fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36
  37 #define NFSDBG_FACILITY         NFSDBG_PNFS
  38 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  39
  40 /* Locking:
  41  *
  42  * pnfs_spinlock:
  43  *      protects pnfs_modules_tbl.
  44  */
  45 static DEFINE_SPINLOCK(pnfs_spinlock);
  46
  47 /*
  48  * pnfs_modules_tbl holds all pnfs modules
  49  */
  50 static LIST_HEAD(pnfs_modules_tbl);
  51
  52 /* Return the registered pnfs layout driver module matching given id */
  53 static struct pnfs_layoutdriver_type *
  54 find_pnfs_driver_locked(u32 id)
  55 {
  56         struct pnfs_layoutdriver_type *local;
  57
  58         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  59                 if (local->id == id)
  60                         goto out;
  61         local = NULL;
  62 out:
  63         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  64         return local;
  65 }
  66
  67 static struct pnfs_layoutdriver_type *
  68 find_pnfs_driver(u32 id)
  69 {
  70         struct pnfs_layoutdriver_type *local;
  71
  72         spin_lock(&pnfs_spinlock);
  73         local = find_pnfs_driver_locked(id);
  74         if (local != NULL && !try_module_get(local->owner)) {
  75                 dprintk("%s: Could not grab reference on module\n", __func__);
  76                 local = NULL;
  77         }
  78         spin_unlock(&pnfs_spinlock);
  79         return local;
  80 }
  81
  82 void
  83 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  84 {
  85         if (nfss->pnfs_curr_ld) {
  86                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  87                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  88                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  89                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  90                         nfs4_deviceid_purge_client(nfss->nfs_client);
  91                 module_put(nfss->pnfs_curr_ld->owner);
  92         }
  93         nfss->pnfs_curr_ld = NULL;
  94 }
  95
  96 /*
  97  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  98  * Currently only one pNFS layout driver per filesystem is supported.
  99  *
 100  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 101  */
 102 void
 103 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 104                       u32 id)
 105 {
 106         struct pnfs_layoutdriver_type *ld_type = NULL;
 107
 108         if (id == 0)
 109                 goto out_no_driver;
 110         if (!(server->nfs_client->cl_exchange_flags &
 111                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 112                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 113                         __func__, id, server->nfs_client->cl_exchange_flags);
 114                 goto out_no_driver;
 115         }
 116         ld_type = find_pnfs_driver(id);
 117         if (!ld_type) {
 118                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 119                 ld_type = find_pnfs_driver(id);
 120                 if (!ld_type) {
 121                         dprintk("%s: No pNFS module found for %u.\n",
 122                                 __func__, id);
 123                         goto out_no_driver;
 124                 }
 125         }
 126         server->pnfs_curr_ld = ld_type;
 127         if (ld_type->set_layoutdriver
 128             && ld_type->set_layoutdriver(server, mntfh)) {
 129                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 130                         "driver %u.\n", __func__, id);
 131                 module_put(ld_type->owner);
 132                 goto out_no_driver;
 133         }
 134         /* Bump the MDS count */
 135         atomic_inc(&server->nfs_client->cl_mds_count);
 136
 137         dprintk("%s: pNFS module for %u set\n", __func__, id);
 138         return;
 139
 140 out_no_driver:
 141         dprintk("%s: Using NFSv4 I/O\n", __func__);
 142         server->pnfs_curr_ld = NULL;
 143 }
 144
 145 int
 146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 147 {
 148         int status = -EINVAL;
 149         struct pnfs_layoutdriver_type *tmp;
 150
 151         if (ld_type->id == 0) {
 152                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 153                 return status;
 154         }
 155         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 156                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 157                        "alloc_lseg and free_lseg.\n", __func__);
 158                 return status;
 159         }
 160
 161         spin_lock(&pnfs_spinlock);
 162         tmp = find_pnfs_driver_locked(ld_type->id);
 163         if (!tmp) {
 164                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 165                 status = 0;
 166                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 167                         ld_type->name);
 168         } else {
 169                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 170                         __func__, ld_type->id);
 171         }
 172         spin_unlock(&pnfs_spinlock);
 173
 174         return status;
 175 }
 176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 177
 178 void
 179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 180 {
 181         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 182         spin_lock(&pnfs_spinlock);
 183         list_del(&ld_type->pnfs_tblid);
 184         spin_unlock(&pnfs_spinlock);
 185 }
 186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 187
 188 /*
 189  * pNFS client layout cache
 190  */
 191
 192 /* Need to hold i_lock if caller does not already hold reference */
 193 void
 194 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 195 {
 196         atomic_inc(&lo->plh_refcount);
 197 }
 198
 199 static struct pnfs_layout_hdr *
 200 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 201 {
 202         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 203         return ld->alloc_layout_hdr(ino, gfp_flags);
 204 }
 205
 206 static void
 207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 208 {
 209         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 210         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 211
 212         if (!list_empty(&lo->plh_layouts)) {
 213                 struct nfs_client *clp = server->nfs_client;
 214
 215                 spin_lock(&clp->cl_lock);
 216                 list_del_init(&lo->plh_layouts);
 217                 spin_unlock(&clp->cl_lock);
 218         }
 219         put_rpccred(lo->plh_lc_cred);
 220         return ld->free_layout_hdr(lo);
 221 }
 222
 223 static void
 224 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 225 {
 226         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 227         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 228         nfsi->layout = NULL;
 229         /* Reset MDS Threshold I/O counters */
 230         nfsi->write_io = 0;
 231         nfsi->read_io = 0;
 232 }
 233
 234 void
 235 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 236 {
 237         struct inode *inode = lo->plh_inode;
 238
 239         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 240                 pnfs_detach_layout_hdr(lo);
 241                 spin_unlock(&inode->i_lock);
 242                 pnfs_free_layout_hdr(lo);
 243         }
 244 }
 245
 246 static int
 247 pnfs_iomode_to_fail_bit(u32 iomode)
 248 {
 249         return iomode == IOMODE_RW ?
