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 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 459
 460 /*
 461  * Called by the state manger to remove all layouts established under an
 462  * expired lease.
 463  */
 464 void
 465 pnfs_destroy_all_layouts(struct nfs_client *clp)
 466 {
 467         struct nfs_server *server;
 468         struct pnfs_layout_hdr *lo;
 469         LIST_HEAD(tmp_list);
 470
 471         nfs4_deviceid_mark_client_invalid(clp);
 472         nfs4_deviceid_purge_client(clp);
 473
 474         spin_lock(&clp->cl_lock);
 475         rcu_read_lock();
 476         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 477                 if (!list_empty(&server->layouts))
 478                         list_splice_init(&server->layouts, &tmp_list);
 479         }
 480         rcu_read_unlock();
 481         spin_unlock(&clp->cl_lock);
 482
 483         while (!list_empty(&tmp_list)) {
 484                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 485                                 plh_layouts);
 486                 dprintk("%s freeing layout for inode %lu\n", __func__,
 487                         lo->plh_inode->i_ino);
 488                 list_del_init(&lo->plh_layouts);
 489                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 490         }
 491 }
 492
 493 /* update lo->plh_stateid with new if is more recent */
 494 void
 495 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 496                         bool update_barrier)
 497 {
 498         u32 oldseq, newseq;
 499
 500         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 501         newseq = be32_to_cpu(new->seqid);
 502         if ((int)(newseq - oldseq) > 0) {
 503                 nfs4_stateid_copy(&lo->plh_stateid, new);
 504                 if (update_barrier) {
 505                         u32 new_barrier = be32_to_cpu(new->seqid);
 506
 507                         if ((int)(new_barrier - lo->plh_barrier))
 508                                 lo->plh_barrier = new_barrier;
 509                 } else {
 510                         /* Because of wraparound, we want to keep the barrier
 511                          * "close" to the current seqids.  It needs to be
 512                          * within 2**31 to count as "behind", so if it
 513                          * gets too near that limit, give us a litle leeway
 514                          * and bring it to within 2**30.
 515                          * NOTE - and yes, this is all unsigned arithmetic.
 516                          */
 517                         if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
 518                                 lo->plh_barrier = newseq - (1 << 30);
 519                 }
 520         }
 521 }
 522
 523 /* lget is set to 1 if called from inside send_layoutget call chain */
 524 static bool
 525 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
 526                         int lget)
 527 {
 528         if ((stateid) &&
 529             (int)(lo->plh_barrier - be32_to_cpu(stateid->seqid)) >= 0)
 530                 return true;
 531         return lo->plh_block_lgets ||
 532                 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
 533                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 534                 (list_empty(&lo->plh_segs) &&
 535                  (atomic_read(&lo->plh_outstanding) > lget));
 536 }
 537
 538 int
 539 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 540                               struct nfs4_state *open_state)
 541 {
 542         int status = 0;
 543
 544         dprintk("--> %s\n", __func__);
 545         spin_lock(&lo->plh_inode->i_lock);
 546         if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
 547                 status = -EAGAIN;
 548         } else if (list_empty(&lo->plh_segs)) {
 549                 int seq;
 550
 551                 do {
 552                         seq = read_seqbegin(&open_state->seqlock);
 553                         nfs4_stateid_copy(dst, &open_state->stateid);
 554                 } while (read_seqretry(&open_state->seqlock, seq));
 555         } else
 556                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 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 = nfs_page_array_len(0, max_resp_sz);
 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 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 697
 698 bool pnfs_roc(struct inode *ino)
 699 {
 700         struct pnfs_layout_hdr *lo;
 701         struct pnfs_layout_segment *lseg, *tmp;
 702         LIST_HEAD(tmp_list);
 703         bool found = false;
 704
 705         spin_lock(&ino->i_lock);
 706         lo = NFS_I(ino)->layout;
 707         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 708             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 709                 goto out_nolayout;
 710         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 711                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 712                         mark_lseg_invalid(lseg, &tmp_list);
 713                         found = true;
 714                 }
 715         if (!found)
 716                 goto out_nolayout;
 717         lo->plh_block_lgets++;
 718         get_layout_hdr(lo); /* matched in pnfs_roc_release */
 719         spin_unlock(&ino->i_lock);
 720         pnfs_free_lseg_list(&tmp_list);
 721         return true;
 722
 723 out_nolayout:
 724         spin_unlock(&ino->i_lock);
