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 <[email protected]>
   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                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 366         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 367 }
 368
 369 void
 370 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 371 {
 372         struct pnfs_layout_hdr *lo;
 373         struct inode *inode;
 374
 375         if (!lseg)
 376                 return;
 377
 378         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 379                 atomic_read(&lseg->pls_refcount),
 380                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 381
 382         lo = lseg->pls_layout;
 383         inode = lo->plh_inode;
 384
 385         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 386                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 387                         spin_unlock(&inode->i_lock);
 388                         return;
 389                 }
 390                 pnfs_get_layout_hdr(lo);
 391                 pnfs_layout_remove_lseg(lo, lseg);
 392                 spin_unlock(&inode->i_lock);
 393                 pnfs_free_lseg(lseg);
 394                 pnfs_put_layout_hdr(lo);
 395         }
 396 }
 397 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 398
 399 static void pnfs_free_lseg_async_work(struct work_struct *work)
 400 {
 401         struct pnfs_layout_segment *lseg;
 402         struct pnfs_layout_hdr *lo;
 403
 404         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
 405         lo = lseg->pls_layout;
 406
 407         pnfs_free_lseg(lseg);
 408         pnfs_put_layout_hdr(lo);
 409 }
 410
 411 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
 412 {
 413         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
 414         schedule_work(&lseg->pls_work);
 415 }
 416
 417 void
 418 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
 419 {
 420         if (!lseg)
 421                 return;
 422
 423         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
 424
 425         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 426                 atomic_read(&lseg->pls_refcount),
 427                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 428         if (atomic_dec_and_test(&lseg->pls_refcount)) {
 429                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
 430                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 431                         return;
 432                 pnfs_get_layout_hdr(lo);
 433                 pnfs_layout_remove_lseg(lo, lseg);
 434                 pnfs_free_lseg_async(lseg);
 435         }
 436 }
 437 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
 438
 439 static u64
 440 end_offset(u64 start, u64 len)
 441 {
 442         u64 end;
 443
 444         end = start + len;
 445         return end >= start ? end : NFS4_MAX_UINT64;
 446 }
 447
 448 /*
 449  * is l2 fully contained in l1?
 450  *   start1                             end1
 451  *   [----------------------------------)
 452  *           start2           end2
 453  *           [----------------)
 454  */
 455 static bool
 456 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 457                  const struct pnfs_layout_range *l2)
 458 {
 459         u64 start1 = l1->offset;
 460         u64 end1 = end_offset(start1, l1->length);
 461         u64 start2 = l2->offset;
 462         u64 end2 = end_offset(start2, l2->length);
 463
 464         return (start1 <= start2) && (end1 >= end2);
 465 }
 466
 467 /*
 468  * is l1 and l2 intersecting?
 469  *   start1                             end1
 470  *   [----------------------------------)
 471  *                              start2           end2
 472  *                              [----------------)
 473  */
 474 static bool
 475 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 476                     const struct pnfs_layout_range *l2)
 477 {
 478         u64 start1 = l1->offset;
 479         u64 end1 = end_offset(start1, l1->length);
 480         u64 start2 = l2->offset;
 481         u64 end2 = end_offset(start2, l2->length);
 482
 483         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 484                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 485 }
 486
 487 static bool
 488 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 489                  const struct pnfs_layout_range *recall_range)
 490 {
 491         return (recall_range->iomode == IOMODE_ANY ||
 492                 lseg_range->iomode == recall_range->iomode) &&
 493                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 494 }
 495
 496 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 497                 struct list_head *tmp_list)
 498 {
 499         if (!atomic_dec_and_test(&lseg->pls_refcount))
 500                 return false;
 501         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 502         list_add(&lseg->pls_list, tmp_list);
 503         return true;
 504 }
 505
 506 /* Returns 1 if lseg is removed from list, 0 otherwise */
 507 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 508                              struct list_head *tmp_list)
 509 {
 510         int rv = 0;
 511
 512         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 513                 /* Remove the reference keeping the lseg in the
 514                  * list.  It will now be removed when all
 515                  * outstanding io is finished.
 516                  */
 517                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 518                         atomic_read(&lseg->pls_refcount));
 519                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 520                         rv = 1;
 521         }
 522         return rv;
 523 }
 524
 525 /*
 526  * Compare 2 layout stateid sequence ids, to see which is newer,
 527  * taking into account wraparound issues.
 528  */
 529 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 530 {
 531         return (s32)(s1 - s2) > 0;
 532 }
 533
 534 /**
 535  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 536  * @lo: layout header containing the lsegs
 537  * @tmp_list: list head where doomed lsegs should go
 538  * @recall_range: optional recall range argument to match (may be NULL)
 539  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 540  *
 541  * Walk the list of lsegs in the layout header, and tear down any that should
 542  * be destroyed. If "recall_range" is specified then the segment must match
 543  * that range. If "seq" is non-zero, then only match segments that were handed
 544  * out at or before that sequence.
 545  *
 546  * Returns number of matching invalid lsegs remaining in list after scanning
 547  * it and purging them.
 548  */
 549 int
 550 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 551                             struct list_head *tmp_list,
 552                             const struct pnfs_layout_range *recall_range,
 553                             u32 seq)
 554 {
 555         struct pnfs_layout_segment *lseg, *next;
 556         int remaining = 0;
 557
 558         dprintk("%s:Begin lo %p\n", __func__, lo);
 559
 560         if (list_empty(&lo->plh_segs))
 561                 return 0;
 562         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 563                 if (!recall_range ||
 564                     should_free_lseg(&lseg->pls_range, recall_range)) {
 565                         if (seq && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 566                                 continue;
 567                         dprintk("%s: freeing lseg %p iomode %d seq %u"
 568                                 "offset %llu length %llu\n", __func__,
 569                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 570                                 lseg->pls_range.offset, lseg->pls_range.length);
 571                         if (!mark_lseg_invalid(lseg, tmp_list))
 572                                 remaining++;
 573                 }
 574         dprintk("%s:Return %i\n", __func__, remaining);
 575         return remaining;
 576 }
 577
 578 /* note free_me must contain lsegs from a single layout_hdr */
 579 void
 580 pnfs_free_lseg_list(struct list_head *free_me)
 581 {
 582         struct pnfs_layout_segment *lseg, *tmp;
 583
 584         if (list_empty(free_me))
 585                 return;
 586
 587         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 588                 list_del(&lseg->pls_list);
 589                 pnfs_free_lseg(lseg);
 590         }
 591 }
 592
 593 void
 594 pnfs_destroy_layout(struct nfs_inode *nfsi)
 595 {
 596         struct pnfs_layout_hdr *lo;
 597         LIST_HEAD(tmp_list);
 598
 599         spin_lock(&nfsi->vfs_inode.i_lock);
 600         lo = nfsi->layout;
 601         if (lo) {
 602                 pnfs_get_layout_hdr(lo);
 603                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 604                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 605                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 606                 spin_unlock(&nfsi->vfs_inode.i_lock);
 607                 pnfs_free_lseg_list(&tmp_list);
 608                 pnfs_put_layout_hdr(lo);
 609         } else
 610                 spin_unlock(&nfsi->vfs_inode.i_lock);
 611 }
 612 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 613
 614 static bool
 615 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 616                 struct list_head *layout_list)
 617 {
 618         struct pnfs_layout_hdr *lo;
 619         bool ret = false;
 620
 621         spin_lock(&inode->i_lock);
 622         lo = NFS_I(inode)->layout;
 623         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 624                 pnfs_get_layout_hdr(lo);
 625                 list_add(&lo->plh_bulk_destroy, layout_list);
 626                 ret = true;
 627         }
 628         spin_unlock(&inode->i_lock);
 629         return ret;
 630 }
 631
 632 /* Caller must hold rcu_read_lock and clp->cl_lock */
 633 static int
 634 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 635                 struct nfs_server *server,
 636                 struct list_head *layout_list)
 637 {
 638         struct pnfs_layout_hdr *lo, *next;
 639         struct inode *inode;
 640
 641         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 642                 inode = igrab(lo->plh_inode);
 643                 if (inode == NULL)
 644                         continue;
 645                 list_del_init(&lo->plh_layouts);
 646                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 647                         continue;
 648                 rcu_read_unlock();
 649                 spin_unlock(&clp->cl_lock);
 650                 iput(inode);
 651                 spin_lock(&clp->cl_lock);
 652                 rcu_read_lock();
 653                 return -EAGAIN;
 654         }
 655         return 0;
 656 }
 657
 658 static int
 659 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 660                 bool is_bulk_recall)
 661 {
 662         struct pnfs_layout_hdr *lo;
 663         struct inode *inode;
 664         LIST_HEAD(lseg_list);
 665         int ret = 0;
 666