 250                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 251 }
 252
 253 static void
 254 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 255 {
 256         lo->plh_retry_timestamp = jiffies;
 257         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 258                 atomic_inc(&lo->plh_refcount);
 259 }
 260
 261 static void
 262 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 263 {
 264         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 265                 atomic_dec(&lo->plh_refcount);
 266 }
 267
 268 static void
 269 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 270 {
 271         struct inode *inode = lo->plh_inode;
 272         struct pnfs_layout_range range = {
 273                 .iomode = iomode,
 274                 .offset = 0,
 275                 .length = NFS4_MAX_UINT64,
 276         };
 277         LIST_HEAD(head);
 278
 279         spin_lock(&inode->i_lock);
 280         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 281         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 282         spin_unlock(&inode->i_lock);
 283         pnfs_free_lseg_list(&head);
 284         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 285                         iomode == IOMODE_RW ?  "RW" : "READ");
 286 }
 287
 288 static bool
 289 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 290 {
 291         unsigned long start, end;
 292         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 293
 294         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 295                 return false;
 296         end = jiffies;
 297         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 298         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 299                 /* It is time to retry the failed layoutgets */
 300                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 301                 return false;
 302         }
 303         return true;
 304 }
 305
 306 static void
 307 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 308 {
 309         INIT_LIST_HEAD(&lseg->pls_list);
 310         INIT_LIST_HEAD(&lseg->pls_lc_list);
 311         atomic_set(&lseg->pls_refcount, 1);
 312         smp_mb();
 313         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 314         lseg->pls_layout = lo;
 315 }
 316
 317 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 318 {
 319         struct inode *ino = lseg->pls_layout->plh_inode;
 320
 321         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 322 }
 323
 324 static void
 325 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 326                 struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *inode = lo->plh_inode;
 329
 330         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 331         list_del_init(&lseg->pls_list);
 332         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 333         atomic_dec(&lo->plh_refcount);
 334         if (list_empty(&lo->plh_segs))
 335                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 336         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 337 }
 338
 339 void
 340 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 341 {
 342         struct pnfs_layout_hdr *lo;
 343         struct inode *inode;
 344
 345         if (!lseg)
 346                 return;
 347
 348         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 349                 atomic_read(&lseg->pls_refcount),
 350                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 351         lo = lseg->pls_layout;
 352         inode = lo->plh_inode;
 353         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 354                 pnfs_get_layout_hdr(lo);
 355                 pnfs_layout_remove_lseg(lo, lseg);
 356                 spin_unlock(&inode->i_lock);
 357                 pnfs_free_lseg(lseg);
 358                 pnfs_put_layout_hdr(lo);
 359         }
 360 }
 361 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 362
 363 static inline u64
 364 end_offset(u64 start, u64 len)
 365 {
 366         u64 end;
 367
 368         end = start + len;
 369         return end >= start ? end : NFS4_MAX_UINT64;
 370 }
 371
 372 /*
 373  * is l2 fully contained in l1?
 374  *   start1                             end1
 375  *   [----------------------------------)
 376  *           start2           end2
 377  *           [----------------)
 378  */
 379 static inline int
 380 lo_seg_contained(struct pnfs_layout_range *l1,
 381                  struct pnfs_layout_range *l2)
 382 {
 383         u64 start1 = l1->offset;
 384         u64 end1 = end_offset(start1, l1->length);
 385         u64 start2 = l2->offset;
 386         u64 end2 = end_offset(start2, l2->length);
 387
 388         return (start1 <= start2) && (end1 >= end2);
 389 }
 390
 391 /*
 392  * is l1 and l2 intersecting?
 393  *   start1                             end1
 394  *   [----------------------------------)
 395  *                              start2           end2
 396  *                              [----------------)
 397  */
 398 static inline int
 399 lo_seg_intersecting(struct pnfs_layout_range *l1,
 400                     struct pnfs_layout_range *l2)
 401 {
 402         u64 start1 = l1->offset;
 403         u64 end1 = end_offset(start1, l1->length);
 404         u64 start2 = l2->offset;
 405         u64 end2 = end_offset(start2, l2->length);
 406
 407         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 408                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 409 }
 410
 411 static bool
 412 should_free_lseg(struct pnfs_layout_range *lseg_range,
 413                  struct pnfs_layout_range *recall_range)
 414 {
 415         return (recall_range->iomode == IOMODE_ANY ||
 416                 lseg_range->iomode == recall_range->iomode) &&
 417                lo_seg_intersecting(lseg_range, recall_range);
 418 }
 419
 420 /* Returns 1 if lseg is removed from list, 0 otherwise */
 421 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 422                              struct list_head *tmp_list)
 423 {
 424         int rv = 0;
 425
 426         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 427                 /* Remove the reference keeping the lseg in the
 428                  * list.  It will now be removed when all
 429                  * outstanding io is finished.
 430                  */
 431                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 432                         atomic_read(&lseg->pls_refcount));
 433                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 434                         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 435                         list_add(&lseg->pls_list, tmp_list);
 436                         rv = 1;
 437                 }
 438         }
 439         return rv;
 440 }
 441
 442 /* Returns count of number of matching invalid lsegs remaining in list
 443  * after call.
 444  */
 445 int
 446 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 447                             struct list_head *tmp_list,
 448                             struct pnfs_layout_range *recall_range)
 449 {
 450         struct pnfs_layout_segment *lseg, *next;
 451         int invalid = 0, removed = 0;
 452
 453         dprintk("%s:Begin lo %p\n", __func__, lo);
 454
 455         if (list_empty(&lo->plh_segs))
 456                 return 0;
 457         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 458                 if (!recall_range ||
 459                     should_free_lseg(&lseg->pls_range, recall_range)) {
 460                         dprintk("%s: freeing lseg %p iomode %d "
 461                                 "offset %llu length %llu\n", __func__,
 462                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 463                                 lseg->pls_range.length);
 464                         invalid++;
 465                         removed += mark_lseg_invalid(lseg, tmp_list);
 466                 }
 467         dprintk("%s:Return %i\n", __func__, invalid - removed);
 468         return invalid - removed;
 469 }
 470
 471 /* note free_me must contain lsegs from a single layout_hdr */
 472 void
 473 pnfs_free_lseg_list(struct list_head *free_me)
 474 {
 475         struct pnfs_layout_segment *lseg, *tmp;
 476
 477         if (list_empty(free_me))
 478                 return;
 479
 480         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 481                 list_del(&lseg->pls_list);
 482                 pnfs_free_lseg(lseg);
 483         }
 484 }
 485
 486 void
 487 pnfs_destroy_layout(struct nfs_inode *nfsi)
 488 {
 489         struct pnfs_layout_hdr *lo;
 490         LIST_HEAD(tmp_list);
 491
 492         spin_lock(&nfsi->vfs_inode.i_lock);
 493         lo = nfsi->layout;
 494         if (lo) {
 495                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 496                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 497                 pnfs_get_layout_hdr(lo);
 498                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 499                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 500                 spin_unlock(&nfsi->vfs_inode.i_lock);
 501                 pnfs_free_lseg_list(&tmp_list);
 502                 pnfs_put_layout_hdr(lo);
 503         } else
 504                 spin_unlock(&nfsi->vfs_inode.i_lock);
 505 }
 506 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 507
 508 /*
 509  * Called by the state manger to remove all layouts established under an
 510  * expired lease.
