fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36 #include "nfs4trace.h"
  37 #include "delegation.h"
  38 #include "nfs42.h"
  39
  40 #define NFSDBG_FACILITY         NFSDBG_PNFS
  41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  42
  43 /* Locking:
  44  *
  45  * pnfs_spinlock:
  46  *      protects pnfs_modules_tbl.
  47  */
  48 static DEFINE_SPINLOCK(pnfs_spinlock);
  49
  50 /*
  51  * pnfs_modules_tbl holds all pnfs modules
  52  */
  53 static LIST_HEAD(pnfs_modules_tbl);
  54
  55 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  56
  57 /* Return the registered pnfs layout driver module matching given id */
  58 static struct pnfs_layoutdriver_type *
  59 find_pnfs_driver_locked(u32 id)
  60 {
  61         struct pnfs_layoutdriver_type *local;
  62
  63         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  64                 if (local->id == id)
  65                         goto out;
  66         local = NULL;
  67 out:
  68         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  69         return local;
  70 }
  71
  72 static struct pnfs_layoutdriver_type *
  73 find_pnfs_driver(u32 id)
  74 {
  75         struct pnfs_layoutdriver_type *local;
  76
  77         spin_lock(&pnfs_spinlock);
  78         local = find_pnfs_driver_locked(id);
  79         if (local != NULL && !try_module_get(local->owner)) {
  80                 dprintk("%s: Could not grab reference on module\n", __func__);
  81                 local = NULL;
  82         }
  83         spin_unlock(&pnfs_spinlock);
  84         return local;
  85 }
  86
  87 void
  88 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  89 {
  90         if (nfss->pnfs_curr_ld) {
  91                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  92                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  93                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  94                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  95                         nfs4_deviceid_purge_client(nfss->nfs_client);
  96                 module_put(nfss->pnfs_curr_ld->owner);
  97         }
  98         nfss->pnfs_curr_ld = NULL;
  99 }
 100
 101 /*
 102  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 103  * Currently only one pNFS layout driver per filesystem is supported.
 104  *
 105  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 106  */
 107 void
 108 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 109                       u32 id)
 110 {
 111         struct pnfs_layoutdriver_type *ld_type = NULL;
 112
 113         if (id == 0)
 114                 goto out_no_driver;
 115         if (!(server->nfs_client->cl_exchange_flags &
 116                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 117                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 118                         __func__, id, server->nfs_client->cl_exchange_flags);
 119                 goto out_no_driver;
 120         }
 121         ld_type = find_pnfs_driver(id);
 122         if (!ld_type) {
 123                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 124                 ld_type = find_pnfs_driver(id);
 125                 if (!ld_type) {
 126                         dprintk("%s: No pNFS module found for %u.\n",
 127                                 __func__, id);
 128                         goto out_no_driver;
 129                 }
 130         }
 131         server->pnfs_curr_ld = ld_type;
 132         if (ld_type->set_layoutdriver
 133             && ld_type->set_layoutdriver(server, mntfh)) {
 134                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 135                         "driver %u.\n", __func__, id);
 136                 module_put(ld_type->owner);
 137                 goto out_no_driver;
 138         }
 139         /* Bump the MDS count */
 140         atomic_inc(&server->nfs_client->cl_mds_count);
 141
 142         dprintk("%s: pNFS module for %u set\n", __func__, id);
 143         return;
 144
 145 out_no_driver:
 146         dprintk("%s: Using NFSv4 I/O\n", __func__);
 147         server->pnfs_curr_ld = NULL;
 148 }
 149
 150 int
 151 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 152 {
 153         int status = -EINVAL;
 154         struct pnfs_layoutdriver_type *tmp;
 155
 156         if (ld_type->id == 0) {
 157                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 158                 return status;
 159         }
 160         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 161                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 162                        "alloc_lseg and free_lseg.\n", __func__);
 163                 return status;
 164         }
 165
 166         spin_lock(&pnfs_spinlock);
 167         tmp = find_pnfs_driver_locked(ld_type->id);
 168         if (!tmp) {
 169                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 170                 status = 0;
 171                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 172                         ld_type->name);
 173         } else {
 174                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 175                         __func__, ld_type->id);
 176         }
 177         spin_unlock(&pnfs_spinlock);
 178
 179         return status;
 180 }
 181 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 182
 183 void
 184 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 185 {
 186         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 187         spin_lock(&pnfs_spinlock);
 188         list_del(&ld_type->pnfs_tblid);
 189         spin_unlock(&pnfs_spinlock);
 190 }
 191 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 192
 193 /*
 194  * pNFS client layout cache
 195  */
 196
 197 /* Need to hold i_lock if caller does not already hold reference */
 198 void
 199 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 200 {
 201         atomic_inc(&lo->plh_refcount);
 202 }
 203
 204 static struct pnfs_layout_hdr *
 205 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 206 {
 207         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 208         return ld->alloc_layout_hdr(ino, gfp_flags);
 209 }
 210
 211 static void
 212 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 213 {
 214         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 215         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 216
 217         if (!list_empty(&lo->plh_layouts)) {
 218                 struct nfs_client *clp = server->nfs_client;
 219
 220                 spin_lock(&clp->cl_lock);
 221                 list_del_init(&lo->plh_layouts);
 222                 spin_unlock(&clp->cl_lock);
 223         }
 224         put_rpccred(lo->plh_lc_cred);
 225         return ld->free_layout_hdr(lo);
 226 }
 227
 228 static void
 229 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 230 {
 231         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 232         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 233         nfsi->layout = NULL;
 234         /* Reset MDS Threshold I/O counters */
 235         nfsi->write_io = 0;
 236         nfsi->read_io = 0;
 237 }
 238
 239 void
 240 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 241 {
 242         struct inode *inode = lo->plh_inode;
 243
 244         pnfs_layoutreturn_before_put_layout_hdr(lo);
 245
 246         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 247                 if (!list_empty(&lo->plh_segs))
 248                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 249                 pnfs_detach_layout_hdr(lo);
 250                 spin_unlock(&inode->i_lock);
 251                 pnfs_free_layout_hdr(lo);
 252         }
 253 }
 254
 255 /*
 256  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 257  *
 258  * In order to continue using the pnfs_layout_hdr, a full recovery
 259  * is required.
 260  * Note that caller must hold inode->i_lock.
 261  */
 262 static int
 263 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 264                 struct list_head *lseg_list)
 265 {
 266         struct pnfs_layout_range range = {
 267                 .iomode = IOMODE_ANY,
 268                 .offset = 0,
 269                 .length = NFS4_MAX_UINT64,
 270         };
 271
 272         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 273         return pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range, 0);
 274 }
 275
 276 static int
 277 pnfs_iomode_to_fail_bit(u32 iomode)
 278 {
 279         return iomode == IOMODE_RW ?
 280                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 281 }
 282
 283 static void
 284 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 285 {
 286         lo->plh_retry_timestamp = jiffies;
 287         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 288                 atomic_inc(&lo->plh_refcount);
 289 }
 290
 291 static void
 292 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 293 {
 294         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 295                 atomic_dec(&lo->plh_refcount);
 296 }
 297
 298 static void
 299 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 300 {
 301         struct inode *inode = lo->plh_inode;
 302         struct pnfs_layout_range range = {
 303                 .iomode = iomode,
 304                 .offset = 0,
 305                 .length = NFS4_MAX_UINT64,
 306         };
 307         LIST_HEAD(head);
 308
 309         spin_lock(&inode->i_lock);
 310         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 311         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 312         spin_unlock(&inode->i_lock);
 313         pnfs_free_lseg_list(&head);
 314         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 315                         iomode == IOMODE_RW ?  "RW" : "READ");
 316 }
 317
 318 static bool
 319 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 320 {
 321         unsigned long start, end;
 322         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 323
 324         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 325                 return false;
 326         end = jiffies;
 327         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 328         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 329                 /* It is time to retry the failed layoutgets */
 330                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 331                 return false;
 332         }
 333         return true;
 334 }
 335
 336 static void
 337 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 338 {
 339         INIT_LIST_HEAD(&lseg->pls_list);
 340         INIT_LIST_HEAD(&lseg->pls_lc_list);
 341         atomic_set(&lseg->pls_refcount, 1);
 342         smp_mb();
 343         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 344         lseg->pls_layout = lo;
 345 }
 346
 347 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 348 {
 349         struct inode *ino = lseg->pls_layout->plh_inode;
 350
 351         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 352 }
 353
 354 static void
 355 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 356                 struct pnfs_layout_segment *lseg)
 357 {
 358         struct inode *inode = lo->plh_inode;
 359
 360         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 361         list_del_init(&lseg->pls_list);
 362         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 363         atomic_dec(&lo->plh_refcount);
 364         if (list_empty(&lo->plh_segs)) {
 365                 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 366                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 367         }
 368         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 369 }
 370
 371 void
 372 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 373 {
 374         struct pnfs_layout_hdr *lo;
 375         struct inode *inode;
 376
 377         if (!lseg)
 378                 return;
 379
 380         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 381                 atomic_read(&lseg->pls_refcount),
 382                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 383
 384         lo = lseg->pls_layout;
 385         inode = lo->plh_inode;
 386
 387         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 388                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 389                         spin_unlock(&inode->i_lock);
 390                         return;
 391                 }
 392                 pnfs_get_layout_hdr(lo);
 393                 pnfs_layout_remove_lseg(lo, lseg);
 394                 spin_unlock(&inode->i_lock);
 395                 pnfs_free_lseg(lseg);
 396                 pnfs_put_layout_hdr(lo);
 397         }
 398 }
 399 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 400
 401 static void pnfs_free_lseg_async_work(struct work_struct *work)
 402 {
 403         struct pnfs_layout_segment *lseg;
 404         struct pnfs_layout_hdr *lo;
 405
 406         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 407         lo = lseg->pls_layout;
 408
 409         pnfs_free_lseg(lseg);
 410         pnfs_put_layout_hdr(lo);
 411 }
 412
 413 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 414 {
 415         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 416         schedule_work(&lseg->pls_work);
 417 }
 418
 419 void
 420 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 421 {
 422         if (!lseg)
 423                 return;
 424
 425         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 426
 427         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 428                 atomic_read(&lseg->pls_refcount),
 429                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 430         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 431                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 432                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 433                         return;
 434                 pnfs_get_layout_hdr(lo);
 435                 pnfs_layout_remove_lseg(lo, lseg);
 436                 pnfs_free_lseg_async(lseg);
 437         }
 438 }
 439 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 440
 441 static u64
 442 end_offset(u64 start, u64 len)
 443 {
 444         u64 end;
 445
 446         end = start + len;
 447         return end >= start ? end : NFS4_MAX_UINT64;
 448 }
 449
 450 /*
 451  * is l2 fully contained in l1?
