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         struct page **pages = NULL;
 587         int i;
 588         u32 max_resp_sz, max_pages;
 589
 590         dprintk("--> %s\n", __func__);
 591
 592         BUG_ON(ctx == NULL);
 593         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 594         if (lgp == NULL)
 595                 return NULL;
 596
 597         /* allocate pages for xdr post processing */
 598         max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
 599         max_pages = nfs_page_array_len(0, max_resp_sz);
 600
 601         pages = kcalloc(max_pages, sizeof(struct page *), gfp_flags);
 602         if (!pages)
 603                 goto out_err_free;
 604
 605         for (i = 0; i < max_pages; i++) {
 606                 pages[i] = alloc_page(gfp_flags);
 607                 if (!pages[i])
 608                         goto out_err_free;
 609         }
 610
 611         lgp->args.minlength = PAGE_CACHE_SIZE;
 612         if (lgp->args.minlength > range->length)
 613                 lgp->args.minlength = range->length;
 614         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 615         lgp->args.range = *range;
 616         lgp->args.type = server->pnfs_curr_ld->id;
 617         lgp->args.inode = ino;
 618         lgp->args.ctx = get_nfs_open_context(ctx);
 619         lgp->args.layout.pages = pages;
 620         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
 621         lgp->lsegpp = &lseg;
 622         lgp->gfp_flags = gfp_flags;
 623
 624         /* Synchronously retrieve layout information from server and
 625          * store in lseg.
 626          */
 627         nfs4_proc_layoutget(lgp);
 628         if (!lseg) {
 629                 /* remember that LAYOUTGET failed and suspend trying */
 630                 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
 631         }
 632
 633         /* free xdr pages */
 634         for (i = 0; i < max_pages; i++)
 635                 __free_page(pages[i]);
 636         kfree(pages);
 637
 638         return lseg;
 639
 640 out_err_free:
 641         /* free any allocated xdr pages, lgp as it's not used */
 642         if (pages) {
 643                 for (i = 0; i < max_pages; i++) {
 644                         if (!pages[i])
 645                                 break;
 646                         __free_page(pages[i]);
 647                 }
 648                 kfree(pages);
 649         }
 650         kfree(lgp);
 651         return NULL;
 652 }
 653
 654 /*
 655  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 656  * when the layout segment list is empty.
 657  *
 658  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 659  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 660  * deviceid is marked invalid.
 661  */
 662 int
 663 _pnfs_return_layout(struct inode *ino)
 664 {
 665         struct pnfs_layout_hdr *lo = NULL;
 666         struct nfs_inode *nfsi = NFS_I(ino);
 667         LIST_HEAD(tmp_list);
 668         struct nfs4_layoutreturn *lrp;
 669         nfs4_stateid stateid;
 670         int status = 0, empty;
 671
 672         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 673
 674         spin_lock(&ino->i_lock);
 675         lo = nfsi->layout;
 676         if (!lo || pnfs_test_layout_returned(lo)) {
 677                 spin_unlock(&ino->i_lock);
 678                 dprintk("NFS: %s no layout to return\n", __func__);
 679                 goto out;
 680         }
 681         stateid = nfsi->layout->plh_stateid;
 682         /* Reference matched in nfs4_layoutreturn_release */
 683         get_layout_hdr(lo);
 684         empty = list_empty(&lo->plh_segs);
 685         mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 686         /* Don't send a LAYOUTRETURN if list was initially empty */
 687         if (empty) {
 688                 spin_unlock(&ino->i_lock);
 689                 put_layout_hdr(lo);
 690                 dprintk("NFS: %s no layout segments to return\n", __func__);
 691                 goto out;
 692         }
 693         lo->plh_block_lgets++;
 694         pnfs_mark_layout_returned(lo);
 695         spin_unlock(&ino->i_lock);
 696         pnfs_free_lseg_list(&tmp_list);
 697
 698         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 699
 700         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 701         if (unlikely(lrp == NULL)) {
 702                 status = -ENOMEM;
 703                 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
 704                 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
 705                 pnfs_clear_layout_returned(lo);
 706                 put_layout_hdr(lo);
 707                 goto out;
 708         }
 709
 710         lrp->args.stateid = stateid;
 711         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 712         lrp->args.inode = ino;