 511  */
 512 void
 513 pnfs_destroy_all_layouts(struct nfs_client *clp)
 514 {
 515         struct nfs_server *server;
 516         struct pnfs_layout_hdr *lo;
 517         LIST_HEAD(tmp_list);
 518
 519         nfs4_deviceid_mark_client_invalid(clp);
 520         nfs4_deviceid_purge_client(clp);
 521
 522         spin_lock(&clp->cl_lock);
 523         rcu_read_lock();
 524         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 525                 if (!list_empty(&server->layouts))
 526                         list_splice_init(&server->layouts, &tmp_list);
 527         }
 528         rcu_read_unlock();
 529         spin_unlock(&clp->cl_lock);
 530
 531         while (!list_empty(&tmp_list)) {
 532                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 533                                 plh_layouts);
 534                 dprintk("%s freeing layout for inode %lu\n", __func__,
 535                         lo->plh_inode->i_ino);
 536                 list_del_init(&lo->plh_layouts);
 537                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 538         }
 539 }
 540
 541 /*
 542  * Compare 2 layout stateid sequence ids, to see which is newer,
 543  * taking into account wraparound issues.
 544  */
 545 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 546 {
 547         return (s32)s1 - (s32)s2 > 0;
 548 }
 549
 550 /* update lo->plh_stateid with new if is more recent */
 551 void
 552 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 553                         bool update_barrier)
 554 {
 555         u32 oldseq, newseq, new_barrier;
 556         int empty = list_empty(&lo->plh_segs);
 557
 558         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 559         newseq = be32_to_cpu(new->seqid);
 560         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 561                 nfs4_stateid_copy(&lo->plh_stateid, new);
 562                 if (update_barrier) {
 563                         new_barrier = be32_to_cpu(new->seqid);
 564                 } else {
 565                         /* Because of wraparound, we want to keep the barrier
 566                          * "close" to the current seqids.
 567                          */
 568                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 569                 }
 570                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 571                         lo->plh_barrier = new_barrier;
 572         }
 573 }
 574
 575 static bool
 576 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 577                 const nfs4_stateid *stateid)
 578 {
 579         u32 seqid = be32_to_cpu(stateid->seqid);
 580
 581         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 582 }
 583
 584 /* lget is set to 1 if called from inside send_layoutget call chain */
 585 static bool
 586 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 587 {
 588         return lo->plh_block_lgets ||
 589                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 590                 (list_empty(&lo->plh_segs) &&
 591                  (atomic_read(&lo->plh_outstanding) > lget));
 592 }
 593
 594 int
 595 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 596                               struct nfs4_state *open_state)
 597 {
 598         int status = 0;
 599
 600         dprintk("--> %s\n", __func__);
 601         spin_lock(&lo->plh_inode->i_lock);
 602         if (pnfs_layoutgets_blocked(lo, 1)) {
 603                 status = -EAGAIN;
 604         } else if (list_empty(&lo->plh_segs)) {
 605                 int seq;
 606
 607                 do {
 608                         seq = read_seqbegin(&open_state->seqlock);
 609                         nfs4_stateid_copy(dst, &open_state->stateid);
 610                 } while (read_seqretry(&open_state->seqlock, seq));
 611         } else
 612                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 613         spin_unlock(&lo->plh_inode->i_lock);
 614         dprintk("<-- %s\n", __func__);
 615         return status;
 616 }
 617
 618 /*
 619 * Get layout from server.
 620 *    for now, assume that whole file layouts are requested.
 621 *    arg->offset: 0
 622 *    arg->length: all ones
 623 */
 624 static struct pnfs_layout_segment *
 625 send_layoutget(struct pnfs_layout_hdr *lo,
 626            struct nfs_open_context *ctx,
 627            struct pnfs_layout_range *range,
 628            gfp_t gfp_flags)
 629 {
 630         struct inode *ino = lo->plh_inode;
 631         struct nfs_server *server = NFS_SERVER(ino);
 632         struct nfs4_layoutget *lgp;
 633         struct pnfs_layout_segment *lseg;
 634
 635         dprintk("--> %s\n", __func__);
 636
 637         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 638         if (lgp == NULL)
 639                 return NULL;
 640
 641         lgp->args.minlength = PAGE_CACHE_SIZE;
 642         if (lgp->args.minlength > range->length)
 643                 lgp->args.minlength = range->length;
 644         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 645         lgp->args.range = *range;
 646         lgp->args.type = server->pnfs_curr_ld->id;
 647         lgp->args.inode = ino;
 648         lgp->args.ctx = get_nfs_open_context(ctx);
 649         lgp->gfp_flags = gfp_flags;
 650
 651         /* Synchronously retrieve layout information from server and
 652          * store in lseg.
 653          */
 654         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 655         if (IS_ERR(lseg)) {
 656                 switch (PTR_ERR(lseg)) {
 657                 case -ENOMEM:
 658                 case -ERESTARTSYS:
 659                         break;
 660                 default:
 661                         /* remember that LAYOUTGET failed and suspend trying */
 662                         pnfs_layout_io_set_failed(lo, range->iomode);
 663                 }
 664                 return NULL;
 665         }
 666
 667         return lseg;
 668 }
 669
 670 /*
 671  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 672  * when the layout segment list is empty.
 673  *
 674  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 675  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 676  * deviceid is marked invalid.