 667         while (!list_empty(layout_list)) {
 668                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 669                                 plh_bulk_destroy);
 670                 dprintk("%s freeing layout for inode %lu\n", __func__,
 671                         lo->plh_inode->i_ino);
 672                 inode = lo->plh_inode;
 673
 674                 pnfs_layoutcommit_inode(inode, false);
 675
 676                 spin_lock(&inode->i_lock);
 677                 list_del_init(&lo->plh_bulk_destroy);
 678                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 679                         if (is_bulk_recall)
 680                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 681                         ret = -EAGAIN;
 682                 }
 683                 spin_unlock(&inode->i_lock);
 684                 pnfs_free_lseg_list(&lseg_list);
 685                 /* Free all lsegs that are attached to commit buckets */
 686                 nfs_commit_inode(inode, 0);
 687                 pnfs_put_layout_hdr(lo);
 688                 iput(inode);
 689         }
 690         return ret;
 691 }
 692
 693 int
 694 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 695                 struct nfs_fsid *fsid,
 696                 bool is_recall)
 697 {
 698         struct nfs_server *server;
 699         LIST_HEAD(layout_list);
 700
 701         spin_lock(&clp->cl_lock);
 702         rcu_read_lock();
 703 restart:
 704         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 705                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 706                         continue;
 707                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 708                                 server,
 709                                 &layout_list) != 0)
 710                         goto restart;
 711         }
 712         rcu_read_unlock();
 713         spin_unlock(&clp->cl_lock);
 714
 715         if (list_empty(&layout_list))
 716                 return 0;
 717         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 718 }
 719
 720 int
 721 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 722                 bool is_recall)
 723 {
 724         struct nfs_server *server;
 725         LIST_HEAD(layout_list);
 726
 727         spin_lock(&clp->cl_lock);
 728         rcu_read_lock();
 729 restart:
 730         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 731                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 732                                         server,
 733                                         &layout_list) != 0)
 734                         goto restart;
 735         }
 736         rcu_read_unlock();
 737         spin_unlock(&clp->cl_lock);
 738
 739         if (list_empty(&layout_list))
 740                 return 0;
 741         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 742 }
 743
 744 /*
 745  * Called by the state manger to remove all layouts established under an
 746  * expired lease.
 747  */
 748 void
 749 pnfs_destroy_all_layouts(struct nfs_client *clp)
 750 {
 751         nfs4_deviceid_mark_client_invalid(clp);
 752         nfs4_deviceid_purge_client(clp);
 753
 754         pnfs_destroy_layouts_byclid(clp, false);
 755 }
 756
 757 /* update lo->plh_stateid with new if is more recent */
 758 void
 759 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 760                         bool update_barrier)
 761 {
 762         u32 oldseq, newseq, new_barrier;
 763         int empty = list_empty(&lo->plh_segs);
 764
 765         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 766         newseq = be32_to_cpu(new->seqid);
 767         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 768                 nfs4_stateid_copy(&lo->plh_stateid, new);
 769                 if (update_barrier) {
 770                         new_barrier = be32_to_cpu(new->seqid);
 771                 } else {
 772                         /* Because of wraparound, we want to keep the barrier
 773                          * "close" to the current seqids.
 774                          */
 775                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 776                 }
 777                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 778                         lo->plh_barrier = new_barrier;
 779         }
 780 }
 781
 782 static bool
 783 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 784                 const nfs4_stateid *stateid)
 785 {
 786         u32 seqid = be32_to_cpu(stateid->seqid);
 787
 788         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 789 }
 790
 791 /* lget is set to 1 if called from inside send_layoutget call chain */
 792 static bool
 793 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 794 {
 795         return lo->plh_block_lgets ||
 796                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 797 }
 798
 799 /*
 800  * Get layout from server.
 801  *    for now, assume that whole file layouts are requested.
 802  *    arg->offset: 0
 803  *    arg->length: all ones
 804  */
 805 static struct pnfs_layout_segment *
 806 send_layoutget(struct pnfs_layout_hdr *lo,
 807            struct nfs_open_context *ctx,
 808            nfs4_stateid *stateid,
 809            const struct pnfs_layout_range *range,
 810            long *timeout, gfp_t gfp_flags)
 811 {
 812         struct inode *ino = lo->plh_inode;
 813         struct nfs_server *server = NFS_SERVER(ino);
 814         struct nfs4_layoutget *lgp;
 815         loff_t i_size;
 816
 817         dprintk("--> %s\n", __func__);
 818
 819         /*
 820          * Synchronously retrieve layout information from server and
 821          * store in lseg. If we race with a concurrent seqid morphing
 822          * op, then re-send the LAYOUTGET.
 823          */
 824         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 825         if (lgp == NULL)
 826                 return ERR_PTR(-ENOMEM);
 827
 828         i_size = i_size_read(ino);
 829
 830         lgp->args.minlength = PAGE_SIZE;
 831         if (lgp->args.minlength > range->length)
 832                 lgp->args.minlength = range->length;
 833         if (range->iomode == IOMODE_READ) {
 834                 if (range->offset >= i_size)
 835                         lgp->args.minlength = 0;
 836                 else if (i_size - range->offset < lgp->args.minlength)
 837                         lgp->args.minlength = i_size - range->offset;
 838         }
 839         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 840         pnfs_copy_range(&lgp->args.range, range);
 841         lgp->args.type = server->pnfs_curr_ld->id;
 842         lgp->args.inode = ino;
 843         lgp->args.ctx = get_nfs_open_context(ctx);
 844         nfs4_stateid_copy(&lgp->args.stateid, stateid);
 845         lgp->gfp_flags = gfp_flags;
 846         lgp->cred = lo->plh_lc_cred;
 847
 848         return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 849 }
 850
 851 static void pnfs_clear_layoutcommit(struct inode *inode,
 852                 struct list_head *head)
 853 {
 854         struct nfs_inode *nfsi = NFS_I(inode);
 855         struct pnfs_layout_segment *lseg, *tmp;
 856
 857         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 858                 return;
 859         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 860                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 861                         continue;
 862                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 863         }
 864 }
 865
 866 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 867 {
 868         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 869         smp_mb__after_atomic();
 870         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 871         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 872 }
 873
 874 static bool
 875 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 876 {
 877         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 878                 return false;
 879         lo->plh_return_iomode = 0;
 880         lo->plh_return_seq = 0;
 881         pnfs_get_layout_hdr(lo);
 882         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 883         return true;
 884 }
 885
 886 static int
 887 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
 888                        enum pnfs_iomode iomode, bool sync)
 889 {
 890         struct inode *ino = lo->plh_inode;
 891         struct nfs4_layoutreturn *lrp;
 892         int status = 0;
 893
 894         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
 895         if (unlikely(lrp == NULL)) {
 896                 status = -ENOMEM;
 897                 spin_lock(&ino->i_lock);
 898                 pnfs_clear_layoutreturn_waitbit(lo);
 899                 spin_unlock(&ino->i_lock);
 900                 pnfs_put_layout_hdr(lo);
 901                 goto out;
 902         }
 903
 904         nfs4_stateid_copy(&lrp->args.stateid, stateid);
 905         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 906         lrp->args.inode = ino;
 907         lrp->args.range.iomode = iomode;
 908         lrp->args.range.offset = 0;
 909         lrp->args.range.length = NFS4_MAX_UINT64;
 910         lrp->args.layout = lo;
 911         lrp->clp = NFS_SERVER(ino)->nfs_client;
 912         lrp->cred = lo->plh_lc_cred;
 913
 914         status = nfs4_proc_layoutreturn(lrp, sync);
 915 out:
 916         dprintk("<-- %s status: %d\n", __func__, status);
 917         return status;
 918 }
 919
 920 /* Return true if layoutreturn is needed */
 921 static bool
 922 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
 923 {
 924         struct pnfs_layout_segment *s;
 925
 926         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 927                 return false;
 928
 929         /* Defer layoutreturn until all lsegs are done */
 930         list_for_each_entry(s, &lo->plh_segs, pls_list) {
 931                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
 932                         return false;
 933         }
 934
 935         return true;
 936 }
 937
 938 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
 939 {
 940         struct inode *inode= lo->plh_inode;
 941
 942         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
 943                 return;
 944         spin_lock(&inode->i_lock);
 945         if (pnfs_layout_need_return(lo)) {
 946                 nfs4_stateid stateid;
 947                 enum pnfs_iomode iomode;
 948                 bool send;
 949
 950                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
 951                 stateid.seqid = cpu_to_be32(lo->plh_return_seq);
 952                 iomode = lo->plh_return_iomode;
 953                 send = pnfs_prepare_layoutreturn(lo);
 954                 spin_unlock(&inode->i_lock);
 955                 if (send) {
 956                         /* Send an async layoutreturn so we dont deadlock */
 957                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
 958                 }
 959         } else
 960                 spin_unlock(&inode->i_lock);
 961 }
 962
 963 /*
 964  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 965  * when the layout segment list is empty.