 725         return false;
 726 }
 727
 728 void pnfs_roc_release(struct inode *ino)
 729 {
 730         struct pnfs_layout_hdr *lo;
 731
 732         spin_lock(&ino->i_lock);
 733         lo = NFS_I(ino)->layout;
 734         lo->plh_block_lgets--;
 735         put_layout_hdr_locked(lo);
 736         spin_unlock(&ino->i_lock);
 737 }
 738
 739 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 740 {
 741         struct pnfs_layout_hdr *lo;
 742
 743         spin_lock(&ino->i_lock);
 744         lo = NFS_I(ino)->layout;
 745         if ((int)(barrier - lo->plh_barrier) > 0)
 746                 lo->plh_barrier = barrier;
 747         spin_unlock(&ino->i_lock);
 748 }
 749
 750 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
 751 {
 752         struct nfs_inode *nfsi = NFS_I(ino);
 753         struct pnfs_layout_segment *lseg;
 754         bool found = false;
 755
 756         spin_lock(&ino->i_lock);
 757         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 758                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 759                         found = true;
 760                         break;
 761                 }
 762         if (!found) {
 763                 struct pnfs_layout_hdr *lo = nfsi->layout;
 764                 u32 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 765
 766                 /* Since close does not return a layout stateid for use as
 767                  * a barrier, we choose the worst-case barrier.
 768                  */
 769                 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 770         }
 771         spin_unlock(&ino->i_lock);
 772         return found;
 773 }
 774
 775 /*
 776  * Compare two layout segments for sorting into layout cache.
 777  * We want to preferentially return RW over RO layouts, so ensure those
 778  * are seen first.
 779  */
 780 static s64
 781 cmp_layout(struct pnfs_layout_range *l1,
 782            struct pnfs_layout_range *l2)
 783 {
 784         s64 d;
 785
 786         /* high offset > low offset */
 787         d = l1->offset - l2->offset;
 788         if (d)
 789                 return d;
 790
 791         /* short length > long length */
 792         d = l2->length - l1->length;
 793         if (d)
 794                 return d;
 795
 796         /* read > read/write */
 797         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 798 }
 799
 800 static void
 801 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
 802                    struct pnfs_layout_segment *lseg)
 803 {
 804         struct pnfs_layout_segment *lp;
 805
 806         dprintk("%s:Begin\n", __func__);
 807
 808         assert_spin_locked(&lo->plh_inode->i_lock);
 809         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 810                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 811                         continue;
 812                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 813                 dprintk("%s: inserted lseg %p "
 814                         "iomode %d offset %llu length %llu before "
 815                         "lp %p iomode %d offset %llu length %llu\n",
 816                         __func__, lseg, lseg->pls_range.iomode,
 817                         lseg->pls_range.offset, lseg->pls_range.length,
 818                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 819                         lp->pls_range.length);
 820                 goto out;
 821         }
 822         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 823         dprintk("%s: inserted lseg %p "
 824                 "iomode %d offset %llu length %llu at tail\n",
 825                 __func__, lseg, lseg->pls_range.iomode,
 826                 lseg->pls_range.offset, lseg->pls_range.length);
 827 out:
 828         get_layout_hdr(lo);
 829
 830         dprintk("%s:Return\n", __func__);
 831 }
 832
 833 static struct pnfs_layout_hdr *
 834 alloc_init_layout_hdr(struct inode *ino,
 835                       struct nfs_open_context *ctx,
 836                       gfp_t gfp_flags)
 837 {
 838         struct pnfs_layout_hdr *lo;
 839
 840         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 841         if (!lo)
 842                 return NULL;
 843         atomic_set(&lo->plh_refcount, 1);
 844         INIT_LIST_HEAD(&lo->plh_layouts);
 845         INIT_LIST_HEAD(&lo->plh_segs);
 846         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 847         lo->plh_inode = ino;
 848         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 849         return lo;
 850 }
 851
 852 static struct pnfs_layout_hdr *
 853 pnfs_find_alloc_layout(struct inode *ino,
 854                        struct nfs_open_context *ctx,
 855                        gfp_t gfp_flags)
 856 {
 857         struct nfs_inode *nfsi = NFS_I(ino);
 858         struct pnfs_layout_hdr *new = NULL;
 859
 860         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 861
 862         assert_spin_locked(&ino->i_lock);
 863         if (nfsi->layout) {
 864                 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
 865                         return NULL;
 866                 else
 867                         return nfsi->layout;
 868         }
 869         spin_unlock(&ino->i_lock);
 870         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 871         spin_lock(&ino->i_lock);
 872
 873         if (likely(nfsi->layout == NULL))       /* Won the race? */
 874                 nfsi->layout = new;
 875         else