 677  */
 678 int
 679 _pnfs_return_layout(struct inode *ino)
 680 {
 681         struct pnfs_layout_hdr *lo = NULL;
 682         struct nfs_inode *nfsi = NFS_I(ino);
 683         LIST_HEAD(tmp_list);
 684         struct nfs4_layoutreturn *lrp;
 685         nfs4_stateid stateid;
 686         int status = 0, empty;
 687
 688         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 689
 690         spin_lock(&ino->i_lock);
 691         lo = nfsi->layout;
 692         if (!lo) {
 693                 spin_unlock(&ino->i_lock);
 694                 dprintk("NFS: %s no layout to return\n", __func__);
 695                 goto out;
 696         }
 697         stateid = nfsi->layout->plh_stateid;
 698         /* Reference matched in nfs4_layoutreturn_release */
 699         pnfs_get_layout_hdr(lo);
 700         empty = list_empty(&lo->plh_segs);
 701         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 702         /* Don't send a LAYOUTRETURN if list was initially empty */
 703         if (empty) {
 704                 spin_unlock(&ino->i_lock);
 705                 pnfs_put_layout_hdr(lo);
 706                 dprintk("NFS: %s no layout segments to return\n", __func__);
 707                 goto out;
 708         }
 709         lo->plh_block_lgets++;
 710         spin_unlock(&ino->i_lock);
 711         pnfs_free_lseg_list(&tmp_list);
 712
 713         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 714
 715         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 716         if (unlikely(lrp == NULL)) {
 717                 status = -ENOMEM;
 718                 spin_lock(&ino->i_lock);
 719                 lo->plh_block_lgets--;
 720                 spin_unlock(&ino->i_lock);
 721                 pnfs_put_layout_hdr(lo);
 722                 goto out;
 723         }
 724
 725         lrp->args.stateid = stateid;
 726         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 727         lrp->args.inode = ino;
 728         lrp->args.layout = lo;
 729         lrp->clp = NFS_SERVER(ino)->nfs_client;
 730
 731         status = nfs4_proc_layoutreturn(lrp);
 732 out:
 733         dprintk("<-- %s status: %d\n", __func__, status);
 734         return status;
 735 }
 736 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 737
 738 bool pnfs_roc(struct inode *ino)
 739 {
 740         struct pnfs_layout_hdr *lo;
 741         struct pnfs_layout_segment *lseg, *tmp;
 742         LIST_HEAD(tmp_list);
 743         bool found = false;
 744
 745         spin_lock(&ino->i_lock);
 746         lo = NFS_I(ino)->layout;
 747         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 748             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 749                 goto out_nolayout;
 750         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 751                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 752                         mark_lseg_invalid(lseg, &tmp_list);
 753                         found = true;
 754                 }
 755         if (!found)
 756                 goto out_nolayout;
 757         lo->plh_block_lgets++;
 758         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 759         spin_unlock(&ino->i_lock);
 760         pnfs_free_lseg_list(&tmp_list);
 761         return true;
 762
 763 out_nolayout:
 764         spin_unlock(&ino->i_lock);
 765         return false;
 766 }
 767
 768 void pnfs_roc_release(struct inode *ino)
 769 {
 770         struct pnfs_layout_hdr *lo;
 771
 772         spin_lock(&ino->i_lock);
 773         lo = NFS_I(ino)->layout;
 774         lo->plh_block_lgets--;
 775         if (atomic_dec_and_test(&lo->plh_refcount)) {
 776                 pnfs_detach_layout_hdr(lo);
 777                 spin_unlock(&ino->i_lock);
 778                 pnfs_free_layout_hdr(lo);
 779         } else
 780                 spin_unlock(&ino->i_lock);
 781 }
 782
 783 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 784 {
 785         struct pnfs_layout_hdr *lo;
 786
 787         spin_lock(&ino->i_lock);
 788         lo = NFS_I(ino)->layout;
 789         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 790                 lo->plh_barrier = barrier;
 791         spin_unlock(&ino->i_lock);
 792 }
 793
 794 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 795 {
 796         struct nfs_inode *nfsi = NFS_I(ino);
 797         struct pnfs_layout_hdr *lo;
 798         struct pnfs_layout_segment *lseg;
 799         u32 current_seqid;
 800         bool found = false;
 801
 802         spin_lock(&ino->i_lock);
 803         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 804                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 805                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 806                         found = true;
 807                         goto out;
 808                 }
 809         lo = nfsi->layout;
 810         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 811
 812         /* Since close does not return a layout stateid for use as
 813          * a barrier, we choose the worst-case barrier.
 814          */
 815         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 816 out:
 817         spin_unlock(&ino->i_lock);
 818         return found;
 819 }
 820
 821 /*
 822  * Compare two layout segments for sorting into layout cache.
 823  * We want to preferentially return RW over RO layouts, so ensure those
 824  * are seen first.
 825  */
 826 static s64
 827 cmp_layout(struct pnfs_layout_range *l1,
 828            struct pnfs_layout_range *l2)
 829 {
 830         s64 d;
 831
 832         /* high offset > low offset */
 833         d = l1->offset - l2->offset;
 834         if (d)
 835                 return d;
 836
 837         /* short length > long length */
 838         d = l2->length - l1->length;
 839         if (d)
 840                 return d;
 841
 842         /* read > read/write */
 843         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 844 }
 845
 846 static void
 847 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
 848                    struct pnfs_layout_segment *lseg)
 849 {
 850         struct pnfs_layout_segment *lp;
 851
 852         dprintk("%s:Begin\n", __func__);
 853
 854         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 855                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 856                         continue;
 857                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 858                 dprintk("%s: inserted lseg %p "
 859                         "iomode %d offset %llu length %llu before "
 860                         "lp %p iomode %d offset %llu length %llu\n",
 861                         __func__, lseg, lseg->pls_range.iomode,
 862                         lseg->pls_range.offset, lseg->pls_range.length,
 863                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 864                         lp->pls_range.length);
 865                 goto out;
 866         }
 867         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 868         dprintk("%s: inserted lseg %p "
