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.seqid);
 500         newseq = be32_to_cpu(new->seqid);
 501         if ((int)(newseq - oldseq) > 0) {
 502                 nfs4_stateid_copy(&lo->plh_stateid, new);
 503                 if (update_barrier) {
 504                         u32 new_barrier = be32_to_cpu(new->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->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                         nfs4_stateid_copy(dst, &open_state->stateid);
 553                 } while (read_seqretry(&open_state->seqlock, seq));
 554         } else
 555                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 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 = kcalloc(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.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(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
 969                 dprintk("%s matches recall, use MDS\n", __func__);
 970                 goto out_unlock;
 971         }
 972
 973         /* if LAYOUTGET already failed once we don't try again */
 974         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
 975                 goto out_unlock;
 976
 977         /* Check to see if the layout for the given range already exists */
 978         lseg = pnfs_find_lseg(lo, &arg);
 979         if (lseg)
 980                 goto out_unlock;
 981
 982         if (pnfs_layoutgets_blocked(lo, NULL, 0))
 983                 goto out_unlock;
 984         atomic_inc(&lo->plh_outstanding);
 985
 986         get_layout_hdr(lo);
 987         if (list_empty(&lo->plh_segs))
 988                 first = true;
 989         spin_unlock(&ino->i_lock);
 990         if (first) {
 991                 /* The lo must be on the clp list if there is any
 992                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
 993                  */
 994                 spin_lock(&clp->cl_lock);
 995                 BUG_ON(!list_empty(&lo->plh_layouts));
 996                 list_add_tail(&lo->plh_layouts, &server->layouts);
 997                 spin_unlock(&clp->cl_lock);
 998         }
 999
1000         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1001         if (pg_offset) {
1002                 arg.offset -= pg_offset;
1003                 arg.length += pg_offset;
1004         }
1005         if (arg.length != NFS4_MAX_UINT64)
1006                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1007
1008         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1009         if (!lseg && first) {
1010                 spin_lock(&clp->cl_lock);
1011                 list_del_init(&lo->plh_layouts);
1012                 spin_unlock(&clp->cl_lock);
1013         }
1014         atomic_dec(&lo->plh_outstanding);
1015         put_layout_hdr(lo);
1016 out:
1017         dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1018                 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1019         return lseg;
1020 out_unlock:
1021         spin_unlock(&ino->i_lock);
1022         goto out;
1023 }
1024 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1025
1026 int
1027 pnfs_layout_process(struct nfs4_layoutget *lgp)
1028 {
1029         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1030         struct nfs4_layoutget_res *res = &lgp->res;
1031         struct pnfs_layout_segment *lseg;
1032         struct inode *ino = lo->plh_inode;
1033         int status = 0;
1034
1035         /* Inject layout blob into I/O device driver */
1036         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1037         if (!lseg || IS_ERR(lseg)) {
1038                 if (!lseg)
1039                         status = -ENOMEM;
1040                 else
1041                         status = PTR_ERR(lseg);
1042                 dprintk("%s: Could not allocate layout: error %d\n",
1043                        __func__, status);
1044                 goto out;
1045         }
1046
1047         spin_lock(&ino->i_lock);
1048         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1049                 dprintk("%s forget reply due to recall\n", __func__);
1050                 goto out_forget_reply;
1051         }
1052
1053         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1054                 dprintk("%s forget reply due to state\n", __func__);
1055                 goto out_forget_reply;
1056         }
1057         init_lseg(lo, lseg);
1058         lseg->pls_range = res->range;
1059         *lgp->lsegpp = get_lseg(lseg);
1060         pnfs_insert_layout(lo, lseg);
1061
1062         if (res->return_on_close) {
1063                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1064                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1065         }
1066
1067         /* Done processing layoutget. Set the layout stateid */
1068         pnfs_set_layout_stateid(lo, &res->stateid, false);
1069         spin_unlock(&ino->i_lock);
1070 out:
1071         return status;
1072
1073 out_forget_reply:
1074         spin_unlock(&ino->i_lock);
1075         lseg->pls_layout = lo;
1076         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1077         goto out;
1078 }
