kernel/fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37 #include "delegation.h"
  38 #include "nfs42.h"
  39
  40 #define NFSDBG_FACILITY         NFSDBG_PNFS
  41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  42
  43 /* Locking:
  44  *
  45  * pnfs_spinlock:
  46  *      protects pnfs_modules_tbl.
  47  */
  48 static DEFINE_SPINLOCK(pnfs_spinlock);
  49
  50 /*
  51  * pnfs_modules_tbl holds all pnfs modules
  52  */
  53 static LIST_HEAD(pnfs_modules_tbl);
  54
  55 static int
  56 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
  57                        enum pnfs_iomode iomode, bool sync);
  58
  59 /* Return the registered pnfs layout driver module matching given id */
  60 static struct pnfs_layoutdriver_type *
  61 find_pnfs_driver_locked(u32 id)
  62 {
  63         struct pnfs_layoutdriver_type *local;
  64
  65         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  66                 if (local->id == id)
  67                         goto out;
  68         local = NULL;
  69 out:
  70         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  71         return local;
  72 }
  73
  74 static struct pnfs_layoutdriver_type *
  75 find_pnfs_driver(u32 id)
  76 {
  77         struct pnfs_layoutdriver_type *local;
  78
  79         spin_lock(&pnfs_spinlock);
  80         local = find_pnfs_driver_locked(id);
  81         if (local != NULL && !try_module_get(local->owner)) {
  82                 dprintk("%s: Could not grab reference on module\n", __func__);
  83                 local = NULL;
  84         }
  85         spin_unlock(&pnfs_spinlock);
  86         return local;
  87 }
  88
  89 void
  90 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  91 {
  92         if (nfss->pnfs_curr_ld) {
  93                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  94                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  95                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  96                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  97                         nfs4_deviceid_purge_client(nfss->nfs_client);
  98                 module_put(nfss->pnfs_curr_ld->owner);
  99         }
 100         nfss->pnfs_curr_ld = NULL;
 101 }
 102
 103 /*
 104  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 105  * Currently only one pNFS layout driver per filesystem is supported.
 106  *
 107  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 108  */
 109 void
 110 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 111                       u32 id)
 112 {
 113         struct pnfs_layoutdriver_type *ld_type = NULL;
 114
 115         if (id == 0)
 116                 goto out_no_driver;
 117         if (!(server->nfs_client->cl_exchange_flags &
 118                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 119                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 120                         __func__, id, server->nfs_client->cl_exchange_flags);
 121                 goto out_no_driver;
 122         }
 123         ld_type = find_pnfs_driver(id);
 124         if (!ld_type) {
 125                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 126                 ld_type = find_pnfs_driver(id);
 127                 if (!ld_type) {
 128                         dprintk("%s: No pNFS module found for %u.\n",
 129                                 __func__, id);
 130                         goto out_no_driver;
 131                 }
 132         }
 133         server->pnfs_curr_ld = ld_type;
 134         if (ld_type->set_layoutdriver
 135             && ld_type->set_layoutdriver(server, mntfh)) {
 136                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 137                         "driver %u.\n", __func__, id);
 138                 module_put(ld_type->owner);
 139                 goto out_no_driver;
 140         }
 141         /* Bump the MDS count */
 142         atomic_inc(&server->nfs_client->cl_mds_count);
 143
 144         dprintk("%s: pNFS module for %u set\n", __func__, id);
 145         return;
 146
 147 out_no_driver:
 148         dprintk("%s: Using NFSv4 I/O\n", __func__);
 149         server->pnfs_curr_ld = NULL;
 150 }
 151
 152 int
 153 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 154 {
 155         int status = -EINVAL;
 156         struct pnfs_layoutdriver_type *tmp;
 157
 158         if (ld_type->id == 0) {
 159                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 160                 return status;
 161         }
 162         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 163                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 164                        "alloc_lseg and free_lseg.\n", __func__);
 165                 return status;
 166         }
 167
 168         spin_lock(&pnfs_spinlock);
 169         tmp = find_pnfs_driver_locked(ld_type->id);
 170         if (!tmp) {
 171                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 172                 status = 0;
 173                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 174                         ld_type->name);
 175         } else {
 176                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 177                         __func__, ld_type->id);
 178         }
 179         spin_unlock(&pnfs_spinlock);
 180
 181         return status;
 182 }
 183 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 184
 185 void
 186 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 187 {
 188         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 189         spin_lock(&pnfs_spinlock);
 190         list_del(&ld_type->pnfs_tblid);
 191         spin_unlock(&pnfs_spinlock);
 192 }
 193 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 194
 195 /*
 196  * pNFS client layout cache
 197  */
 198
 199 /* Need to hold i_lock if caller does not already hold reference */
 200 void
 201 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 202 {
 203         atomic_inc(&lo->plh_refcount);
 204 }
 205
 206 static struct pnfs_layout_hdr *
 207 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 208 {
 209         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 210         return ld->alloc_layout_hdr(ino, gfp_flags);
 211 }
 212
 213 static void
 214 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 215 {
 216         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 217         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 218
 219         if (!list_empty(&lo->plh_layouts)) {
 220                 struct nfs_client *clp = server->nfs_client;
 221
 222                 spin_lock(&clp->cl_lock);
 223                 list_del_init(&lo->plh_layouts);
 224                 spin_unlock(&clp->cl_lock);
 225         }
 226         put_rpccred(lo->plh_lc_cred);
 227         return ld->free_layout_hdr(lo);
 228 }
 229
 230 static void
 231 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 232 {
 233         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 234         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 235         nfsi->layout = NULL;
 236         /* Reset MDS Threshold I/O counters */
 237         nfsi->write_io = 0;
 238         nfsi->read_io = 0;
 239 }
 240
 241 void
 242 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 243 {
 244         struct inode *inode = lo->plh_inode;
 245
 246         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 247                 if (!list_empty(&lo->plh_segs))
 248                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 249                 pnfs_detach_layout_hdr(lo);
 250                 spin_unlock(&inode->i_lock);
 251                 pnfs_free_layout_hdr(lo);
 252         }
 253 }
 254
 255 static int
 256 pnfs_iomode_to_fail_bit(u32 iomode)
 257 {
 258         return iomode == IOMODE_RW ?