 869                 "iomode %d offset %llu length %llu at tail\n",
 870                 __func__, lseg, lseg->pls_range.iomode,
 871                 lseg->pls_range.offset, lseg->pls_range.length);
 872 out:
 873         pnfs_get_layout_hdr(lo);
 874
 875         dprintk("%s:Return\n", __func__);
 876 }
 877
 878 static struct pnfs_layout_hdr *
 879 alloc_init_layout_hdr(struct inode *ino,
 880                       struct nfs_open_context *ctx,
 881                       gfp_t gfp_flags)
 882 {
 883         struct pnfs_layout_hdr *lo;
 884
 885         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 886         if (!lo)
 887                 return NULL;
 888         atomic_set(&lo->plh_refcount, 1);
 889         INIT_LIST_HEAD(&lo->plh_layouts);
 890         INIT_LIST_HEAD(&lo->plh_segs);
 891         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 892         lo->plh_inode = ino;
 893         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 894         return lo;
 895 }
 896
 897 static struct pnfs_layout_hdr *
 898 pnfs_find_alloc_layout(struct inode *ino,
 899                        struct nfs_open_context *ctx,
 900                        gfp_t gfp_flags)
 901 {
 902         struct nfs_inode *nfsi = NFS_I(ino);
 903         struct pnfs_layout_hdr *new = NULL;
 904
 905         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 906
 907         if (nfsi->layout != NULL)
 908                 goto out_existing;
 909         spin_unlock(&ino->i_lock);
 910         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 911         spin_lock(&ino->i_lock);
 912
 913         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
 914                 nfsi->layout = new;
 915                 return new;
 916         } else if (new != NULL)
 917                 pnfs_free_layout_hdr(new);
 918 out_existing:
 919         pnfs_get_layout_hdr(nfsi->layout);
 920         return nfsi->layout;
 921 }
 922
 923 /*
 924  * iomode matching rules:
 925  * iomode       lseg    match
 926  * -----        -----   -----
 927  * ANY          READ    true
 928  * ANY          RW      true
 929  * RW           READ    false
 930  * RW           RW      true
 931  * READ         READ    true
 932  * READ         RW      true
 933  */
 934 static int
 935 is_matching_lseg(struct pnfs_layout_range *ls_range,
 936                  struct pnfs_layout_range *range)
 937 {
 938         struct pnfs_layout_range range1;
 939
 940         if ((range->iomode == IOMODE_RW &&
 941              ls_range->iomode != IOMODE_RW) ||
 942             !lo_seg_intersecting(ls_range, range))
 943                 return 0;
 944
 945         /* range1 covers only the first byte in the range */
 946         range1 = *range;
 947         range1.length = 1;
 948         return lo_seg_contained(ls_range, &range1);
 949 }
 950
 951 /*
 952  * lookup range in layout
 953  */
 954 static struct pnfs_layout_segment *
 955 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 956                 struct pnfs_layout_range *range)
 957 {
 958         struct pnfs_layout_segment *lseg, *ret = NULL;
 959
 960         dprintk("%s:Begin\n", __func__);
 961
 962         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 963                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 964                     is_matching_lseg(&lseg->pls_range, range)) {
 965                         ret = pnfs_get_lseg(lseg);
 966                         break;
 967                 }
 968                 if (lseg->pls_range.offset > range->offset)
 969                         break;
 970         }
 971
 972         dprintk("%s:Return lseg %p ref %d\n",
 973                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 974         return ret;
 975 }
 976
 977 /*
 978  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 979  * to the MDS or over pNFS
 980  *
 981  * The nfs_inode read_io and write_io fields are cumulative counters reset
 982  * when there are no layout segments. Note that in pnfs_update_layout iomode
 983  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 984  * WRITE request.
 985  *
 986  * A return of true means use MDS I/O.
 987  *
 988  * From rfc 5661:
 989  * If a file's size is smaller than the file size threshold, data accesses
 990  * SHOULD be sent to the metadata server.  If an I/O request has a length that
 991  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
 992  * server.  If both file size and I/O size are provided, the client SHOULD
 993  * reach or exceed  both thresholds before sending its read or write
 994  * requests to the data server.
 995  */
 996 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
 997                                      struct inode *ino, int iomode)
 998 {
 999         struct nfs4_threshold *t = ctx->mdsthreshold;
1000         struct nfs_inode *nfsi = NFS_I(ino);
1001         loff_t fsize = i_size_read(ino);
1002         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1003
1004         if (t == NULL)
1005                 return ret;
1006
1007         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1008                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1009
1010         switch (iomode) {
1011         case IOMODE_READ:
1012                 if (t->bm & THRESHOLD_RD) {
1013                         dprintk("%s fsize %llu\n", __func__, fsize);
1014                         size_set = true;
1015                         if (fsize < t->rd_sz)
1016                                 size = true;
1017                 }
1018                 if (t->bm & THRESHOLD_RD_IO) {
1019                         dprintk("%s nfsi->read_io %llu\n", __func__,
1020                                 nfsi->read_io);
1021                         io_set = true;
1022                         if (nfsi->read_io < t->rd_io_sz)
1023                                 io = true;
1024                 }
1025                 break;
1026         case IOMODE_RW:
1027                 if (t->bm & THRESHOLD_WR) {
1028                         dprintk("%s fsize %llu\n", __func__, fsize);
1029                         size_set = true;
1030                         if (fsize < t->wr_sz)
1031                                 size = true;
1032                 }
1033                 if (t->bm & THRESHOLD_WR_IO) {
1034                         dprintk("%s nfsi->write_io %llu\n", __func__,
1035                                 nfsi->write_io);
1036                         io_set = true;
1037                         if (nfsi->write_io < t->wr_io_sz)
1038                                 io = true;
1039                 }
1040                 break;
1041         }
1042         if (size_set && io_set) {
1043                 if (size && io)
1044                         ret = true;
1045         } else if (size || io)
1046                 ret = true;
1047
1048         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1049         return ret;
1050 }
1051
1052 /*
1053  * Layout segment is retreived from the server if not cached.
1054  * The appropriate layout segment is referenced and returned to the caller.
