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 "internal.h"
  33 #include "pnfs.h"
  34 #include "iostat.h"
  35
  36 #define NFSDBG_FACILITY         NFSDBG_PNFS
  37
  38 /* Locking:
  39  *
  40  * pnfs_spinlock:
  41  *      protects pnfs_modules_tbl.
  42  */
  43 static DEFINE_SPINLOCK(pnfs_spinlock);
  44
  45 /*
  46  * pnfs_modules_tbl holds all pnfs modules
  47  */
  48 static LIST_HEAD(pnfs_modules_tbl);
  49
  50 /* Return the registered pnfs layout driver module matching given id */
  51 static struct pnfs_layoutdriver_type *
  52 find_pnfs_driver_locked(u32 id)
  53 {
  54         struct pnfs_layoutdriver_type *local;
  55
  56         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  57                 if (local->id == id)
  58                         goto out;
  59         local = NULL;
  60 out:
  61         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  62         return local;
  63 }
  64
  65 static struct pnfs_layoutdriver_type *
  66 find_pnfs_driver(u32 id)
  67 {
  68         struct pnfs_layoutdriver_type *local;
  69
  70         spin_lock(&pnfs_spinlock);
  71         local = find_pnfs_driver_locked(id);
  72         spin_unlock(&pnfs_spinlock);
  73         return local;
  74 }
  75
  76 void
  77 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  78 {
  79         if (nfss->pnfs_curr_ld) {
  80                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  81                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  82                 module_put(nfss->pnfs_curr_ld->owner);
  83         }
  84         nfss->pnfs_curr_ld = NULL;
  85 }
  86
  87 /*
  88  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  89  * Currently only one pNFS layout driver per filesystem is supported.
  90  *
  91  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
  92  */
  93 void
  94 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
  95                       u32 id)
  96 {
  97         struct pnfs_layoutdriver_type *ld_type = NULL;
  98
  99         if (id == 0)
 100                 goto out_no_driver;
 101         if (!(server->nfs_client->cl_exchange_flags &
 102                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 103                 printk(KERN_ERR "%s: id %u cl_exchange_flags 0x%x\n", __func__,
 104                        id, server->nfs_client->cl_exchange_flags);
 105                 goto out_no_driver;
 106         }
 107         ld_type = find_pnfs_driver(id);
 108         if (!ld_type) {
 109                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 110                 ld_type = find_pnfs_driver(id);
 111                 if (!ld_type) {
 112                         dprintk("%s: No pNFS module found for %u.\n",
 113                                 __func__, id);
 114                         goto out_no_driver;
 115                 }
 116         }
 117         if (!try_module_get(ld_type->owner)) {
 118                 dprintk("%s: Could not grab reference on module\n", __func__);
 119                 goto out_no_driver;
 120         }
 121         server->pnfs_curr_ld = ld_type;
 122         if (ld_type->set_layoutdriver
 123             && ld_type->set_layoutdriver(server, mntfh)) {
 124                 printk(KERN_ERR "%s: Error initializing pNFS layout driver %u.\n",
 125                                 __func__, id);
 126                 module_put(ld_type->owner);
 127                 goto out_no_driver;
 128         }
 129
 130         dprintk("%s: pNFS module for %u set\n", __func__, id);
 131         return;
 132
 133 out_no_driver:
 134         dprintk("%s: Using NFSv4 I/O\n", __func__);
 135         server->pnfs_curr_ld = NULL;
 136 }
 137
 138 int
 139 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 140 {
 141         int status = -EINVAL;
 142         struct pnfs_layoutdriver_type *tmp;
 143
 144         if (ld_type->id == 0) {
 145                 printk(KERN_ERR "%s id 0 is reserved\n", __func__);
 146                 return status;
 147         }
 148         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 149                 printk(KERN_ERR "%s Layout driver must provide "
 150                        "alloc_lseg and free_lseg.\n", __func__);
 151                 return status;
 152         }
 153
 154         spin_lock(&pnfs_spinlock);
 155         tmp = find_pnfs_driver_locked(ld_type->id);
 156         if (!tmp) {
 157                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 158                 status = 0;
 159                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 160                         ld_type->name);
 161         } else {
 162                 printk(KERN_ERR "%s Module with id %d already loaded!\n",
 163                         __func__, ld_type->id);
 164         }
 165         spin_unlock(&pnfs_spinlock);
 166
 167         return status;
 168 }
 169 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 170
 171 void
 172 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 173 {
 174         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 175         spin_lock(&pnfs_spinlock);
 176         list_del(&ld_type->pnfs_tblid);
 177         spin_unlock(&pnfs_spinlock);
 178 }
 179 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 180
 181 /*
 182  * pNFS client layout cache
 183  */
 184
 185 /* Need to hold i_lock if caller does not already hold reference */
 186 void
 187 get_layout_hdr(struct pnfs_layout_hdr *lo)
 188 {
 189         atomic_inc(&lo->plh_refcount);
 190 }
 191
 192 static struct pnfs_layout_hdr *
 193 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 194 {
 195         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 196         return ld->alloc_layout_hdr ? ld->alloc_layout_hdr(ino, gfp_flags) :
 197                 kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
 198 }
 199
 200 static void
 201 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 202 {
 203         struct pnfs_layoutdriver_type *ld = NFS_SERVER(lo->plh_inode)->pnfs_curr_ld;
 204         put_rpccred(lo->plh_lc_cred);
 205         return ld->alloc_layout_hdr ? ld->free_layout_hdr(lo) : kfree(lo);