 259                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 260 }
 261
 262 static void
 263 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 264 {
 265         lo->plh_retry_timestamp = jiffies;
 266         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 267                 atomic_inc(&lo->plh_refcount);
 268 }
 269
 270 static void
 271 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 272 {
 273         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 274                 atomic_dec(&lo->plh_refcount);
 275 }
 276
 277 static void
 278 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 279 {
 280         struct inode *inode = lo->plh_inode;
 281         struct pnfs_layout_range range = {
 282                 .iomode = iomode,
 283                 .offset = 0,
 284                 .length = NFS4_MAX_UINT64,
 285         };
 286         LIST_HEAD(head);
 287
 288         spin_lock(&inode->i_lock);
 289         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 290         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 291         spin_unlock(&inode->i_lock);
 292         pnfs_free_lseg_list(&head);
 293         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 294                         iomode == IOMODE_RW ?  "RW" : "READ");
 295 }
 296
 297 static bool
 298 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 299 {
 300         unsigned long start, end;
 301         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 302
 303         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 304                 return false;
 305         end = jiffies;
 306         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 307         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 308                 /* It is time to retry the failed layoutgets */
 309                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 310                 return false;
 311         }
 312         return true;
 313 }
 314
 315 static void
 316 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 317 {
 318         INIT_LIST_HEAD(&lseg->pls_list);
 319         INIT_LIST_HEAD(&lseg->pls_lc_list);
 320         atomic_set(&lseg->pls_refcount, 1);
 321         smp_mb();
 322         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 323         lseg->pls_layout = lo;
 324 }
 325
 326 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *ino = lseg->pls_layout->plh_inode;
 329
 330         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 331 }
 332
 333 static void
 334 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 335                 struct pnfs_layout_segment *lseg)
 336 {
 337         struct inode *inode = lo->plh_inode;
 338
 339         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 340         list_del_init(&lseg->pls_list);
 341         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 342         atomic_dec(&lo->plh_refcount);
 343         if (list_empty(&lo->plh_segs))
 344                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 345         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 346 }
 347
 348 /* Return true if layoutreturn is needed */
 349 static bool
 350 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 351                         struct pnfs_layout_segment *lseg)
 352 {
 353         struct pnfs_layout_segment *s;
 354
 355         if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 356                 return false;
 357
 358         list_for_each_entry(s, &lo->plh_segs, pls_list)
 359                 if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 360                         return false;
 361
 362         return true;
 363 }
 364
 365 static bool
 366 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 367 {
 368         /* Serialise LAYOUTGET/LAYOUTRETURN */
 369         if (atomic_read(&lo->plh_outstanding) != 0)
 370                 return false;
 371         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 372                 return false;
 373         lo->plh_return_iomode = 0;
 374         pnfs_get_layout_hdr(lo);
 375         clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
 376         return true;
 377 }
 378
 379 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
 380                 struct pnfs_layout_hdr *lo, struct inode *inode)
 381 {
 382         lo = lseg->pls_layout;
 383         inode = lo->plh_inode;
 384
 385         spin_lock(&inode->i_lock);
 386         if (pnfs_layout_need_return(lo, lseg)) {
 387                 nfs4_stateid stateid;
 388                 enum pnfs_iomode iomode;
 389                 bool send;
 390
 391                 stateid = lo->plh_stateid;
 392                 iomode = lo->plh_return_iomode;
 393                 send = pnfs_prepare_layoutreturn(lo);
 394                 spin_unlock(&inode->i_lock);
 395                 if (send) {
 396                         /* Send an async layoutreturn so we dont deadlock */
 397                         pnfs_send_layoutreturn(lo, stateid, iomode, false);
 398                 }
 399         } else
 400                 spin_unlock(&inode->i_lock);
 401 }
 402
 403 void
 404 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 405 {
 406         struct pnfs_layout_hdr *lo;
 407         struct inode *inode;
 408
 409         if (!lseg)
 410                 return;
 411
 412         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 413                 atomic_read(&lseg->pls_refcount),
 414                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 415
 416         /* Handle the case where refcount != 1 */
 417         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
 418                 return;
 419
 420         lo = lseg->pls_layout;
 421         inode = lo->plh_inode;
 422         /* Do we need a layoutreturn? */
 423         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 424                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
 425
 426         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 427                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 428                         spin_unlock(&inode->i_lock);
 429                         return;
 430                 }
 431                 pnfs_get_layout_hdr(lo);
 432                 pnfs_layout_remove_lseg(lo, lseg);
 433                 spin_unlock(&inode->i_lock);
 434                 pnfs_free_lseg(lseg);
 435                 pnfs_put_layout_hdr(lo);
 436         }
 437 }
 438 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 439
 440 static void pnfs_free_lseg_async_work(struct work_struct *work)
 441 {
 442         struct pnfs_layout_segment *lseg;
 443         struct pnfs_layout_hdr *lo;
 444
 445         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 446         lo = lseg->pls_layout;
 447
 448         pnfs_free_lseg(lseg);
 449         pnfs_put_layout_hdr(lo);
 450 }
 451
 452 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 453 {
 454         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 455         schedule_work(&lseg->pls_work);
 456 }
 457
 458 void
 459 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 460 {
 461         if (!lseg)
 462                 return;
 463
 464         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 465
 466         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 467                 atomic_read(&lseg->pls_refcount),
 468                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 469         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 470                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 471                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 472                         return;
 473                 pnfs_get_layout_hdr(lo);
 474                 pnfs_layout_remove_lseg(lo, lseg);
 475                 pnfs_free_lseg_async(lseg);
 476         }
 477 }
 478 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 479
 480 static u64
 481 end_offset(u64 start, u64 len)
 482 {
 483         u64 end;
 484
 485         end = start + len;
 486         return end >= start ? end : NFS4_MAX_UINT64;
 487 }
 488
 489 /*
 490  * is l2 fully contained in l1?
 491  *   start1                             end1
 492  *   [----------------------------------)
 493  *           start2           end2
 494  *           [----------------)
 495  */
 496 static bool
 497 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 498                  const struct pnfs_layout_range *l2)
 499 {
 500         u64 start1 = l1->offset;
 501         u64 end1 = end_offset(start1, l1->length);
 502         u64 start2 = l2->offset;
 503         u64 end2 = end_offset(start2, l2->length);
 504
 505         return (start1 <= start2) && (end1 >= end2);
 506 }
 507
 508 /*
 509  * is l1 and l2 intersecting?
 510  *   start1                             end1
 511  *   [----------------------------------)
 512  *                              start2           end2
 513  *                              [----------------)
 514  */
 515 static bool
 516 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 517                     const struct pnfs_layout_range *l2)
 518 {
 519         u64 start1 = l1->offset;
 520         u64 end1 = end_offset(start1, l1->length);
 521         u64 start2 = l2->offset;
 522         u64 end2 = end_offset(start2, l2->length);
 523
 524         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 525                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 526 }
 527
 528 static bool
 529 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 530                  const struct pnfs_layout_range *recall_range)
 531 {
 532         return (recall_range->iomode == IOMODE_ANY ||
 533                 lseg_range->iomode == recall_range->iomode) &&
 534                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 535 }
 536
 537 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 538                 struct list_head *tmp_list)
 539 {
 540         if (!atomic_dec_and_test(&lseg->pls_refcount))
 541                 return false;
 542         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 543         list_add(&lseg->pls_list, tmp_list);
 544         return true;
 545 }
 546
 547 /* Returns 1 if lseg is removed from list, 0 otherwise */
 548 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 549                              struct list_head *tmp_list)
 550 {
 551         int rv = 0;
 552
 553         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 554                 /* Remove the reference keeping the lseg in the
 555                  * list.  It will now be removed when all
 556                  * outstanding io is finished.
 557                  */
 558                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 559                         atomic_read(&lseg->pls_refcount));
 560                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 561                         rv = 1;
 562         }
 563         return rv;
 564 }
 565
 566 /* Returns count of number of matching invalid lsegs remaining in list
 567  * after call.
