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