 206 }
 207
 208 static void
 209 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
 210 {
 211         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 212         BUG_ON(!list_empty(&lo->plh_layouts));
 213         NFS_I(lo->plh_inode)->layout = NULL;
 214         pnfs_free_layout_hdr(lo);
 215 }
 216
 217 static void
 218 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
 219 {
 220         if (atomic_dec_and_test(&lo->plh_refcount))
 221                 destroy_layout_hdr(lo);
 222 }
 223
 224 void
 225 put_layout_hdr(struct pnfs_layout_hdr *lo)
 226 {
 227         struct inode *inode = lo->plh_inode;
 228
 229         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 230                 destroy_layout_hdr(lo);
 231                 spin_unlock(&inode->i_lock);
 232         }
 233 }
 234
 235 static void
 236 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 237 {
 238         INIT_LIST_HEAD(&lseg->pls_list);
 239         INIT_LIST_HEAD(&lseg->pls_lc_list);
 240         atomic_set(&lseg->pls_refcount, 1);
 241         smp_mb();
 242         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 243         lseg->pls_layout = lo;
 244 }
 245
 246 static void free_lseg(struct pnfs_layout_segment *lseg)
 247 {
 248         struct inode *ino = lseg->pls_layout->plh_inode;
 249
 250         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 251         /* Matched by get_layout_hdr in pnfs_insert_layout */
 252         put_layout_hdr(NFS_I(ino)->layout);
 253 }
 254
 255 static void
 256 put_lseg_common(struct pnfs_layout_segment *lseg)
 257 {
 258         struct inode *inode = lseg->pls_layout->plh_inode;
 259
 260         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 261         list_del_init(&lseg->pls_list);
 262         if (list_empty(&lseg->pls_layout->plh_segs)) {
 263                 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
 264                 /* Matched by initial refcount set in alloc_init_layout_hdr */
 265                 put_layout_hdr_locked(lseg->pls_layout);
 266         }
 267         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 268 }
 269
 270 void
 271 put_lseg(struct pnfs_layout_segment *lseg)
 272 {
 273         struct inode *inode;
 274
 275         if (!lseg)
 276                 return;
 277
 278         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 279                 atomic_read(&lseg->pls_refcount),
 280                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 281         inode = lseg->pls_layout->plh_inode;
 282         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 283                 LIST_HEAD(free_me);
 284
 285                 put_lseg_common(lseg);
 286                 list_add(&lseg->pls_list, &free_me);
 287                 spin_unlock(&inode->i_lock);
 288                 pnfs_free_lseg_list(&free_me);
 289         }
 290 }
 291 EXPORT_SYMBOL_GPL(put_lseg);
 292
 293 static inline u64
 294 end_offset(u64 start, u64 len)
 295 {
 296         u64 end;
 297
 298         end = start + len;
 299         return end >= start ? end : NFS4_MAX_UINT64;
 300 }
 301
 302 /* last octet in a range */
 303 static inline u64
 304 last_byte_offset(u64 start, u64 len)
 305 {
 306         u64 end;
 307
 308         BUG_ON(!len);
 309         end = start + len;
 310         return end > start ? end - 1 : NFS4_MAX_UINT64;
 311 }
 312
 313 /*
 314  * is l2 fully contained in l1?
 315  *   start1                             end1
 316  *   [----------------------------------)
 317  *           start2           end2
 318  *           [----------------)
 319  */
 320 static inline int
 321 lo_seg_contained(struct pnfs_layout_range *l1,
 322                  struct pnfs_layout_range *l2)
 323 {
 324         u64 start1 = l1->offset;
 325         u64 end1 = end_offset(start1, l1->length);
 326         u64 start2 = l2->offset;
 327         u64 end2 = end_offset(start2, l2->length);
 328
 329         return (start1 <= start2) && (end1 >= end2);
 330 }
 331
 332 /*
 333  * is l1 and l2 intersecting?
 334  *   start1                             end1
 335  *   [----------------------------------)
 336  *                              start2           end2
 337  *                              [----------------)
 338  */
 339 static inline int
 340 lo_seg_intersecting(struct pnfs_layout_range *l1,
 341                     struct pnfs_layout_range *l2)
 342 {
 343         u64 start1 = l1->offset;
 344         u64 end1 = end_offset(start1, l1->length);
 345         u64 start2 = l2->offset;
 346         u64 end2 = end_offset(start2, l2->length);
 347
 348         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 349                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 350 }
 351
 352 static bool
 353 should_free_lseg(struct pnfs_layout_range *lseg_range,
 354                  struct pnfs_layout_range *recall_range)
 355 {
 356         return (recall_range->iomode == IOMODE_ANY ||
 357                 lseg_range->iomode == recall_range->iomode) &&
 358                lo_seg_intersecting(lseg_range, recall_range);
 359 }
 360
 361 /* Returns 1 if lseg is removed from list, 0 otherwise */
 362 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 363                              struct list_head *tmp_list)
 364 {
 365         int rv = 0;
 366
 367         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 368                 /* Remove the reference keeping the lseg in the
 369                  * list.  It will now be removed when all
 370                  * outstanding io is finished.
 371                  */
 372                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 373                         atomic_read(&lseg->pls_refcount));
 374                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 375                         put_lseg_common(lseg);
 376                         list_add(&lseg->pls_list, tmp_list);
 377                         rv = 1;
 378                 }
 379         }
 380         return rv;
 381 }
 382
 383 /* Returns count of number of matching invalid lsegs remaining in list
 384  * after call.