 966  *
 967  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 968  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 969  * deviceid is marked invalid.
 970  */
 971 int
 972 _pnfs_return_layout(struct inode *ino)
 973 {
 974         struct pnfs_layout_hdr *lo = NULL;
 975         struct nfs_inode *nfsi = NFS_I(ino);
 976         LIST_HEAD(tmp_list);
 977         nfs4_stateid stateid;
 978         int status = 0, empty;
 979         bool send;
 980
 981         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 982
 983         spin_lock(&ino->i_lock);
 984         lo = nfsi->layout;
 985         if (!lo) {
 986                 spin_unlock(&ino->i_lock);
 987                 dprintk("NFS: %s no layout to return\n", __func__);
 988                 goto out;
 989         }
 990         nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
 991         /* Reference matched in nfs4_layoutreturn_release */
 992         pnfs_get_layout_hdr(lo);
 993         empty = list_empty(&lo->plh_segs);
 994         pnfs_clear_layoutcommit(ino, &tmp_list);
 995         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
 996
 997         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
 998                 struct pnfs_layout_range range = {
 999                         .iomode         = IOMODE_ANY,
1000                         .offset         = 0,
1001                         .length         = NFS4_MAX_UINT64,
1002                 };
1003                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1004         }
1005
1006         /* Don't send a LAYOUTRETURN if list was initially empty */
1007         if (empty) {
1008                 spin_unlock(&ino->i_lock);
1009                 dprintk("NFS: %s no layout segments to return\n", __func__);
1010                 goto out_put_layout_hdr;
1011         }
1012
1013         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1014         send = pnfs_prepare_layoutreturn(lo);
1015         spin_unlock(&ino->i_lock);
1016         pnfs_free_lseg_list(&tmp_list);
1017         if (send)
1018                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1019 out_put_layout_hdr:
1020         pnfs_put_layout_hdr(lo);
1021 out:
1022         dprintk("<-- %s status: %d\n", __func__, status);
1023         return status;
1024 }
1025 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1026
1027 int
1028 pnfs_commit_and_return_layout(struct inode *inode)
1029 {
1030         struct pnfs_layout_hdr *lo;
1031         int ret;
1032
1033         spin_lock(&inode->i_lock);
1034         lo = NFS_I(inode)->layout;
1035         if (lo == NULL) {
1036                 spin_unlock(&inode->i_lock);
1037                 return 0;
1038         }
1039         pnfs_get_layout_hdr(lo);
1040         /* Block new layoutgets and read/write to ds */
1041         lo->plh_block_lgets++;
1042         spin_unlock(&inode->i_lock);
1043         filemap_fdatawait(inode->i_mapping);
1044         ret = pnfs_layoutcommit_inode(inode, true);
1045         if (ret == 0)
1046                 ret = _pnfs_return_layout(inode);
1047         spin_lock(&inode->i_lock);
1048         lo->plh_block_lgets--;
1049         spin_unlock(&inode->i_lock);
1050         pnfs_put_layout_hdr(lo);
1051         return ret;
1052 }
1053
1054 bool pnfs_roc(struct inode *ino)
1055 {
1056         struct nfs_inode *nfsi = NFS_I(ino);
1057         struct nfs_open_context *ctx;
1058         struct nfs4_state *state;
1059         struct pnfs_layout_hdr *lo;
1060         struct pnfs_layout_segment *lseg, *tmp;
1061         nfs4_stateid stateid;
1062         LIST_HEAD(tmp_list);
1063         bool found = false, layoutreturn = false, roc = false;
1064
1065         spin_lock(&ino->i_lock);
1066         lo = nfsi->layout;
1067         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1068                 goto out_noroc;
1069
1070         /* no roc if we hold a delegation */
1071         if (nfs4_check_delegation(ino, FMODE_READ))
1072                 goto out_noroc;
1073
1074         list_for_each_entry(ctx, &nfsi->open_files, list) {
1075                 state = ctx->state;
1076                 /* Don't return layout if there is open file state */
1077                 if (state != NULL && state->state != 0)
1078                         goto out_noroc;
1079         }
1080
1081         nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1082         /* always send layoutreturn if being marked so */
1083         if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
1084                                    &lo->plh_flags))
1085                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1086
1087         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1088                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1089                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1090                         mark_lseg_invalid(lseg, &tmp_list);
1091                         found = true;
1092                 }
1093         /* ROC in two conditions:
1094          * 1. there are ROC lsegs
1095          * 2. we don't send layoutreturn
1096          */
1097         if (found && !layoutreturn) {
1098                 /* lo ref dropped in pnfs_roc_release() */
1099                 pnfs_get_layout_hdr(lo);
1100                 roc = true;
1101         }
1102
1103 out_noroc:
1104         spin_unlock(&ino->i_lock);
1105         pnfs_free_lseg_list(&tmp_list);
1106         pnfs_layoutcommit_inode(ino, true);
1107         if (layoutreturn)
1108                 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1109         return roc;
1110 }
1111
1112 void pnfs_roc_release(struct inode *ino)
1113 {
1114         struct pnfs_layout_hdr *lo;
1115
1116         spin_lock(&ino->i_lock);
1117         lo = NFS_I(ino)->layout;
1118         pnfs_clear_layoutreturn_waitbit(lo);
1119         if (atomic_dec_and_test(&lo->plh_refcount)) {
1120                 pnfs_detach_layout_hdr(lo);
1121                 spin_unlock(&ino->i_lock);
1122                 pnfs_free_layout_hdr(lo);
1123         } else
1124                 spin_unlock(&ino->i_lock);
1125 }
1126
1127 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1128 {
1129         struct pnfs_layout_hdr *lo;
1130
1131         spin_lock(&ino->i_lock);
1132         lo = NFS_I(ino)->layout;
1133         pnfs_mark_layout_returned_if_empty(lo);
1134         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1135                 lo->plh_barrier = barrier;
1136         spin_unlock(&ino->i_lock);
1137         trace_nfs4_layoutreturn_on_close(ino, 0);
1138 }
1139
1140 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1141 {
1142         struct nfs_inode *nfsi = NFS_I(ino);
1143         struct pnfs_layout_hdr *lo;
1144         u32 current_seqid;
1145
1146         spin_lock(&ino->i_lock);
1147         lo = nfsi->layout;
1148         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1149
1150         /* Since close does not return a layout stateid for use as
1151          * a barrier, we choose the worst-case barrier.