 876                 pnfs_free_layout_hdr(new);
 877         return nfsi->layout;
 878 }
 879
 880 /*
 881  * iomode matching rules:
 882  * iomode       lseg    match
 883  * -----        -----   -----
 884  * ANY          READ    true
 885  * ANY          RW      true
 886  * RW           READ    false
 887  * RW           RW      true
 888  * READ         READ    true
 889  * READ         RW      true
 890  */
 891 static int
 892 is_matching_lseg(struct pnfs_layout_range *ls_range,
 893                  struct pnfs_layout_range *range)
 894 {
 895         struct pnfs_layout_range range1;
 896
 897         if ((range->iomode == IOMODE_RW &&
 898              ls_range->iomode != IOMODE_RW) ||
 899             !lo_seg_intersecting(ls_range, range))
 900                 return 0;
 901
 902         /* range1 covers only the first byte in the range */
 903         range1 = *range;
 904         range1.length = 1;
 905         return lo_seg_contained(ls_range, &range1);
 906 }
 907
 908 /*
 909  * lookup range in layout
 910  */
 911 static struct pnfs_layout_segment *
 912 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 913                 struct pnfs_layout_range *range)
 914 {
 915         struct pnfs_layout_segment *lseg, *ret = NULL;
 916
 917         dprintk("%s:Begin\n", __func__);
 918
 919         assert_spin_locked(&lo->plh_inode->i_lock);
 920         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 921                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 922                     is_matching_lseg(&lseg->pls_range, range)) {
 923                         ret = get_lseg(lseg);
 924                         break;
 925                 }
 926                 if (lseg->pls_range.offset > range->offset)
 927                         break;
 928         }
 929
 930         dprintk("%s:Return lseg %p ref %d\n",
 931                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 932         return ret;
 933 }
 934
 935 /*
 936  * Layout segment is retreived from the server if not cached.
 937  * The appropriate layout segment is referenced and returned to the caller.
 938  */
 939 struct pnfs_layout_segment *
 940 pnfs_update_layout(struct inode *ino,
 941                    struct nfs_open_context *ctx,
 942                    loff_t pos,
 943                    u64 count,
 944                    enum pnfs_iomode iomode,
 945                    gfp_t gfp_flags)
 946 {
 947         struct pnfs_layout_range arg = {
 948                 .iomode = iomode,
 949                 .offset = pos,
 950                 .length = count,
 951         };
 952         unsigned pg_offset;
 953         struct nfs_inode *nfsi = NFS_I(ino);
 954         struct nfs_server *server = NFS_SERVER(ino);
 955         struct nfs_client *clp = server->nfs_client;
 956         struct pnfs_layout_hdr *lo;
 957         struct pnfs_layout_segment *lseg = NULL;
 958         bool first = false;
 959
 960         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
 961                 return NULL;
 962         spin_lock(&ino->i_lock);
 963         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
 964         if (lo == NULL) {
 965                 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
 966                 goto out_unlock;
 967         }
 968
 969         /* Do we even need to bother with this? */
 970         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
 971                 dprintk("%s matches recall, use MDS\n", __func__);
 972                 goto out_unlock;
 973         }
 974
 975         /* if LAYOUTGET already failed once we don't try again */
 976         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
 977                 goto out_unlock;
 978
 979         /* Check to see if the layout for the given range already exists */
 980         lseg = pnfs_find_lseg(lo, &arg);
 981         if (lseg)
 982                 goto out_unlock;
 983
 984         if (pnfs_layoutgets_blocked(lo, NULL, 0))
 985                 goto out_unlock;
 986         atomic_inc(&lo->plh_outstanding);
 987
 988         get_layout_hdr(lo);
 989         if (list_empty(&lo->plh_segs))
 990                 first = true;
 991         spin_unlock(&ino->i_lock);
 992         if (first) {
 993                 /* The lo must be on the clp list if there is any
 994                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
 995                  */
 996                 spin_lock(&clp->cl_lock);
 997                 BUG_ON(!list_empty(&lo->plh_layouts));
 998                 list_add_tail(&lo->plh_layouts, &server->layouts);
 999                 spin_unlock(&clp->cl_lock);
1000         }
1001
1002         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1003         if (pg_offset) {
1004                 arg.offset -= pg_offset;
1005                 arg.length += pg_offset;
1006         }
1007         if (arg.length != NFS4_MAX_UINT64)
1008                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1009
1010         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1011         if (!lseg && first) {
1012                 spin_lock(&clp->cl_lock);
1013                 list_del_init(&lo->plh_layouts);
1014                 spin_unlock(&clp->cl_lock);
1015         }
1016         atomic_dec(&lo->plh_outstanding);
1017         put_layout_hdr(lo);
1018 out:
1019         dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1020                 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1021         return lseg;