 385  */
 386 int
 387 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 388                             struct list_head *tmp_list,
 389                             struct pnfs_layout_range *recall_range)
 390 {
 391         struct pnfs_layout_segment *lseg, *next;
 392         int invalid = 0, removed = 0;
 393
 394         dprintk("%s:Begin lo %p\n", __func__, lo);
 395
 396         if (list_empty(&lo->plh_segs)) {
 397                 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
 398                         put_layout_hdr_locked(lo);
 399                 return 0;
 400         }
 401         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 402                 if (!recall_range ||
 403                     should_free_lseg(&lseg->pls_range, recall_range)) {
 404                         dprintk("%s: freeing lseg %p iomode %d "
 405                                 "offset %llu length %llu\n", __func__,
 406                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 407                                 lseg->pls_range.length);
 408                         invalid++;
 409                         removed += mark_lseg_invalid(lseg, tmp_list);
 410                 }
 411         dprintk("%s:Return %i\n", __func__, invalid - removed);
 412         return invalid - removed;
 413 }
 414
 415 /* note free_me must contain lsegs from a single layout_hdr */
 416 void
 417 pnfs_free_lseg_list(struct list_head *free_me)
 418 {
 419         struct pnfs_layout_segment *lseg, *tmp;
 420         struct pnfs_layout_hdr *lo;
 421
 422         if (list_empty(free_me))
 423                 return;
 424
 425         lo = list_first_entry(free_me, struct pnfs_layout_segment,
 426                               pls_list)->pls_layout;
 427
 428         if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
 429                 struct nfs_client *clp;
 430
 431                 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
 432                 spin_lock(&clp->cl_lock);
 433                 list_del_init(&lo->plh_layouts);
 434                 spin_unlock(&clp->cl_lock);
 435         }
 436         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 437                 list_del(&lseg->pls_list);
 438                 free_lseg(lseg);
 439         }
 440 }
 441
 442 void
 443 pnfs_destroy_layout(struct nfs_inode *nfsi)
 444 {
 445         struct pnfs_layout_hdr *lo;
 446         LIST_HEAD(tmp_list);
 447
 448         spin_lock(&nfsi->vfs_inode.i_lock);
 449         lo = nfsi->layout;
 450         if (lo) {
 451                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 452                 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 453         }
 454         spin_unlock(&nfsi->vfs_inode.i_lock);
 455         pnfs_free_lseg_list(&tmp_list);
 456 }
 457
 458 /*
 459  * Called by the state manger to remove all layouts established under an
 460  * expired lease.
 461  */
 462 void
 463 pnfs_destroy_all_layouts(struct nfs_client *clp)
 464 {
 465         struct nfs_server *server;
 466         struct pnfs_layout_hdr *lo;
 467         LIST_HEAD(tmp_list);
 468
 469         nfs4_deviceid_mark_client_invalid(clp);
 470         nfs4_deviceid_purge_client(clp);
 471
 472         spin_lock(&clp->cl_lock);
 473         rcu_read_lock();
 474         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 475                 if (!list_empty(&server->layouts))
 476                         list_splice_init(&server->layouts, &tmp_list);
 477         }
 478         rcu_read_unlock();
 479         spin_unlock(&clp->cl_lock);
 480
 481         while (!list_empty(&tmp_list)) {
 482                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 483                                 plh_layouts);
 484                 dprintk("%s freeing layout for inode %lu\n", __func__,
 485                         lo->plh_inode->i_ino);
 486                 list_del_init(&lo->plh_layouts);
 487                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 488         }
 489 }
 490
 491 /* update lo->plh_stateid with new if is more recent */
 492 void
 493 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 494                         bool update_barrier)
 495 {
 496         u32 oldseq, newseq;
 497
 498         oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
 499         newseq = be32_to_cpu(new->stateid.seqid);
 500         if ((int)(newseq - oldseq) > 0) {
 501                 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
 502                 if (update_barrier) {
 503                         u32 new_barrier = be32_to_cpu(new->stateid.seqid);
 504
 505                         if ((int)(new_barrier - lo->plh_barrier))
 506                                 lo->plh_barrier = new_barrier;
 507                 } else {
 508                         /* Because of wraparound, we want to keep the barrier
 509                          * "close" to the current seqids.  It needs to be
 510                          * within 2**31 to count as "behind", so if it
 511                          * gets too near that limit, give us a litle leeway
 512                          * and bring it to within 2**30.
 513                          * NOTE - and yes, this is all unsigned arithmetic.
 514                          */
 515                         if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
 516                                 lo->plh_barrier = newseq - (1 << 30);
 517                 }
 518         }
 519 }
 520
 521 /* lget is set to 1 if called from inside send_layoutget call chain */
 522 static bool
 523 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
 524                         int lget)
 525 {
 526         if ((stateid) &&
 527             (int)(lo->plh_barrier - be32_to_cpu(stateid->stateid.seqid)) >= 0)
 528                 return true;
 529         return lo->plh_block_lgets ||
 530                 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
 531                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 532                 (list_empty(&lo->plh_segs) &&
 533                  (atomic_read(&lo->plh_outstanding) > lget));
 534 }
 535
 536 int
 537 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 538                               struct nfs4_state *open_state)
 539 {
 540         int status = 0;
 541
 542         dprintk("--> %s\n", __func__);
 543         spin_lock(&lo->plh_inode->i_lock);
 544         if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
 545                 status = -EAGAIN;
 546         } else if (list_empty(&lo->plh_segs)) {
 547                 int seq;
 548
 549                 do {
 550                         seq = read_seqbegin(&open_state->seqlock);
 551                         memcpy(dst->data, open_state->stateid.data,
 552                                sizeof(open_state->stateid.data));
 553                 } while (read_seqretry(&open_state->seqlock, seq));
 554         } else
 555                 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
 556         spin_unlock(&lo->plh_inode->i_lock);
 557         dprintk("<-- %s\n", __func__);
 558         return status;
 559 }
 560
 561 /*
 562 * Get layout from server.
 563 *    for now, assume that whole file layouts are requested.
