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