 568  */
 569 int
 570 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 571                             struct list_head *tmp_list,
 572                             struct pnfs_layout_range *recall_range)
 573 {
 574         struct pnfs_layout_segment *lseg, *next;
 575         int invalid = 0, removed = 0;
 576
 577         dprintk("%s:Begin lo %p\n", __func__, lo);
 578
 579         if (list_empty(&lo->plh_segs))
 580                 return 0;
 581         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 582                 if (!recall_range ||
 583                     should_free_lseg(&lseg->pls_range, recall_range)) {
 584                         dprintk("%s: freeing lseg %p iomode %d "
 585                                 "offset %llu length %llu\n", __func__,
 586                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 587                                 lseg->pls_range.length);
 588                         invalid++;
 589                         removed += mark_lseg_invalid(lseg, tmp_list);
 590                 }
 591         dprintk("%s:Return %i\n", __func__, invalid - removed);
 592         return invalid - removed;
 593 }
 594
 595 /* note free_me must contain lsegs from a single layout_hdr */
 596 void
 597 pnfs_free_lseg_list(struct list_head *free_me)
 598 {
 599         struct pnfs_layout_segment *lseg, *tmp;
 600
 601         if (list_empty(free_me))
 602                 return;
 603
 604         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 605                 list_del(&lseg->pls_list);
 606                 pnfs_free_lseg(lseg);
 607         }
 608 }
 609
 610 void
 611 pnfs_destroy_layout(struct nfs_inode *nfsi)
 612 {
 613         struct pnfs_layout_hdr *lo;
 614         LIST_HEAD(tmp_list);
 615
 616         spin_lock(&nfsi->vfs_inode.i_lock);
 617         lo = nfsi->layout;
 618         if (lo) {
 619                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 620                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 621                 pnfs_get_layout_hdr(lo);
 622                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 623                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 624                 pnfs_clear_retry_layoutget(lo);
 625                 spin_unlock(&nfsi->vfs_inode.i_lock);
 626                 pnfs_free_lseg_list(&tmp_list);
 627                 pnfs_put_layout_hdr(lo);
 628         } else
 629                 spin_unlock(&nfsi->vfs_inode.i_lock);
 630 }
 631 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 632
 633 static bool
 634 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 635                 struct list_head *layout_list)
 636 {
 637         struct pnfs_layout_hdr *lo;
 638         bool ret = false;
 639
 640         spin_lock(&inode->i_lock);
 641         lo = NFS_I(inode)->layout;
 642         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 643                 pnfs_get_layout_hdr(lo);
 644                 list_add(&lo->plh_bulk_destroy, layout_list);
 645                 ret = true;
 646         }
 647         spin_unlock(&inode->i_lock);
 648         return ret;
 649 }
 650
 651 /* Caller must hold rcu_read_lock and clp->cl_lock */
 652 static int
 653 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 654                 struct nfs_server *server,
 655                 struct list_head *layout_list)
 656 {
 657         struct pnfs_layout_hdr *lo, *next;
 658         struct inode *inode;
 659
 660         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 661                 inode = igrab(lo->plh_inode);
 662                 if (inode == NULL)
 663                         continue;
 664                 list_del_init(&lo->plh_layouts);
 665                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 666                         continue;
 667                 rcu_read_unlock();
 668                 spin_unlock(&clp->cl_lock);
 669                 iput(inode);
 670                 spin_lock(&clp->cl_lock);
 671                 rcu_read_lock();
 672                 return -EAGAIN;
 673         }
 674         return 0;
 675 }
 676
 677 static int
 678 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 679                 bool is_bulk_recall)
 680 {
 681         struct pnfs_layout_hdr *lo;
 682         struct inode *inode;
 683         struct pnfs_layout_range range = {
 684                 .iomode = IOMODE_ANY,
 685                 .offset = 0,
 686                 .length = NFS4_MAX_UINT64,
 687         };
 688         LIST_HEAD(lseg_list);
 689         int ret = 0;
 690
 691         while (!list_empty(layout_list)) {
 692                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 693                                 plh_bulk_destroy);
 694                 dprintk("%s freeing layout for inode %lu\n", __func__,
 695                         lo->plh_inode->i_ino);
 696                 inode = lo->plh_inode;
 697
 698                 pnfs_layoutcommit_inode(inode, false);
 699
 700                 spin_lock(&inode->i_lock);
 701                 list_del_init(&lo->plh_bulk_destroy);
 702                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 703                 if (is_bulk_recall)
 704                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 705                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 706                         ret = -EAGAIN;
 707                 spin_unlock(&inode->i_lock);
 708                 pnfs_free_lseg_list(&lseg_list);
 709                 pnfs_put_layout_hdr(lo);
 710                 iput(inode);
 711         }
 712         return ret;
 713 }
 714
 715 int
 716 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 717                 struct nfs_fsid *fsid,
 718                 bool is_recall)
 719 {
 720         struct nfs_server *server;
 721         LIST_HEAD(layout_list);
 722
 723         spin_lock(&clp->cl_lock);
 724         rcu_read_lock();
 725 restart:
 726         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 727                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 728                         continue;
 729                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 730                                 server,
 731                                 &layout_list) != 0)
 732                         goto restart;
 733         }
 734         rcu_read_unlock();
 735         spin_unlock(&clp->cl_lock);
 736
 737         if (list_empty(&layout_list))
 738                 return 0;
 739         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 740 }
 741
 742 int
 743 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 744                 bool is_recall)
 745 {
 746         struct nfs_server *server;
 747         LIST_HEAD(layout_list);
 748
 749         spin_lock(&clp->cl_lock);
 750         rcu_read_lock();
 751 restart:
 752         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 753                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 754                                         server,
 755                                         &layout_list) != 0)
 756                         goto restart;
 757         }
 758         rcu_read_unlock();
 759         spin_unlock(&clp->cl_lock);
 760
 761         if (list_empty(&layout_list))
 762                 return 0;
 763         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 764 }
 765
 766 /*
 767  * Called by the state manger to remove all layouts established under an
 768  * expired lease.
 769  */
 770 void
 771 pnfs_destroy_all_layouts(struct nfs_client *clp)
 772 {
 773         nfs4_deviceid_mark_client_invalid(clp);
 774         nfs4_deviceid_purge_client(clp);
 775
 776         pnfs_destroy_layouts_byclid(clp, false);
 777 }
 778
 779 /*
 780  * Compare 2 layout stateid sequence ids, to see which is newer,
 781  * taking into account wraparound issues.
 782  */
 783 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 784 {
 785         return (s32)(s1 - s2) > 0;
 786 }
 787
 788 /* update lo->plh_stateid with new if is more recent */
 789 void
 790 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 791                         bool update_barrier)
 792 {
 793         u32 oldseq, newseq, new_barrier;
 794         int empty = list_empty(&lo->plh_segs);
 795
 796         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 797         newseq = be32_to_cpu(new->seqid);
 798         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 799                 nfs4_stateid_copy(&lo->plh_stateid, new);
 800                 if (update_barrier) {
 801                         new_barrier = be32_to_cpu(new->seqid);
 802                 } else {
 803                         /* Because of wraparound, we want to keep the barrier
 804                          * "close" to the current seqids.