 713         lrp->args.layout = lo;
 714         lrp->clp = NFS_SERVER(ino)->nfs_client;
 715
 716         status = nfs4_proc_layoutreturn(lrp);
 717 out:
 718         dprintk("<-- %s status: %d\n", __func__, status);
 719         return status;
 720 }
 721 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 722
 723 bool pnfs_roc(struct inode *ino)
 724 {
 725         struct pnfs_layout_hdr *lo;
 726         struct pnfs_layout_segment *lseg, *tmp;
 727         LIST_HEAD(tmp_list);
 728         bool found = false;
 729
 730         spin_lock(&ino->i_lock);
 731         lo = NFS_I(ino)->layout;
 732         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 733             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 734                 goto out_nolayout;
 735         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 736                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 737                         mark_lseg_invalid(lseg, &tmp_list);
 738                         found = true;
 739                 }
 740         if (!found)
 741                 goto out_nolayout;
 742         lo->plh_block_lgets++;
 743         get_layout_hdr(lo); /* matched in pnfs_roc_release */
 744         spin_unlock(&ino->i_lock);
 745         pnfs_free_lseg_list(&tmp_list);
 746         return true;
 747
 748 out_nolayout:
 749         spin_unlock(&ino->i_lock);
 750         return false;
 751 }
 752
 753 void pnfs_roc_release(struct inode *ino)
 754 {
 755         struct pnfs_layout_hdr *lo;
 756
 757         spin_lock(&ino->i_lock);
 758         lo = NFS_I(ino)->layout;
 759         lo->plh_block_lgets--;
 760         put_layout_hdr_locked(lo);
 761         spin_unlock(&ino->i_lock);
 762 }
 763
 764 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 765 {
 766         struct pnfs_layout_hdr *lo;
 767
 768         spin_lock(&ino->i_lock);
 769         lo = NFS_I(ino)->layout;
 770         if ((int)(barrier - lo->plh_barrier) > 0)
 771                 lo->plh_barrier = barrier;
 772         spin_unlock(&ino->i_lock);
 773 }
 774
 775 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
 776 {
 777         struct nfs_inode *nfsi = NFS_I(ino);
 778         struct pnfs_layout_segment *lseg;
 779         bool found = false;
 780
 781         spin_lock(&ino->i_lock);
 782         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 783                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 784                         found = true;
 785                         break;
 786                 }
 787         if (!found) {
 788                 struct pnfs_layout_hdr *lo = nfsi->layout;
 789                 u32 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 790
 791                 /* Since close does not return a layout stateid for use as
 792                  * a barrier, we choose the worst-case barrier.
 793                  */
 794                 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 795         }
 796         spin_unlock(&ino->i_lock);
 797         return found;
 798 }
 799
 800 /*
 801  * Compare two layout segments for sorting into layout cache.
 802  * We want to preferentially return RW over RO layouts, so ensure those
 803  * are seen first.
 804  */
 805 static s64
 806 cmp_layout(struct pnfs_layout_range *l1,
 807            struct pnfs_layout_range *l2)
 808 {
 809         s64 d;
 810
 811         /* high offset > low offset */
 812         d = l1->offset - l2->offset;
 813         if (d)
 814                 return d;
 815
 816         /* short length > long length */
 817         d = l2->length - l1->length;
 818         if (d)
 819                 return d;
 820
 821         /* read > read/write */
 822         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 823 }
 824
 825 static void
 826 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
 827                    struct pnfs_layout_segment *lseg)
 828 {
 829         struct pnfs_layout_segment *lp;
 830
 831         dprintk("%s:Begin\n", __func__);
 832
 833         assert_spin_locked(&lo->plh_inode->i_lock);
 834         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 835                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 836                         continue;
 837                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 838                 dprintk("%s: inserted lseg %p "
 839                         "iomode %d offset %llu length %llu before "
 840                         "lp %p iomode %d offset %llu length %llu\n",
 841                         __func__, lseg, lseg->pls_range.iomode,
 842                         lseg->pls_range.offset, lseg->pls_range.length,
 843                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 844                         lp->pls_range.length);
 845                 goto out;
 846         }