 452  *   start1                             end1
 453  *   [----------------------------------)
 454  *           start2           end2
 455  *           [----------------)
 456  */
 457 static bool
 458 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 459                  const struct pnfs_layout_range *l2)
 460 {
 461         u64 start1 = l1->offset;
 462         u64 end1 = end_offset(start1, l1->length);
 463         u64 start2 = l2->offset;
 464         u64 end2 = end_offset(start2, l2->length);
 465
 466         return (start1 <= start2) && (end1 >= end2);
 467 }
 468
 469 /*
 470  * is l1 and l2 intersecting?
 471  *   start1                             end1
 472  *   [----------------------------------)
 473  *                              start2           end2
 474  *                              [----------------)
 475  */
 476 static bool
 477 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 478                     const struct pnfs_layout_range *l2)
 479 {
 480         u64 start1 = l1->offset;
 481         u64 end1 = end_offset(start1, l1->length);
 482         u64 start2 = l2->offset;
 483         u64 end2 = end_offset(start2, l2->length);
 484
 485         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 486                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 487 }
 488
 489 static bool
 490 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 491                  const struct pnfs_layout_range *recall_range)
 492 {
 493         return (recall_range->iomode == IOMODE_ANY ||
 494                 lseg_range->iomode == recall_range->iomode) &&
 495                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 496 }
 497
 498 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 499                 struct list_head *tmp_list)
 500 {
 501         if (!atomic_dec_and_test(&lseg->pls_refcount))
 502                 return false;
 503         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 504         list_add(&lseg->pls_list, tmp_list);
 505         return true;
 506 }
 507
 508 /* Returns 1 if lseg is removed from list, 0 otherwise */
 509 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 510                              struct list_head *tmp_list)
 511 {
 512         int rv = 0;
 513
 514         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 515                 /* Remove the reference keeping the lseg in the
 516                  * list.  It will now be removed when all
 517                  * outstanding io is finished.
 518                  */
 519                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 520                         atomic_read(&lseg->pls_refcount));
 521                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 522                         rv = 1;
 523         }
 524         return rv;
 525 }
 526
 527 /*
 528  * Compare 2 layout stateid sequence ids, to see which is newer,
 529  * taking into account wraparound issues.
 530  */
 531 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 532 {
 533         return (s32)(s1 - s2) > 0;
 534 }
 535
 536 /**
 537  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 538  * @lo: layout header containing the lsegs
 539  * @tmp_list: list head where doomed lsegs should go
 540  * @recall_range: optional recall range argument to match (may be NULL)
 541  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 542  *
 543  * Walk the list of lsegs in the layout header, and tear down any that should
 544  * be destroyed. If "recall_range" is specified then the segment must match
 545  * that range. If "seq" is non-zero, then only match segments that were handed
 546  * out at or before that sequence.
 547  *
 548  * Returns number of matching invalid lsegs remaining in list after scanning
 549  * it and purging them.
 550  */
 551 int
 552 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 553                             struct list_head *tmp_list,
 554                             const struct pnfs_layout_range *recall_range,
 555                             u32 seq)
 556 {
 557         struct pnfs_layout_segment *lseg, *next;
 558         int remaining = 0;
 559
 560         dprintk("%s:Begin lo %p\n", __func__, lo);
 561
 562         if (list_empty(&lo->plh_segs))
 563                 return 0;
 564         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 565                 if (!recall_range ||
 566                     should_free_lseg(&lseg->pls_range, recall_range)) {
 567                         if (seq && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 568                                 continue;
 569                         dprintk("%s: freeing lseg %p iomode %d seq %u"
 570                                 "offset %llu length %llu\n", __func__,
 571                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 572                                 lseg->pls_range.offset, lseg->pls_range.length);
 573                         if (!mark_lseg_invalid(lseg, tmp_list))
 574                                 remaining++;
 575                 }
 576         dprintk("%s:Return %i\n", __func__, remaining);
 577         return remaining;
 578 }
 579
 580 /* note free_me must contain lsegs from a single layout_hdr */
 581 void
 582 pnfs_free_lseg_list(struct list_head *free_me)
 583 {
 584         struct pnfs_layout_segment *lseg, *tmp;
 585
 586         if (list_empty(free_me))
 587                 return;
 588
 589         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 590                 list_del(&lseg->pls_list);
 591                 pnfs_free_lseg(lseg);
 592         }
 593 }
 594
 595 void
 596 pnfs_destroy_layout(struct nfs_inode *nfsi)
 597 {
 598         struct pnfs_layout_hdr *lo;
 599         LIST_HEAD(tmp_list);
 600
 601         spin_lock(&nfsi->vfs_inode.i_lock);
 602         lo = nfsi->layout;
 603         if (lo) {
 604                 pnfs_get_layout_hdr(lo);
 605                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 606                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 607                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 608                 spin_unlock(&nfsi->vfs_inode.i_lock);
 609                 pnfs_free_lseg_list(&tmp_list);
 610                 pnfs_put_layout_hdr(lo);
 611         } else
 612                 spin_unlock(&nfsi->vfs_inode.i_lock);
 613 }
 614 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 615
 616 static bool
 617 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 618                 struct list_head *layout_list)
 619 {
 620         struct pnfs_layout_hdr *lo;
 621         bool ret = false;
 622
 623         spin_lock(&inode->i_lock);
 624         lo = NFS_I(inode)->layout;
 625         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 626                 pnfs_get_layout_hdr(lo);
 627                 list_add(&lo->plh_bulk_destroy, layout_list);
 628                 ret = true;
 629         }
 630         spin_unlock(&inode->i_lock);
 631         return ret;
 632 }
 633
 634 /* Caller must hold rcu_read_lock and clp->cl_lock */
 635 static int
 636 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 637                 struct nfs_server *server,
 638                 struct list_head *layout_list)
 639 {
 640         struct pnfs_layout_hdr *lo, *next;
 641         struct inode *inode;
 642
 643         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 644                 inode = igrab(lo->plh_inode);
 645                 if (inode == NULL)
 646                         continue;
 647                 list_del_init(&lo->plh_layouts);
 648                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 649                         continue;
 650                 rcu_read_unlock();
 651                 spin_unlock(&clp->cl_lock);
 652                 iput(inode);
 653                 spin_lock(&clp->cl_lock);
 654                 rcu_read_lock();
 655                 return -EAGAIN;
 656         }
 657         return 0;
 658 }
 659
 660 static int
 661 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 662                 bool is_bulk_recall)
 663 {
 664         struct pnfs_layout_hdr *lo;
 665         struct inode *inode;
 666         LIST_HEAD(lseg_list);
 667         int ret = 0;
 668
 669         while (!list_empty(layout_list)) {
 670                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 671                                 plh_bulk_destroy);
 672                 dprintk("%s freeing layout for inode %lu\n", __func__,
 673                         lo->plh_inode->i_ino);
 674                 inode = lo->plh_inode;
 675
 676                 pnfs_layoutcommit_inode(inode, false);
 677
 678                 spin_lock(&inode->i_lock);
 679                 list_del_init(&lo->plh_bulk_destroy);
 680                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 681                         if (is_bulk_recall)
 682                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 683                         ret = -EAGAIN;
 684                 }
 685                 spin_unlock(&inode->i_lock);
 686                 pnfs_free_lseg_list(&lseg_list);
 687                 /* Free all lsegs that are attached to commit buckets */
 688                 nfs_commit_inode(inode, 0);
 689                 pnfs_put_layout_hdr(lo);
 690                 iput(inode);
 691         }
 692         return ret;
 693 }
 694
 695 int
 696 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 697                 struct nfs_fsid *fsid,
 698                 bool is_recall)
 699 {
 700         struct nfs_server *server;
 701         LIST_HEAD(layout_list);
 702
 703         spin_lock(&clp->cl_lock);
 704         rcu_read_lock();
 705 restart:
 706         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 707                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 708                         continue;
 709                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 710                                 server,
 711                                 &layout_list) != 0)
 712                         goto restart;
 713         }
 714         rcu_read_unlock();
 715         spin_unlock(&clp->cl_lock);
 716
 717         if (list_empty(&layout_list))
 718                 return 0;
 719         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 720 }
 721
 722 int
 723 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 724                 bool is_recall)
 725 {
 726         struct nfs_server *server;
 727         LIST_HEAD(layout_list);
 728
 729         spin_lock(&clp->cl_lock);
 730         rcu_read_lock();
 731 restart:
 732         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 733                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 734                                         server,
 735                                         &layout_list) != 0)
 736                         goto restart;
 737         }
 738         rcu_read_unlock();
 739         spin_unlock(&clp->cl_lock);
 740
 741         if (list_empty(&layout_list))
 742                 return 0;
 743         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 744 }
 745
 746 /*
 747  * Called by the state manger to remove all layouts established under an
 748  * expired lease.