1022 out_unlock:
1023         spin_unlock(&ino->i_lock);
1024         goto out;
1025 }
1026 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1027
1028 int
1029 pnfs_layout_process(struct nfs4_layoutget *lgp)
1030 {
1031         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1032         struct nfs4_layoutget_res *res = &lgp->res;
1033         struct pnfs_layout_segment *lseg;
1034         struct inode *ino = lo->plh_inode;
1035         int status = 0;
1036
1037         /* Inject layout blob into I/O device driver */
1038         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1039         if (!lseg || IS_ERR(lseg)) {
1040                 if (!lseg)
1041                         status = -ENOMEM;
1042                 else
1043                         status = PTR_ERR(lseg);
1044                 dprintk("%s: Could not allocate layout: error %d\n",
1045                        __func__, status);
1046                 goto out;
1047         }
1048
1049         spin_lock(&ino->i_lock);
1050         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1051                 dprintk("%s forget reply due to recall\n", __func__);
1052                 goto out_forget_reply;
1053         }
1054
1055         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1056                 dprintk("%s forget reply due to state\n", __func__);
1057                 goto out_forget_reply;
1058         }
1059         init_lseg(lo, lseg);
1060         lseg->pls_range = res->range;
1061         *lgp->lsegpp = get_lseg(lseg);
1062         pnfs_insert_layout(lo, lseg);
1063
1064         if (res->return_on_close) {
1065                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1066                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1067         }
1068
1069         /* Done processing layoutget. Set the layout stateid */
1070         pnfs_set_layout_stateid(lo, &res->stateid, false);
1071         spin_unlock(&ino->i_lock);
1072 out:
1073         return status;
1074
1075 out_forget_reply:
1076         spin_unlock(&ino->i_lock);
1077         lseg->pls_layout = lo;
1078         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1079         goto out;
1080 }
1081
1082 void
1083 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1084 {
1085         BUG_ON(pgio->pg_lseg != NULL);
1086
1087         if (req->wb_offset != req->wb_pgbase) {
1088                 nfs_pageio_reset_read_mds(pgio);
1089                 return;
1090         }
1091         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1092                                            req->wb_context,
1093                                            req_offset(req),
1094                                            req->wb_bytes,
1095                                            IOMODE_READ,
1096                                            GFP_KERNEL);
1097         /* If no lseg, fall back to read through mds */
1098         if (pgio->pg_lseg == NULL)
1099                 nfs_pageio_reset_read_mds(pgio);
1100
1101 }
1102 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1103
1104 void
1105 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1106 {
1107         BUG_ON(pgio->pg_lseg != NULL);
1108
1109         if (req->wb_offset != req->wb_pgbase) {
1110                 nfs_pageio_reset_write_mds(pgio);
1111                 return;
1112         }
1113         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1114                                            req->wb_context,
1115                                            req_offset(req),
1116                                            req->wb_bytes,
1117                                            IOMODE_RW,
1118                                            GFP_NOFS);
1119         /* If no lseg, fall back to write through mds */
1120         if (pgio->pg_lseg == NULL)
1121                 nfs_pageio_reset_write_mds(pgio);
1122 }
1123 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1124
1125 bool
1126 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1127                       const struct nfs_pgio_completion_ops *compl_ops)
1128 {
1129         struct nfs_server *server = NFS_SERVER(inode);
1130         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1131
1132         if (ld == NULL)
1133                 return false;
1134         nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops,
1135                         server->rsize, 0);
1136         return true;
1137 }
1138
1139 bool
1140 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1141                        int ioflags,
1142                        const struct nfs_pgio_completion_ops *compl_ops)
1143 {
1144         struct nfs_server *server = NFS_SERVER(inode);
1145         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1146
1147         if (ld == NULL)
1148                 return false;
1149         nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops,
1150                         server->wsize, ioflags);
1151         return true;
1152 }
1153
1154 bool
1155 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1156                      struct nfs_page *req)
1157 {
1158         if (pgio->pg_lseg == NULL)
1159                 return nfs_generic_pg_test(pgio, prev, req);
1160
1161         /*
1162          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1163          * Note that this test makes several assumptions:
1164          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1165          *   is known to lie within the range.