 564 *    arg->offset: 0
 565 *    arg->length: all ones
 566 */
 567 static struct pnfs_layout_segment *
 568 send_layoutget(struct pnfs_layout_hdr *lo,
 569            struct nfs_open_context *ctx,
 570            struct pnfs_layout_range *range,
 571            gfp_t gfp_flags)
 572 {
 573         struct inode *ino = lo->plh_inode;
 574         struct nfs_server *server = NFS_SERVER(ino);
 575         struct nfs4_layoutget *lgp;
 576         struct pnfs_layout_segment *lseg = NULL;
 577         struct page **pages = NULL;
 578         int i;
 579         u32 max_resp_sz, max_pages;
 580
 581         dprintk("--> %s\n", __func__);
 582
 583         BUG_ON(ctx == NULL);
 584         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 585         if (lgp == NULL)
 586                 return NULL;
 587
 588         /* allocate pages for xdr post processing */
 589         max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
 590         max_pages = max_resp_sz >> PAGE_SHIFT;
 591
 592         pages = kzalloc(max_pages * sizeof(struct page *), gfp_flags);
 593         if (!pages)
 594                 goto out_err_free;
 595
 596         for (i = 0; i < max_pages; i++) {
 597                 pages[i] = alloc_page(gfp_flags);
 598                 if (!pages[i])
 599                         goto out_err_free;
 600         }
 601
 602         lgp->args.minlength = PAGE_CACHE_SIZE;
 603         if (lgp->args.minlength > range->length)
 604                 lgp->args.minlength = range->length;
 605         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 606         lgp->args.range = *range;
 607         lgp->args.type = server->pnfs_curr_ld->id;
 608         lgp->args.inode = ino;
 609         lgp->args.ctx = get_nfs_open_context(ctx);
 610         lgp->args.layout.pages = pages;
 611         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
 612         lgp->lsegpp = &lseg;
 613         lgp->gfp_flags = gfp_flags;
 614
 615         /* Synchronously retrieve layout information from server and
 616          * store in lseg.
 617          */
 618         nfs4_proc_layoutget(lgp);
 619         if (!lseg) {
 620                 /* remember that LAYOUTGET failed and suspend trying */
 621                 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
 622         }
 623
 624         /* free xdr pages */
 625         for (i = 0; i < max_pages; i++)
 626                 __free_page(pages[i]);
 627         kfree(pages);
 628
 629         return lseg;
 630
 631 out_err_free:
 632         /* free any allocated xdr pages, lgp as it's not used */
 633         if (pages) {
 634                 for (i = 0; i < max_pages; i++) {
 635                         if (!pages[i])
 636                                 break;
 637                         __free_page(pages[i]);
 638                 }
 639                 kfree(pages);
 640         }
 641         kfree(lgp);
 642         return NULL;
 643 }
 644
 645 /* Initiates a LAYOUTRETURN(FILE) */
 646 int
 647 _pnfs_return_layout(struct inode *ino)
 648 {
 649         struct pnfs_layout_hdr *lo = NULL;
 650         struct nfs_inode *nfsi = NFS_I(ino);
 651         LIST_HEAD(tmp_list);
 652         struct nfs4_layoutreturn *lrp;
 653         nfs4_stateid stateid;
 654         int status = 0;
 655
 656         dprintk("--> %s\n", __func__);
 657
 658         spin_lock(&ino->i_lock);
 659         lo = nfsi->layout;
 660         if (!lo) {
 661                 spin_unlock(&ino->i_lock);
 662                 dprintk("%s: no layout to return\n", __func__);
 663                 return status;
 664         }
 665         stateid = nfsi->layout->plh_stateid;
 666         /* Reference matched in nfs4_layoutreturn_release */
 667         get_layout_hdr(lo);
 668         mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 669         lo->plh_block_lgets++;
 670         spin_unlock(&ino->i_lock);
 671         pnfs_free_lseg_list(&tmp_list);
 672
 673         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 674
 675         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 676         if (unlikely(lrp == NULL)) {
 677                 status = -ENOMEM;
 678                 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
 679                 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
 680                 put_layout_hdr(lo);
 681                 goto out;
 682         }
 683
 684         lrp->args.stateid = stateid;
 685         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 686         lrp->args.inode = ino;
 687         lrp->args.layout = lo;
 688         lrp->clp = NFS_SERVER(ino)->nfs_client;
 689
 690         status = nfs4_proc_layoutreturn(lrp);
 691 out:
 692         dprintk("<-- %s status: %d\n", __func__, status);
 693         return status;
 694 }
 695
 696 bool pnfs_roc(struct inode *ino)
 697 {
 698         struct pnfs_layout_hdr *lo;
 699         struct pnfs_layout_segment *lseg, *tmp;
 700         LIST_HEAD(tmp_list);
 701         bool found = false;
 702
 703         spin_lock(&ino->i_lock);
 704         lo = NFS_I(ino)->layout;
 705         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 706             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 707                 goto out_nolayout;
 708         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 709                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 710                         mark_lseg_invalid(lseg, &tmp_list);
 711                         found = true;
 712                 }
 713         if (!found)
 714                 goto out_nolayout;
 715         lo->plh_block_lgets++;
 716         get_layout_hdr(lo); /* matched in pnfs_roc_release */
 717         spin_unlock(&ino->i_lock);
 718         pnfs_free_lseg_list(&tmp_list);
 719         return true;
 720
 721 out_nolayout:
 722         spin_unlock(&ino->i_lock);
 723         return false;
 724 }
 725
 726 void pnfs_roc_release(struct inode *ino)
 727 {
 728         struct pnfs_layout_hdr *lo;
 729
 730         spin_lock(&ino->i_lock);
 731         lo = NFS_I(ino)->layout;
 732         lo->plh_block_lgets--;
 733         put_layout_hdr_locked(lo);
 734         spin_unlock(&ino->i_lock);
 735 }
 736
 737 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 738 {
 739         struct pnfs_layout_hdr *lo;
 740
 741         spin_lock(&ino->i_lock);
 742         lo = NFS_I(ino)->layout;
 743         if ((int)(barrier - lo->plh_barrier) > 0)
 744                 lo->plh_barrier = barrier;
 745         spin_unlock(&ino->i_lock);
 746 }
 747
 748 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
 749 {
 750         struct nfs_inode *nfsi = NFS_I(ino);
 751         struct pnfs_layout_segment *lseg;
 752         bool found = false;
 753
 754         spin_lock(&ino->i_lock);
 755         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 756                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 757                         found = true;
 758                         break;
 759                 }
 760         if (!found) {
 761                 struct pnfs_layout_hdr *lo = nfsi->layout;
 762                 u32 current_seqid = be32_to_cpu(lo->plh_stateid.stateid.seqid);
 763
 764                 /* Since close does not return a layout stateid for use as
 765                  * a barrier, we choose the worst-case barrier.
