;
2379         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2380         status = ld->prepare_layoutstats(&data->args);
2381         if (status)
2382                 goto out_free;
2383
2384         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2385
2386 out:
2387         dprintk("%s returns %d\n", __func__, status);
2388         return status;
2389
2390 out_free:
2391         kfree(data);
2392 out_put:
2393         pnfs_put_layout_hdr(hdr);
2394         smp_mb__before_atomic();
2395         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2396         smp_mb__after_atomic();
2397         goto out;
2398 }
2399 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2400 #endif
2401
2402 unsigned int layoutstats_timer;
2403 module_param(layoutstats_timer, uint, 0644);
2404 EXPORT_SYMBOL_GPL(layoutstats_timer);