1166          *   - that the nfs_page being tested is known to be contiguous with the
1167          *   previous nfs_page.
1168          *   - Layout ranges are page aligned, so we only have to test the
1169          *   start offset of the request.
1170          *
1171          * Please also note that 'end_offset' is actually the offset of the
1172          * first byte that lies outside the pnfs_layout_range. FIXME?
1173          *
1174          */
1175         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1176                                          pgio->pg_lseg->pls_range.length);
1177 }
1178 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1179
1180 int pnfs_write_done_resend_to_mds(struct inode *inode,
1181                                 struct list_head *head,
1182                                 const struct nfs_pgio_completion_ops *compl_ops)
1183 {
1184         struct nfs_pageio_descriptor pgio;
1185         LIST_HEAD(failed);
1186
1187         /* Resend all requests through the MDS */
1188         nfs_pageio_init_write_mds(&pgio, inode, FLUSH_STABLE, compl_ops);
1189         while (!list_empty(head)) {
1190                 struct nfs_page *req = nfs_list_entry(head->next);
1191
1192                 nfs_list_remove_request(req);
1193                 if (!nfs_pageio_add_request(&pgio, req))
1194                         nfs_list_add_request(req, &failed);
1195         }
1196         nfs_pageio_complete(&pgio);
1197
1198         if (!list_empty(&failed)) {
1199                 /* For some reason our attempt to resend pages. Mark the
1200                  * overall send request as having failed, and let
1201                  * nfs_writeback_release_full deal with the error.
1202                  */
1203                 list_move(&failed, head);
1204                 return -EIO;
1205         }
1206         return 0;
1207 }
1208 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1209
1210 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1211 {
1212         struct nfs_pgio_header *hdr = data->header;
1213
1214         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1215         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1216             PNFS_LAYOUTRET_ON_ERROR) {
1217                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1218                 pnfs_return_layout(hdr->inode);
1219         }
1220         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1221                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1222                                                         &hdr->pages,
1223                                                         hdr->completion_ops);
1224 }
1225
1226 /*
1227  * Called by non rpc-based layout drivers
1228  */
1229 void pnfs_ld_write_done(struct nfs_write_data *data)
1230 {
1231         struct nfs_pgio_header *hdr = data->header;
1232
1233         if (!hdr->pnfs_error) {
1234                 pnfs_set_layoutcommit(data);
1235                 hdr->mds_ops->rpc_call_done(&data->task, data);
1236         } else
1237                 pnfs_ld_handle_write_error(data);
1238         hdr->mds_ops->rpc_release(data);
1239 }
1240 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1241
1242 static void
1243 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1244                 struct nfs_write_data *data)
1245 {
1246         struct nfs_pgio_header *hdr = data->header;
1247
1248         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1249                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1250                 nfs_pageio_reset_write_mds(desc);
1251                 desc->pg_recoalesce = 1;
1252         }
1253         nfs_writedata_release(data);
1254 }
1255
1256 static enum pnfs_try_status
1257 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1258                         const struct rpc_call_ops *call_ops,
1259                         struct pnfs_layout_segment *lseg,
1260                         int how)
1261 {
1262         struct nfs_pgio_header *hdr = wdata->header;
1263         struct inode *inode = hdr->inode;
1264         enum pnfs_try_status trypnfs;
1265         struct nfs_server *nfss = NFS_SERVER(inode);
1266
1267         hdr->mds_ops = call_ops;
1268
1269         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1270                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1271         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1272         if (trypnfs != PNFS_NOT_ATTEMPTED)
1273                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1274         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1275         return trypnfs;
1276 }
1277
1278 static void
1279 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1280 {
1281         struct nfs_write_data *data;