 766                  */
 767                 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 768         }
 769         spin_unlock(&ino->i_lock);
 770         return found;
 771 }
 772
 773 /*
 774  * Compare two layout segments for sorting into layout cache.
 775  * We want to preferentially return RW over RO layouts, so ensure those
 776  * are seen first.
 777  */
 778 static s64
 779 cmp_layout(struct pnfs_layout_range *l1,
 780            struct pnfs_layout_range *l2)
 781 {
 782         s64 d;
 783
 784         /* high offset > low offset */
 785         d = l1->offset - l2->offset;
 786         if (d)
 787                 return d;
 788
 789         /* short length > long length */
 790         d = l2->length - l1->length;
 791         if (d)
 792                 return d;
 793
 794         /* read > read/write */
 795         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 796 }
 797
 798 static void
 799 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
 800                    struct pnfs_layout_segment *lseg)
 801 {
 802         struct pnfs_layout_segment *lp;
 803
 804         dprintk("%s:Begin\n", __func__);
 805
 806         assert_spin_locked(&lo->plh_inode->i_lock);
 807         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 808                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 809                         continue;
 810                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 811                 dprintk("%s: inserted lseg %p "
 812                         "iomode %d offset %llu length %llu before "
 813                         "lp %p iomode %d offset %llu length %llu\n",
 814                         __func__, lseg, lseg->pls_range.iomode,
 815                         lseg->pls_range.offset, lseg->pls_range.length,
 816                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 817                         lp->pls_range.length);
 818                 goto out;
 819         }
 820         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 821         dprintk("%s: inserted lseg %p "
 822                 "iomode %d offset %llu length %llu at tail\n",
 823                 __func__, lseg, lseg->pls_range.iomode,
 824                 lseg->pls_range.offset, lseg->pls_range.length);
 825 out:
 826         get_layout_hdr(lo);
 827
 828         dprintk("%s:Return\n", __func__);
 829 }
 830
 831 static struct pnfs_layout_hdr *
 832 alloc_init_layout_hdr(struct inode *ino,
 833                       struct nfs_open_context *ctx,
 834                       gfp_t gfp_flags)
 835 {
 836         struct pnfs_layout_hdr *lo;
 837
 838         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 839         if (!lo)
 840                 return NULL;
 841         atomic_set(&lo->plh_refcount, 1);
 842         INIT_LIST_HEAD(&lo->plh_layouts);
 843         INIT_LIST_HEAD(&lo->plh_segs);
 844         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 845         lo->plh_inode = ino;
 846         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 847         return lo;
 848 }
 849
 850 static struct pnfs_layout_hdr *
 851 pnfs_find_alloc_layout(struct inode *ino,
 852                        struct nfs_open_context *ctx,
 853                        gfp_t gfp_flags)
 854 {
 855         struct nfs_inode *nfsi = NFS_I(ino);
 856         struct pnfs_layout_hdr *new = NULL;
 857
 858         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 859
 860         assert_spin_locked(&ino->i_lock);
 861         if (nfsi->layout) {
 862                 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
 863                         return NULL;
 864                 else
 865                         return nfsi->layout;
 866         }
 867         spin_unlock(&ino->i_lock);
 868         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 869         spin_lock(&ino->i_lock);
 870
 871         if (likely(nfsi->layout == NULL))       /* Won the race? */
 872                 nfsi->layout = new;
 873         else
 874                 pnfs_free_layout_hdr(new);
 875         return nfsi->layout;
 876 }
 877
 878 /*
 879  * iomode matching rules:
 880  * iomode       lseg    match
 881  * -----        -----   -----
 882  * ANY          READ    true
 883  * ANY          RW      true
 884  * RW           READ    false
 885  * RW           RW      true
 886  * READ         READ    true
 887  * READ         RW      true
 888  */
 889 static int
 890 is_matching_lseg(struct pnfs_layout_range *ls_range,
 891                  struct pnfs_layout_range *range)
 892 {
 893         struct pnfs_layout_range range1;
 894
 895         if ((range->iomode == IOMODE_RW &&
 896              ls_range->iomode != IOMODE_RW) ||
 897             !lo_seg_intersecting(ls_range, range))
 898                 return 0;
 899
 900         /* range1 covers only the first byte in the range */
 901         range1 = *range;
 902         range1.length = 1;
 903         return lo_seg_contained(ls_range, &range1);
 904 }
 905
 906 /*
 907  * lookup range in layout
 908  */
 909 static struct pnfs_layout_segment *
 910 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 911                 struct pnfs_layout_range *range)
 912 {
 913         struct pnfs_layout_segment *lseg, *ret = NULL;
 914
 915         dprintk("%s:Begin\n", __func__);
 916
 917         assert_spin_locked(&lo->plh_inode->i_lock);
 918         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 919                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 920                     is_matching_lseg(&lseg->pls_range, range)) {
 921                         ret = get_lseg(lseg);
 922                         break;
 923                 }
 924                 if (lseg->pls_range.offset > range->offset)
 925                         break;
 926         }
 927
 928         dprintk("%s:Return lseg %p ref %d\n",
 929                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 930         return ret;
 931 }
 932
 933 /*
 934  * Layout segment is retreived from the server if not cached.