 805                          */
 806                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 807                 }
 808                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 809                         lo->plh_barrier = new_barrier;
 810         }
 811 }
 812
 813 static bool
 814 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 815                 const nfs4_stateid *stateid)
 816 {
 817         u32 seqid = be32_to_cpu(stateid->seqid);
 818
 819         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 820 }
 821
 822 /* lget is set to 1 if called from inside send_layoutget call chain */
 823 static bool
 824 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 825 {
 826         return lo->plh_block_lgets ||
 827                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 828 }
 829
 830 int
 831 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 832                               struct pnfs_layout_range *range,
 833                               struct nfs4_state *open_state)
 834 {
 835         int status = 0;
 836
 837         dprintk("--> %s\n", __func__);
 838         spin_lock(&lo->plh_inode->i_lock);
 839         if (pnfs_layoutgets_blocked(lo)) {
 840                 status = -EAGAIN;
 841         } else if (!nfs4_valid_open_stateid(open_state)) {
 842                 status = -EBADF;
 843         } else if (list_empty(&lo->plh_segs) ||
 844                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
 845                 int seq;
 846
 847                 do {
 848                         seq = read_seqbegin(&open_state->seqlock);
 849                         nfs4_stateid_copy(dst, &open_state->stateid);
 850                 } while (read_seqretry(&open_state->seqlock, seq));
 851         } else
 852                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 853         spin_unlock(&lo->plh_inode->i_lock);
 854         dprintk("<-- %s\n", __func__);
 855         return status;
 856 }
 857
 858 /*
 859 * Get layout from server.
 860 *    for now, assume that whole file layouts are requested.
 861 *    arg->offset: 0
 862 *    arg->length: all ones
 863 */
 864 static struct pnfs_layout_segment *
 865 send_layoutget(struct pnfs_layout_hdr *lo,
 866            struct nfs_open_context *ctx,
 867            struct pnfs_layout_range *range,
 868            gfp_t gfp_flags)
 869 {
 870         struct inode *ino = lo->plh_inode;
 871         struct nfs_server *server = NFS_SERVER(ino);
 872         struct nfs4_layoutget *lgp;
 873         struct pnfs_layout_segment *lseg;
 874         loff_t i_size;
 875
 876         dprintk("--> %s\n", __func__);
 877
 878         /*
 879          * Synchronously retrieve layout information from server and
 880          * store in lseg. If we race with a concurrent seqid morphing
 881          * op, then re-send the LAYOUTGET.
 882          */
 883         do {
 884                 lgp = kzalloc(sizeof(*lgp), gfp_flags);
 885                 if (lgp == NULL)
 886                         return NULL;
 887
 888                 i_size = i_size_read(ino);
 889
 890                 lgp->args.minlength = PAGE_CACHE_SIZE;
 891                 if (lgp->args.minlength > range->length)
 892                         lgp->args.minlength = range->length;
 893                 if (range->iomode == IOMODE_READ) {
 894                         if (range->offset >= i_size)
 895                                 lgp->args.minlength = 0;
 896                         else if (i_size - range->offset < lgp->args.minlength)
 897                                 lgp->args.minlength = i_size - range->offset;
 898                 }
 899                 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 900                 lgp->args.range = *range;
 901                 lgp->args.type = server->pnfs_curr_ld->id;
 902                 lgp->args.inode = ino;
 903                 lgp->args.ctx = get_nfs_open_context(ctx);
 904                 lgp->gfp_flags = gfp_flags;
 905                 lgp->cred = lo->plh_lc_cred;
 906
 907                 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 908         } while (lseg == ERR_PTR(-EAGAIN));
 909
 910         if (IS_ERR(lseg)) {
 911                 switch (PTR_ERR(lseg)) {
 912                 case -ENOMEM:
 913                 case -ERESTARTSYS:
 914                         break;
 915                 default:
 916                         /* remember that LAYOUTGET failed and suspend trying */
 917                         pnfs_layout_io_set_failed(lo, range->iomode);
 918                 }
 919                 return NULL;
 920         } else
 921                 pnfs_layout_clear_fail_bit(lo,
 922                                 pnfs_iomode_to_fail_bit(range->iomode));
 923
 924         return lseg;
 925 }
 926
 927 static void pnfs_clear_layoutcommit(struct inode *inode,
 928                 struct list_head *head)
 929 {
 930         struct nfs_inode *nfsi = NFS_I(inode);
 931         struct pnfs_layout_segment *lseg, *tmp;
 932
 933         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 934                 return;
 935         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 936                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 937                         continue;
 938                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 939         }
 940 }
 941
 942 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 943 {
 944         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 945         smp_mb__after_atomic();
 946         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 947         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 948 }
 949
 950 static int
 951 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
 952                        enum pnfs_iomode iomode, bool sync)
 953 {
 954         struct inode *ino = lo->plh_inode;
 955         struct nfs4_layoutreturn *lrp;
 956         int status = 0;
 957
 958         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 959         if (unlikely(lrp == NULL)) {
 960                 status = -ENOMEM;
 961                 spin_lock(&ino->i_lock);
 962                 pnfs_clear_layoutreturn_waitbit(lo);
 963                 spin_unlock(&ino->i_lock);
 964                 pnfs_put_layout_hdr(lo);
 965                 goto out;
 966         }
 967
 968         lrp->args.stateid = stateid;
 969         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 970         lrp->args.inode = ino;
 971         lrp->args.range.iomode = iomode;
 972         lrp->args.range.offset = 0;
 973         lrp->args.range.length = NFS4_MAX_UINT64;
 974         lrp->args.layout = lo;
 975         lrp->clp = NFS_SERVER(ino)->nfs_client;
 976         lrp->cred = lo->plh_lc_cred;
 977
 978         status = nfs4_proc_layoutreturn(lrp, sync);
 979 out:
 980         dprintk("<-- %s status: %d\n", __func__, status);
 981         return status;
 982 }
 983
 984 /*
 985  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 986  * when the layout segment list is empty.
 987  *
 988  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 989  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 990  * deviceid is marked invalid.