 847         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 848         dprintk("%s: inserted lseg %p "
 849                 "iomode %d offset %llu length %llu at tail\n",
 850                 __func__, lseg, lseg->pls_range.iomode,
 851                 lseg->pls_range.offset, lseg->pls_range.length);
 852 out:
 853         get_layout_hdr(lo);
 854
 855         dprintk("%s:Return\n", __func__);
 856 }
 857
 858 static struct pnfs_layout_hdr *
 859 alloc_init_layout_hdr(struct inode *ino,
 860                       struct nfs_open_context *ctx,
 861                       gfp_t gfp_flags)
 862 {
 863         struct pnfs_layout_hdr *lo;
 864
 865         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 866         if (!lo)
 867                 return NULL;
 868         atomic_set(&lo->plh_refcount, 1);
 869         INIT_LIST_HEAD(&lo->plh_layouts);
 870         INIT_LIST_HEAD(&lo->plh_segs);
 871         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 872         lo->plh_inode = ino;
 873         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 874         return lo;
 875 }
 876
 877 static struct pnfs_layout_hdr *
 878 pnfs_find_alloc_layout(struct inode *ino,
 879                        struct nfs_open_context *ctx,
 880                        gfp_t gfp_flags)
 881 {
 882         struct nfs_inode *nfsi = NFS_I(ino);
 883         struct pnfs_layout_hdr *new = NULL;
 884
 885         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 886
 887         assert_spin_locked(&ino->i_lock);
 888         if (nfsi->layout) {
 889                 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
 890                         return NULL;
 891                 else
 892                         return nfsi->layout;
 893         }
 894         spin_unlock(&ino->i_lock);
 895         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 896         spin_lock(&ino->i_lock);
 897
 898         if (likely(nfsi->layout == NULL))       /* Won the race? */
 899                 nfsi->layout = new;
 900         else
 901                 pnfs_free_layout_hdr(new);
 902         return nfsi->layout;
 903 }
 904
 905 /*
 906  * iomode matching rules:
 907  * iomode       lseg    match
 908  * -----        -----   -----
 909  * ANY          READ    true
 910  * ANY          RW      true
 911  * RW           READ    false
 912  * RW           RW      true
 913  * READ         READ    true
 914  * READ         RW      true
 915  */
 916 static int
 917 is_matching_lseg(struct pnfs_layout_range *ls_range,
 918                  struct pnfs_layout_range *range)
 919 {
 920         struct pnfs_layout_range range1;
 921
 922         if ((range->iomode == IOMODE_RW &&
 923              ls_range->iomode != IOMODE_RW) ||
 924             !lo_seg_intersecting(ls_range, range))
 925                 return 0;
 926
 927         /* range1 covers only the first byte in the range */
 928         range1 = *range;
 929         range1.length = 1;
 930         return lo_seg_contained(ls_range, &range1);
 931 }
 932
 933 /*
 934  * lookup range in layout
 935  */
 936 static struct pnfs_layout_segment *
 937 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 938                 struct pnfs_layout_range *range)
 939 {
 940         struct pnfs_layout_segment *lseg, *ret = NULL;
 941
 942         dprintk("%s:Begin\n", __func__);
 943
 944         assert_spin_locked(&lo->plh_inode->i_lock);
 945         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 946                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 947                     is_matching_lseg(&lseg->pls_range, range)) {
 948                         ret = get_lseg(lseg);
 949                         break;
 950                 }
 951                 if (lseg->pls_range.offset > range->offset)
 952                         break;
 953         }
 954
 955         dprintk("%s:Return lseg %p ref %d\n",
 956                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 957         return ret;
 958 }
 959
 960 /*
 961  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 962  * to the MDS or over pNFS
 963  *
 964  * The nfs_inode read_io and write_io fields are cumulative counters reset
 965  * when there are no layout segments. Note that in pnfs_update_layout iomode
 966  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 967  * WRITE request.
 968  *
 969  * A return of true means use MDS I/O.
 970  *
 971  * From rfc 5661:
 972  * If a file's size is smaller than the file size threshold, data accesses
 973  * SHOULD be sent to the metadata server.  If an I/O request has a length that
 974  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
 975  * server.  If both file size and I/O size are provided, the client SHOULD
 976  * reach or exceed  both thresholds before sending its read or write
 977  * requests to the data server.