1079
1080 void
1081 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1082 {
1083         BUG_ON(pgio->pg_lseg != NULL);
1084
1085         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1086                                            req->wb_context,
1087                                            req_offset(req),
1088                                            req->wb_bytes,
1089                                            IOMODE_READ,
1090                                            GFP_KERNEL);
1091         /* If no lseg, fall back to read through mds */
1092         if (pgio->pg_lseg == NULL)
1093                 nfs_pageio_reset_read_mds(pgio);
1094
1095 }
1096 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1097
1098 void
1099 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1100 {
1101         BUG_ON(pgio->pg_lseg != NULL);
1102
1103         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1104                                            req->wb_context,
1105                                            req_offset(req),
1106                                            req->wb_bytes,
1107                                            IOMODE_RW,
1108                                            GFP_NOFS);
1109         /* If no lseg, fall back to write through mds */
1110         if (pgio->pg_lseg == NULL)
1111                 nfs_pageio_reset_write_mds(pgio);
1112 }
1113 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1114
1115 bool
1116 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
1117 {
1118         struct nfs_server *server = NFS_SERVER(inode);
1119         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1120
1121         if (ld == NULL)
1122                 return false;
1123         nfs_pageio_init(pgio, inode, ld->pg_read_ops, server->rsize, 0);
1124         return true;
1125 }
1126
1127 bool
1128 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode, int ioflags)
1129 {
1130         struct nfs_server *server = NFS_SERVER(inode);
1131         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1132
1133         if (ld == NULL)
1134                 return false;
1135         nfs_pageio_init(pgio, inode, ld->pg_write_ops, server->wsize, ioflags);
1136         return true;
1137 }
1138
1139 bool
1140 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1141                      struct nfs_page *req)
1142 {
1143         if (pgio->pg_lseg == NULL)
1144                 return nfs_generic_pg_test(pgio, prev, req);
1145
1146         /*
1147          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1148          * Note that this test makes several assumptions:
1149          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1150          *   is known to lie within the range.
1151          *   - that the nfs_page being tested is known to be contiguous with the
1152          *   previous nfs_page.
1153          *   - Layout ranges are page aligned, so we only have to test the
1154          *   start offset of the request.
1155          *
1156          * Please also note that 'end_offset' is actually the offset of the
1157          * first byte that lies outside the pnfs_layout_range. FIXME?
1158          *
1159          */
1160         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1161                                          pgio->pg_lseg->pls_range.length);
1162 }
1163 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1164
1165 static int pnfs_write_done_resend_to_mds(struct inode *inode, struct list_head *head)
1166 {
1167         struct nfs_pageio_descriptor pgio;
1168         LIST_HEAD(failed);
1169
1170         /* Resend all requests through the MDS */
1171         nfs_pageio_init_write_mds(&pgio, inode, FLUSH_STABLE);
1172         while (!list_empty(head)) {
1173                 struct nfs_page *req = nfs_list_entry(head->next);
1174
1175                 nfs_list_remove_request(req);
1176                 if (!nfs_pageio_add_request(&pgio, req))
1177                         nfs_list_add_request(req, &failed);
1178         }
1179         nfs_pageio_complete(&pgio);
1180
1181         if (!list_empty(&failed)) {
1182                 /* For some reason our attempt to resend pages. Mark the
1183                  * overall send request as having failed, and let
1184                  * nfs_writeback_release_full deal with the error.
1185                  */
1186                 list_move(&failed, head);
1187                 return -EIO;
1188         }
1189         return 0;
1190 }
1191
1192 /*
1193  * Called by non rpc-based layout drivers
1194  */
1195 void pnfs_ld_write_done(struct nfs_write_data *data)
1196 {
1197         if (likely(!data->pnfs_error)) {
1198                 pnfs_set_layoutcommit(data);
1199                 data->mds_ops->rpc_call_done(&data->task, data);
1200         } else {
1201                 dprintk("pnfs write error = %d\n", data->pnfs_error);
1202                 if (NFS_SERVER(data->inode)->pnfs_curr_ld->flags &
1203                                                 PNFS_LAYOUTRET_ON_ERROR) {
1204                         /* Don't lo_commit on error, Server will needs to
1205                          * preform a file recovery.