 749  */
 750 void
 751 pnfs_destroy_all_layouts(struct nfs_client *clp)
 752 {
 753         nfs4_deviceid_mark_client_invalid(clp);
 754         nfs4_deviceid_purge_client(clp);
 755
 756         pnfs_destroy_layouts_byclid(clp, false);
 757 }
 758
 759 /* update lo->plh_stateid with new if is more recent */
 760 void
 761 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 762                         bool update_barrier)
 763 {
 764         u32 oldseq, newseq, new_barrier;
 765         int empty = list_empty(&lo->plh_segs);
 766
 767         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 768         newseq = be32_to_cpu(new->seqid);
 769         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 770                 nfs4_stateid_copy(&lo->plh_stateid, new);
 771                 if (update_barrier) {
 772                         new_barrier = be32_to_cpu(new->seqid);
 773                 } else {
 774                         /* Because of wraparound, we want to keep the barrier
 775                          * "close" to the current seqids.
 776                          */
 777                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 778                 }
 779                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 780                         lo->plh_barrier = new_barrier;
 781         }
 782 }
 783
 784 static bool
 785 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 786                 const nfs4_stateid *stateid)
 787 {
 788         u32 seqid = be32_to_cpu(stateid->seqid);
 789
 790         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 791 }
 792
 793 /* lget is set to 1 if called from inside send_layoutget call chain */
 794 static bool
 795 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 796 {
 797         return lo->plh_block_lgets ||
 798                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 799 }
 800
 801 /*
 802  * Get layout from server.
 803  *    for now, assume that whole file layouts are requested.
 804  *    arg->offset: 0
 805  *    arg->length: all ones
 806  */
 807 static struct pnfs_layout_segment *
 808 send_layoutget(struct pnfs_layout_hdr *lo,
 809            struct nfs_open_context *ctx,
 810            nfs4_stateid *stateid,
 811            const struct pnfs_layout_range *range,
 812            long *timeout, gfp_t gfp_flags)
 813 {
 814         struct inode *ino = lo->plh_inode;
 815         struct nfs_server *server = NFS_SERVER(ino);
 816         struct nfs4_layoutget *lgp;
 817         loff_t i_size;
 818
 819         dprintk("--> %s\n", __func__);
 820
 821         /*
 822          * Synchronously retrieve layout information from server and
 823          * store in lseg. If we race with a concurrent seqid morphing
 824          * op, then re-send the LAYOUTGET.
 825          */
 826         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 827         if (lgp == NULL)
 828                 return ERR_PTR(-ENOMEM);
 829
 830         i_size = i_size_read(ino);
 831
 832         lgp->args.minlength = PAGE_SIZE;
 833         if (lgp->args.minlength > range->length)
 834                 lgp->args.minlength = range->length;
 835         if (range->iomode == IOMODE_READ) {
 836                 if (range->offset >= i_size)
 837                         lgp->args.minlength = 0;
 838                 else if (i_size - range->offset < lgp->args.minlength)
 839                         lgp->args.minlength = i_size - range->offset;
 840         }
 841         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 842         pnfs_copy_range(&lgp->args.range, range);
 843         lgp->args.type = server->pnfs_curr_ld->id;
 844         lgp->args.inode = ino;
 845         lgp->args.ctx = get_nfs_open_context(ctx);
 846         nfs4_stateid_copy(&lgp->args.stateid, stateid);
 847         lgp->gfp_flags = gfp_flags;
 848         lgp->cred = lo->plh_lc_cred;
 849
 850         return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 851 }
 852
 853 static void pnfs_clear_layoutcommit(struct inode *inode,
 854                 struct list_head *head)
 855 {
 856         struct nfs_inode *nfsi = NFS_I(inode);
 857         struct pnfs_layout_segment *lseg, *tmp;
 858
 859         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 860                 return;
 861         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 862                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 863                         continue;
 864                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 865         }
 866 }
 867
 868 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 869 {
 870         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 871         smp_mb__after_atomic();
 872         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 873         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 874 }
 875
 876 static bool
 877 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 878 {
 879         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 880                 return false;
 881         lo->plh_return_iomode = 0;
 882         lo->plh_return_seq = 0;
 883         pnfs_get_layout_hdr(lo);
 884         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 885         return true;
 886 }
 887
 888 static int
 889 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
 890                        enum pnfs_iomode iomode, bool sync)
 891 {
 892         struct inode *ino = lo->plh_inode;
 893         struct nfs4_layoutreturn *lrp;
 894         int status = 0;
 895
 896         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 897         if (unlikely(lrp == NULL)) {
 898                 status = -ENOMEM;
 899                 spin_lock(&ino->i_lock);
 900                 pnfs_clear_layoutreturn_waitbit(lo);
 901                 spin_unlock(&ino->i_lock);
 902                 pnfs_put_layout_hdr(lo);
 903                 goto out;
 904         }
 905
 906         nfs4_stateid_copy(&lrp->args.stateid, stateid);
 907         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 908         lrp->args.inode = ino;
 909         lrp->args.range.iomode = iomode;
 910         lrp->args.range.offset = 0;
 911         lrp->args.range.length = NFS4_MAX_UINT64;
 912         lrp->args.layout = lo;
 913         lrp->clp = NFS_SERVER(ino)->nfs_client;
 914         lrp->cred = lo->plh_lc_cred;
 915
 916         status = nfs4_proc_layoutreturn(lrp, sync);
 917 out:
 918         dprintk("<-- %s status: %d\n", __func__, status);
 919         return status;
 920 }
 921
 922 /* Return true if layoutreturn is needed */
 923 static bool
 924 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
 925 {
 926         struct pnfs_layout_segment *s;
 927
 928         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 929                 return false;
 930
 931         /* Defer layoutreturn until all lsegs are done */
 932         list_for_each_entry(s, &lo->plh_segs, pls_list) {
 933                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 934                         return false;
 935         }
 936
 937         return true;
 938 }
 939
 940 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
 941 {
 942         struct inode *inode= lo->plh_inode;
 943
 944         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 945                 return;
 946         spin_lock(&inode->i_lock);
 947         if (pnfs_layout_need_return(lo)) {
 948                 nfs4_stateid stateid;
 949                 enum pnfs_iomode iomode;
 950                 bool send;
 951
 952                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
 953                 stateid.seqid = cpu_to_be32(lo->plh_return_seq);
 954                 iomode = lo->plh_return_iomode;
 955                 send = pnfs_prepare_layoutreturn(lo);
 956                 spin_unlock(&inode->i_lock);
 957                 if (send) {
 958                         /* Send an async layoutreturn so we dont deadlock */
 959                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
 960                 }
 961         } else
 962                 spin_unlock(&inode->i_lock);
 963 }
 964
 965 /*
 966  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 967  * when the layout segment list is empty.
 968  *
 969  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 970  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 971  * deviceid is marked invalid.