 978  */
 979 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
 980                                      struct inode *ino, int iomode)
 981 {
 982         struct nfs4_threshold *t = ctx->mdsthreshold;
 983         struct nfs_inode *nfsi = NFS_I(ino);
 984         loff_t fsize = i_size_read(ino);
 985         bool size = false, size_set = false, io = false, io_set = false, ret = false;
 986
 987         if (t == NULL)
 988                 return ret;
 989
 990         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
 991                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
 992
 993         switch (iomode) {
 994         case IOMODE_READ:
 995                 if (t->bm & THRESHOLD_RD) {
 996                         dprintk("%s fsize %llu\n", __func__, fsize);
 997                         size_set = true;
 998                         if (fsize < t->rd_sz)
 999                                 size = true;
1000                 }
1001                 if (t->bm & THRESHOLD_RD_IO) {
1002                         dprintk("%s nfsi->read_io %llu\n", __func__,
1003                                 nfsi->read_io);
1004                         io_set = true;
1005                         if (nfsi->read_io < t->rd_io_sz)
1006                                 io = true;
1007                 }
1008                 break;
1009         case IOMODE_RW:
1010                 if (t->bm & THRESHOLD_WR) {
1011                         dprintk("%s fsize %llu\n", __func__, fsize);
1012                         size_set = true;
1013                         if (fsize < t->wr_sz)
1014                                 size = true;
1015                 }
1016                 if (t->bm & THRESHOLD_WR_IO) {
1017                         dprintk("%s nfsi->write_io %llu\n", __func__,
1018                                 nfsi->write_io);
1019                         io_set = true;
1020                         if (nfsi->write_io < t->wr_io_sz)
1021                                 io = true;
1022                 }
1023                 break;
1024         }
1025         if (size_set && io_set) {
1026                 if (size && io)
1027                         ret = true;
1028         } else if (size || io)
1029                 ret = true;
1030
1031         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1032         return ret;
1033 }
1034
1035 /*
1036  * Layout segment is retreived from the server if not cached.
1037  * The appropriate layout segment is referenced and returned to the caller.
1038  */
1039 struct pnfs_layout_segment *
1040 pnfs_update_layout(struct inode *ino,
1041                    struct nfs_open_context *ctx,
1042                    loff_t pos,
1043                    u64 count,
1044                    enum pnfs_iomode iomode,
1045                    gfp_t gfp_flags)
1046 {
1047         struct pnfs_layout_range arg = {
1048                 .iomode = iomode,
1049                 .offset = pos,
1050                 .length = count,
1051         };
1052         unsigned pg_offset;
1053         struct nfs_inode *nfsi = NFS_I(ino);
1054         struct nfs_server *server = NFS_SERVER(ino);
1055         struct nfs_client *clp = server->nfs_client;
1056         struct pnfs_layout_hdr *lo;
1057         struct pnfs_layout_segment *lseg = NULL;
1058         bool first = false;
1059
1060         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1061                 return NULL;
1062
1063         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1064                 return NULL;
1065
1066         spin_lock(&ino->i_lock);
1067         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1068         if (lo == NULL) {
1069                 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
1070                 goto out_unlock;
1071         }
1072
1073         /* Do we even need to bother with this? */
1074         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1075                 dprintk("%s matches recall, use MDS\n", __func__);
1076                 goto out_unlock;
1077         }
1078
1079         /* if LAYOUTGET already failed once we don't try again */
1080         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
1081                 goto out_unlock;
1082
1083         /* Check to see if the layout for the given range already exists */
1084         lseg = pnfs_find_lseg(lo, &arg);
1085         if (lseg)
1086                 goto out_unlock;
1087
1088         if (pnfs_layoutgets_blocked(lo, NULL, 0))
1089                 goto out_unlock;
1090         atomic_inc(&lo->plh_outstanding);
1091
1092         get_layout_hdr(lo);
1093         if (list_empty(&lo->plh_segs))
1094                 first = true;
1095
1096         /* Enable LAYOUTRETURNs */
1097         pnfs_clear_layout_returned(lo);
1098
1099         spin_unlock(&ino->i_lock);
1100         if (first) {
1101                 /* The lo must be on the clp list if there is any
1102                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1103                  */
1104                 spin_lock(&clp->cl_lock);
1105                 BUG_ON(!list_empty(&lo->plh_layouts));
1106                 list_add_tail(&lo->plh_layouts, &server->layouts);
1107                 spin_unlock(&clp->cl_lock);
1108         }
1109
1110         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1111         if (pg_offset) {
1112                 arg.offset -= pg_offset;
1113                 arg.length += pg_offset;
1114         }
1115         if (arg.length != NFS4_MAX_UINT64)
1116                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1117
1118         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1119         if (!lseg && first) {
1120                 spin_lock(&clp->cl_lock);
1121                 list_del_init(&lo->plh_layouts);
1122                 spin_unlock(&clp->cl_lock);
1123         }
1124         atomic_dec(&lo->plh_outstanding);
1125         put_layout_hdr(lo);
1126 out:
1127         dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1128                 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1129         return lseg;
1130 out_unlock:
1131         spin_unlock(&ino->i_lock);
1132         goto out;
1133 }
1134 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1135
1136 int
1137 pnfs_layout_process(struct nfs4_layoutget *lgp)
1138 {
1139         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1140         struct nfs4_layoutget_res *res = &lgp->res;
1141         struct pnfs_layout_segment *lseg;
1142         struct inode *ino = lo->plh_inode;
1143         int status = 0;
1144
1145         /* Inject layout blob into I/O device driver */
1146         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1147         if (!lseg || IS_ERR(lseg)) {
1148                 if (!lseg)
1149                         status = -ENOMEM;
1150                 else
1151                         status = PTR_ERR(lseg);
1152                 dprintk("%s: Could not allocate layout: error %d\n",
1153                        __func__, status);
1154                 goto out;
1155         }
1156
1157         spin_lock(&ino->i_lock);
1158         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1159                 dprintk("%s forget reply due to recall\n", __func__);
1160                 goto out_forget_reply;
1161         }
1162
1163         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1164                 dprintk("%s forget reply due to state\n", __func__);
1165                 goto out_forget_reply;
1166         }
1167         init_lseg(lo, lseg);
1168         lseg->pls_range = res->range;
1169         *lgp->lsegpp = get_lseg(lseg);
1170         pnfs_insert_layout(lo, lseg);
1171
1172         if (res->return_on_close) {
1173                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1174                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1175         }
1176
1177         /* Done processing layoutget. Set the layout stateid */
1178         pnfs_set_layout_stateid(lo, &res->stateid, false);
1179         spin_unlock(&ino->i_lock);
1180 out:
1181         return status;
1182
1183 out_forget_reply:
1184         spin_unlock(&ino->i_lock);
1185         lseg->pls_layout = lo;
1186         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1187         goto out;
1188 }
1189
1190 void
1191 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1192 {
1193         BUG_ON(pgio->pg_lseg != NULL);
1194
1195         if (req->wb_offset != req->wb_pgbase) {
1196                 nfs_pageio_reset_read_mds(pgio);
1197                 return;
1198         }
1199         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1200                                            req->wb_context,
1201                                            req_offset(req),
1202                                            req->wb_bytes,
1203                                            IOMODE_READ,
1204                                            GFP_KERNEL);
1205         /* If no lseg, fall back to read through mds */
1206         if (pgio->pg_lseg == NULL)
1207                 nfs_pageio_reset_read_mds(pgio);
1208
1209 }
1210 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1211
1212 void
1213 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1214 {
1215         BUG_ON(pgio->pg_lseg != NULL);
1216
1217         if (req->wb_offset != req->wb_pgbase) {
1218                 nfs_pageio_reset_write_mds(pgio);
1219                 return;
1220         }
1221         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1222                                            req->wb_context,
1223                                            req_offset(req),
1224                                            req->wb_bytes,
1225                                            IOMODE_RW,
1226                                            GFP_NOFS);
1227         /* If no lseg, fall back to write through mds */
1228         if (pgio->pg_lseg == NULL)
1229                 nfs_pageio_reset_write_mds(pgio);
1230 }
1231 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1232
1233 void
1234 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1235                       const struct nfs_pgio_completion_ops *compl_ops)
1236 {
1237         struct nfs_server *server = NFS_SERVER(inode);
1238         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1239
1240         if (ld == NULL)
1241                 nfs_pageio_init_read(pgio, inode, compl_ops);
1242         else
1243                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1244 }
1245
1246 void
1247 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1248                        int ioflags,
1249                        const struct nfs_pgio_completion_ops *compl_ops)
1250 {
1251         struct nfs_server *server = NFS_SERVER(inode);
1252         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1253
1254         if (ld == NULL)
1255                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1256         else
1257                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1258 }
1259
1260 bool
1261 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1262                      struct nfs_page *req)
1263 {
1264         if (pgio->pg_lseg == NULL)
1265                 return nfs_generic_pg_test(pgio, prev, req);
1266
1267         /*
1268          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1269          * Note that this test makes several assumptions:
1270          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1271          *   is known to lie within the range.
1272          *   - that the nfs_page being tested is known to be contiguous with the
1273          *   previous nfs_page.
1274          *   - Layout ranges are page aligned, so we only have to test the
1275          *   start offset of the request.
1276          *
1277          * Please also note that 'end_offset' is actually the offset of the
1278          * first byte that lies outside the pnfs_layout_range. FIXME?