1055  */
1056 struct pnfs_layout_segment *
1057 pnfs_update_layout(struct inode *ino,
1058                    struct nfs_open_context *ctx,
1059                    loff_t pos,
1060                    u64 count,
1061                    enum pnfs_iomode iomode,
1062                    gfp_t gfp_flags)
1063 {
1064         struct pnfs_layout_range arg = {
1065                 .iomode = iomode,
1066                 .offset = pos,
1067                 .length = count,
1068         };
1069         unsigned pg_offset;
1070         struct nfs_server *server = NFS_SERVER(ino);
1071         struct nfs_client *clp = server->nfs_client;
1072         struct pnfs_layout_hdr *lo;
1073         struct pnfs_layout_segment *lseg = NULL;
1074         bool first = false;
1075
1076         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1077                 goto out;
1078
1079         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1080                 goto out;
1081
1082         spin_lock(&ino->i_lock);
1083         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1084         if (lo == NULL) {
1085                 spin_unlock(&ino->i_lock);
1086                 goto out;
1087         }
1088
1089         /* Do we even need to bother with this? */
1090         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1091                 dprintk("%s matches recall, use MDS\n", __func__);
1092                 goto out_unlock;
1093         }
1094
1095         /* if LAYOUTGET already failed once we don't try again */
1096         if (pnfs_layout_io_test_failed(lo, iomode))
1097                 goto out_unlock;
1098
1099         /* Check to see if the layout for the given range already exists */
1100         lseg = pnfs_find_lseg(lo, &arg);
1101         if (lseg)
1102                 goto out_unlock;
1103
1104         if (pnfs_layoutgets_blocked(lo, 0))
1105                 goto out_unlock;
1106         atomic_inc(&lo->plh_outstanding);
1107
1108         if (list_empty(&lo->plh_segs))
1109                 first = true;
1110
1111         spin_unlock(&ino->i_lock);
1112         if (first) {
1113                 /* The lo must be on the clp list if there is any
1114                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1115                  */
1116                 spin_lock(&clp->cl_lock);
1117                 list_add_tail(&lo->plh_layouts, &server->layouts);
1118                 spin_unlock(&clp->cl_lock);
1119         }
1120
1121         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1122         if (pg_offset) {
1123                 arg.offset -= pg_offset;
1124                 arg.length += pg_offset;
1125         }
1126         if (arg.length != NFS4_MAX_UINT64)
1127                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1128
1129         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1130         atomic_dec(&lo->plh_outstanding);
1131 out_put_layout_hdr:
1132         pnfs_put_layout_hdr(lo);
1133 out:
1134         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1135                         "(%s, offset: %llu, length: %llu)\n",
1136                         __func__, ino->i_sb->s_id,
1137                         (unsigned long long)NFS_FILEID(ino),
1138                         lseg == NULL ? "not found" : "found",
1139                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1140                         (unsigned long long)pos,
1141                         (unsigned long long)count);
1142         return lseg;
1143 out_unlock:
1144         spin_unlock(&ino->i_lock);
1145         goto out_put_layout_hdr;
1146 }
1147 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1148
1149 struct pnfs_layout_segment *
1150 pnfs_layout_process(struct nfs4_layoutget *lgp)
1151 {
1152         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1153         struct nfs4_layoutget_res *res = &lgp->res;
1154         struct pnfs_layout_segment *lseg;
1155         struct inode *ino = lo->plh_inode;
1156         int status = 0;
1157
1158         /* Inject layout blob into I/O device driver */
1159         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1160         if (!lseg || IS_ERR(lseg)) {
1161                 if (!lseg)
1162                         status = -ENOMEM;
1163                 else
1164                         status = PTR_ERR(lseg);
1165                 dprintk("%s: Could not allocate layout: error %d\n",
1166                        __func__, status);
1167                 goto out;
1168         }
1169
1170         spin_lock(&ino->i_lock);
1171         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1172                 dprintk("%s forget reply due to recall\n", __func__);
1173                 goto out_forget_reply;
1174         }
1175
1176         if (pnfs_layoutgets_blocked(lo, 1) ||
1177             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1178                 dprintk("%s forget reply due to state\n", __func__);
1179                 goto out_forget_reply;
1180         }
1181
1182         /* Done processing layoutget. Set the layout stateid */
1183         pnfs_set_layout_stateid(lo, &res->stateid, false);
1184
1185         init_lseg(lo, lseg);
1186         lseg->pls_range = res->range;
1187         pnfs_get_lseg(lseg);
1188         pnfs_layout_insert_lseg(lo, lseg);
1189
1190         if (res->return_on_close) {
1191                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1192                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1193         }
1194
1195         spin_unlock(&ino->i_lock);
1196         return lseg;
1197 out:
1198         return ERR_PTR(status);
1199
1200 out_forget_reply:
1201         spin_unlock(&ino->i_lock);
1202         lseg->pls_layout = lo;
1203         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1204         goto out;
1205 }
1206
1207 void
1208 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1209 {
1210         u64 rd_size = req->wb_bytes;
1211
1212         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1213
1214         if (req->wb_offset != req->wb_pgbase) {
1215                 nfs_pageio_reset_read_mds(pgio);
1216                 return;
1217         }
1218
1219         if (pgio->pg_dreq == NULL)
1220                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1221         else
1222                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1223
1224         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1225                                            req->wb_context,
1226                                            req_offset(req),
1227                                            rd_size,
1228                                            IOMODE_READ,
1229                                            GFP_KERNEL);
1230         /* If no lseg, fall back to read through mds */
1231         if (pgio->pg_lseg == NULL)
1232                 nfs_pageio_reset_read_mds(pgio);
1233
1234 }
1235 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1236
1237 void
1238 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1239                            struct nfs_page *req, u64 wb_size)
1240 {
1241         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1242
1243         if (req->wb_offset != req->wb_pgbase) {
1244                 nfs_pageio_reset_write_mds(pgio);
1245                 return;
1246         }
1247
1248         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1249                                            req->wb_context,
1250                                            req_offset(req),
1251                                            wb_size,
1252                                            IOMODE_RW,
1253                                            GFP_NOFS);
1254         /* If no lseg, fall back to write through mds */
1255         if (pgio->pg_lseg == NULL)
1256                 nfs_pageio_reset_write_mds(pgio);
1257 }
1258 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1259
1260 void
1261 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1262                       const struct nfs_pgio_completion_ops *compl_ops)
1263 {
1264         struct nfs_server *server = NFS_SERVER(inode);
1265         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1266
1267         if (ld == NULL)
1268                 nfs_pageio_init_read(pgio, inode, compl_ops);
1269         else
1270                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1271 }
1272
1273 void
1274 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1275                        int ioflags,
1276                        const struct nfs_pgio_completion_ops *compl_ops)
1277 {
1278         struct nfs_server *server = NFS_SERVER(inode);
1279         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1280
1281         if (ld == NULL)
1282                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1283         else
1284                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1285 }
1286
1287 bool
1288 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1289                      struct nfs_page *req)
1290 {
1291         if (pgio->pg_lseg == NULL)
1292                 return nfs_generic_pg_test(pgio, prev, req);
1293
1294         /*
1295          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1296          * Note that this test makes several assumptions:
1297          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1298          *   is known to lie within the range.
1299          *   - that the nfs_page being tested is known to be contiguous with the
1300          *   previous nfs_page.
1301          *   - Layout ranges are page aligned, so we only have to test the
1302          *   start offset of the request.
1303          *
1304          * Please also note that 'end_offset' is actually the offset of the
1305          * first byte that lies outside the pnfs_layout_range. FIXME?