 991  */
 992 int
 993 _pnfs_return_layout(struct inode *ino)
 994 {
 995         struct pnfs_layout_hdr *lo = NULL;
 996         struct nfs_inode *nfsi = NFS_I(ino);
 997         LIST_HEAD(tmp_list);
 998         nfs4_stateid stateid;
 999         int status = 0, empty;
1000         bool send;
1001
1002         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1003
1004         spin_lock(&ino->i_lock);
1005         lo = nfsi->layout;
1006         if (!lo) {
1007                 spin_unlock(&ino->i_lock);
1008                 dprintk("NFS: %s no layout to return\n", __func__);
1009                 goto out;
1010         }
1011         stateid = nfsi->layout->plh_stateid;
1012         /* Reference matched in nfs4_layoutreturn_release */
1013         pnfs_get_layout_hdr(lo);
1014         empty = list_empty(&lo->plh_segs);
1015         pnfs_clear_layoutcommit(ino, &tmp_list);
1016         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1017
1018         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1019                 struct pnfs_layout_range range = {
1020                         .iomode         = IOMODE_ANY,
1021                         .offset         = 0,
1022                         .length         = NFS4_MAX_UINT64,
1023                 };
1024                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1025         }
1026
1027         /* Don't send a LAYOUTRETURN if list was initially empty */
1028         if (empty) {
1029                 spin_unlock(&ino->i_lock);
1030                 dprintk("NFS: %s no layout segments to return\n", __func__);
1031                 goto out_put_layout_hdr;
1032         }
1033
1034         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1035         send = pnfs_prepare_layoutreturn(lo);
1036         spin_unlock(&ino->i_lock);
1037         pnfs_free_lseg_list(&tmp_list);
1038         if (send)
1039                 status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1040 out_put_layout_hdr:
1041         pnfs_put_layout_hdr(lo);
1042 out:
1043         dprintk("<-- %s status: %d\n", __func__, status);
1044         return status;
1045 }
1046 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1047
1048 int
1049 pnfs_commit_and_return_layout(struct inode *inode)
1050 {
1051         struct pnfs_layout_hdr *lo;
1052         int ret;
1053
1054         spin_lock(&inode->i_lock);
1055         lo = NFS_I(inode)->layout;
1056         if (lo == NULL) {
1057                 spin_unlock(&inode->i_lock);
1058                 return 0;
1059         }
1060         pnfs_get_layout_hdr(lo);
1061         /* Block new layoutgets and read/write to ds */
1062         lo->plh_block_lgets++;
1063         spin_unlock(&inode->i_lock);
1064         filemap_fdatawait(inode->i_mapping);
1065         ret = pnfs_layoutcommit_inode(inode, true);
1066         if (ret == 0)
1067                 ret = _pnfs_return_layout(inode);
1068         spin_lock(&inode->i_lock);
1069         lo->plh_block_lgets--;
1070         spin_unlock(&inode->i_lock);
1071         pnfs_put_layout_hdr(lo);
1072         return ret;
1073 }
1074
1075 bool pnfs_roc(struct inode *ino)
1076 {
1077         struct nfs_inode *nfsi = NFS_I(ino);
1078         struct nfs_open_context *ctx;
1079         struct nfs4_state *state;
1080         struct pnfs_layout_hdr *lo;
1081         struct pnfs_layout_segment *lseg, *tmp;
1082         nfs4_stateid stateid;
1083         LIST_HEAD(tmp_list);
1084         bool found = false, layoutreturn = false, roc = false;
1085
1086         spin_lock(&ino->i_lock);
1087         lo = nfsi->layout;
1088         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1089                 goto out_noroc;
1090
1091         /* no roc if we hold a delegation */
1092         if (nfs4_check_delegation(ino, FMODE_READ))
1093                 goto out_noroc;
1094
1095         list_for_each_entry(ctx, &nfsi->open_files, list) {
1096                 state = ctx->state;
1097                 /* Don't return layout if there is open file state */
1098                 if (state != NULL && state->state != 0)
1099                         goto out_noroc;
1100         }
1101
1102         stateid = lo->plh_stateid;
1103         /* always send layoutreturn if being marked so */
1104         if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1105                                    &lo->plh_flags))
1106                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1107
1108         pnfs_clear_retry_layoutget(lo);
1109         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1110                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1111                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1112                         mark_lseg_invalid(lseg, &tmp_list);
1113                         found = true;
1114                 }
1115         /* ROC in two conditions:
1116          * 1. there are ROC lsegs
1117          * 2. we don't send layoutreturn
1118          */
1119         if (found && !layoutreturn) {
1120                 /* lo ref dropped in pnfs_roc_release() */
1121                 pnfs_get_layout_hdr(lo);
1122                 roc = true;
1123         }
1124
1125 out_noroc:
1126         spin_unlock(&ino->i_lock);
1127         pnfs_free_lseg_list(&tmp_list);
1128         pnfs_layoutcommit_inode(ino, true);
1129         if (layoutreturn)
1130                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1131         return roc;
1132 }
1133
1134 void pnfs_roc_release(struct inode *ino)
1135 {
1136         struct pnfs_layout_hdr *lo;
1137
1138         spin_lock(&ino->i_lock);
1139         lo = NFS_I(ino)->layout;
1140         pnfs_clear_layoutreturn_waitbit(lo);
1141         if (atomic_dec_and_test(&lo->plh_refcount)) {
1142                 pnfs_detach_layout_hdr(lo);
1143                 spin_unlock(&ino->i_lock);
1144                 pnfs_free_layout_hdr(lo);
1145         } else
1146                 spin_unlock(&ino->i_lock);
1147 }
1148
1149 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1150 {
1151         struct pnfs_layout_hdr *lo;
1152
1153         spin_lock(&ino->i_lock);
1154         lo = NFS_I(ino)->layout;
1155         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1156                 lo->plh_barrier = barrier;
1157         spin_unlock(&ino->i_lock);
1158         trace_nfs4_layoutreturn_on_close(ino, 0);
1159 }
1160
1161 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1162 {
1163         struct nfs_inode *nfsi = NFS_I(ino);
1164         struct pnfs_layout_hdr *lo;
1165         u32 current_seqid;
1166
1167         spin_lock(&ino->i_lock);
1168         lo = nfsi->layout;
1169         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1170
1171         /* Since close does not return a layout stateid for use as
1172          * a barrier, we choose the worst-case barrier.
1173          */
1174         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1175         spin_unlock(&ino->i_lock);
1176 }
1177
1178 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1179 {
1180         struct nfs_inode *nfsi = NFS_I(ino);
1181         struct pnfs_layout_hdr *lo;
1182         bool sleep = false;
1183
1184         /* we might not have grabbed lo reference. so need to check under
1185          * i_lock */
1186         spin_lock(&ino->i_lock);
1187         lo = nfsi->layout;
1188         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1189                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1190                 sleep = true;
1191         }
1192         spin_unlock(&ino->i_lock);
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                 goto out;
1526
1527         if (iomode == IOMODE_READ && i_size_read(ino) == 0)
1528                 goto out;
1529
1530         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1531                 goto out;
1532
1533 lookup_again:
1534         nfs4_client_recover_expired_lease(clp);
1535         first = false;
1536         spin_lock(&ino->i_lock);
1537         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1538         if (lo == NULL) {
1539                 spin_unlock(&ino->i_lock);
1540                 goto out;
1541         }
1542
1543         /* Do we even need to bother with this? */
1544         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1545                 dprintk("%s matches recall, use MDS\n", __func__);
1546                 goto out_unlock;
1547         }
1548
1549         /* if LAYOUTGET already failed once we don't try again */
1550         if (pnfs_layout_io_test_failed(lo, iomode) &&
1551             !pnfs_should_retry_layoutget(lo))
1552                 goto out_unlock;
1553
1554         first = list_empty(&lo->plh_segs);
1555         if (first) {
1556                 /* The first layoutget for the file. Need to serialize per
1557                  * RFC 5661 Errata 3208.
1558                  */
1559                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1560                                      &lo->plh_flags)) {
1561                         spin_unlock(&ino->i_lock);
1562                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1563                                     TASK_UNINTERRUPTIBLE);
1564                         pnfs_put_layout_hdr(lo);
1565                         goto lookup_again;
1566                 }
1567         } else {
1568                 /* Check to see if the layout for the given range
1569                  * already exists
1570                  */
1571                 lseg = pnfs_find_lseg(lo, &arg);
1572                 if (lseg)
1573                         goto out_unlock;
1574         }
1575
1576         /*
1577          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1578          * for LAYOUTRETURN even if first is true.