1279          *
1280          */
1281         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1282                                          pgio->pg_lseg->pls_range.length);
1283 }
1284 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1285
1286 int pnfs_write_done_resend_to_mds(struct inode *inode,
1287                                 struct list_head *head,
1288                                 const struct nfs_pgio_completion_ops *compl_ops)
1289 {
1290         struct nfs_pageio_descriptor pgio;
1291         LIST_HEAD(failed);
1292
1293         /* Resend all requests through the MDS */
1294         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1295         while (!list_empty(head)) {
1296                 struct nfs_page *req = nfs_list_entry(head->next);
1297
1298                 nfs_list_remove_request(req);
1299                 if (!nfs_pageio_add_request(&pgio, req))
1300                         nfs_list_add_request(req, &failed);
1301         }
1302         nfs_pageio_complete(&pgio);
1303
1304         if (!list_empty(&failed)) {
1305                 /* For some reason our attempt to resend pages. Mark the
1306                  * overall send request as having failed, and let
1307                  * nfs_writeback_release_full deal with the error.
1308                  */
1309                 list_move(&failed, head);
1310                 return -EIO;
1311         }
1312         return 0;
1313 }
1314 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1315
1316 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1317 {
1318         struct nfs_pgio_header *hdr = data->header;
1319
1320         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1321         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1322             PNFS_LAYOUTRET_ON_ERROR) {
1323                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1324                 pnfs_return_layout(hdr->inode);
1325         }
1326         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1327                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1328                                                         &hdr->pages,
1329                                                         hdr->completion_ops);
1330 }
1331
1332 /*
1333  * Called by non rpc-based layout drivers
1334  */
1335 void pnfs_ld_write_done(struct nfs_write_data *data)
1336 {
1337         struct nfs_pgio_header *hdr = data->header;
1338
1339         if (!hdr->pnfs_error) {
1340                 pnfs_set_layoutcommit(data);
1341                 hdr->mds_ops->rpc_call_done(&data->task, data);
1342         } else
1343                 pnfs_ld_handle_write_error(data);
1344         hdr->mds_ops->rpc_release(data);
1345 }
1346 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1347
1348 static void
1349 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1350                 struct nfs_write_data *data)
1351 {
1352         struct nfs_pgio_header *hdr = data->header;
1353
1354         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1355                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1356                 nfs_pageio_reset_write_mds(desc);
1357                 desc->pg_recoalesce = 1;
1358         }
1359         nfs_writedata_release(data);
1360 }
1361
1362 static enum pnfs_try_status
1363 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1364                         const struct rpc_call_ops *call_ops,
1365                         struct pnfs_layout_segment *lseg,
1366                         int how)
1367 {
1368         struct nfs_pgio_header *hdr = wdata->header;
1369         struct inode *inode = hdr->inode;
1370         enum pnfs_try_status trypnfs;
1371         struct nfs_server *nfss = NFS_SERVER(inode);
1372
1373         hdr->mds_ops = call_ops;
1374
1375         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1376                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1377         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1378         if (trypnfs != PNFS_NOT_ATTEMPTED)
1379                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1380         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1381         return trypnfs;
1382 }
1383
1384 static void
1385 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1386 {
1387         struct nfs_write_data *data;
1388         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1389         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1390
1391         desc->pg_lseg = NULL;
1392         while (!list_empty(head)) {
1393                 enum pnfs_try_status trypnfs;
1394
1395                 data = list_first_entry(head, struct nfs_write_data, list);
1396                 list_del_init(&data->list);
1397
1398                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1399                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1400                         pnfs_write_through_mds(desc, data);
1401         }
1402         put_lseg(lseg);
1403 }
1404
1405 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1406 {
1407         put_lseg(hdr->lseg);
1408         nfs_writehdr_free(hdr);
1409 }
1410 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1411
1412 int
1413 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1414 {
1415         struct nfs_write_header *whdr;
1416         struct nfs_pgio_header *hdr;
1417         int ret;
1418
1419         whdr = nfs_writehdr_alloc();
1420         if (!whdr) {
1421                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1422                 put_lseg(desc->pg_lseg);
1423                 desc->pg_lseg = NULL;
1424                 return -ENOMEM;
1425         }
1426         hdr = &whdr->header;
1427         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1428         hdr->lseg = get_lseg(desc->pg_lseg);
1429         atomic_inc(&hdr->refcnt);
1430         ret = nfs_generic_flush(desc, hdr);
1431         if (ret != 0) {
1432                 put_lseg(desc->pg_lseg);
1433                 desc->pg_lseg = NULL;
1434         } else
1435                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1436         if (atomic_dec_and_test(&hdr->refcnt))