1282         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1283         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1284
1285         desc->pg_lseg = NULL;
1286         while (!list_empty(head)) {
1287                 enum pnfs_try_status trypnfs;
1288
1289                 data = list_first_entry(head, struct nfs_write_data, list);
1290                 list_del_init(&data->list);
1291
1292                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1293                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1294                         pnfs_write_through_mds(desc, data);
1295         }
1296         put_lseg(lseg);
1297 }
1298
1299 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1300 {
1301         put_lseg(hdr->lseg);
1302         nfs_writehdr_free(hdr);
1303 }
1304
1305 int
1306 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1307 {
1308         struct nfs_write_header *whdr;
1309         struct nfs_pgio_header *hdr;
1310         int ret;
1311
1312         whdr = nfs_writehdr_alloc();
1313         if (!whdr) {
1314                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1315                 put_lseg(desc->pg_lseg);
1316                 desc->pg_lseg = NULL;
1317                 return -ENOMEM;
1318         }
1319         hdr = &whdr->header;
1320         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1321         hdr->lseg = get_lseg(desc->pg_lseg);
1322         atomic_inc(&hdr->refcnt);
1323         ret = nfs_generic_flush(desc, hdr);
1324         if (ret != 0) {
1325                 put_lseg(desc->pg_lseg);
1326                 desc->pg_lseg = NULL;
1327         } else
1328                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1329         if (atomic_dec_and_test(&hdr->refcnt))
1330                 hdr->completion_ops->completion(hdr);
1331         return ret;
1332 }
1333 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1334
1335 int pnfs_read_done_resend_to_mds(struct inode *inode,
1336                                 struct list_head *head,
1337                                 const struct nfs_pgio_completion_ops *compl_ops)
1338 {
1339         struct nfs_pageio_descriptor pgio;
1340         LIST_HEAD(failed);
1341
1342         /* Resend all requests through the MDS */
1343         nfs_pageio_init_read_mds(&pgio, inode, compl_ops);
1344         while (!list_empty(head)) {
1345                 struct nfs_page *req = nfs_list_entry(head->next);
1346
1347                 nfs_list_remove_request(req);
1348                 if (!nfs_pageio_add_request(&pgio, req))
1349                         nfs_list_add_request(req, &failed);
1350         }
1351         nfs_pageio_complete(&pgio);
1352
1353         if (!list_empty(&failed)) {
1354                 list_move(&failed, head);
1355                 return -EIO;
1356         }
1357         return 0;
1358 }
1359 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1360
1361 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1362 {
1363         struct nfs_pgio_header *hdr = data->header;
1364
1365         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1366         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1367             PNFS_LAYOUTRET_ON_ERROR) {
1368                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1369                 pnfs_return_layout(hdr->inode);
1370         }
1371         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1372                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1373                                                         &hdr->pages,
1374                                                         hdr->completion_ops);
1375 }
1376
1377 /*
1378  * Called by non rpc-based layout drivers
1379  */
1380 void pnfs_ld_read_done(struct nfs_read_data *data)
1381 {
1382         struct nfs_pgio_header *hdr = data->header;
1383
1384         if (likely(!hdr->pnfs_error)) {
1385                 __nfs4_read_done_cb(data);
1386                 hdr->mds_ops->rpc_call_done(&data->task, data);
1387         } else
1388                 pnfs_ld_handle_read_error(data);
1389         hdr->mds_ops->rpc_release(data);
1390 }
1391 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1392
1393 static void
1394 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1395                 struct nfs_read_data *data)
1396 {
1397         struct nfs_pgio_header *hdr = data->header;
1398
1399         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1400                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1401                 nfs_pageio_reset_read_mds(desc);
1402                 desc->pg_recoalesce = 1;
1403         }
1404         nfs_readdata_release(data);
1405 }
1406
1407 /*
1408  * Call the appropriate parallel I/O subsystem read function.