1206                          */
1207                         clear_bit(NFS_INO_LAYOUTCOMMIT,
1208                                   &NFS_I(data->inode)->flags);
1209                         pnfs_return_layout(data->inode);
1210                 }
1211                 data->task.tk_status = pnfs_write_done_resend_to_mds(data->inode, &data->pages);
1212         }
1213         put_lseg(data->lseg);
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         put_lseg(data->lseg);
1228         nfs_writedata_release(data);
1229 }
1230
1231 static enum pnfs_try_status
1232 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1233                         const struct rpc_call_ops *call_ops,
1234                         struct pnfs_layout_segment *lseg,
1235                         int how)
1236 {
1237         struct inode *inode = wdata->inode;
1238         enum pnfs_try_status trypnfs;
1239         struct nfs_server *nfss = NFS_SERVER(inode);
1240
1241         wdata->mds_ops = call_ops;
1242         wdata->lseg = get_lseg(lseg);
1243
1244         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1245                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1246
1247         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1248         if (trypnfs == PNFS_NOT_ATTEMPTED) {
1249                 put_lseg(wdata->lseg);
1250                 wdata->lseg = NULL;
1251         } else
1252                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1253
1254         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1255         return trypnfs;
1256 }
1257
1258 static void
1259 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1260 {
1261         struct nfs_write_data *data;
1262         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1263         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1264
1265         desc->pg_lseg = NULL;
1266         while (!list_empty(head)) {
1267                 enum pnfs_try_status trypnfs;
1268
1269                 data = list_entry(head->next, struct nfs_write_data, list);
1270                 list_del_init(&data->list);
1271
1272                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1273                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1274                         pnfs_write_through_mds(desc, data);
1275         }
1276         put_lseg(lseg);
1277 }
1278
1279 int
1280 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1281 {
1282         LIST_HEAD(head);
1283         int ret;
1284
1285         ret = nfs_generic_flush(desc, &head);
1286         if (ret != 0) {
1287                 put_lseg(desc->pg_lseg);
1288                 desc->pg_lseg = NULL;
1289                 return ret;
1290         }
1291         pnfs_do_multiple_writes(desc, &head, desc->pg_ioflags);
1292         return 0;
1293 }
1294 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1295
1296 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1297 {
1298         struct nfs_pageio_descriptor pgio;
1299
1300         put_lseg(data->lseg);
1301         data->lseg = NULL;
1302         dprintk("pnfs write error = %d\n", data->pnfs_error);
1303         if (NFS_SERVER(data->inode)->pnfs_curr_ld->flags &
1304                                                 PNFS_LAYOUTRET_ON_ERROR)
1305                 pnfs_return_layout(data->inode);
1306
1307         nfs_pageio_init_read_mds(&pgio, data->inode);
1308
1309         while (!list_empty(&data->pages)) {
1310                 struct nfs_page *req = nfs_list_entry(data->pages.next);
1311
1312                 nfs_list_remove_request(req);
1313                 nfs_pageio_add_request(&pgio, req);
1314         }
1315         nfs_pageio_complete(&pgio);
1316 }
1317
1318 /*
1319  * Called by non rpc-based layout drivers
1320  */
1321 void pnfs_ld_read_done(struct nfs_read_data *data)
1322 {
1323         if (likely(!data->pnfs_error)) {
1324                 __nfs4_read_done_cb(data);
1325                 data->mds_ops->rpc_call_done(&data->task, data);
1326         } else
1327                 pnfs_ld_handle_read_error(data);
1328         put_lseg(data->lseg);
1329         data->mds_ops->rpc_release(data);
1330 }
1331 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1332
1333 static void
1334 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1335                 struct nfs_read_data *data)
1336 {
1337         list_splice_tail_init(&data->pages, &desc->pg_list);
1338         if (data->req && list_empty(&data->req->wb_list))
1339                 nfs_list_add_request(data->req, &desc->pg_list);
1340         nfs_pageio_reset_read_mds(desc);
1341         desc->pg_recoalesce = 1;
1342         nfs_readdata_release(data);
1343 }
1344
1345 /*
1346  * Call the appropriate parallel I/O subsystem read function.
1347  */
1348 static enum pnfs_try_status
1349 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1350                        const struct rpc_call_ops *call_ops,
1351                        struct pnfs_layout_segment *lseg)
1352 {
1353         struct inode *inode = rdata->inode;
1354         struct nfs_server *nfss = NFS_SERVER(inode);
1355         enum pnfs_try_status trypnfs;
1356
1357         rdata->mds_ops = call_ops;
1358         rdata->lseg = get_lseg(lseg);
1359
1360         dprintk("%s: Reading ino:%lu %u@%llu\n",
1361                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1362
1363         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1364         if (trypnfs == PNFS_NOT_ATTEMPTED) {
1365                 put_lseg(rdata->lseg);
1366                 rdata->lseg = NULL;
1367         } else {
1368                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1369         }
1370         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1371         return trypnfs;
1372 }
1373
1374 static void
1375 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1376 {
1377         struct nfs_read_data *data;
1378         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1379         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1380
1381         desc->pg_lseg = NULL;
1382         while (!list_empty(head)) {
1383                 enum pnfs_try_status trypnfs;
1384
1385                 data = list_entry(head->next, struct nfs_read_data, list);
1386                 list_del_init(&data->list);
1387
1388                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1389                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1390                         pnfs_read_through_mds(desc, data);
1391         }
1392         put_lseg(lseg);
1393 }
1394
1395 int
1396 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1397 {
1398         LIST_HEAD(head);
1399         int ret;
1400
1401         ret = nfs_generic_pagein(desc, &head);
1402         if (ret != 0) {
1403                 put_lseg(desc->pg_lseg);
1404                 desc->pg_lseg = NULL;
1405                 return ret;
1406         }
1407         pnfs_do_multiple_reads(desc, &head);
1408         return 0;
1409 }
1410 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1411
1412 /*
1413  * There can be multiple RW segments.
