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