1306          *
1307          */
1308         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1309                                          pgio->pg_lseg->pls_range.length);
1310 }
1311 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1312
1313 int pnfs_write_done_resend_to_mds(struct inode *inode,
1314                                 struct list_head *head,
1315                                 const struct nfs_pgio_completion_ops *compl_ops)
1316 {
1317         struct nfs_pageio_descriptor pgio;
1318         LIST_HEAD(failed);
1319
1320         /* Resend all requests through the MDS */
1321         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1322         while (!list_empty(head)) {
1323                 struct nfs_page *req = nfs_list_entry(head->next);
1324
1325                 nfs_list_remove_request(req);
1326                 if (!nfs_pageio_add_request(&pgio, req))
1327                         nfs_list_add_request(req, &failed);
1328         }
1329         nfs_pageio_complete(&pgio);
1330
1331         if (!list_empty(&failed)) {
1332                 /* For some reason our attempt to resend pages. Mark the
1333                  * overall send request as having failed, and let
1334                  * nfs_writeback_release_full deal with the error.
1335                  */
1336                 list_move(&failed, head);
1337                 return -EIO;
1338         }
1339         return 0;
1340 }
1341 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1342
1343 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1344 {
1345         struct nfs_pgio_header *hdr = data->header;
1346
1347         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1348         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1349             PNFS_LAYOUTRET_ON_ERROR) {
1350                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1351                 pnfs_return_layout(hdr->inode);
1352         }
1353         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1354                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1355                                                         &hdr->pages,
1356                                                         hdr->completion_ops);
1357 }
1358
1359 /*
1360  * Called by non rpc-based layout drivers
1361  */
1362 void pnfs_ld_write_done(struct nfs_write_data *data)
1363 {
1364         struct nfs_pgio_header *hdr = data->header;
1365
1366         if (!hdr->pnfs_error) {
1367                 pnfs_set_layoutcommit(data);
1368                 hdr->mds_ops->rpc_call_done(&data->task, data);
1369         } else
1370                 pnfs_ld_handle_write_error(data);
1371         hdr->mds_ops->rpc_release(data);
1372 }
1373 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1374
1375 static void
1376 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1377                 struct nfs_write_data *data)
1378 {
1379         struct nfs_pgio_header *hdr = data->header;
1380
1381         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1382                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1383                 nfs_pageio_reset_write_mds(desc);
1384                 desc->pg_recoalesce = 1;
1385         }
1386         nfs_writedata_release(data);
1387 }
1388
1389 static enum pnfs_try_status
1390 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1391                         const struct rpc_call_ops *call_ops,
1392                         struct pnfs_layout_segment *lseg,
1393                         int how)
1394 {
1395         struct nfs_pgio_header *hdr = wdata->header;
1396         struct inode *inode = hdr->inode;
1397         enum pnfs_try_status trypnfs;
1398         struct nfs_server *nfss = NFS_SERVER(inode);
1399
1400         hdr->mds_ops = call_ops;
1401
1402         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1403                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1404         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1405         if (trypnfs != PNFS_NOT_ATTEMPTED)
1406                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1407         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1408         return trypnfs;
1409 }
1410
1411 static void
1412 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1413 {
1414         struct nfs_write_data *data;
1415         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1416         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1417
1418         desc->pg_lseg = NULL;
1419         while (!list_empty(head)) {
1420                 enum pnfs_try_status trypnfs;
1421
1422                 data = list_first_entry(head, struct nfs_write_data, list);
1423                 list_del_init(&data->list);
1424
1425                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1426                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1427                         pnfs_write_through_mds(desc, data);
1428         }
1429         pnfs_put_lseg(lseg);
1430 }
1431
1432 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1433 {
1434         pnfs_put_lseg(hdr->lseg);
1435         nfs_writehdr_free(hdr);
1436 }
1437 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1438
1439 int
1440 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1441 {
1442         struct nfs_write_header *whdr;
1443         struct nfs_pgio_header *hdr;
1444         int ret;
1445
1446         whdr = nfs_writehdr_alloc();
1447         if (!whdr) {
1448                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1449                 pnfs_put_lseg(desc->pg_lseg);
1450                 desc->pg_lseg = NULL;
1451                 return -ENOMEM;
1452         }
1453         hdr = &whdr->header;
1454         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1455         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1456         atomic_inc(&hdr->refcnt);
1457         ret = nfs_generic_flush(desc, hdr);
1458         if (ret != 0) {
1459                 pnfs_put_lseg(desc->pg_lseg);
1460                 desc->pg_lseg = NULL;
1461         } else
1462                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1463         if (atomic_dec_and_test(&hdr->refcnt))
1464                 hdr->completion_ops->completion(hdr);
1465         return ret;
1466 }
1467 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1468
1469 int pnfs_read_done_resend_to_mds(struct inode *inode,
1470                                 struct list_head *head,
1471                                 const struct nfs_pgio_completion_ops *compl_ops)
1472 {
1473         struct nfs_pageio_descriptor pgio;
1474         LIST_HEAD(failed);
1475
1476         /* Resend all requests through the MDS */
1477         nfs_pageio_init_read(&pgio, inode, compl_ops);
1478         while (!list_empty(head)) {
1479                 struct nfs_page *req = nfs_list_entry(head->next);
1480
1481                 nfs_list_remove_request(req);
1482                 if (!nfs_pageio_add_request(&pgio, req))
1483                         nfs_list_add_request(req, &failed);
1484         }
1485         nfs_pageio_complete(&pgio);
1486
1487         if (!list_empty(&failed)) {
1488                 list_move(&failed, head);
1489                 return -EIO;
1490         }
1491         return 0;
1492 }
1493 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1494
1495 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1496 {
1497         struct nfs_pgio_header *hdr = data->header;
1498
1499         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1500         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1501             PNFS_LAYOUTRET_ON_ERROR) {
1502                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1503                 pnfs_return_layout(hdr->inode);
1504         }
1505         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1506                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1507                                                         &hdr->pages,
1508                                                         hdr->completion_ops);
1509 }
1510
1511 /*
1512  * Called by non rpc-based layout drivers
1513  */
1514 void pnfs_ld_read_done(struct nfs_read_data *data)
1515 {
1516         struct nfs_pgio_header *hdr = data->header;
1517
1518         if (likely(!hdr->pnfs_error)) {
1519                 __nfs4_read_done_cb(data);
1520                 hdr->mds_ops->rpc_call_done(&data->task, data);
1521         } else
1522                 pnfs_ld_handle_read_error(data);
1523         hdr->mds_ops->rpc_release(data);
1524 }
1525 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1526
1527 static void
1528 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1529                 struct nfs_read_data *data)
1530 {
1531         struct nfs_pgio_header *hdr = data->header;
1532
1533         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1534                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1535                 nfs_pageio_reset_read_mds(desc);
1536                 desc->pg_recoalesce = 1;
1537         }
1538         nfs_readdata_release(data);
1539 }
1540
1541 /*
1542  * Call the appropriate parallel I/O subsystem read function.