1579          */
1580         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1581                 spin_unlock(&ino->i_lock);
1582                 dprintk("%s wait for layoutreturn\n", __func__);
1583                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1584                         if (first)
1585                                 pnfs_clear_first_layoutget(lo);
1586                         pnfs_put_layout_hdr(lo);
1587                         dprintk("%s retrying\n", __func__);
1588                         goto lookup_again;
1589                 }
1590                 goto out_put_layout_hdr;
1591         }
1592
1593         if (pnfs_layoutgets_blocked(lo))
1594                 goto out_unlock;
1595         atomic_inc(&lo->plh_outstanding);
1596         spin_unlock(&ino->i_lock);
1597
1598         if (list_empty(&lo->plh_layouts)) {
1599                 /* The lo must be on the clp list if there is any
1600                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1601                  */
1602                 spin_lock(&clp->cl_lock);
1603                 if (list_empty(&lo->plh_layouts))
1604                         list_add_tail(&lo->plh_layouts, &server->layouts);
1605                 spin_unlock(&clp->cl_lock);
1606         }
1607
1608         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1609         if (pg_offset) {
1610                 arg.offset -= pg_offset;
1611                 arg.length += pg_offset;
1612         }
1613         if (arg.length != NFS4_MAX_UINT64)
1614                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1615
1616         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1617         pnfs_clear_retry_layoutget(lo);
1618         atomic_dec(&lo->plh_outstanding);
1619 out_put_layout_hdr:
1620         if (first)
1621                 pnfs_clear_first_layoutget(lo);
1622         pnfs_put_layout_hdr(lo);
1623 out:
1624         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1625                         "(%s, offset: %llu, length: %llu)\n",
1626                         __func__, ino->i_sb->s_id,
1627                         (unsigned long long)NFS_FILEID(ino),
1628                         lseg == NULL ? "not found" : "found",
1629                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1630                         (unsigned long long)pos,
1631                         (unsigned long long)count);
1632         return lseg;
1633 out_unlock:
1634         spin_unlock(&ino->i_lock);
1635         goto out_put_layout_hdr;
1636 }
1637 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1638
1639 static bool
1640 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1641 {
1642         switch (range->iomode) {
1643         case IOMODE_READ:
1644         case IOMODE_RW:
1645                 break;
1646         default:
1647                 return false;
1648         }
1649         if (range->offset == NFS4_MAX_UINT64)
1650                 return false;
1651         if (range->length == 0)
1652                 return false;
1653         if (range->length != NFS4_MAX_UINT64 &&
1654             range->length > NFS4_MAX_UINT64 - range->offset)
1655                 return false;
1656         return true;
1657 }
1658
1659 struct pnfs_layout_segment *
1660 pnfs_layout_process(struct nfs4_layoutget *lgp)
1661 {
1662         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1663         struct nfs4_layoutget_res *res = &lgp->res;
1664         struct pnfs_layout_segment *lseg;
1665         struct inode *ino = lo->plh_inode;
1666         LIST_HEAD(free_me);
1667         int status = -EINVAL;
1668
1669         if (!pnfs_sanity_check_layout_range(&res->range))
1670                 goto out;
1671
1672         /* Inject layout blob into I/O device driver */
1673         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1674         if (!lseg || IS_ERR(lseg)) {
1675                 if (!lseg)
1676                         status = -ENOMEM;
1677                 else
1678                         status = PTR_ERR(lseg);
1679                 dprintk("%s: Could not allocate layout: error %d\n",
1680                        __func__, status);
1681                 goto out;
1682         }
1683
1684         init_lseg(lo, lseg);
1685         lseg->pls_range = res->range;
1686
1687         spin_lock(&ino->i_lock);
1688         if (pnfs_layoutgets_blocked(lo)) {
1689                 dprintk("%s forget reply due to state\n", __func__);
1690                 goto out_forget_reply;
1691         }
1692
1693         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1694                 /* existing state ID, make sure the sequence number matches. */
1695                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1696                         dprintk("%s forget reply due to sequence\n", __func__);
1697                         status = -EAGAIN;
1698                         goto out_forget_reply;
1699                 }
1700                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1701         } else {
1702                 /*
1703                  * We got an entirely new state ID.  Mark all segments for the
1704                  * inode invalid, and don't bother validating the stateid
1705                  * sequence number.
1706                  */
1707                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1708
1709                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1710                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1711         }
1712
1713         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1714
1715         pnfs_get_lseg(lseg);
1716         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1717
1718         if (res->return_on_close)
1719                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1720
1721         spin_unlock(&ino->i_lock);
1722         pnfs_free_lseg_list(&free_me);
1723         return lseg;
1724 out:
1725         return ERR_PTR(status);
1726
1727 out_forget_reply:
1728         spin_unlock(&ino->i_lock);
1729         lseg->pls_layout = lo;
1730         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1731         goto out;
1732 }
1733
1734 static void
1735 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1736                                 struct list_head *tmp_list,
1737                                 struct pnfs_layout_range *return_range)
1738 {
1739         struct pnfs_layout_segment *lseg, *next;
1740
1741         dprintk("%s:Begin lo %p\n", __func__, lo);
1742
1743         if (list_empty(&lo->plh_segs))
1744                 return;
1745
1746         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1747                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1748                         dprintk("%s: marking lseg %p iomode %d "
1749                                 "offset %llu length %llu\n", __func__,
1750                                 lseg, lseg->pls_range.iomode,
1751                                 lseg->pls_range.offset,
1752                                 lseg->pls_range.length);
1753                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1754                         mark_lseg_invalid(lseg, tmp_list);
1755                         set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1756                                         &lo->plh_flags);
1757                 }
1758 }
1759
1760 void pnfs_error_mark_layout_for_return(struct inode *inode,
1761                                        struct pnfs_layout_segment *lseg)
1762 {
1763         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1764         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1765         struct pnfs_layout_range range = {
1766                 .iomode = lseg->pls_range.iomode,
1767                 .offset = 0,
1768                 .length = NFS4_MAX_UINT64,
1769         };
1770         LIST_HEAD(free_me);
1771
1772         spin_lock(&inode->i_lock);
1773         /* set failure bit so that pnfs path will be retried later */
1774         pnfs_layout_set_fail_bit(lo, iomode);
1775         if (lo->plh_return_iomode == 0)
1776                 lo->plh_return_iomode = range.iomode;
1777         else if (lo->plh_return_iomode != range.iomode)
1778                 lo->plh_return_iomode = IOMODE_ANY;
1779         /*
1780          * mark all matching lsegs so that we are sure to have no live
1781          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1782          * for how it works.
1783          */
1784         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1785         spin_unlock(&inode->i_lock);
1786         pnfs_free_lseg_list(&free_me);
1787 }
1788 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1789
1790 void
1791 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1792 {
1793         u64 rd_size = req->wb_bytes;
1794
1795         if (pgio->pg_lseg == NULL) {
1796                 if (pgio->pg_dreq == NULL)
1797                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1798                 else
1799                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1800
1801                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1802                                                    req->wb_context,
1803                                                    req_offset(req),
1804                                                    rd_size,
1805                                                    IOMODE_READ,
1806                                                    GFP_KERNEL);
1807         }
1808         /* If no lseg, fall back to read through mds */
1809         if (pgio->pg_lseg == NULL)
1810                 nfs_pageio_reset_read_mds(pgio);
1811
1812 }
1813 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1814
1815 void
1816 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1817                            struct nfs_page *req, u64 wb_size)
1818 {
1819         if (pgio->pg_lseg == NULL)
1820                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1821                                                    req->wb_context,
1822                                                    req_offset(req),
1823                                                    wb_size,
1824                                                    IOMODE_RW,
1825                                                    GFP_NOFS);
1826         /* If no lseg, fall back to write through mds */
1827         if (pgio->pg_lseg == NULL)
1828                 nfs_pageio_reset_write_mds(pgio);
1829 }
1830 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1831
1832 void
1833 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1834 {
1835         if (desc->pg_lseg) {
1836                 pnfs_put_lseg(desc->pg_lseg);
1837                 desc->pg_lseg = NULL;
1838         }
1839 }
1840 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1841
1842 /*
1843  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1844  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1845  */
1846 size_t
1847 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1848                      struct nfs_page *prev, struct nfs_page *req)
1849 {
1850         unsigned int size;
1851         u64 seg_end, req_start, seg_left;
1852
1853         size = nfs_generic_pg_test(pgio, prev, req);
1854         if (!size)
1855                 return 0;
1856
1857         /*
1858          * 'size' contains the number of bytes left in the current page (up
1859          * to the original size asked for in @req->wb_bytes).