1152          */
1153         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1154         spin_unlock(&ino->i_lock);
1155 }
1156
1157 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1158 {
1159         struct nfs_inode *nfsi = NFS_I(ino);
1160         struct pnfs_layout_hdr *lo;
1161         bool sleep = false;
1162
1163         /* we might not have grabbed lo reference. so need to check under
1164          * i_lock */
1165         spin_lock(&ino->i_lock);
1166         lo = nfsi->layout;
1167         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1168                 sleep = true;
1169         spin_unlock(&ino->i_lock);
1170
1171         if (sleep)
1172                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1173
1174         return sleep;
1175 }
1176
1177 /*
1178  * Compare two layout segments for sorting into layout cache.
1179  * We want to preferentially return RW over RO layouts, so ensure those
1180  * are seen first.
1181  */
1182 static s64
1183 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1184            const struct pnfs_layout_range *l2)
1185 {
1186         s64 d;
1187
1188         /* high offset > low offset */
1189         d = l1->offset - l2->offset;
1190         if (d)
1191                 return d;
1192
1193         /* short length > long length */
1194         d = l2->length - l1->length;
1195         if (d)
1196                 return d;
1197
1198         /* read > read/write */
1199         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1200 }
1201
1202 static bool
1203 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1204                 const struct pnfs_layout_range *l2)
1205 {
1206         return pnfs_lseg_range_cmp(l1, l2) > 0;
1207 }
1208
1209 static bool
1210 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1211                 struct pnfs_layout_segment *old)
1212 {
1213         return false;
1214 }
1215
1216 void
1217 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1218                    struct pnfs_layout_segment *lseg,
1219                    bool (*is_after)(const struct pnfs_layout_range *,
1220                            const struct pnfs_layout_range *),
1221                    bool (*do_merge)(struct pnfs_layout_segment *,
1222                            struct pnfs_layout_segment *),
1223                    struct list_head *free_me)
1224 {
1225         struct pnfs_layout_segment *lp, *tmp;
1226
1227         dprintk("%s:Begin\n", __func__);
1228
1229         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1230                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1231                         continue;
1232                 if (do_merge(lseg, lp)) {
1233                         mark_lseg_invalid(lp, free_me);
1234                         continue;
1235                 }
1236                 if (is_after(&lseg->pls_range, &lp->pls_range))
1237                         continue;
1238                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1239                 dprintk("%s: inserted lseg %p "
1240                         "iomode %d offset %llu length %llu before "
1241                         "lp %p iomode %d offset %llu length %llu\n",
1242                         __func__, lseg, lseg->pls_range.iomode,
1243                         lseg->pls_range.offset, lseg->pls_range.length,
1244                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1245                         lp->pls_range.length);
1246                 goto out;
1247         }
1248         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1249         dprintk("%s: inserted lseg %p "
1250                 "iomode %d offset %llu length %llu at tail\n",
1251                 __func__, lseg, lseg->pls_range.iomode,
1252                 lseg->pls_range.offset, lseg->pls_range.length);
1253 out:
1254         pnfs_get_layout_hdr(lo);
1255
1256         dprintk("%s:Return\n", __func__);
1257 }
1258 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1259
1260 static void
1261 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1262                    struct pnfs_layout_segment *lseg,
1263                    struct list_head *free_me)
1264 {
1265         struct inode *inode = lo->plh_inode;
1266         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1267
1268         if (ld->add_lseg != NULL)
1269                 ld->add_lseg(lo, lseg, free_me);
1270         else
1271                 pnfs_generic_layout_insert_lseg(lo, lseg,
1272                                 pnfs_lseg_range_is_after,
1273                                 pnfs_lseg_no_merge,
1274                                 free_me);
1275 }
1276
1277 static struct pnfs_layout_hdr *
1278 alloc_init_layout_hdr(struct inode *ino,
1279                       struct nfs_open_context *ctx,
1280                       gfp_t gfp_flags)
1281 {
1282         struct pnfs_layout_hdr *lo;
1283
1284         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1285         if (!lo)
1286                 return NULL;
1287         atomic_set(&lo->plh_refcount, 1);
1288         INIT_LIST_HEAD(&lo->plh_layouts);
1289         INIT_LIST_HEAD(&lo->plh_segs);
1290         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1291         lo->plh_inode = ino;
1292         lo->plh_lc_cred = get_rpccred(ctx->cred);
1293         return lo;
1294 }
1295
1296 static struct pnfs_layout_hdr *
1297 pnfs_find_alloc_layout(struct inode *ino,
1298                        struct nfs_open_context *ctx,
1299                        gfp_t gfp_flags)
1300 {
1301         struct nfs_inode *nfsi = NFS_I(ino);
1302         struct pnfs_layout_hdr *new = NULL;
1303
1304         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1305
1306         if (nfsi->layout != NULL)
1307                 goto out_existing;
1308         spin_unlock(&ino->i_lock);
1309         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1310         spin_lock(&ino->i_lock);
1311
1312         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1313                 nfsi->layout = new;
1314                 return new;
1315         } else if (new != NULL)
1316                 pnfs_free_layout_hdr(new);
1317 out_existing:
1318         pnfs_get_layout_hdr(nfsi->layout);
1319         return nfsi->layout;
1320 }
1321
1322 /*
1323  * iomode matching rules:
1324  * iomode       lseg    strict match
1325  *                      iomode
1326  * -----        -----   ------ -----
1327  * ANY          READ    N/A    true
1328  * ANY          RW      N/A    true
1329  * RW           READ    N/A    false
1330  * RW           RW      N/A    true
1331  * READ         READ    N/A    true
1332  * READ         RW      true   false
1333  * READ         RW      false  true
1334  */
1335 static bool
1336 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1337                  const struct pnfs_layout_range *range,
1338                  bool strict_iomode)
1339 {
1340         struct pnfs_layout_range range1;
1341
1342         if ((range->iomode == IOMODE_RW &&
1343              ls_range->iomode != IOMODE_RW) ||
1344             (range->iomode != ls_range->iomode &&
1345              strict_iomode == true) ||
1346             !pnfs_lseg_range_intersecting(ls_range, range))
1347                 return 0;
1348
1349         /* range1 covers only the first byte in the range */
1350         range1 = *range;
1351         range1.length = 1;
1352         return pnfs_lseg_range_contained(ls_range, &range1);
1353 }
1354
1355 /*
1356  * lookup range in layout
1357  */
1358 static struct pnfs_layout_segment *
1359 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1360                 struct pnfs_layout_range *range,
1361                 bool strict_iomode)
1362 {
1363         struct pnfs_layout_segment *lseg, *ret = NULL;
1364
1365         dprintk("%s:Begin\n", __func__);
1366
1367         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1368                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1369                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1370                     pnfs_lseg_range_match(&lseg->pls_range, range,
1371                                           strict_iomode)) {
1372                         ret = pnfs_get_lseg(lseg);
1373                         break;
1374                 }
1375         }
1376
1377         dprintk("%s:Return lseg %p ref %d\n",
1378                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1379         return ret;
1380 }
1381
1382 /*
1383  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1384  * to the MDS or over pNFS
1385  *
1386  * The nfs_inode read_io and write_io fields are cumulative counters reset
1387  * when there are no layout segments. Note that in pnfs_update_layout iomode
1388  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1389  * WRITE request.
1390  *
1391  * A return of true means use MDS I/O.