1543  */
1544 static enum pnfs_try_status
1545 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1546                        const struct rpc_call_ops *call_ops,
1547                        struct pnfs_layout_segment *lseg)
1548 {
1549         struct nfs_pgio_header *hdr = rdata->header;
1550         struct inode *inode = hdr->inode;
1551         struct nfs_server *nfss = NFS_SERVER(inode);
1552         enum pnfs_try_status trypnfs;
1553
1554         hdr->mds_ops = call_ops;
1555
1556         dprintk("%s: Reading ino:%lu %u@%llu\n",
1557                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1558
1559         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1560         if (trypnfs != PNFS_NOT_ATTEMPTED)
1561                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1562         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1563         return trypnfs;
1564 }
1565
1566 static void
1567 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1568 {
1569         struct nfs_read_data *data;
1570         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1571         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1572
1573         desc->pg_lseg = NULL;
1574         while (!list_empty(head)) {
1575                 enum pnfs_try_status trypnfs;
1576
1577                 data = list_first_entry(head, struct nfs_read_data, list);
1578                 list_del_init(&data->list);
1579
1580                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1581                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1582                         pnfs_read_through_mds(desc, data);
1583         }
1584         pnfs_put_lseg(lseg);
1585 }
1586
1587 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1588 {
1589         pnfs_put_lseg(hdr->lseg);
1590         nfs_readhdr_free(hdr);
1591 }
1592 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1593
1594 int
1595 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1596 {
1597         struct nfs_read_header *rhdr;
1598         struct nfs_pgio_header *hdr;
1599         int ret;
1600
1601         rhdr = nfs_readhdr_alloc();
1602         if (!rhdr) {
1603                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1604                 ret = -ENOMEM;
1605                 pnfs_put_lseg(desc->pg_lseg);
1606                 desc->pg_lseg = NULL;
1607                 return ret;
1608         }
1609         hdr = &rhdr->header;
1610         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1611         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1612         atomic_inc(&hdr->refcnt);
1613         ret = nfs_generic_pagein(desc, hdr);
1614         if (ret != 0) {
1615                 pnfs_put_lseg(desc->pg_lseg);
1616                 desc->pg_lseg = NULL;
1617         } else
1618                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1619         if (atomic_dec_and_test(&hdr->refcnt))
1620                 hdr->completion_ops->completion(hdr);
1621         return ret;
1622 }
1623 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1624
1625 /*
1626  * There can be multiple RW segments.
1627  */
1628 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1629 {
1630         struct pnfs_layout_segment *lseg;
1631
1632         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1633                 if (lseg->pls_range.iomode == IOMODE_RW &&
1634                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1635                         list_add(&lseg->pls_lc_list, listp);
1636         }
1637 }
1638
1639 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1640 {
1641         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1642 }
1643 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1644
1645 void
1646 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1647 {
1648         struct nfs_pgio_header *hdr = wdata->header;
1649         struct inode *inode = hdr->inode;
1650         struct nfs_inode *nfsi = NFS_I(inode);
1651         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1652         bool mark_as_dirty = false;
1653
1654         spin_lock(&inode->i_lock);
1655         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1656                 mark_as_dirty = true;
1657                 dprintk("%s: Set layoutcommit for inode %lu ",
1658                         __func__, inode->i_ino);
1659         }
1660         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1661                 /* references matched in nfs4_layoutcommit_release */
1662                 pnfs_get_lseg(hdr->lseg);
1663         }
1664         if (end_pos > nfsi->layout->plh_lwb)
1665                 nfsi->layout->plh_lwb = end_pos;
1666         spin_unlock(&inode->i_lock);
1667         dprintk("%s: lseg %p end_pos %llu\n",
1668                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1669
1670         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1671          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1672         if (mark_as_dirty)
1673                 mark_inode_dirty_sync(inode);
1674 }
1675 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1676
1677 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1678 {
1679         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1680
1681         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1682                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1683 }
1684
1685 /*
1686  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1687  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1688  * data to disk to allow the server to recover the data if it crashes.
1689  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1690  * is off, and a COMMIT is sent to a data server, or
1691  * if WRITEs to a data server return NFS_DATA_SYNC.
1692  */
1693 int
1694 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1695 {
1696         struct nfs4_layoutcommit_data *data;
1697         struct nfs_inode *nfsi = NFS_I(inode);
1698         loff_t end_pos;
1699         int status = 0;
1700
1701         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1702
1703         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1704                 return 0;
1705
1706         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1707         data = kzalloc(sizeof(*data), GFP_NOFS);
1708         if (!data) {
1709                 status = -ENOMEM;
1710                 goto out;
1711         }
1712
1713         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1714                 goto out_free;
1715
1716         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1717                 if (!sync) {
1718                         status = -EAGAIN;
1719                         goto out_free;
1720                 }
1721                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1722                                         nfs_wait_bit_killable, TASK_KILLABLE);
1723                 if (status)
1724                         goto out_free;
1725         }
1726
1727         INIT_LIST_HEAD(&data->lseg_list);
1728         spin_lock(&inode->i_lock);
1729         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1730                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1731                 spin_unlock(&inode->i_lock);
1732                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1733                 goto out_free;
1734         }
1735
1736         pnfs_list_write_lseg(inode, &data->lseg_list);
1737
1738         end_pos = nfsi->layout->plh_lwb;
1739         nfsi->layout->plh_lwb = 0;
1740
1741         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1742         spin_unlock(&inode->i_lock);
1743
1744         data->args.inode = inode;
1745         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1746         nfs_fattr_init(&data->fattr);
1747         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1748         data->res.fattr = &data->fattr;
1749         data->args.lastbytewritten = end_pos - 1;
1750         data->res.server = NFS_SERVER(inode);
1751
1752         status = nfs4_proc_layoutcommit(data, sync);
1753 out:
1754         if (status)
1755                 mark_inode_dirty_sync(inode);
1756         dprintk("<-- %s status %d\n", __func__, status);
1757         return status;
1758 out_free:
1759         kfree(data);
1760         goto out;
1761 }
1762
1763 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1764 {
1765         struct nfs4_threshold *thp;
1766
1767         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1768         if (!thp) {
1769                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1770                 return NULL;
1771         }
1772         return thp;
1773 }