1437                 hdr->completion_ops->completion(hdr);
1438         return ret;
1439 }
1440 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1441
1442 int pnfs_read_done_resend_to_mds(struct inode *inode,
1443                                 struct list_head *head,
1444                                 const struct nfs_pgio_completion_ops *compl_ops)
1445 {
1446         struct nfs_pageio_descriptor pgio;
1447         LIST_HEAD(failed);
1448
1449         /* Resend all requests through the MDS */
1450         nfs_pageio_init_read(&pgio, inode, compl_ops);
1451         while (!list_empty(head)) {
1452                 struct nfs_page *req = nfs_list_entry(head->next);
1453
1454                 nfs_list_remove_request(req);
1455                 if (!nfs_pageio_add_request(&pgio, req))
1456                         nfs_list_add_request(req, &failed);
1457         }
1458         nfs_pageio_complete(&pgio);
1459
1460         if (!list_empty(&failed)) {
1461                 list_move(&failed, head);
1462                 return -EIO;
1463         }
1464         return 0;
1465 }
1466 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1467
1468 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1469 {
1470         struct nfs_pgio_header *hdr = data->header;
1471
1472         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1473         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1474             PNFS_LAYOUTRET_ON_ERROR) {
1475                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1476                 pnfs_return_layout(hdr->inode);
1477         }
1478         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1479                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1480                                                         &hdr->pages,
1481                                                         hdr->completion_ops);
1482 }
1483
1484 /*
1485  * Called by non rpc-based layout drivers
1486  */
1487 void pnfs_ld_read_done(struct nfs_read_data *data)
1488 {
1489         struct nfs_pgio_header *hdr = data->header;
1490
1491         if (likely(!hdr->pnfs_error)) {
1492                 __nfs4_read_done_cb(data);
1493                 hdr->mds_ops->rpc_call_done(&data->task, data);
1494         } else
1495                 pnfs_ld_handle_read_error(data);
1496         hdr->mds_ops->rpc_release(data);
1497 }
1498 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1499
1500 static void
1501 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1502                 struct nfs_read_data *data)
1503 {
1504         struct nfs_pgio_header *hdr = data->header;
1505
1506         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1507                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1508                 nfs_pageio_reset_read_mds(desc);
1509                 desc->pg_recoalesce = 1;
1510         }
1511         nfs_readdata_release(data);
1512 }
1513
1514 /*
1515  * Call the appropriate parallel I/O subsystem read function.
1516  */
1517 static enum pnfs_try_status
1518 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1519                        const struct rpc_call_ops *call_ops,
1520                        struct pnfs_layout_segment *lseg)
1521 {
1522         struct nfs_pgio_header *hdr = rdata->header;
1523         struct inode *inode = hdr->inode;
1524         struct nfs_server *nfss = NFS_SERVER(inode);
1525         enum pnfs_try_status trypnfs;
1526
1527         hdr->mds_ops = call_ops;
1528
1529         dprintk("%s: Reading ino:%lu %u@%llu\n",
1530                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1531
1532         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1533         if (trypnfs != PNFS_NOT_ATTEMPTED)
1534                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1535         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1536         return trypnfs;
1537 }
1538
1539 static void
1540 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1541 {
1542         struct nfs_read_data *data;
1543         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1544         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1545
1546         desc->pg_lseg = NULL;
1547         while (!list_empty(head)) {
1548                 enum pnfs_try_status trypnfs;
1549
1550                 data = list_first_entry(head, struct nfs_read_data, list);
1551                 list_del_init(&data->list);
1552
1553                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1554                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1555                         pnfs_read_through_mds(desc, data);
1556         }
1557         put_lseg(lseg);
1558 }
1559
1560 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1561 {
1562         put_lseg(hdr->lseg);
1563         nfs_readhdr_free(hdr);
1564 }
1565 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1566
1567 int
1568 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1569 {
1570         struct nfs_read_header *rhdr;
1571         struct nfs_pgio_header *hdr;
1572         int ret;
1573
1574         rhdr = nfs_readhdr_alloc();
1575         if (!rhdr) {
1576                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1577                 ret = -ENOMEM;
1578                 put_lseg(desc->pg_lseg);
1579                 desc->pg_lseg = NULL;
1580                 return ret;
1581         }
1582         hdr = &rhdr->header;
1583         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1584         hdr->lseg = get_lseg(desc->pg_lseg);
1585         atomic_inc(&hdr->refcnt);
1586         ret = nfs_generic_pagein(desc, hdr);
1587         if (ret != 0) {
1588                 put_lseg(desc->pg_lseg);
1589                 desc->pg_lseg = NULL;
1590         } else
1591                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1592         if (atomic_dec_and_test(&hdr->refcnt))
1593                 hdr->completion_ops->completion(hdr);
1594         return ret;
1595 }
1596 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1597
1598 /*
1599  * There can be multiple RW segments.