1392  *
1393  * From rfc 5661:
1394  * If a file's size is smaller than the file size threshold, data accesses
1395  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1396  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1397  * server.  If both file size and I/O size are provided, the client SHOULD
1398  * reach or exceed  both thresholds before sending its read or write
1399  * requests to the data server.
1400  */
1401 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1402                                      struct inode *ino, int iomode)
1403 {
1404         struct nfs4_threshold *t = ctx->mdsthreshold;
1405         struct nfs_inode *nfsi = NFS_I(ino);
1406         loff_t fsize = i_size_read(ino);
1407         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1408
1409         if (t == NULL)
1410                 return ret;
1411
1412         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1413                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1414
1415         switch (iomode) {
1416         case IOMODE_READ:
1417                 if (t->bm & THRESHOLD_RD) {
1418                         dprintk("%s fsize %llu\n", __func__, fsize);
1419                         size_set = true;
1420                         if (fsize < t->rd_sz)
1421                                 size = true;
1422                 }
1423                 if (t->bm & THRESHOLD_RD_IO) {
1424                         dprintk("%s nfsi->read_io %llu\n", __func__,
1425                                 nfsi->read_io);
1426                         io_set = true;
1427                         if (nfsi->read_io < t->rd_io_sz)
1428                                 io = true;
1429                 }
1430                 break;
1431         case IOMODE_RW:
1432                 if (t->bm & THRESHOLD_WR) {
1433                         dprintk("%s fsize %llu\n", __func__, fsize);
1434                         size_set = true;
1435                         if (fsize < t->wr_sz)
1436                                 size = true;
1437                 }
1438                 if (t->bm & THRESHOLD_WR_IO) {
1439                         dprintk("%s nfsi->write_io %llu\n", __func__,
1440                                 nfsi->write_io);
1441                         io_set = true;
1442                         if (nfsi->write_io < t->wr_io_sz)
1443                                 io = true;
1444                 }
1445                 break;
1446         }
1447         if (size_set && io_set) {
1448                 if (size && io)
1449                         ret = true;
1450         } else if (size || io)
1451                 ret = true;
1452
1453         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1454         return ret;
1455 }
1456
1457 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1458 {
1459         /*
1460          * send layoutcommit as it can hold up layoutreturn due to lseg
1461          * reference
1462          */
1463         pnfs_layoutcommit_inode(lo->plh_inode, false);
1464         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1465                                    nfs_wait_bit_killable,
1466                                    TASK_UNINTERRUPTIBLE);
1467 }
1468
1469 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1470 {
1471         unsigned long *bitlock = &lo->plh_flags;
1472
1473         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1474         smp_mb__after_atomic();
1475         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1476 }
1477
1478 /*
1479  * Layout segment is retreived from the server if not cached.
1480  * The appropriate layout segment is referenced and returned to the caller.
1481  */
1482 struct pnfs_layout_segment *
1483 pnfs_update_layout(struct inode *ino,
1484                    struct nfs_open_context *ctx,
1485                    loff_t pos,
1486                    u64 count,
1487                    enum pnfs_iomode iomode,
1488                    bool strict_iomode,
1489                    gfp_t gfp_flags)
1490 {
1491         struct pnfs_layout_range arg = {
1492                 .iomode = iomode,
1493                 .offset = pos,
1494                 .length = count,
1495         };
1496         unsigned pg_offset, seq;
1497         struct nfs_server *server = NFS_SERVER(ino);
1498         struct nfs_client *clp = server->nfs_client;
1499         struct pnfs_layout_hdr *lo = NULL;
1500         struct pnfs_layout_segment *lseg = NULL;
1501         nfs4_stateid stateid;
1502         long timeout = 0;
1503         unsigned long giveup = jiffies + rpc_get_timeout(server->client);
1504         bool first;
1505
1506         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1507                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1508                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1509                 goto out;
1510         }
1511
1512         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1513                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1514                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1515                 goto out;
1516         }
1517
1518         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1519                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1520                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1521                 goto out;
1522         }
1523
1524 lookup_again:
1525         first = false;
1526         spin_lock(&ino->i_lock);
1527         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1528         if (lo == NULL) {
1529                 spin_unlock(&ino->i_lock);
1530                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1531                                  PNFS_UPDATE_LAYOUT_NOMEM);
1532                 goto out;
1533         }
1534
1535         /* Do we even need to bother with this? */
1536         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1537                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1538                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1539                 dprintk("%s matches recall, use MDS\n", __func__);
1540                 goto out_unlock;
1541         }
1542
1543         /* if LAYOUTGET already failed once we don't try again */
1544         if (pnfs_layout_io_test_failed(lo, iomode)) {
1545                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1546                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1547                 goto out_unlock;
1548         }
1549
1550         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1551         if (lseg) {
1552                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1553                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1554                 goto out_unlock;
1555         }
1556
1557         if (!nfs4_valid_open_stateid(ctx->state)) {
1558                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1559                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1560                 goto out_unlock;
1561         }
1562
1563         /*
1564          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1565          * stateid, or it has been invalidated, then we must use the open
1566          * stateid.
1567          */
1568         if (lo->plh_stateid.seqid == 0 ||
1569             test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1570
1571                 /*
1572                  * The first layoutget for the file. Need to serialize per
1573                  * RFC 5661 Errata 3208.
1574                  */
1575                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1576                                      &lo->plh_flags)) {
1577                         spin_unlock(&ino->i_lock);
1578                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1579                                     TASK_UNINTERRUPTIBLE);
1580                         pnfs_put_layout_hdr(lo);
1581                         dprintk("%s retrying\n", __func__);
1582                         goto lookup_again;
1583                 }
1584
1585                 first = true;
1586                 do {
1587                         seq = read_seqbegin(&ctx->state->seqlock);
1588                         nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1589                 } while (read_seqretry(&ctx->state->seqlock, seq));
1590         } else {
1591                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1592         }
1593
1594         /*
1595          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1596          * for LAYOUTRETURN even if first is true.