 972  */
 973 int
 974 _pnfs_return_layout(struct inode *ino)
 975 {
 976         struct pnfs_layout_hdr *lo = NULL;
 977         struct nfs_inode *nfsi = NFS_I(ino);
 978         LIST_HEAD(tmp_list);
 979         nfs4_stateid stateid;
 980         int status = 0, empty;
 981         bool send;
 982
 983         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 984
 985         spin_lock(&ino->i_lock);
 986         lo = nfsi->layout;
 987         if (!lo) {
 988                 spin_unlock(&ino->i_lock);
 989                 dprintk("NFS: %s no layout to return\n", __func__);
 990                 goto out;
 991         }
 992         nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
 993         /* Reference matched in nfs4_layoutreturn_release */
 994         pnfs_get_layout_hdr(lo);
 995         empty = list_empty(&lo->plh_segs);
 996         pnfs_clear_layoutcommit(ino, &tmp_list);
 997         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
 998
 999         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1000                 struct pnfs_layout_range range = {
1001                         .iomode         = IOMODE_ANY,
1002                         .offset         = 0,
1003                         .length         = NFS4_MAX_UINT64,
1004                 };
1005                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1006         }
1007
1008         /* Don't send a LAYOUTRETURN if list was initially empty */
1009         if (empty) {
1010                 spin_unlock(&ino->i_lock);
1011                 dprintk("NFS: %s no layout segments to return\n", __func__);
1012                 goto out_put_layout_hdr;
1013         }
1014
1015         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1016         send = pnfs_prepare_layoutreturn(lo);
1017         spin_unlock(&ino->i_lock);
1018         pnfs_free_lseg_list(&tmp_list);
1019         if (send)
1020                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1021 out_put_layout_hdr:
1022         pnfs_put_layout_hdr(lo);
1023 out:
1024         dprintk("<-- %s status: %d\n", __func__, status);
1025         return status;
1026 }
1027 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1028
1029 int
1030 pnfs_commit_and_return_layout(struct inode *inode)
1031 {
1032         struct pnfs_layout_hdr *lo;
1033         int ret;
1034
1035         spin_lock(&inode->i_lock);
1036         lo = NFS_I(inode)->layout;
1037         if (lo == NULL) {
1038                 spin_unlock(&inode->i_lock);
1039                 return 0;
1040         }
1041         pnfs_get_layout_hdr(lo);
1042         /* Block new layoutgets and read/write to ds */
1043         lo->plh_block_lgets++;
1044         spin_unlock(&inode->i_lock);
1045         filemap_fdatawait(inode->i_mapping);
1046         ret = pnfs_layoutcommit_inode(inode, true);
1047         if (ret == 0)
1048                 ret = _pnfs_return_layout(inode);
1049         spin_lock(&inode->i_lock);
1050         lo->plh_block_lgets--;
1051         spin_unlock(&inode->i_lock);
1052         pnfs_put_layout_hdr(lo);
1053         return ret;
1054 }
1055
1056 bool pnfs_roc(struct inode *ino)
1057 {
1058         struct nfs_inode *nfsi = NFS_I(ino);
1059         struct nfs_open_context *ctx;
1060         struct nfs4_state *state;
1061         struct pnfs_layout_hdr *lo;
1062         struct pnfs_layout_segment *lseg, *tmp;
1063         nfs4_stateid stateid;
1064         LIST_HEAD(tmp_list);
1065         bool found = false, layoutreturn = false, roc = false;
1066
1067         spin_lock(&ino->i_lock);
1068         lo = nfsi->layout;
1069         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1070                 goto out_noroc;
1071
1072         /* no roc if we hold a delegation */
1073         if (nfs4_check_delegation(ino, FMODE_READ))
1074                 goto out_noroc;
1075
1076         list_for_each_entry(ctx, &nfsi->open_files, list) {
1077                 state = ctx->state;
1078                 /* Don't return layout if there is open file state */
1079                 if (state != NULL && state->state != 0)
1080                         goto out_noroc;
1081         }
1082
1083         nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1084         /* always send layoutreturn if being marked so */
1085         if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
1086                                    &lo->plh_flags))
1087                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1088
1089         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1090                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1091                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1092                         mark_lseg_invalid(lseg, &tmp_list);
1093                         found = true;
1094                 }
1095         /* ROC in two conditions:
1096          * 1. there are ROC lsegs
1097          * 2. we don't send layoutreturn
1098          */
1099         if (found && !layoutreturn) {
1100                 /* lo ref dropped in pnfs_roc_release() */
1101                 pnfs_get_layout_hdr(lo);
1102                 roc = true;
1103         }
1104
1105 out_noroc:
1106         spin_unlock(&ino->i_lock);
1107         pnfs_free_lseg_list(&tmp_list);
1108         pnfs_layoutcommit_inode(ino, true);
1109         if (layoutreturn)
1110                 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1111         return roc;
1112 }
1113
1114 void pnfs_roc_release(struct inode *ino)
1115 {
1116         struct pnfs_layout_hdr *lo;
1117
1118         spin_lock(&ino->i_lock);
1119         lo = NFS_I(ino)->layout;
1120         pnfs_clear_layoutreturn_waitbit(lo);
1121         if (atomic_dec_and_test(&lo->plh_refcount)) {
1122                 pnfs_detach_layout_hdr(lo);
1123                 spin_unlock(&ino->i_lock);
1124                 pnfs_free_layout_hdr(lo);
1125         } else
1126                 spin_unlock(&ino->i_lock);
1127 }
1128
1129 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1130 {
1131         struct pnfs_layout_hdr *lo;
1132
1133         spin_lock(&ino->i_lock);
1134         lo = NFS_I(ino)->layout;
1135         pnfs_mark_layout_returned_if_empty(lo);
1136         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1137                 lo->plh_barrier = barrier;
1138         spin_unlock(&ino->i_lock);
1139         trace_nfs4_layoutreturn_on_close(ino, 0);
1140 }
1141
1142 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1143 {
1144         struct nfs_inode *nfsi = NFS_I(ino);
1145         struct pnfs_layout_hdr *lo;
1146         u32 current_seqid;
1147
1148         spin_lock(&ino->i_lock);
1149         lo = nfsi->layout;
1150         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1151
1152         /* Since close does not return a layout stateid for use as
1153          * a barrier, we choose the worst-case barrier.
1154          */
1155         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1156         spin_unlock(&ino->i_lock);
1157 }
1158
1159 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1160 {
1161         struct nfs_inode *nfsi = NFS_I(ino);
1162         struct pnfs_layout_hdr *lo;
1163         bool sleep = false;
1164
1165         /* we might not have grabbed lo reference. so need to check under
1166          * i_lock */
1167         spin_lock(&ino->i_lock);
1168         lo = nfsi->layout;
1169         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1170                 sleep = true;
1171         spin_unlock(&ino->i_lock);
1172
1173         if (sleep)
1174                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1175
1176         return sleep;
1177 }
1178
1179 /*
1180  * Compare two layout segments for sorting into layout cache.
1181  * We want to preferentially return RW over RO layouts, so ensure those
1182  * are seen first.