1860          *
1861          * Calculate how many bytes are left in the layout segment
1862          * and if there are less bytes than 'size', return that instead.
1863          *
1864          * Please also note that 'end_offset' is actually the offset of the
1865          * first byte that lies outside the pnfs_layout_range. FIXME?
1866          *
1867          */
1868         if (pgio->pg_lseg) {
1869                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1870                                      pgio->pg_lseg->pls_range.length);
1871                 req_start = req_offset(req);
1872                 WARN_ON_ONCE(req_start >= seg_end);
1873                 /* start of request is past the last byte of this segment */
1874                 if (req_start >= seg_end) {
1875                         /* reference the new lseg */
1876                         if (pgio->pg_ops->pg_cleanup)
1877                                 pgio->pg_ops->pg_cleanup(pgio);
1878                         if (pgio->pg_ops->pg_init)
1879                                 pgio->pg_ops->pg_init(pgio, req);
1880                         return 0;
1881                 }
1882
1883                 /* adjust 'size' iff there are fewer bytes left in the
1884                  * segment than what nfs_generic_pg_test returned */
1885                 seg_left = seg_end - req_start;
1886                 if (seg_left < size)
1887                         size = (unsigned int)seg_left;
1888         }
1889
1890         return size;
1891 }
1892 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1893
1894 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1895 {
1896         struct nfs_pageio_descriptor pgio;
1897
1898         /* Resend all requests through the MDS */
1899         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1900                               hdr->completion_ops);
1901         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1902         return nfs_pageio_resend(&pgio, hdr);
1903 }
1904 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1905
1906 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1907 {
1908
1909         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1910         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1911             PNFS_LAYOUTRET_ON_ERROR) {
1912                 pnfs_return_layout(hdr->inode);
1913         }
1914         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1915                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1916 }
1917
1918 /*
1919  * Called by non rpc-based layout drivers
1920  */
1921 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1922 {
1923         if (likely(!hdr->pnfs_error)) {
1924                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1925                                 hdr->mds_offset + hdr->res.count);
1926                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1927         }
1928         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1929         if (unlikely(hdr->pnfs_error))
1930                 pnfs_ld_handle_write_error(hdr);
1931         hdr->mds_ops->rpc_release(hdr);
1932 }
1933 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1934
1935 static void
1936 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1937                 struct nfs_pgio_header *hdr)
1938 {
1939         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1940
1941         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1942                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1943                 nfs_pageio_reset_write_mds(desc);
1944                 mirror->pg_recoalesce = 1;
1945         }
1946         nfs_pgio_data_destroy(hdr);
1947         hdr->release(hdr);
1948 }
1949
1950 static enum pnfs_try_status
1951 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1952                         const struct rpc_call_ops *call_ops,
1953                         struct pnfs_layout_segment *lseg,
1954                         int how)
1955 {
1956         struct inode *inode = hdr->inode;
1957         enum pnfs_try_status trypnfs;
1958         struct nfs_server *nfss = NFS_SERVER(inode);
1959
1960         hdr->mds_ops = call_ops;
1961
1962         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1963                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1964         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1965         if (trypnfs != PNFS_NOT_ATTEMPTED)
1966                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1967         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1968         return trypnfs;
1969 }
1970
1971 static void
1972 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1973               struct nfs_pgio_header *hdr, int how)
1974 {
1975         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1976         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1977         enum pnfs_try_status trypnfs;
1978
1979         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1980         if (trypnfs == PNFS_NOT_ATTEMPTED)
1981                 pnfs_write_through_mds(desc, hdr);
1982 }
1983
1984 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1985 {
1986         pnfs_put_lseg(hdr->lseg);
1987         nfs_pgio_header_free(hdr);
1988 }
1989
1990 int
1991 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1992 {
1993         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1994
1995         struct nfs_pgio_header *hdr;
1996         int ret;
1997
1998         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1999         if (!hdr) {
2000                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2001                 return -ENOMEM;
2002         }
2003         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2004
2005         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2006         ret = nfs_generic_pgio(desc, hdr);
2007         if (!ret)
2008                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2009
2010         return ret;
2011 }
2012 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2013
2014 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2015 {
2016         struct nfs_pageio_descriptor pgio;
2017
2018         /* Resend all requests through the MDS */
2019         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2020         return nfs_pageio_resend(&pgio, hdr);
2021 }
2022 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2023
2024 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2025 {
2026         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2027         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2028             PNFS_LAYOUTRET_ON_ERROR) {
2029                 pnfs_return_layout(hdr->inode);
2030         }
2031         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2032                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2033 }
2034
2035 /*
2036  * Called by non rpc-based layout drivers
2037  */
2038 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2039 {
2040         if (likely(!hdr->pnfs_error)) {
2041                 __nfs4_read_done_cb(hdr);
2042                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2043         }
2044         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2045         if (unlikely(hdr->pnfs_error))
2046                 pnfs_ld_handle_read_error(hdr);
2047         hdr->mds_ops->rpc_release(hdr);
2048 }
2049 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2050
2051 static void
2052 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2053                 struct nfs_pgio_header *hdr)
2054 {
2055         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2056
2057         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2058                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2059                 nfs_pageio_reset_read_mds(desc);
2060                 mirror->pg_recoalesce = 1;
2061         }
2062         nfs_pgio_data_destroy(hdr);
2063         hdr->release(hdr);
2064 }
2065
2066 /*
2067  * Call the appropriate parallel I/O subsystem read function.
2068  */
2069 static enum pnfs_try_status
2070 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2071                        const struct rpc_call_ops *call_ops,
2072                        struct pnfs_layout_segment *lseg)
2073 {
2074         struct inode *inode = hdr->inode;
2075         struct nfs_server *nfss = NFS_SERVER(inode);
2076         enum pnfs_try_status trypnfs;
2077
2078         hdr->mds_ops = call_ops;
2079
2080         dprintk("%s: Reading ino:%lu %u@%llu\n",
2081                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2082
2083         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2084         if (trypnfs != PNFS_NOT_ATTEMPTED)
2085                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2086         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2087         return trypnfs;
2088 }
2089
2090 /* Resend all requests through pnfs. */
2091 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2092 {
2093         struct nfs_pageio_descriptor pgio;
2094
2095         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2096         return nfs_pageio_resend(&pgio, hdr);
2097 }
2098 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2099
2100 static void
2101 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2102 {
2103         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2104         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2105         enum pnfs_try_status trypnfs;
2106         int err = 0;
2107
2108         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2109         if (trypnfs == PNFS_TRY_AGAIN)
2110                 err = pnfs_read_resend_pnfs(hdr);
2111         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2112                 pnfs_read_through_mds(desc, hdr);
2113 }
2114
2115 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2116 {
2117         pnfs_put_lseg(hdr->lseg);
2118         nfs_pgio_header_free(hdr);
2119 }
2120
2121 int
2122 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2123 {
2124         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2125
2126         struct nfs_pgio_header *hdr;
2127         int ret;
2128
2129         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2130         if (!hdr) {
2131                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2132                 return -ENOMEM;
2133         }
2134         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2135         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2136         ret = nfs_generic_pgio(desc, hdr);
2137         if (!ret)
2138                 pnfs_do_read(desc, hdr);
2139         return ret;
2140 }
2141 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2142
2143 static void pnfs_clear_layoutcommitting(struct inode *inode)
2144 {
2145         unsigned long *bitlock = &NFS_I(inode)->flags;
2146
2147         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2148         smp_mb__after_atomic();
2149         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2150 }
2151
2152 /*
2153  * There can be multiple RW segments.