1597          */
1598         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1599                 spin_unlock(&ino->i_lock);
1600                 dprintk("%s wait for layoutreturn\n", __func__);
1601                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1602                         if (first)
1603                                 pnfs_clear_first_layoutget(lo);
1604                         pnfs_put_layout_hdr(lo);
1605                         dprintk("%s retrying\n", __func__);
1606                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1607                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1608                         goto lookup_again;
1609                 }
1610                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1611                                 PNFS_UPDATE_LAYOUT_RETURN);
1612                 goto out_put_layout_hdr;
1613         }
1614
1615         if (pnfs_layoutgets_blocked(lo)) {
1616                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1617                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1618                 goto out_unlock;
1619         }
1620         atomic_inc(&lo->plh_outstanding);
1621         spin_unlock(&ino->i_lock);
1622
1623         if (list_empty(&lo->plh_layouts)) {
1624                 /* The lo must be on the clp list if there is any
1625                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1626                  */
1627                 spin_lock(&clp->cl_lock);
1628                 if (list_empty(&lo->plh_layouts))
1629                         list_add_tail(&lo->plh_layouts, &server->layouts);
1630                 spin_unlock(&clp->cl_lock);
1631         }
1632
1633         pg_offset = arg.offset & ~PAGE_MASK;
1634         if (pg_offset) {
1635                 arg.offset -= pg_offset;
1636                 arg.length += pg_offset;
1637         }
1638         if (arg.length != NFS4_MAX_UINT64)
1639                 arg.length = PAGE_ALIGN(arg.length);
1640
1641         lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1642         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1643                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1644         if (IS_ERR(lseg)) {
1645                 switch(PTR_ERR(lseg)) {
1646                 case -ERECALLCONFLICT:
1647                         if (time_after(jiffies, giveup))
1648                                 lseg = NULL;
1649                         /* Fallthrough */
1650                 case -EAGAIN:
1651                         pnfs_put_layout_hdr(lo);
1652                         if (first)
1653                                 pnfs_clear_first_layoutget(lo);
1654                         if (lseg) {
1655                                 trace_pnfs_update_layout(ino, pos, count,
1656                                         iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1657                                 goto lookup_again;
1658                         }
1659                         /* Fallthrough */
1660                 default:
1661                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1662                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1663                                 lseg = NULL;
1664                         }
1665                 }
1666         } else {
1667                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1668         }
1669
1670         atomic_dec(&lo->plh_outstanding);
1671 out_put_layout_hdr:
1672         if (first)
1673                 pnfs_clear_first_layoutget(lo);
1674         pnfs_put_layout_hdr(lo);
1675 out:
1676         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1677                         "(%s, offset: %llu, length: %llu)\n",
1678                         __func__, ino->i_sb->s_id,
1679                         (unsigned long long)NFS_FILEID(ino),
1680                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1681                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1682                         (unsigned long long)pos,
1683                         (unsigned long long)count);
1684         return lseg;
1685 out_unlock:
1686         spin_unlock(&ino->i_lock);
1687         goto out_put_layout_hdr;
1688 }
1689 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1690
1691 static bool
1692 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1693 {
1694         switch (range->iomode) {
1695         case IOMODE_READ:
1696         case IOMODE_RW:
1697                 break;
1698         default:
1699                 return false;
1700         }
1701         if (range->offset == NFS4_MAX_UINT64)
1702                 return false;
1703         if (range->length == 0)
1704                 return false;
1705         if (range->length != NFS4_MAX_UINT64 &&
1706             range->length > NFS4_MAX_UINT64 - range->offset)
1707                 return false;
1708         return true;
1709 }
1710
1711 struct pnfs_layout_segment *
1712 pnfs_layout_process(struct nfs4_layoutget *lgp)
1713 {
1714         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1715         struct nfs4_layoutget_res *res = &lgp->res;
1716         struct pnfs_layout_segment *lseg;
1717         struct inode *ino = lo->plh_inode;
1718         LIST_HEAD(free_me);
1719
1720         if (!pnfs_sanity_check_layout_range(&res->range))
1721                 return ERR_PTR(-EINVAL);
1722
1723         /* Inject layout blob into I/O device driver */
1724         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1725         if (IS_ERR_OR_NULL(lseg)) {
1726                 if (!lseg)
1727                         lseg = ERR_PTR(-ENOMEM);
1728
1729                 dprintk("%s: Could not allocate layout: error %ld\n",
1730                        __func__, PTR_ERR(lseg));
1731                 return lseg;
1732         }
1733
1734         init_lseg(lo, lseg);
1735         lseg->pls_range = res->range;
1736         lseg->pls_seq = be32_to_cpu(res->stateid.seqid);
1737
1738         spin_lock(&ino->i_lock);
1739         if (pnfs_layoutgets_blocked(lo)) {
1740                 dprintk("%s forget reply due to state\n", __func__);
1741                 goto out_forget;
1742         }
1743
1744         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1745                 /* existing state ID, make sure the sequence number matches. */
1746                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1747                         dprintk("%s forget reply due to sequence\n", __func__);
1748                         goto out_forget;
1749                 }
1750                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1751         } else {
1752                 /*
1753                  * We got an entirely new state ID.  Mark all segments for the
1754                  * inode invalid, and don't bother validating the stateid
1755                  * sequence number.
1756                  */
1757                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL, 0);
1758
1759                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1760                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1761         }
1762
1763         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1764
1765         pnfs_get_lseg(lseg);
1766         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1767
1768         if (res->return_on_close)
1769                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1770
1771         spin_unlock(&ino->i_lock);
1772         pnfs_free_lseg_list(&free_me);
1773         return lseg;
1774
1775 out_forget:
1776         spin_unlock(&ino->i_lock);
1777         lseg->pls_layout = lo;
1778         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1779         return ERR_PTR(-EAGAIN);
1780 }
1781
1782 static void
1783 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
1784                          u32 seq)
1785 {
1786         if (lo->plh_return_iomode == iomode)
1787                 return;
1788         if (lo->plh_return_iomode != 0)
1789                 iomode = IOMODE_ANY;
1790         lo->plh_return_iomode = iomode;
1791         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
1792         if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
1793                 lo->plh_return_seq = seq;
1794 }
1795
1796 /**
1797  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1798  * @lo: pointer to layout header
1799  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1800  * @return_range: describe layout segment ranges to be returned
1801  *
1802  * This function is mainly intended for use by layoutrecall. It attempts
1803  * to free the layout segment immediately, or else to mark it for return
1804  * as soon as its reference count drops to zero.
1805  */
1806 int
1807 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1808                                 struct list_head *tmp_list,
1809                                 const struct pnfs_layout_range *return_range,
1810                                 u32 seq)
1811 {
1812         struct pnfs_layout_segment *lseg, *next;
1813         int remaining = 0;
1814
1815         dprintk("%s:Begin lo %p\n", __func__, lo);
1816
1817         if (list_empty(&lo->plh_segs))
1818                 return 0;
1819
1820         assert_spin_locked(&lo->plh_inode->i_lock);
1821
1822         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1823                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1824                         dprintk("%s: marking lseg %p iomode %d "
1825                                 "offset %llu length %llu\n", __func__,
1826                                 lseg, lseg->pls_range.iomode,
1827                                 lseg->pls_range.offset,
1828                                 lseg->pls_range.length);
1829                         if (mark_lseg_invalid(lseg, tmp_list))
1830                                 continue;
1831                         remaining++;
1832                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1833                 }
1834
1835         if (remaining)
1836                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
1837
1838         return remaining;
1839 }
1840
1841 void pnfs_error_mark_layout_for_return(struct inode *inode,
1842                                        struct pnfs_layout_segment *lseg)
1843 {
1844         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1845         struct pnfs_layout_range range = {
1846                 .iomode = lseg->pls_range.iomode,
1847                 .offset = 0,
1848                 .length = NFS4_MAX_UINT64,
1849         };
1850         LIST_HEAD(free_me);
1851         bool return_now = false;
1852
1853         spin_lock(&inode->i_lock);
1854         pnfs_set_plh_return_info(lo, range.iomode, lseg->pls_seq);
1855         /*
1856          * mark all matching lsegs so that we are sure to have no live
1857          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1858          * for how it works.
1859          */
1860         if (!pnfs_mark_matching_lsegs_return(lo, &free_me,
1861                                                 &range, lseg->pls_seq)) {
1862                 nfs4_stateid stateid;
1863                 enum pnfs_iomode iomode = lo->plh_return_iomode;
1864
1865                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1866                 return_now = pnfs_prepare_layoutreturn(lo);
1867                 spin_unlock(&inode->i_lock);
1868                 if (return_now)
1869                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1870         } else {
1871                 spin_unlock(&inode->i_lock);
1872                 nfs_commit_inode(inode, 0);
1873         }
1874         pnfs_free_lseg_list(&free_me);
1875 }
1876 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1877
1878 void
1879 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1880 {
1881         u64 rd_size = req->wb_bytes;
1882
1883         if (pgio->pg_lseg == NULL) {
1884                 if (pgio->pg_dreq == NULL)
1885                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1886                 else
1887                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1888
1889                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1890                                                    req->wb_context,
1891                                                    req_offset(req),
1892                                                    rd_size,
1893                                                    IOMODE_READ,
1894                                                    false,
1895                                                    GFP_KERNEL);
1896                 if (IS_ERR(pgio->pg_lseg)) {
1897                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1898                         pgio->pg_lseg = NULL;
1899                         return;
1900                 }
1901         }
1902         /* If no lseg, fall back to read through mds */
1903         if (pgio->pg_lseg == NULL)
1904                 nfs_pageio_reset_read_mds(pgio);
1905
1906 }
1907 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1908
1909 void
1910 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1911                            struct nfs_page *req, u64 wb_size)
1912 {
1913         if (pgio->pg_lseg == NULL) {
1914                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1915                                                    req->wb_context,
1916                                                    req_offset(req),
1917                                                    wb_size,
1918                                                    IOMODE_RW,
1919                                                    false,
1920                                                    GFP_NOFS);
1921                 if (IS_ERR(pgio->pg_lseg)) {
1922                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1923                         pgio->pg_lseg = NULL;
1924                         return;
1925                 }
1926         }
1927         /* If no lseg, fall back to write through mds */
1928         if (pgio->pg_lseg == NULL)
1929                 nfs_pageio_reset_write_mds(pgio);
1930 }
1931 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1932
1933 void
1934 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1935 {
1936         if (desc->pg_lseg) {
1937                 pnfs_put_lseg(desc->pg_lseg);
1938                 desc->pg_lseg = NULL;
1939         }
1940 }
1941 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1942
1943 /*
1944  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1945  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1946  */
1947 size_t
1948 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1949                      struct nfs_page *prev, struct nfs_page *req)
1950 {
1951         unsigned int size;
1952         u64 seg_end, req_start, seg_left;
1953
1954         size = nfs_generic_pg_test(pgio, prev, req);
1955         if (!size)
1956                 return 0;
1957
1958         /*
1959          * 'size' contains the number of bytes left in the current page (up
1960          * to the original size asked for in @req->wb_bytes).