1183  */
1184 static s64
1185 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1186            const struct pnfs_layout_range *l2)
1187 {
1188         s64 d;
1189
1190         /* high offset > low offset */
1191         d = l1->offset - l2->offset;
1192         if (d)
1193                 return d;
1194
1195         /* short length > long length */
1196         d = l2->length - l1->length;
1197         if (d)
1198                 return d;
1199
1200         /* read > read/write */
1201         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1202 }
1203
1204 static bool
1205 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1206                 const struct pnfs_layout_range *l2)
1207 {
1208         return pnfs_lseg_range_cmp(l1, l2) > 0;
1209 }
1210
1211 static bool
1212 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1213                 struct pnfs_layout_segment *old)
1214 {
1215         return false;
1216 }
1217
1218 void
1219 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1220                    struct pnfs_layout_segment *lseg,
1221                    bool (*is_after)(const struct pnfs_layout_range *,
1222                            const struct pnfs_layout_range *),
1223                    bool (*do_merge)(struct pnfs_layout_segment *,
1224                            struct pnfs_layout_segment *),
1225                    struct list_head *free_me)
1226 {
1227         struct pnfs_layout_segment *lp, *tmp;
1228
1229         dprintk("%s:Begin\n", __func__);
1230
1231         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1232                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1233                         continue;
1234                 if (do_merge(lseg, lp)) {
1235                         mark_lseg_invalid(lp, free_me);
1236                         continue;
1237                 }
1238                 if (is_after(&lseg->pls_range, &lp->pls_range))
1239                         continue;
1240                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1241                 dprintk("%s: inserted lseg %p "
1242                         "iomode %d offset %llu length %llu before "
1243                         "lp %p iomode %d offset %llu length %llu\n",
1244                         __func__, lseg, lseg->pls_range.iomode,
1245                         lseg->pls_range.offset, lseg->pls_range.length,
1246                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1247                         lp->pls_range.length);
1248                 goto out;
1249         }
1250         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1251         dprintk("%s: inserted lseg %p "
1252                 "iomode %d offset %llu length %llu at tail\n",
1253                 __func__, lseg, lseg->pls_range.iomode,
1254                 lseg->pls_range.offset, lseg->pls_range.length);
1255 out:
1256         pnfs_get_layout_hdr(lo);
1257
1258         dprintk("%s:Return\n", __func__);
1259 }
1260 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1261
1262 static void
1263 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1264                    struct pnfs_layout_segment *lseg,
1265                    struct list_head *free_me)
1266 {
1267         struct inode *inode = lo->plh_inode;
1268         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1269
1270         if (ld->add_lseg != NULL)
1271                 ld->add_lseg(lo, lseg, free_me);
1272         else
1273                 pnfs_generic_layout_insert_lseg(lo, lseg,
1274                                 pnfs_lseg_range_is_after,
1275                                 pnfs_lseg_no_merge,
1276                                 free_me);
1277 }
1278
1279 static struct pnfs_layout_hdr *
1280 alloc_init_layout_hdr(struct inode *ino,
1281                       struct nfs_open_context *ctx,
1282                       gfp_t gfp_flags)
1283 {
1284         struct pnfs_layout_hdr *lo;
1285
1286         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1287         if (!lo)
1288                 return NULL;
1289         atomic_set(&lo->plh_refcount, 1);
1290         INIT_LIST_HEAD(&lo->plh_layouts);
1291         INIT_LIST_HEAD(&lo->plh_segs);
1292         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1293         lo->plh_inode = ino;
1294         lo->plh_lc_cred = get_rpccred(ctx->cred);
1295         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1296         return lo;
1297 }
1298
1299 static struct pnfs_layout_hdr *
1300 pnfs_find_alloc_layout(struct inode *ino,
1301                        struct nfs_open_context *ctx,
1302                        gfp_t gfp_flags)
1303         __releases(&ino->i_lock)
1304         __acquires(&ino->i_lock)
1305 {
1306         struct nfs_inode *nfsi = NFS_I(ino);
1307         struct pnfs_layout_hdr *new = NULL;
1308
1309         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1310
1311         if (nfsi->layout != NULL)
1312                 goto out_existing;
1313         spin_unlock(&ino->i_lock);
1314         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1315         spin_lock(&ino->i_lock);
1316
1317         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1318                 nfsi->layout = new;
1319                 return new;
1320         } else if (new != NULL)
1321                 pnfs_free_layout_hdr(new);
1322 out_existing:
1323         pnfs_get_layout_hdr(nfsi->layout);
1324         return nfsi->layout;
1325 }
1326
1327 /*
1328  * iomode matching rules:
1329  * iomode       lseg    strict match
1330  *                      iomode
1331  * -----        -----   ------ -----
1332  * ANY          READ    N/A    true
1333  * ANY          RW      N/A    true
1334  * RW           READ    N/A    false
1335  * RW           RW      N/A    true
1336  * READ         READ    N/A    true
1337  * READ         RW      true   false
1338  * READ         RW      false  true
1339  */
1340 static bool
1341 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1342                  const struct pnfs_layout_range *range,
1343                  bool strict_iomode)
1344 {
1345         struct pnfs_layout_range range1;
1346
1347         if ((range->iomode == IOMODE_RW &&
1348              ls_range->iomode != IOMODE_RW) ||
1349             (range->iomode != ls_range->iomode &&
1350              strict_iomode == true) ||
1351             !pnfs_lseg_range_intersecting(ls_range, range))
1352                 return 0;
1353
1354         /* range1 covers only the first byte in the range */
1355         range1 = *range;
1356         range1.length = 1;
1357         return pnfs_lseg_range_contained(ls_range, &range1);
1358 }
1359
1360 /*
1361  * lookup range in layout
1362  */
1363 static struct pnfs_layout_segment *
1364 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1365                 struct pnfs_layout_range *range,
1366                 bool strict_iomode)
1367 {
1368         struct pnfs_layout_segment *lseg, *ret = NULL;
1369
1370         dprintk("%s:Begin\n", __func__);
1371
1372         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1373                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1374                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1375                     pnfs_lseg_range_match(&lseg->pls_range, range,
1376                                           strict_iomode)) {
1377                         ret = pnfs_get_lseg(lseg);
1378                         break;
1379                 }
1380         }
1381
1382         dprintk("%s:Return lseg %p ref %d\n",
1383                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1384         return ret;
1385 }
1386
1387 /*
1388  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1389  * to the MDS or over pNFS
1390  *
1391  * The nfs_inode read_io and write_io fields are cumulative counters reset
1392  * when there are no layout segments. Note that in pnfs_update_layout iomode
1393  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1394  * WRITE request.
1395  *
1396  * A return of true means use MDS I/O.
1397  *
1398  * From rfc 5661:
1399  * If a file's size is smaller than the file size threshold, data accesses
1400  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1401  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1402  * server.  If both file size and I/O size are provided, the client SHOULD
1403  * reach or exceed  both thresholds before sending its read or write
1404  * requests to the data server.
1405  */
1406 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1407                                      struct inode *ino, int iomode)
1408 {
1409         struct nfs4_threshold *t = ctx->mdsthreshold;
1410         struct nfs_inode *nfsi = NFS_I(ino);
1411         loff_t fsize = i_size_read(ino);
1412         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1413
1414         if (t == NULL)
1415                 return ret;
1416
1417         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1418                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1419
1420         switch (iomode) {
1421         case IOMODE_READ:
1422                 if (t->bm & THRESHOLD_RD) {
1423                         dprintk("%s fsize %llu\n", __func__, fsize);
1424                         size_set = true;
1425                         if (fsize < t->rd_sz)
1426                                 size = true;
1427                 }
1428                 if (t->bm & THRESHOLD_RD_IO) {
1429                         dprintk("%s nfsi->read_io %llu\n", __func__,
1430                                 nfsi->read_io);
1431                         io_set = true;
1432                         if (nfsi->read_io < t->rd_io_sz)
1433                                 io = true;
1434                 }
1435                 break;
1436         case IOMODE_RW:
1437                 if (t->bm & THRESHOLD_WR) {
1438                         dprintk("%s fsize %llu\n", __func__, fsize);
1439                         size_set = true;
1440                         if (fsize < t->wr_sz)
1441                                 size = true;
1442                 }
1443                 if (t->bm & THRESHOLD_WR_IO) {
1444                         dprintk("%s nfsi->write_io %llu\n", __func__,
1445                                 nfsi->write_io);
1446                         io_set = true;
1447                         if (nfsi->write_io < t->wr_io_sz)
1448                                 io = true;
1449                 }
1450                 break;
1451         }
1452         if (size_set && io_set) {
1453                 if (size && io)
1454                         ret = true;
1455         } else if (size || io)
1456                 ret = true;
1457
1458         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1459         return ret;
1460 }
1461
1462 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1463 {
1464         /*
1465          * send layoutcommit as it can hold up layoutreturn due to lseg
1466          * reference
1467          */
1468         pnfs_layoutcommit_inode(lo->plh_inode, false);
1469         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1470                                    nfs_wait_bit_killable,
1471                                    TASK_UNINTERRUPTIBLE);
1472 }
1473
1474 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1475 {
1476         unsigned long *bitlock = &lo->plh_flags;
1477
1478         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1479         smp_mb__after_atomic();
1480         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1481 }
1482
1483 /*
1484  * Layout segment is retreived from the server if not cached.
1485  * The appropriate layout segment is referenced and returned to the caller.
1486  */
1487 struct pnfs_layout_segment *
1488 pnfs_update_layout(struct inode *ino,
1489                    struct nfs_open_context *ctx,
1490                    loff_t pos,
1491                    u64 count,
1492                    enum pnfs_iomode iomode,
1493                    bool strict_iomode,
1494                    gfp_t gfp_flags)
1495 {
1496         struct pnfs_layout_range arg = {
1497                 .iomode = iomode,
1498                 .offset = pos,
1499                 .length = count,
1500         };
1501         unsigned pg_offset, seq;
1502         struct nfs_server *server = NFS_SERVER(ino);
1503         struct nfs_client *clp = server->nfs_client;
1504         struct pnfs_layout_hdr *lo = NULL;
1505         struct pnfs_layout_segment *lseg = NULL;
1506         nfs4_stateid stateid;
1507         long timeout = 0;
1508         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1509         bool first;
1510
1511         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1512                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1513                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1514                 goto out;
1515         }
1516
1517         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1518                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1519                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1520                 goto out;
1521         }
1522
1523         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1524                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1525                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1526                 goto out;
1527         }
1528
1529 lookup_again:
1530         first = false;
1531         spin_lock(&ino->i_lock);
1532         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1533         if (lo == NULL) {
1534                 spin_unlock(&ino->i_lock);
1535                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1536                                  PNFS_UPDATE_LAYOUT_NOMEM);
1537                 goto out;
1538         }
1539
1540         /* Do we even need to bother with this? */
1541         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1542                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1543                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1544                 dprintk("%s matches recall, use MDS\n", __func__);
1545                 goto out_unlock;
1546         }
1547
1548         /* if LAYOUTGET already failed once we don't try again */
1549         if (pnfs_layout_io_test_failed(lo, iomode)) {
1550                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1551                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1552                 goto out_unlock;
1553         }
1554
1555         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1556         if (lseg) {
1557                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1558                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1559                 goto out_unlock;
1560         }
1561
1562         if (!nfs4_valid_open_stateid(ctx->state)) {
1563                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1564                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1565                 goto out_unlock;
1566         }
1567
1568         /*
1569          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1570          * stateid, or it has been invalidated, then we must use the open
1571          * stateid.