2154  */
2155 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2156 {
2157         struct pnfs_layout_segment *lseg;
2158
2159         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2160                 if (lseg->pls_range.iomode == IOMODE_RW &&
2161                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2162                         list_add(&lseg->pls_lc_list, listp);
2163         }
2164 }
2165
2166 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2167 {
2168         struct pnfs_layout_segment *lseg, *tmp;
2169
2170         /* Matched by references in pnfs_set_layoutcommit */
2171         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2172                 list_del_init(&lseg->pls_lc_list);
2173                 pnfs_put_lseg(lseg);
2174         }
2175
2176         pnfs_clear_layoutcommitting(inode);
2177 }
2178
2179 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2180 {
2181         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2182 }
2183 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2184
2185 void
2186 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2187                 loff_t end_pos)
2188 {
2189         struct nfs_inode *nfsi = NFS_I(inode);
2190         bool mark_as_dirty = false;
2191
2192         spin_lock(&inode->i_lock);
2193         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2194                 nfsi->layout->plh_lwb = end_pos;
2195                 mark_as_dirty = true;
2196                 dprintk("%s: Set layoutcommit for inode %lu ",
2197                         __func__, inode->i_ino);
2198         } else if (end_pos > nfsi->layout->plh_lwb)
2199                 nfsi->layout->plh_lwb = end_pos;
2200         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2201                 /* references matched in nfs4_layoutcommit_release */
2202                 pnfs_get_lseg(lseg);
2203         }
2204         spin_unlock(&inode->i_lock);
2205         dprintk("%s: lseg %p end_pos %llu\n",
2206                 __func__, lseg, nfsi->layout->plh_lwb);
2207
2208         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2209          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2210         if (mark_as_dirty)
2211                 mark_inode_dirty_sync(inode);
2212 }
2213 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2214
2215 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2216 {
2217         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2218
2219         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2220                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2221         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2222 }
2223
2224 /*
2225  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2226  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2227  * data to disk to allow the server to recover the data if it crashes.
2228  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2229  * is off, and a COMMIT is sent to a data server, or
2230  * if WRITEs to a data server return NFS_DATA_SYNC.
2231  */
2232 int
2233 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2234 {
2235         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2236         struct nfs4_layoutcommit_data *data;
2237         struct nfs_inode *nfsi = NFS_I(inode);
2238         loff_t end_pos;
2239         int status;
2240
2241         if (!pnfs_layoutcommit_outstanding(inode))
2242                 return 0;
2243
2244         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2245
2246         status = -EAGAIN;
2247         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2248                 if (!sync)
2249                         goto out;
2250                 status = wait_on_bit_lock_action(&nfsi->flags,
2251                                 NFS_INO_LAYOUTCOMMITTING,
2252                                 nfs_wait_bit_killable,
2253                                 TASK_KILLABLE);
2254                 if (status)
2255                         goto out;
2256         }
2257
2258         status = -ENOMEM;
2259         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2260         data = kzalloc(sizeof(*data), GFP_NOFS);
2261         if (!data)
2262                 goto clear_layoutcommitting;
2263
2264         status = 0;
2265         spin_lock(&inode->i_lock);
2266         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2267                 goto out_unlock;
2268
2269         INIT_LIST_HEAD(&data->lseg_list);
2270         pnfs_list_write_lseg(inode, &data->lseg_list);
2271
2272         end_pos = nfsi->layout->plh_lwb;
2273
2274         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2275         spin_unlock(&inode->i_lock);
2276
2277         data->args.inode = inode;
2278         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2279         nfs_fattr_init(&data->fattr);
2280         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2281         data->res.fattr = &data->fattr;
2282         data->args.lastbytewritten = end_pos - 1;
2283         data->res.server = NFS_SERVER(inode);
2284
2285         if (ld->prepare_layoutcommit) {
2286                 status = ld->prepare_layoutcommit(&data->args);
2287                 if (status) {
2288                         put_rpccred(data->cred);
2289                         spin_lock(&inode->i_lock);
2290                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2291                         if (end_pos > nfsi->layout->plh_lwb)
2292                                 nfsi->layout->plh_lwb = end_pos;
2293                         goto out_unlock;
2294                 }
2295         }
2296
2297
2298         status = nfs4_proc_layoutcommit(data, sync);
2299 out:
2300         if (status)
2301                 mark_inode_dirty_sync(inode);
2302         dprintk("<-- %s status %d\n", __func__, status);
2303         return status;
2304 out_unlock:
2305         spin_unlock(&inode->i_lock);
2306         kfree(data);
2307 clear_layoutcommitting:
2308         pnfs_clear_layoutcommitting(inode);
2309         goto out;
2310 }
2311 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2312
2313 int
2314 pnfs_generic_sync(struct inode *inode, bool datasync)
2315 {
2316         return pnfs_layoutcommit_inode(inode, true);
2317 }
2318 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2319
2320 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2321 {
2322         struct nfs4_threshold *thp;
2323
2324         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2325         if (!thp) {
2326                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2327                 return NULL;
2328         }
2329         return thp;
2330 }
2331
2332 #if IS_ENABLED(CONFIG_NFS_V4_2)
2333 int
2334 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2335 {
2336         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2337         struct nfs_server *server = NFS_SERVER(inode);
2338         struct nfs_inode *nfsi = NFS_I(inode);
2339         struct nfs42_layoutstat_data *data;
2340         struct pnfs_layout_hdr *hdr;
2341         int status = 0;
2342
2343         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2344                 goto out;
2345
2346         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2347                 goto out;
2348
2349         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2350                 goto out;
2351
2352         spin_lock(&inode->i_lock);
2353         if (!NFS_I(inode)->layout) {
2354                 spin_unlock(&inode->i_lock);
2355                 goto out;
2356         }
2357         hdr = NFS_I(inode)->layout;
2358         pnfs_get_layout_hdr(hdr);
2359         spin_unlock(&inode->i_lock);
2360
2361         data = kzalloc(sizeof(*data), gfp_flags);
2362         if (!data) {
2363                 status = -ENOMEM;
2364                 goto out_put;
2365         }
2366
2367         data->args.fh = NFS_FH(inode);
2368         data->args.inode = inode;
2369         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2370         status = ld->prepare_layoutstats(&data->args);
2371         if (status)
2372                 goto out_free;
2373
2374         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2375
2376 out:
2377         dprintk("%s returns %d\n", __func__, status);
2378         return status;
2379
2380 out_free:
2381         kfree(data);
2382 out_put:
2383         pnfs_put_layout_hdr(hdr);
2384         smp_mb__before_atomic();
2385         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2386         smp_mb__after_atomic();
2387         goto out;
2388 }
2389 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2390 #endif
2391
2392 unsigned int layoutstats_timer;
2393 module_param(layoutstats_timer, uint, 0644);
2394 EXPORT_SYMBOL_GPL(layoutstats_timer);