1414  */
1415 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1416 {
1417         struct pnfs_layout_segment *lseg;
1418
1419         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1420                 if (lseg->pls_range.iomode == IOMODE_RW &&
1421                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1422                         list_add(&lseg->pls_lc_list, listp);
1423         }
1424 }
1425
1426 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1427 {
1428         if (lseg->pls_range.iomode == IOMODE_RW) {
1429                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1430                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1431         } else {
1432                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1433                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1434         }
1435 }
1436 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1437
1438 void
1439 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1440 {
1441         struct nfs_inode *nfsi = NFS_I(wdata->inode);
1442         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1443         bool mark_as_dirty = false;
1444
1445         spin_lock(&nfsi->vfs_inode.i_lock);
1446         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1447                 mark_as_dirty = true;
1448                 dprintk("%s: Set layoutcommit for inode %lu ",
1449                         __func__, wdata->inode->i_ino);
1450         }
1451         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &wdata->lseg->pls_flags)) {
1452                 /* references matched in nfs4_layoutcommit_release */
1453                 get_lseg(wdata->lseg);
1454         }
1455         if (end_pos > nfsi->layout->plh_lwb)
1456                 nfsi->layout->plh_lwb = end_pos;
1457         spin_unlock(&nfsi->vfs_inode.i_lock);
1458         dprintk("%s: lseg %p end_pos %llu\n",
1459                 __func__, wdata->lseg, nfsi->layout->plh_lwb);
1460
1461         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1462          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1463         if (mark_as_dirty)
1464                 mark_inode_dirty_sync(wdata->inode);
1465 }
1466 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1467
1468 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1469 {
1470         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1471
1472         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1473                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1474 }
1475
1476 /*
1477  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1478  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1479  * data to disk to allow the server to recover the data if it crashes.
1480  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1481  * is off, and a COMMIT is sent to a data server, or
1482  * if WRITEs to a data server return NFS_DATA_SYNC.
1483  */
1484 int
1485 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1486 {
1487         struct nfs4_layoutcommit_data *data;
1488         struct nfs_inode *nfsi = NFS_I(inode);
1489         loff_t end_pos;
1490         int status = 0;
1491
1492         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1493
1494         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1495                 return 0;
1496
1497         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1498         data = kzalloc(sizeof(*data), GFP_NOFS);
1499         if (!data) {
1500                 status = -ENOMEM;
1501                 goto out;
1502         }
1503
1504         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1505                 goto out_free;
1506
1507         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1508                 if (!sync) {
1509                         status = -EAGAIN;
1510                         goto out_free;
1511                 }
1512                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1513                                         nfs_wait_bit_killable, TASK_KILLABLE);
1514                 if (status)
1515                         goto out_free;
1516         }
1517
1518         INIT_LIST_HEAD(&data->lseg_list);
1519         spin_lock(&inode->i_lock);
1520         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1521                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1522                 spin_unlock(&inode->i_lock);
1523                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1524                 goto out_free;
1525         }
1526
1527         pnfs_list_write_lseg(inode, &data->lseg_list);
1528
1529         end_pos = nfsi->layout->plh_lwb;
1530         nfsi->layout->plh_lwb = 0;
1531
1532         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1533         spin_unlock(&inode->i_lock);
1534
1535         data->args.inode = inode;
1536         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1537         nfs_fattr_init(&data->fattr);
1538         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1539         data->res.fattr = &data->fattr;
1540         data->args.lastbytewritten = end_pos - 1;
1541         data->res.server = NFS_SERVER(inode);
1542
1543         status = nfs4_proc_layoutcommit(data, sync);
1544 out:
1545         if (status)
1546                 mark_inode_dirty_sync(inode);
1547         dprintk("<-- %s status %d\n", __func__, status);
1548         return status;
1549 out_free:
1550         kfree(data);
1551         goto out;
1552 }