1572          */
1573         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1574
1575                 /*
1576                  * The first layoutget for the file. Need to serialize per
1577                  * RFC 5661 Errata 3208.
1578                  */
1579                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1580                                      &lo->plh_flags)) {
1581                         spin_unlock(&ino->i_lock);
1582                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1583                                     TASK_UNINTERRUPTIBLE);
1584                         pnfs_put_layout_hdr(lo);
1585                         dprintk("%s retrying\n", __func__);
1586                         goto lookup_again;
1587                 }
1588
1589                 first = true;
1590                 do {
1591                         seq = read_seqbegin(&ctx->state->seqlock);
1592                         nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1593                 } while (read_seqretry(&ctx->state->seqlock, seq));
1594         } else {
1595                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1596         }
1597
1598         /*
1599          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1600          * for LAYOUTRETURN even if first is true.
1601          */
1602         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1603                 spin_unlock(&ino->i_lock);
1604                 dprintk("%s wait for layoutreturn\n", __func__);
1605                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1606                         if (first)
1607                                 pnfs_clear_first_layoutget(lo);
1608                         pnfs_put_layout_hdr(lo);
1609                         dprintk("%s retrying\n", __func__);
1610                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1611                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1612                         goto lookup_again;
1613                 }
1614                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1615                                 PNFS_UPDATE_LAYOUT_RETURN);
1616                 goto out_put_layout_hdr;
1617         }
1618
1619         if (pnfs_layoutgets_blocked(lo)) {
1620                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1621                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1622                 goto out_unlock;
1623         }
1624         atomic_inc(&lo->plh_outstanding);
1625         spin_unlock(&ino->i_lock);
1626
1627         if (list_empty(&lo->plh_layouts)) {
1628                 /* The lo must be on the clp list if there is any
1629                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1630                  */
1631                 spin_lock(&clp->cl_lock);
1632                 if (list_empty(&lo->plh_layouts))
1633                         list_add_tail(&lo->plh_layouts, &server->layouts);
1634                 spin_unlock(&clp->cl_lock);
1635         }
1636
1637         pg_offset = arg.offset & ~PAGE_MASK;
1638         if (pg_offset) {
1639                 arg.offset -= pg_offset;
1640                 arg.length += pg_offset;
1641         }
1642         if (arg.length != NFS4_MAX_UINT64)
1643                 arg.length = PAGE_ALIGN(arg.length);
1644
1645         lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1646         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1647                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1648         atomic_dec(&lo->plh_outstanding);
1649         if (IS_ERR(lseg)) {
1650                 switch(PTR_ERR(lseg)) {
1651                 case -EBUSY:
1652                         if (time_after(jiffies, giveup))
1653                                 lseg = NULL;
1654                         break;
1655                 case -ERECALLCONFLICT:
1656                         /* Huh? We hold no layouts, how is there a recall? */
1657                         if (first) {
1658                                 lseg = NULL;
1659                                 break;
1660                         }
1661                         /* Destroy the existing layout and start over */
1662                         if (time_after(jiffies, giveup))
1663                                 pnfs_destroy_layout(NFS_I(ino));
1664                         /* Fallthrough */
1665                 case -EAGAIN:
1666                         break;
1667                 default:
1668                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1669                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1670                                 lseg = NULL;
1671                         }
1672                         goto out_put_layout_hdr;
1673                 }
1674                 if (lseg) {
1675                         if (first)
1676                                 pnfs_clear_first_layoutget(lo);
1677                         trace_pnfs_update_layout(ino, pos, count,
1678                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1679                         pnfs_put_layout_hdr(lo);
1680                         goto lookup_again;
1681                 }
1682         } else {
1683                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1684         }
1685
1686 out_put_layout_hdr:
1687         if (first)
1688                 pnfs_clear_first_layoutget(lo);
1689         pnfs_put_layout_hdr(lo);
1690 out:
1691         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1692                         "(%s, offset: %llu, length: %llu)\n",
1693                         __func__, ino->i_sb->s_id,
1694                         (unsigned long long)NFS_FILEID(ino),
1695                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1696                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1697                         (unsigned long long)pos,
1698                         (unsigned long long)count);
1699         return lseg;
1700 out_unlock:
1701         spin_unlock(&ino->i_lock);
1702         goto out_put_layout_hdr;
1703 }
1704 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1705
1706 static bool
1707 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1708 {
1709         switch (range->iomode) {
1710         case IOMODE_READ:
1711         case IOMODE_RW:
1712                 break;
1713         default:
1714                 return false;
1715         }
1716         if (range->offset == NFS4_MAX_UINT64)
1717                 return false;
1718         if (range->length == 0)
1719                 return false;
1720         if (range->length != NFS4_MAX_UINT64 &&
1721             range->length > NFS4_MAX_UINT64 - range->offset)
1722                 return false;
1723         return true;
1724 }
1725
1726 struct pnfs_layout_segment *
1727 pnfs_layout_process(struct nfs4_layoutget *lgp)
1728 {
1729         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1730         struct nfs4_layoutget_res *res = &lgp->res;
1731         struct pnfs_layout_segment *lseg;
1732         struct inode *ino = lo->plh_inode;
1733         LIST_HEAD(free_me);
1734
1735         if (!pnfs_sanity_check_layout_range(&res->range))
1736                 return ERR_PTR(-EINVAL);
1737
1738         /* Inject layout blob into I/O device driver */
1739         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1740         if (IS_ERR_OR_NULL(lseg)) {
1741                 if (!lseg)
1742                         lseg = ERR_PTR(-ENOMEM);
1743
1744                 dprintk("%s: Could not allocate layout: error %ld\n",
1745                        __func__, PTR_ERR(lseg));
1746                 return lseg;
1747         }
1748
1749         init_lseg(lo, lseg);
1750         lseg->pls_range = res->range;
1751         lseg->pls_seq = be32_to_cpu(res->stateid.seqid);
1752
1753         spin_lock(&ino->i_lock);
1754         if (pnfs_layoutgets_blocked(lo)) {
1755                 dprintk("%s forget reply due to state\n", __func__);
1756                 goto out_forget;
1757         }
1758
1759         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1760                 /* existing state ID, make sure the sequence number matches. */
1761                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1762                         dprintk("%s forget reply due to sequence\n", __func__);
1763                         goto out_forget;
1764                 }
1765                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1766         } else {
1767                 /*
1768                  * We got an entirely new state ID.  Mark all segments for the
1769                  * inode invalid, and don't bother validating the stateid
1770                  * sequence number.
1771                  */
1772                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL, 0);
1773
1774                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1775                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1776         }
1777
1778         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1779
1780         pnfs_get_lseg(lseg);
1781         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1782
1783         if (res->return_on_close)
1784                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1785
1786         spin_unlock(&ino->i_lock);
1787         pnfs_free_lseg_list(&free_me);
1788         return lseg;
1789
1790 out_forget:
1791         spin_unlock(&ino->i_lock);
1792         lseg->pls_layout = lo;
1793         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1794         return ERR_PTR(-EAGAIN);
1795 }
1796
1797 static void
1798 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
1799                          u32 seq)
1800 {
1801         if (lo->plh_return_iomode == iomode)
1802                 return;
1803         if (lo->plh_return_iomode != 0)
1804                 iomode = IOMODE_ANY;
1805         lo->plh_return_iomode = iomode;
1806         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
1807         if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
1808                 lo->plh_return_seq = seq;
1809 }
1810
1811 /**
1812  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1813  * @lo: pointer to layout header
1814  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1815  * @return_range: describe layout segment ranges to be returned
1816  *
1817  * This function is mainly intended for use by layoutrecall. It attempts
1818  * to free the layout segment immediately, or else to mark it for return
1819  * as soon as its reference count drops to zero.
1820  */
1821 int
1822 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1823                                 struct list_head *tmp_list,
1824                                 const struct pnfs_layout_range *return_range,
1825                                 u32 seq)
1826 {
1827         struct pnfs_layout_segment *lseg, *next;
1828         int remaining = 0;
1829
1830         dprintk("%s:Begin lo %p\n", __func__, lo);
1831
1832         if (list_empty(&lo->plh_segs))
1833                 return 0;
1834
1835         assert_spin_locked(&lo->plh_inode->i_lock);
1836
1837         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1838                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1839                         dprintk("%s: marking lseg %p iomode %d "
1840                                 "offset %llu length %llu\n", __func__,
1841                                 lseg, lseg->pls_range.iomode,
1842                                 lseg->pls_range.offset,
1843                                 lseg->pls_range.length);
1844                         if (mark_lseg_invalid(lseg, tmp_list))
1845                                 continue;
1846                         remaining++;
1847                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1848                 }
1849
1850         if (remaining)
1851                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
1852
1853         return remaining;
1854 }
1855
1856 void pnfs_error_mark_layout_for_return(struct inode *inode,
1857                                        struct pnfs_layout_segment *lseg)
1858 {
1859         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1860         struct pnfs_layout_range range = {
1861                 .iomode = lseg->pls_range.iomode,
1862                 .offset = 0,
1863                 .length = NFS4_MAX_UINT64,
1864         };
1865         LIST_HEAD(free_me);
1866         bool return_now = false;
1867
1868         spin_lock(&inode->i_lock);
1869         pnfs_set_plh_return_info(lo, range.iomode, lseg->pls_seq);
1870         /*
1871          * mark all matching lsegs so that we are sure to have no live
1872          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1873          * for how it works.