1961          *
1962          * Calculate how many bytes are left in the layout segment
1963          * and if there are less bytes than 'size', return that instead.
1964          *
1965          * Please also note that 'end_offset' is actually the offset of the
1966          * first byte that lies outside the pnfs_layout_range. FIXME?
1967          *
1968          */
1969         if (pgio->pg_lseg) {
1970                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1971                                      pgio->pg_lseg->pls_range.length);
1972                 req_start = req_offset(req);
1973                 WARN_ON_ONCE(req_start >= seg_end);
1974                 /* start of request is past the last byte of this segment */
1975                 if (req_start >= seg_end) {
1976                         /* reference the new lseg */
1977                         if (pgio->pg_ops->pg_cleanup)
1978                                 pgio->pg_ops->pg_cleanup(pgio);
1979                         if (pgio->pg_ops->pg_init)
1980                                 pgio->pg_ops->pg_init(pgio, req);
1981                         return 0;
1982                 }
1983
1984                 /* adjust 'size' iff there are fewer bytes left in the
1985                  * segment than what nfs_generic_pg_test returned */
1986                 seg_left = seg_end - req_start;
1987                 if (seg_left < size)
1988                         size = (unsigned int)seg_left;
1989         }
1990
1991         return size;
1992 }
1993 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1994
1995 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1996 {
1997         struct nfs_pageio_descriptor pgio;
1998
1999         /* Resend all requests through the MDS */
2000         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2001                               hdr->completion_ops);
2002         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2003         return nfs_pageio_resend(&pgio, hdr);
2004 }
2005 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2006
2007 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2008 {
2009
2010         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2011         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2012             PNFS_LAYOUTRET_ON_ERROR) {
2013                 pnfs_return_layout(hdr->inode);
2014         }
2015         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2016                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2017 }
2018
2019 /*
2020  * Called by non rpc-based layout drivers
2021  */
2022 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2023 {
2024         if (likely(!hdr->pnfs_error)) {
2025                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2026                                 hdr->mds_offset + hdr->res.count);
2027                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2028         }
2029         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2030         if (unlikely(hdr->pnfs_error))
2031                 pnfs_ld_handle_write_error(hdr);
2032         hdr->mds_ops->rpc_release(hdr);
2033 }
2034 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2035
2036 static void
2037 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2038                 struct nfs_pgio_header *hdr)
2039 {
2040         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2041
2042         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2043                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2044                 nfs_pageio_reset_write_mds(desc);
2045                 mirror->pg_recoalesce = 1;
2046         }
2047         nfs_pgio_data_destroy(hdr);
2048         hdr->release(hdr);
2049 }
2050
2051 static enum pnfs_try_status
2052 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2053                         const struct rpc_call_ops *call_ops,
2054                         struct pnfs_layout_segment *lseg,
2055                         int how)
2056 {
2057         struct inode *inode = hdr->inode;
2058         enum pnfs_try_status trypnfs;
2059         struct nfs_server *nfss = NFS_SERVER(inode);
2060
2061         hdr->mds_ops = call_ops;
2062
2063         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2064                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2065         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2066         if (trypnfs != PNFS_NOT_ATTEMPTED)
2067                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2068         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2069         return trypnfs;
2070 }
2071
2072 static void
2073 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2074               struct nfs_pgio_header *hdr, int how)
2075 {
2076         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2077         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2078         enum pnfs_try_status trypnfs;
2079
2080         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2081         if (trypnfs == PNFS_NOT_ATTEMPTED)
2082                 pnfs_write_through_mds(desc, hdr);
2083 }
2084
2085 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2086 {
2087         pnfs_put_lseg(hdr->lseg);
2088         nfs_pgio_header_free(hdr);
2089 }
2090
2091 int
2092 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2093 {
2094         struct nfs_pgio_header *hdr;
2095         int ret;
2096
2097         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2098         if (!hdr) {
2099                 desc->pg_error = -ENOMEM;
2100                 return desc->pg_error;
2101         }
2102         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2103
2104         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2105         ret = nfs_generic_pgio(desc, hdr);
2106         if (!ret)
2107                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2108
2109         return ret;
2110 }
2111 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2112
2113 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2114 {
2115         struct nfs_pageio_descriptor pgio;
2116
2117         /* Resend all requests through the MDS */
2118         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2119         return nfs_pageio_resend(&pgio, hdr);
2120 }
2121 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2122
2123 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2124 {
2125         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2126         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2127             PNFS_LAYOUTRET_ON_ERROR) {
2128                 pnfs_return_layout(hdr->inode);
2129         }
2130         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2131                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2132 }
2133
2134 /*
2135  * Called by non rpc-based layout drivers
2136  */
2137 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2138 {
2139         if (likely(!hdr->pnfs_error)) {
2140                 __nfs4_read_done_cb(hdr);
2141                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2142         }
2143         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2144         if (unlikely(hdr->pnfs_error))
2145                 pnfs_ld_handle_read_error(hdr);
2146         hdr->mds_ops->rpc_release(hdr);
2147 }
2148 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2149
2150 static void
2151 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2152                 struct nfs_pgio_header *hdr)
2153 {
2154         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2155
2156         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2157                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2158                 nfs_pageio_reset_read_mds(desc);
2159                 mirror->pg_recoalesce = 1;
2160         }
2161         nfs_pgio_data_destroy(hdr);
2162         hdr->release(hdr);
2163 }
2164
2165 /*
2166  * Call the appropriate parallel I/O subsystem read function.
2167  */
2168 static enum pnfs_try_status
2169 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2170                        const struct rpc_call_ops *call_ops,
2171                        struct pnfs_layout_segment *lseg)
2172 {
2173         struct inode *inode = hdr->inode;
2174         struct nfs_server *nfss = NFS_SERVER(inode);
2175         enum pnfs_try_status trypnfs;
2176
2177         hdr->mds_ops = call_ops;
2178
2179         dprintk("%s: Reading ino:%lu %u@%llu\n",
2180                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2181
2182         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2183         if (trypnfs != PNFS_NOT_ATTEMPTED)
2184                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2185         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2186         return trypnfs;
2187 }
2188
2189 /* Resend all requests through pnfs. */
2190 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2191 {
2192         struct nfs_pageio_descriptor pgio;
2193
2194         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2195                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2196                                         hdr->completion_ops);
2197                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2198         }
2199 }
2200 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2201
2202 static void
2203 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2204 {
2205         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2206         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2207         enum pnfs_try_status trypnfs;
2208
2209         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2210         if (trypnfs == PNFS_TRY_AGAIN)
2211                 pnfs_read_resend_pnfs(hdr);
2212         if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status)
2213                 pnfs_read_through_mds(desc, hdr);
2214 }
2215
2216 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2217 {
2218         pnfs_put_lseg(hdr->lseg);
2219         nfs_pgio_header_free(hdr);
2220 }
2221
2222 int
2223 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2224 {
2225         struct nfs_pgio_header *hdr;
2226         int ret;
2227
2228         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2229         if (!hdr) {
2230                 desc->pg_error = -ENOMEM;
2231                 return desc->pg_error;
2232         }
2233         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2234         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2235         ret = nfs_generic_pgio(desc, hdr);
2236         if (!ret)
2237                 pnfs_do_read(desc, hdr);
2238         return ret;
2239 }
2240 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2241
2242 static void pnfs_clear_layoutcommitting(struct inode *inode)
2243 {
2244         unsigned long *bitlock = &NFS_I(inode)->flags;
2245
2246         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2247         smp_mb__after_atomic();
2248         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2249 }
2250
2251 /*
2252  * There can be multiple RW segments.