 935  * The appropriate layout segment is referenced and returned to the caller.
 936  */
 937 struct pnfs_layout_segment *
 938 pnfs_update_layout(struct inode *ino,
 939                    struct nfs_open_context *ctx,
 940                    loff_t pos,
 941                    u64 count,
 942                    enum pnfs_iomode iomode,
 943                    gfp_t gfp_flags)
 944 {
 945         struct pnfs_layout_range arg = {
 946                 .iomode = iomode,
 947                 .offset = pos,
 948                 .length = count,
 949         };
 950         unsigned pg_offset;
 951         struct nfs_inode *nfsi = NFS_I(ino);
 952         struct nfs_server *server = NFS_SERVER(ino);
 953         struct nfs_client *clp = server->nfs_client;
 954         struct pnfs_layout_hdr *lo;
 955         struct pnfs_layout_segment *lseg = NULL;
 956         bool first = false;
 957
 958         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
 959                 return NULL;
 960         spin_lock(&ino->i_lock);
 961         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
 962         if (lo == NULL) {
 963                 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
 964                 goto out_unlock;
 965         }
 966
 967         /* Do we even need to bother with this? */
 968         if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
 969             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
 970                 dprintk("%s matches recall, use MDS\n", __func__);
 971                 goto out_unlock;
 972         }
 973
 974         /* if LAYOUTGET already failed once we don't try again */
 975         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
 976                 goto out_unlock;
 977
 978         /* Check to see if the layout for the given range already exists */
 979         lseg = pnfs_find_lseg(lo, &arg);
 980         if (lseg)
 981                 goto out_unlock;
 982
 983         if (pnfs_layoutgets_blocked(lo, NULL, 0))
 984                 goto out_unlock;
 985         atomic_inc(&lo->plh_outstanding);
 986
 987         get_layout_hdr(lo);
 988         if (list_empty(&lo->plh_segs))
 989                 first = true;
 990         spin_unlock(&ino->i_lock);
 991         if (first) {
 992                 /* The lo must be on the clp list if there is any
 993                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
 994                  */
 995                 spin_lock(&clp->cl_lock);
 996                 BUG_ON(!list_empty(&lo->plh_layouts));
 997                 list_add_tail(&lo->plh_layouts, &server->layouts);
 998                 spin_unlock(&clp->cl_lock);
 999         }
1000
1001         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1002         if (pg_offset) {
1003                 arg.offset -= pg_offset;
1004                 arg.length += pg_offset;
1005         }
1006         if (arg.length != NFS4_MAX_UINT64)
1007                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1008
1009         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1010         if (!lseg && first) {
1011                 spin_lock(&clp->cl_lock);
1012                 list_del_init(&lo->plh_layouts);
1013                 spin_unlock(&clp->cl_lock);
1014         }
1015         atomic_dec(&lo->plh_outstanding);
1016         put_layout_hdr(lo);
1017 out:
1018         dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1019                 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1020         return lseg;
1021 out_unlock:
1022         spin_unlock(&ino->i_lock);
1023         goto out;
1024 }
1025 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1026
1027 int
1028 pnfs_layout_process(struct nfs4_layoutget *lgp)
1029 {
1030         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1031         struct nfs4_layoutget_res *res = &lgp->res;
1032         struct pnfs_layout_segment *lseg;
1033         struct inode *ino = lo->plh_inode;
1034         struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
1035         int status = 0;
1036
1037         /* Inject layout blob into I/O device driver */
1038         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1039         if (!lseg || IS_ERR(lseg)) {
1040                 if (!lseg)
1041                         status = -ENOMEM;
1042                 else
1043                         status = PTR_ERR(lseg);
1044                 dprintk("%s: Could not allocate layout: error %d\n",
1045                        __func__, status);
1046                 goto out;
1047         }
1048
1049         spin_lock(&ino->i_lock);
1050         if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
1051             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1052                 dprintk("%s forget reply due to recall\n", __func__);
1053                 goto out_forget_reply;
1054         }
1055
1056         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1057                 dprintk("%s forget reply due to state\n", __func__);
1058                 goto out_forget_reply;
1059         }
1060         init_lseg(lo, lseg);
1061         lseg->pls_range = res->range;
1062         *lgp->lsegpp = get_lseg(lseg);
1063         pnfs_insert_layout(lo, lseg);
1064
1065         if (res->return_on_close) {
1066                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1067                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1068         }
1069
1070         /* Done processing layoutget. Set the layout stateid */
1071         pnfs_set_layout_stateid(lo, &res->stateid, false);
1072         spin_unlock(&ino->i_lock);
1073 out:
1074         return status;
1075
1076 out_forget_reply:
1077         spin_unlock(&ino->i_lock);
1078         lseg->pls_layout = lo;
1079         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1080         goto out;
1081 }
1082
1083 void
1084 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1085 {
1086         BUG_ON(pgio->pg_lseg != NULL);
1087
1088         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1089                                            req->wb_context,
1090                                            req_offset(req),
1091                                            req->wb_bytes,
1092                                            IOMODE_READ,
1093                                            GFP_KERNEL);
1094         /* If no lseg, fall back to read through mds */
1095         if (pgio->pg_lseg == NULL)
1096                 nfs_pageio_reset_read_mds(pgio);
1097
1098 }
1099 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1100
1101 void
1102 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1103 {
1104         BUG_ON(pgio->pg_lseg != NULL);
1105
1106         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1107                                            req->wb_context,
1108                                            req_offset(req),
1109                                            req->wb_bytes,
1110                                            IOMODE_RW,
1111                                            GFP_NOFS);
1112         /* If no lseg, fall back to write through mds */
1113         if (pgio->pg_lseg == NULL)
1114                 nfs_pageio_reset_write_mds(pgio);
1115 }
1116 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1117
1118 bool
1119 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
1120 {
1121         struct nfs_server *server = NFS_SERVER(inode);
1122         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1123
1124         if (ld == NULL)
1125                 return false;
1126         nfs_pageio_init(pgio, inode, ld->pg_read_ops, server->rsize, 0);
1127         return true;
1128 }
1129
1130 bool
1131 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode, int ioflags)
1132 {
1133         struct nfs_server *server = NFS_SERVER(inode);
1134         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1135
1136         if (ld == NULL)
1137                 return false;
1138         nfs_pageio_init(pgio, inode, ld->pg_write_ops, server->wsize, ioflags);
1139         return true;
1140 }
1141
1142 bool
1143 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1144                      struct nfs_page *req)
1145 {
1146         if (pgio->pg_lseg == NULL)
1147                 return nfs_generic_pg_test(pgio, prev, req);
1148
1149         /*
1150          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1151          * Note that this test makes several assumptions:
1152          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1153          *   is known to lie within the range.