1874          */
1875         if (!pnfs_mark_matching_lsegs_return(lo, &free_me,
1876                                                 &range, lseg->pls_seq)) {
1877                 nfs4_stateid stateid;
1878                 enum pnfs_iomode iomode = lo->plh_return_iomode;
1879
1880                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1881                 return_now = pnfs_prepare_layoutreturn(lo);
1882                 spin_unlock(&inode->i_lock);
1883                 if (return_now)
1884                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1885         } else {
1886                 spin_unlock(&inode->i_lock);
1887                 nfs_commit_inode(inode, 0);
1888         }
1889         pnfs_free_lseg_list(&free_me);
1890 }
1891 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1892
1893 void
1894 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1895 {
1896         u64 rd_size = req->wb_bytes;
1897
1898         if (pgio->pg_lseg == NULL) {
1899                 if (pgio->pg_dreq == NULL)
1900                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1901                 else
1902                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1903
1904                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1905                                                    req->wb_context,
1906                                                    req_offset(req),
1907                                                    rd_size,
1908                                                    IOMODE_READ,
1909                                                    false,
1910                                                    GFP_KERNEL);
1911                 if (IS_ERR(pgio->pg_lseg)) {
1912                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1913                         pgio->pg_lseg = NULL;
1914                         return;
1915                 }
1916         }
1917         /* If no lseg, fall back to read through mds */
1918         if (pgio->pg_lseg == NULL)
1919                 nfs_pageio_reset_read_mds(pgio);
1920
1921 }
1922 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1923
1924 void
1925 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1926                            struct nfs_page *req, u64 wb_size)
1927 {
1928         if (pgio->pg_lseg == NULL) {
1929                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1930                                                    req->wb_context,
1931                                                    req_offset(req),
1932                                                    wb_size,
1933                                                    IOMODE_RW,
1934                                                    false,
1935                                                    GFP_NOFS);
1936                 if (IS_ERR(pgio->pg_lseg)) {
1937                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1938                         pgio->pg_lseg = NULL;
1939                         return;
1940                 }
1941         }
1942         /* If no lseg, fall back to write through mds */
1943         if (pgio->pg_lseg == NULL)
1944                 nfs_pageio_reset_write_mds(pgio);
1945 }
1946 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1947
1948 void
1949 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1950 {
1951         if (desc->pg_lseg) {
1952                 pnfs_put_lseg(desc->pg_lseg);
1953                 desc->pg_lseg = NULL;
1954         }
1955 }
1956 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1957
1958 /*
1959  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1960  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1961  */
1962 size_t
1963 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1964                      struct nfs_page *prev, struct nfs_page *req)
1965 {
1966         unsigned int size;
1967         u64 seg_end, req_start, seg_left;
1968
1969         size = nfs_generic_pg_test(pgio, prev, req);
1970         if (!size)
1971                 return 0;
1972
1973         /*
1974          * 'size' contains the number of bytes left in the current page (up
1975          * to the original size asked for in @req->wb_bytes).
1976          *
1977          * Calculate how many bytes are left in the layout segment
1978          * and if there are less bytes than 'size', return that instead.
1979          *
1980          * Please also note that 'end_offset' is actually the offset of the
1981          * first byte that lies outside the pnfs_layout_range. FIXME?
1982          *
1983          */
1984         if (pgio->pg_lseg) {
1985                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1986                                      pgio->pg_lseg->pls_range.length);
1987                 req_start = req_offset(req);
1988                 WARN_ON_ONCE(req_start >= seg_end);
1989                 /* start of request is past the last byte of this segment */
1990                 if (req_start >= seg_end) {
1991                         /* reference the new lseg */
1992                         if (pgio->pg_ops->pg_cleanup)
1993                                 pgio->pg_ops->pg_cleanup(pgio);
1994                         if (pgio->pg_ops->pg_init)
1995                                 pgio->pg_ops->pg_init(pgio, req);
1996                         return 0;
1997                 }
1998
1999                 /* adjust 'size' iff there are fewer bytes left in the
2000                  * segment than what nfs_generic_pg_test returned */
2001                 seg_left = seg_end - req_start;
2002                 if (seg_left < size)
2003                         size = (unsigned int)seg_left;
2004         }
2005
2006         return size;
2007 }
2008 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2009
2010 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2011 {
2012         struct nfs_pageio_descriptor pgio;
2013
2014         /* Resend all requests through the MDS */
2015         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2016                               hdr->completion_ops);
2017         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2018         return nfs_pageio_resend(&pgio, hdr);
2019 }
2020 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2021
2022 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2023 {
2024
2025         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2026         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2027             PNFS_LAYOUTRET_ON_ERROR) {
2028                 pnfs_return_layout(hdr->inode);
2029         }
2030         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2031                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2032 }
2033
2034 /*
2035  * Called by non rpc-based layout drivers
2036  */
2037 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2038 {
2039         if (likely(!hdr->pnfs_error)) {
2040                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2041                                 hdr->mds_offset + hdr->res.count);
2042                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2043         }
2044         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2045         if (unlikely(hdr->pnfs_error))
2046                 pnfs_ld_handle_write_error(hdr);
2047         hdr->mds_ops->rpc_release(hdr);
2048 }
2049 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2050
2051 static void
2052 pnfs_write_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_write_mds(desc);
2060                 mirror->pg_recoalesce = 1;
2061         }
2062         nfs_pgio_data_destroy(hdr);
2063         hdr->release(hdr);
2064 }
2065
2066 static enum pnfs_try_status
2067 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2068                         const struct rpc_call_ops *call_ops,
2069                         struct pnfs_layout_segment *lseg,
2070                         int how)
2071 {
2072         struct inode *inode = hdr->inode;
2073         enum pnfs_try_status trypnfs;
2074         struct nfs_server *nfss = NFS_SERVER(inode);
2075
2076         hdr->mds_ops = call_ops;
2077
2078         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2079                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2080         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2081         if (trypnfs != PNFS_NOT_ATTEMPTED)
2082                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2083         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2084         return trypnfs;
2085 }
2086
2087 static void
2088 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2089               struct nfs_pgio_header *hdr, int how)
2090 {
2091         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2092         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2093         enum pnfs_try_status trypnfs;
2094
2095         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2096         if (trypnfs == PNFS_NOT_ATTEMPTED)
2097                 pnfs_write_through_mds(desc, hdr);
2098 }
2099
2100 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2101 {
2102         pnfs_put_lseg(hdr->lseg);
2103         nfs_pgio_header_free(hdr);
2104 }
2105
2106 int
2107 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2108 {
2109         struct nfs_pgio_header *hdr;
2110         int ret;
2111
2112         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2113         if (!hdr) {
2114                 desc->pg_error = -ENOMEM;
2115                 return desc->pg_error;
2116         }
2117         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2118
2119         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2120         ret = nfs_generic_pgio(desc, hdr);
2121         if (!ret)
2122                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2123
2124         return ret;
2125 }
2126 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2127
2128 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2129 {
2130         struct nfs_pageio_descriptor pgio;
2131
2132         /* Resend all requests through the MDS */
2133         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2134         return nfs_pageio_resend(&pgio, hdr);
2135 }
2136 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2137
2138 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2139 {
2140         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2141         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2142             PNFS_LAYOUTRET_ON_ERROR) {
2143                 pnfs_return_layout(hdr->inode);
2144         }
2145         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2146                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2147 }
2148
2149 /*
2150  * Called by non rpc-based layout drivers
2151  */
2152 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2153 {
2154         if (likely(!hdr->pnfs_error)) {
2155                 __nfs4_read_done_cb(hdr);
2156                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2157         }
2158         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2159         if (unlikely(hdr->pnfs_error))
2160                 pnfs_ld_handle_read_error(hdr);
2161         hdr->mds_ops->rpc_release(hdr);
2162 }
2163 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2164
2165 static void
2166 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2167                 struct nfs_pgio_header *hdr)
2168 {
2169         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2170
2171         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2172                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2173                 nfs_pageio_reset_read_mds(desc);
2174                 mirror->pg_recoalesce = 1;
2175         }
2176         nfs_pgio_data_destroy(hdr);
2177         hdr->release(hdr);
2178 }
2179
2180 /*
2181  * Call the appropriate parallel I/O subsystem read function.
