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