1600  */
1601 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1602 {
1603         struct pnfs_layout_segment *lseg;
1604
1605         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1606                 if (lseg->pls_range.iomode == IOMODE_RW &&
1607                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1608                         list_add(&lseg->pls_lc_list, listp);
1609         }
1610 }
1611
1612 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1613 {
1614         if (lseg->pls_range.iomode == IOMODE_RW) {
1615                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1616                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1617         } else {
1618                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1619                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1620         }
1621 }
1622 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1623
1624 void
1625 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1626 {
1627         struct nfs_pgio_header *hdr = wdata->header;
1628         struct inode *inode = hdr->inode;
1629         struct nfs_inode *nfsi = NFS_I(inode);
1630         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1631         bool mark_as_dirty = false;
1632
1633         spin_lock(&inode->i_lock);
1634         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1635                 mark_as_dirty = true;
1636                 dprintk("%s: Set layoutcommit for inode %lu ",
1637                         __func__, inode->i_ino);
1638         }
1639         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1640                 /* references matched in nfs4_layoutcommit_release */
1641                 get_lseg(hdr->lseg);
1642         }
1643         if (end_pos > nfsi->layout->plh_lwb)
1644                 nfsi->layout->plh_lwb = end_pos;
1645         spin_unlock(&inode->i_lock);
1646         dprintk("%s: lseg %p end_pos %llu\n",
1647                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1648
1649         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1650          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1651         if (mark_as_dirty)
1652                 mark_inode_dirty_sync(inode);
1653 }
1654 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1655
1656 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1657 {
1658         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1659
1660         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1661                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1662 }
1663
1664 /*
1665  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1666  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1667  * data to disk to allow the server to recover the data if it crashes.
1668  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1669  * is off, and a COMMIT is sent to a data server, or
1670  * if WRITEs to a data server return NFS_DATA_SYNC.
1671  */
1672 int
1673 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1674 {
1675         struct nfs4_layoutcommit_data *data;
1676         struct nfs_inode *nfsi = NFS_I(inode);
1677         loff_t end_pos;
1678         int status = 0;
1679
1680         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1681
1682         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1683                 return 0;
1684
1685         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1686         data = kzalloc(sizeof(*data), GFP_NOFS);
1687         if (!data) {
1688                 status = -ENOMEM;
1689                 goto out;
1690         }
1691
1692         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1693                 goto out_free;
1694
1695         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1696                 if (!sync) {
1697                         status = -EAGAIN;
1698                         goto out_free;
1699                 }
1700                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1701                                         nfs_wait_bit_killable, TASK_KILLABLE);
1702                 if (status)
1703                         goto out_free;
1704         }
1705
1706         INIT_LIST_HEAD(&data->lseg_list);
1707         spin_lock(&inode->i_lock);
1708         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1709                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1710                 spin_unlock(&inode->i_lock);
1711                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1712                 goto out_free;
1713         }
1714
1715         pnfs_list_write_lseg(inode, &data->lseg_list);
1716
1717         end_pos = nfsi->layout->plh_lwb;
1718         nfsi->layout->plh_lwb = 0;
1719
1720         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1721         spin_unlock(&inode->i_lock);
1722
1723         data->args.inode = inode;
1724         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1725         nfs_fattr_init(&data->fattr);
1726         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1727         data->res.fattr = &data->fattr;
1728         data->args.lastbytewritten = end_pos - 1;
1729         data->res.server = NFS_SERVER(inode);
1730
1731         status = nfs4_proc_layoutcommit(data, sync);
1732 out:
1733         if (status)
1734                 mark_inode_dirty_sync(inode);
1735         dprintk("<-- %s status %d\n", __func__, status);
1736         return status;
1737 out_free:
1738         kfree(data);
1739         goto out;
1740 }
1741
1742 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1743 {
1744         struct nfs4_threshold *thp;
1745
1746         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1747         if (!thp) {
1748                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1749                 return NULL;
1750         }
1751         return thp;
1752 }