2182  */
2183 static enum pnfs_try_status
2184 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2185                        const struct rpc_call_ops *call_ops,
2186                        struct pnfs_layout_segment *lseg)
2187 {
2188         struct inode *inode = hdr->inode;
2189         struct nfs_server *nfss = NFS_SERVER(inode);
2190         enum pnfs_try_status trypnfs;
2191
2192         hdr->mds_ops = call_ops;
2193
2194         dprintk("%s: Reading ino:%lu %u@%llu\n",
2195                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2196
2197         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2198         if (trypnfs != PNFS_NOT_ATTEMPTED)
2199                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2200         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2201         return trypnfs;
2202 }
2203
2204 /* Resend all requests through pnfs. */
2205 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2206 {
2207         struct nfs_pageio_descriptor pgio;
2208
2209         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2210                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2211                                         hdr->completion_ops);
2212                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2213         }
2214 }
2215 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2216
2217 static void
2218 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2219 {
2220         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2221         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2222         enum pnfs_try_status trypnfs;
2223
2224         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2225         if (trypnfs == PNFS_TRY_AGAIN)
2226                 pnfs_read_resend_pnfs(hdr);
2227         if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status)
2228                 pnfs_read_through_mds(desc, hdr);
2229 }
2230
2231 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2232 {
2233         pnfs_put_lseg(hdr->lseg);
2234         nfs_pgio_header_free(hdr);
2235 }
2236
2237 int
2238 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2239 {
2240         struct nfs_pgio_header *hdr;
2241         int ret;
2242
2243         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2244         if (!hdr) {
2245                 desc->pg_error = -ENOMEM;
2246                 return desc->pg_error;
2247         }
2248         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2249         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2250         ret = nfs_generic_pgio(desc, hdr);
2251         if (!ret)
2252                 pnfs_do_read(desc, hdr);
2253         return ret;
2254 }
2255 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2256
2257 static void pnfs_clear_layoutcommitting(struct inode *inode)
2258 {
2259         unsigned long *bitlock = &NFS_I(inode)->flags;
2260
2261         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2262         smp_mb__after_atomic();
2263         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2264 }
2265
2266 /*
2267  * There can be multiple RW segments.
2268  */
2269 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2270 {
2271         struct pnfs_layout_segment *lseg;
2272
2273         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2274                 if (lseg->pls_range.iomode == IOMODE_RW &&
2275                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2276                         list_add(&lseg->pls_lc_list, listp);
2277         }
2278 }
2279
2280 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2281 {
2282         struct pnfs_layout_segment *lseg, *tmp;
2283
2284         /* Matched by references in pnfs_set_layoutcommit */
2285         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2286                 list_del_init(&lseg->pls_lc_list);
2287                 pnfs_put_lseg(lseg);
2288         }
2289
2290         pnfs_clear_layoutcommitting(inode);
2291 }
2292
2293 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2294 {
2295         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2296 }
2297 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2298
2299 void
2300 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2301                 loff_t end_pos)
2302 {
2303         struct nfs_inode *nfsi = NFS_I(inode);
2304         bool mark_as_dirty = false;
2305
2306         spin_lock(&inode->i_lock);
2307         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2308                 nfsi->layout->plh_lwb = end_pos;
2309                 mark_as_dirty = true;
2310                 dprintk("%s: Set layoutcommit for inode %lu ",
2311                         __func__, inode->i_ino);
2312         } else if (end_pos > nfsi->layout->plh_lwb)
2313                 nfsi->layout->plh_lwb = end_pos;
2314         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2315                 /* references matched in nfs4_layoutcommit_release */
2316                 pnfs_get_lseg(lseg);
2317         }
2318         spin_unlock(&inode->i_lock);
2319         dprintk("%s: lseg %p end_pos %llu\n",
2320                 __func__, lseg, nfsi->layout->plh_lwb);
2321
2322         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2323          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2324         if (mark_as_dirty)
2325                 mark_inode_dirty_sync(inode);
2326 }
2327 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2328
2329 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2330 {
2331         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2332
2333         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2334                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2335         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2336 }
2337
2338 /*
2339  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2340  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2341  * data to disk to allow the server to recover the data if it crashes.
2342  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2343  * is off, and a COMMIT is sent to a data server, or
2344  * if WRITEs to a data server return NFS_DATA_SYNC.
2345  */
2346 int
2347 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2348 {
2349         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2350         struct nfs4_layoutcommit_data *data;
2351         struct nfs_inode *nfsi = NFS_I(inode);
2352         loff_t end_pos;
2353         int status;
2354
2355         if (!pnfs_layoutcommit_outstanding(inode))
2356                 return 0;
2357
2358         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2359
2360         status = -EAGAIN;
2361         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2362                 if (!sync)
2363                         goto out;
2364                 status = wait_on_bit_lock_action(&nfsi->flags,
2365                                 NFS_INO_LAYOUTCOMMITTING,
2366                                 nfs_wait_bit_killable,
2367                                 TASK_KILLABLE);
2368                 if (status)
2369                         goto out;
2370         }
2371
2372         status = -ENOMEM;
2373         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2374         data = kzalloc(sizeof(*data), GFP_NOFS);
2375         if (!data)
2376                 goto clear_layoutcommitting;
2377
2378         status = 0;
2379         spin_lock(&inode->i_lock);
2380         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2381                 goto out_unlock;
2382
2383         INIT_LIST_HEAD(&data->lseg_list);
2384         pnfs_list_write_lseg(inode, &data->lseg_list);
2385
2386         end_pos = nfsi->layout->plh_lwb;
2387
2388         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2389         spin_unlock(&inode->i_lock);
2390
2391         data->args.inode = inode;
2392         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2393         nfs_fattr_init(&data->fattr);
2394         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2395         data->res.fattr = &data->fattr;
2396         if (end_pos != 0)
2397                 data->args.lastbytewritten = end_pos - 1;
2398         else
2399                 data->args.lastbytewritten = U64_MAX;
2400         data->res.server = NFS_SERVER(inode);
2401
2402         if (ld->prepare_layoutcommit) {
2403                 status = ld->prepare_layoutcommit(&data->args);
2404                 if (status) {
2405                         put_rpccred(data->cred);
2406                         spin_lock(&inode->i_lock);
2407                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2408                         if (end_pos > nfsi->layout->plh_lwb)
2409                                 nfsi->layout->plh_lwb = end_pos;
2410                         goto out_unlock;
2411                 }
2412         }
2413
2414
2415         status = nfs4_proc_layoutcommit(data, sync);
2416 out:
2417         if (status)
2418                 mark_inode_dirty_sync(inode);
2419         dprintk("<-- %s status %d\n", __func__, status);
2420         return status;
2421 out_unlock:
2422         spin_unlock(&inode->i_lock);
2423         kfree(data);
2424 clear_layoutcommitting:
2425         pnfs_clear_layoutcommitting(inode);
2426         goto out;
2427 }
2428 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2429
2430 int
2431 pnfs_generic_sync(struct inode *inode, bool datasync)
2432 {
2433         return pnfs_layoutcommit_inode(inode, true);
2434 }
2435 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2436
2437 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2438 {
2439         struct nfs4_threshold *thp;
2440
2441         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2442         if (!thp) {
2443                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2444                 return NULL;
2445         }
2446         return thp;
2447 }
2448
2449 #if IS_ENABLED(CONFIG_NFS_V4_2)
2450 int
2451 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2452 {
2453         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2454         struct nfs_server *server = NFS_SERVER(inode);
2455         struct nfs_inode *nfsi = NFS_I(inode);
2456         struct nfs42_layoutstat_data *data;
2457         struct pnfs_layout_hdr *hdr;
2458         int status = 0;
2459
2460         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2461                 goto out;
2462
2463         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2464                 goto out;
2465
2466         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2467                 goto out;
2468
2469         spin_lock(&inode->i_lock);
2470         if (!NFS_I(inode)->layout) {
2471                 spin_unlock(&inode->i_lock);
2472                 goto out_clear_layoutstats;
2473         }
2474         hdr = NFS_I(inode)->layout;
2475         pnfs_get_layout_hdr(hdr);
2476         spin_unlock(&inode->i_lock);
2477
2478         data = kzalloc(sizeof(*data), gfp_flags);
2479         if (!data) {
2480                 status = -ENOMEM;
2481                 goto out_put;
2482         }
2483
2484         data->args.fh = NFS_FH(inode);
2485         data->args.inode = inode;
2486         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2487         status = ld->prepare_layoutstats(&data->args);
2488         if (status)
2489                 goto out_free;
2490
2491         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2492
2493 out:
2494         dprintk("%s returns %d\n", __func__, status);
2495         return status;
2496
2497 out_free:
2498         kfree(data);
2499 out_put:
2500         pnfs_put_layout_hdr(hdr);
2501 out_clear_layoutstats:
2502         smp_mb__before_atomic();
2503         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2504         smp_mb__after_atomic();
2505         goto out;
2506 }
2507 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2508 #endif
2509
2510 unsigned int layoutstats_timer;
2511 module_param(layoutstats_timer, uint, 0644);
2512 EXPORT_SYMBOL_GPL(layoutstats_timer);