1154          *   - that the nfs_page being tested is known to be contiguous with the
1155          *   previous nfs_page.
1156          *   - Layout ranges are page aligned, so we only have to test the
1157          *   start offset of the request.
1158          *
1159          * Please also note that 'end_offset' is actually the offset of the
1160          * first byte that lies outside the pnfs_layout_range. FIXME?
1161          *
1162          */
1163         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1164                                          pgio->pg_lseg->pls_range.length);
1165 }
1166 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1167
1168 /*
1169  * Called by non rpc-based layout drivers
1170  */
1171 void pnfs_ld_write_done(struct nfs_write_data *data)
1172 {
1173         if (likely(!data->pnfs_error)) {
1174                 pnfs_set_layoutcommit(data);
1175                 data->mds_ops->rpc_call_done(&data->task, data);
1176         } else {
1177                 put_lseg(data->lseg);
1178                 data->lseg = NULL;
1179                 dprintk("pnfs write error = %d\n", data->pnfs_error);
1180         }
1181         data->mds_ops->rpc_release(data);
1182 }
1183 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1184
1185 static void
1186 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1187                 struct nfs_write_data *data)
1188 {
1189         list_splice_tail_init(&data->pages, &desc->pg_list);
1190         if (data->req && list_empty(&data->req->wb_list))
1191                 nfs_list_add_request(data->req, &desc->pg_list);
1192         nfs_pageio_reset_write_mds(desc);
1193         desc->pg_recoalesce = 1;
1194         nfs_writedata_release(data);
1195 }
1196
1197 static enum pnfs_try_status
1198 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1199                         const struct rpc_call_ops *call_ops,
1200                         struct pnfs_layout_segment *lseg,
1201                         int how)
1202 {
1203         struct inode *inode = wdata->inode;
1204         enum pnfs_try_status trypnfs;
1205         struct nfs_server *nfss = NFS_SERVER(inode);
1206
1207         wdata->mds_ops = call_ops;
1208         wdata->lseg = get_lseg(lseg);
1209
1210         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1211                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1212
1213         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1214         if (trypnfs == PNFS_NOT_ATTEMPTED) {
1215                 put_lseg(wdata->lseg);
1216                 wdata->lseg = NULL;
1217         } else
1218                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1219
1220         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1221         return trypnfs;
1222 }
1223
1224 static void
1225 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1226 {
1227         struct nfs_write_data *data;
1228         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1229         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1230
1231         desc->pg_lseg = NULL;
1232         while (!list_empty(head)) {
1233                 enum pnfs_try_status trypnfs;
1234
1235                 data = list_entry(head->next, struct nfs_write_data, list);
1236                 list_del_init(&data->list);
1237
1238                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1239                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1240                         pnfs_write_through_mds(desc, data);
1241         }
1242         put_lseg(lseg);
1243 }
1244
1245 int
1246 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1247 {
1248         LIST_HEAD(head);
1249         int ret;
1250
1251         ret = nfs_generic_flush(desc, &head);
1252         if (ret != 0) {
1253                 put_lseg(desc->pg_lseg);
1254                 desc->pg_lseg = NULL;
1255                 return ret;
1256         }
1257         pnfs_do_multiple_writes(desc, &head, desc->pg_ioflags);
1258         return 0;
1259 }
1260 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1261
1262 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1263 {
1264         struct nfs_pageio_descriptor pgio;
1265
1266         put_lseg(data->lseg);
1267         data->lseg = NULL;
1268         dprintk("pnfs write error = %d\n", data->pnfs_error);
1269
1270         nfs_pageio_init_read_mds(&pgio, data->inode);
1271
1272         while (!list_empty(&data->pages)) {
1273                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1274
1275                 nfs_list_remove_request(req);
1276                 nfs_pageio_add_request(&pgio, req);
1277         }
1278         nfs_pageio_complete(&pgio);
1279 }
1280
1281 /*
1282  * Called by non rpc-based layout drivers
1283  */
1284 void pnfs_ld_read_done(struct nfs_read_data *data)
1285 {
1286         if (likely(!data->pnfs_error)) {
1287                 __nfs4_read_done_cb(data);
1288                 data->mds_ops->rpc_call_done(&data->task, data);
1289         } else
1290                 pnfs_ld_handle_read_error(data);
1291         data->mds_ops->rpc_release(data);
1292 }
1293 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1294
1295 static void
1296 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1297                 struct nfs_read_data *data)
1298 {
1299         list_splice_tail_init(&data->pages, &desc->pg_list);
1300         if (data->req && list_empty(&data->req->wb_list))
1301                 nfs_list_add_request(data->req, &desc->pg_list);
1302         nfs_pageio_reset_read_mds(desc);
1303         desc->pg_recoalesce = 1;
1304         nfs_readdata_release(data);
1305 }
1306
1307 /*
1308  * Call the appropriate parallel I/O subsystem read function.