2253  */
2254 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2255 {
2256         struct pnfs_layout_segment *lseg;
2257
2258         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2259                 if (lseg->pls_range.iomode == IOMODE_RW &&
2260                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2261                         list_add(&lseg->pls_lc_list, listp);
2262         }
2263 }
2264
2265 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2266 {
2267         struct pnfs_layout_segment *lseg, *tmp;
2268
2269         /* Matched by references in pnfs_set_layoutcommit */
2270         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2271                 list_del_init(&lseg->pls_lc_list);
2272                 pnfs_put_lseg(lseg);
2273         }
2274
2275         pnfs_clear_layoutcommitting(inode);
2276 }
2277
2278 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2279 {
2280         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2281 }
2282 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2283
2284 void
2285 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2286                 loff_t end_pos)
2287 {
2288         struct nfs_inode *nfsi = NFS_I(inode);
2289         bool mark_as_dirty = false;
2290
2291         spin_lock(&inode->i_lock);
2292         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2293                 nfsi->layout->plh_lwb = end_pos;
2294                 mark_as_dirty = true;
2295                 dprintk("%s: Set layoutcommit for inode %lu ",
2296                         __func__, inode->i_ino);
2297         } else if (end_pos > nfsi->layout->plh_lwb)
2298                 nfsi->layout->plh_lwb = end_pos;
2299         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2300                 /* references matched in nfs4_layoutcommit_release */
2301                 pnfs_get_lseg(lseg);
2302         }
2303         spin_unlock(&inode->i_lock);
2304         dprintk("%s: lseg %p end_pos %llu\n",
2305                 __func__, lseg, nfsi->layout->plh_lwb);
2306
2307         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2308          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2309         if (mark_as_dirty)
2310                 mark_inode_dirty_sync(inode);
2311 }
2312 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2313
2314 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2315 {
2316         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2317
2318         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2319                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2320         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2321 }
2322
2323 /*
2324  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2325  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2326  * data to disk to allow the server to recover the data if it crashes.
2327  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2328  * is off, and a COMMIT is sent to a data server, or
2329  * if WRITEs to a data server return NFS_DATA_SYNC.
2330  */
2331 int
2332 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2333 {
2334         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2335         struct nfs4_layoutcommit_data *data;
2336         struct nfs_inode *nfsi = NFS_I(inode);
2337         loff_t end_pos;
2338         int status;
2339
2340         if (!pnfs_layoutcommit_outstanding(inode))
2341                 return 0;
2342
2343         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2344
2345         status = -EAGAIN;
2346         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2347                 if (!sync)
2348                         goto out;
2349                 status = wait_on_bit_lock_action(&nfsi->flags,
2350                                 NFS_INO_LAYOUTCOMMITTING,
2351                                 nfs_wait_bit_killable,
2352                                 TASK_KILLABLE);
2353                 if (status)
2354                         goto out;
2355         }
2356
2357         status = -ENOMEM;
2358         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2359         data = kzalloc(sizeof(*data), GFP_NOFS);
2360         if (!data)
2361                 goto clear_layoutcommitting;
2362
2363         status = 0;
2364         spin_lock(&inode->i_lock);
2365         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2366                 goto out_unlock;
2367
2368         INIT_LIST_HEAD(&data->lseg_list);
2369         pnfs_list_write_lseg(inode, &data->lseg_list);
2370
2371         end_pos = nfsi->layout->plh_lwb;
2372
2373         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2374         spin_unlock(&inode->i_lock);
2375
2376         data->args.inode = inode;
2377         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2378         nfs_fattr_init(&data->fattr);
2379         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2380         data->res.fattr = &data->fattr;
2381         data->args.lastbytewritten = end_pos - 1;
2382         data->res.server = NFS_SERVER(inode);
2383
2384         if (ld->prepare_layoutcommit) {
2385                 status = ld->prepare_layoutcommit(&data->args);
2386                 if (status) {
2387                         put_rpccred(data->cred);
2388                         spin_lock(&inode->i_lock);
2389                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2390                         if (end_pos > nfsi->layout->plh_lwb)
2391                                 nfsi->layout->plh_lwb = end_pos;
2392                         goto out_unlock;
2393                 }
2394         }
2395
2396
2397         status = nfs4_proc_layoutcommit(data, sync);
2398 out:
2399         if (status)
2400                 mark_inode_dirty_sync(inode);
2401         dprintk("<-- %s status %d\n", __func__, status);
2402         return status;
2403 out_unlock:
2404         spin_unlock(&inode->i_lock);
2405         kfree(data);
2406 clear_layoutcommitting:
2407         pnfs_clear_layoutcommitting(inode);
2408         goto out;
2409 }
2410 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2411
2412 int
2413 pnfs_generic_sync(struct inode *inode, bool datasync)
2414 {
2415         return pnfs_layoutcommit_inode(inode, true);
2416 }
2417 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2418
2419 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2420 {
2421         struct nfs4_threshold *thp;
2422
2423         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2424         if (!thp) {
2425                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2426                 return NULL;
2427         }
2428         return thp;
2429 }
2430
2431 #if IS_ENABLED(CONFIG_NFS_V4_2)
2432 int
2433 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2434 {
2435         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2436         struct nfs_server *server = NFS_SERVER(inode);
2437         struct nfs_inode *nfsi = NFS_I(inode);
2438         struct nfs42_layoutstat_data *data;
2439         struct pnfs_layout_hdr *hdr;
2440         int status = 0;
2441
2442         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2443                 goto out;
2444
2445         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2446                 goto out;
2447
2448         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2449                 goto out;
2450
2451         spin_lock(&inode->i_lock);
2452         if (!NFS_I(inode)->layout) {
2453                 spin_unlock(&inode->i_lock);
2454                 goto out_clear_layoutstats;
2455         }
2456         hdr = NFS_I(inode)->layout;
2457         pnfs_get_layout_hdr(hdr);
2458         spin_unlock(&inode->i_lock);
2459
2460         data = kzalloc(sizeof(*data), gfp_flags);
2461         if (!data) {
2462                 status = -ENOMEM;
2463                 goto out_put;
2464         }
2465
2466         data->args.fh = NFS_FH(inode);
2467         data->args.inode = inode;
2468         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2469         status = ld->prepare_layoutstats(&data->args);
2470         if (status)
2471                 goto out_free;
2472
2473         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2474
2475 out:
2476         dprintk("%s returns %d\n", __func__, status);
2477         return status;
2478
2479 out_free:
2480         kfree(data);
2481 out_put:
2482         pnfs_put_layout_hdr(hdr);
2483 out_clear_layoutstats:
2484         smp_mb__before_atomic();
2485         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2486         smp_mb__after_atomic();
2487         goto out;
2488 }
2489 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2490 #endif
2491
2492 unsigned int layoutstats_timer;
2493 module_param(layoutstats_timer, uint, 0644);
2494 EXPORT_SYMBOL_GPL(layoutstats_timer);