1309  */
1310 static enum pnfs_try_status
1311 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1312                        const struct rpc_call_ops *call_ops,
1313                        struct pnfs_layout_segment *lseg)
1314 {
1315         struct inode *inode = rdata->inode;
1316         struct nfs_server *nfss = NFS_SERVER(inode);
1317         enum pnfs_try_status trypnfs;
1318
1319         rdata->mds_ops = call_ops;
1320         rdata->lseg = get_lseg(lseg);
1321
1322         dprintk("%s: Reading ino:%lu %u@%llu\n",
1323                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1324
1325         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1326         if (trypnfs == PNFS_NOT_ATTEMPTED) {
1327                 put_lseg(rdata->lseg);
1328                 rdata->lseg = NULL;
1329         } else {
1330                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1331         }
1332         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1333         return trypnfs;
1334 }
1335
1336 static void
1337 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1338 {
1339         struct nfs_read_data *data;
1340         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1341         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1342
1343         desc->pg_lseg = NULL;
1344         while (!list_empty(head)) {
1345                 enum pnfs_try_status trypnfs;
1346
1347                 data = list_entry(head->next, struct nfs_read_data, list);
1348                 list_del_init(&data->list);
1349
1350                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1351                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1352                         pnfs_read_through_mds(desc, data);
1353         }
1354         put_lseg(lseg);
1355 }
1356
1357 int
1358 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1359 {
1360         LIST_HEAD(head);
1361         int ret;
1362
1363         ret = nfs_generic_pagein(desc, &head);
1364         if (ret != 0) {
1365                 put_lseg(desc->pg_lseg);
1366                 desc->pg_lseg = NULL;
1367                 return ret;
1368         }
1369         pnfs_do_multiple_reads(desc, &head);
1370         return 0;
1371 }
1372 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1373
1374 /*
1375  * There can be multiple RW segments.
1376  */
1377 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1378 {
1379         struct pnfs_layout_segment *lseg;
1380
1381         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1382                 if (lseg->pls_range.iomode == IOMODE_RW &&
1383                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1384                         list_add(&lseg->pls_lc_list, listp);
1385         }
1386 }
1387
1388 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1389 {
1390         if (lseg->pls_range.iomode == IOMODE_RW) {
1391                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1392                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1393         } else {
1394                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1395                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1396         }
1397 }
1398 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1399
1400 void
1401 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1402 {
1403         struct nfs_inode *nfsi = NFS_I(wdata->inode);
1404         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1405         bool mark_as_dirty = false;
1406
1407         spin_lock(&nfsi->vfs_inode.i_lock);
1408         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1409                 mark_as_dirty = true;
1410                 dprintk("%s: Set layoutcommit for inode %lu ",
1411                         __func__, wdata->inode->i_ino);
1412         }
1413         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &wdata->lseg->pls_flags)) {
1414                 /* references matched in nfs4_layoutcommit_release */
1415                 get_lseg(wdata->lseg);
1416         }
1417         if (end_pos > nfsi->layout->plh_lwb)
1418                 nfsi->layout->plh_lwb = end_pos;
1419         spin_unlock(&nfsi->vfs_inode.i_lock);
1420         dprintk("%s: lseg %p end_pos %llu\n",
1421                 __func__, wdata->lseg, nfsi->layout->plh_lwb);
1422
1423         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1424          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1425         if (mark_as_dirty)
1426                 mark_inode_dirty_sync(wdata->inode);
1427 }
1428 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1429
1430 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1431 {
1432         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1433
1434         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1435                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1436 }
1437
1438 /*
1439  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1440  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1441  * data to disk to allow the server to recover the data if it crashes.
1442  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1443  * is off, and a COMMIT is sent to a data server, or
1444  * if WRITEs to a data server return NFS_DATA_SYNC.
1445  */
1446 int
1447 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1448 {
1449         struct nfs4_layoutcommit_data *data;
1450         struct nfs_inode *nfsi = NFS_I(inode);
1451         loff_t end_pos;
1452         int status = 0;
1453
1454         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1455
1456         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1457                 return 0;
1458
1459         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1460         data = kzalloc(sizeof(*data), GFP_NOFS);
1461         if (!data) {
1462                 status = -ENOMEM;
1463                 goto out;
1464         }
1465
1466         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1467                 goto out_free;
1468
1469         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1470                 if (!sync) {
1471                         status = -EAGAIN;
1472                         goto out_free;
1473                 }
1474                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1475                                         nfs_wait_bit_killable, TASK_KILLABLE);
1476                 if (status)
1477                         goto out_free;
1478         }
1479
1480         INIT_LIST_HEAD(&data->lseg_list);
1481         spin_lock(&inode->i_lock);
1482         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1483                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1484                 spin_unlock(&inode->i_lock);
1485                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1486                 goto out_free;
1487         }
1488
1489         pnfs_list_write_lseg(inode, &data->lseg_list);
1490
1491         end_pos = nfsi->layout->plh_lwb;
1492         nfsi->layout->plh_lwb = 0;
1493
1494         memcpy(&data->args.stateid.data, nfsi->layout->plh_stateid.data,
1495                 sizeof(nfsi->layout->plh_stateid.data));
1496         spin_unlock(&inode->i_lock);
1497
1498         data->args.inode = inode;
1499         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1500         nfs_fattr_init(&data->fattr);
1501         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1502         data->res.fattr = &data->fattr;
1503         data->args.lastbytewritten = end_pos - 1;
1504         data->res.server = NFS_SERVER(inode);
1505
1506         status = nfs4_proc_layoutcommit(data, sync);
1507 out:
1508         if (status)
1509                 mark_inode_dirty_sync(inode);
1510         dprintk("<-- %s status %d\n", __func__, status);
1511         return status;
1512 out_free:
1513         kfree(data);
1514         goto out;
1515 }