1409  */
1410 static enum pnfs_try_status
1411 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1412                        const struct rpc_call_ops *call_ops,
1413                        struct pnfs_layout_segment *lseg)
1414 {
1415         struct nfs_pgio_header *hdr = rdata->header;
1416         struct inode *inode = hdr->inode;
1417         struct nfs_server *nfss = NFS_SERVER(inode);
1418         enum pnfs_try_status trypnfs;
1419
1420         hdr->mds_ops = call_ops;
1421
1422         dprintk("%s: Reading ino:%lu %u@%llu\n",
1423                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1424
1425         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1426         if (trypnfs != PNFS_NOT_ATTEMPTED)
1427                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1428         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1429         return trypnfs;
1430 }
1431
1432 static void
1433 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1434 {
1435         struct nfs_read_data *data;
1436         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1437         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1438
1439         desc->pg_lseg = NULL;
1440         while (!list_empty(head)) {
1441                 enum pnfs_try_status trypnfs;
1442
1443                 data = list_first_entry(head, struct nfs_read_data, list);
1444                 list_del_init(&data->list);
1445
1446                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1447                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1448                         pnfs_read_through_mds(desc, data);
1449         }
1450         put_lseg(lseg);
1451 }
1452
1453 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1454 {
1455         put_lseg(hdr->lseg);
1456         nfs_readhdr_free(hdr);
1457 }
1458
1459 int
1460 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1461 {
1462         struct nfs_read_header *rhdr;
1463         struct nfs_pgio_header *hdr;
1464         int ret;
1465
1466         rhdr = nfs_readhdr_alloc();
1467         if (!rhdr) {
1468                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1469                 ret = -ENOMEM;
1470                 put_lseg(desc->pg_lseg);
1471                 desc->pg_lseg = NULL;
1472                 return ret;
1473         }
1474         hdr = &rhdr->header;
1475         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1476         hdr->lseg = get_lseg(desc->pg_lseg);
1477         atomic_inc(&hdr->refcnt);
1478         ret = nfs_generic_pagein(desc, hdr);
1479         if (ret != 0) {
1480                 put_lseg(desc->pg_lseg);
1481                 desc->pg_lseg = NULL;
1482         } else
1483                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1484         if (atomic_dec_and_test(&hdr->refcnt))
1485                 hdr->completion_ops->completion(hdr);
1486         return ret;
1487 }
1488 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1489
1490 /*
1491  * There can be multiple RW segments.
1492  */
1493 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1494 {
1495         struct pnfs_layout_segment *lseg;
1496
1497         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1498                 if (lseg->pls_range.iomode == IOMODE_RW &&
1499                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1500                         list_add(&lseg->pls_lc_list, listp);
1501         }
1502 }
1503
1504 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1505 {
1506         if (lseg->pls_range.iomode == IOMODE_RW) {
1507                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1508                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1509         } else {
1510                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1511                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1512         }
1513 }
1514 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1515
1516 void
1517 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1518 {
1519         struct nfs_pgio_header *hdr = wdata->header;
1520         struct inode *inode = hdr->inode;
1521         struct nfs_inode *nfsi = NFS_I(inode);
1522         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1523         bool mark_as_dirty = false;
1524
1525         spin_lock(&inode->i_lock);
1526         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1527                 mark_as_dirty = true;
1528                 dprintk("%s: Set layoutcommit for inode %lu ",
1529                         __func__, inode->i_ino);
1530         }
1531         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1532                 /* references matched in nfs4_layoutcommit_release */
1533                 get_lseg(hdr->lseg);
1534         }
1535         if (end_pos > nfsi->layout->plh_lwb)
1536                 nfsi->layout->plh_lwb = end_pos;
1537         spin_unlock(&inode->i_lock);
1538         dprintk("%s: lseg %p end_pos %llu\n",
1539                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1540
1541         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1542          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1543         if (mark_as_dirty)
1544                 mark_inode_dirty_sync(inode);
1545 }
1546 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1547
1548 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1549 {
1550         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1551
1552         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1553                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1554 }
1555
1556 /*
1557  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1558  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1559  * data to disk to allow the server to recover the data if it crashes.
1560  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1561  * is off, and a COMMIT is sent to a data server, or
1562  * if WRITEs to a data server return NFS_DATA_SYNC.
1563  */
1564 int
1565 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1566 {
1567         struct nfs4_layoutcommit_data *data;
1568         struct nfs_inode *nfsi = NFS_I(inode);
1569         loff_t end_pos;
1570         int status = 0;
1571
1572         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1573
1574         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1575                 return 0;
1576
1577         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1578         data = kzalloc(sizeof(*data), GFP_NOFS);
1579         if (!data) {
1580                 status = -ENOMEM;
1581                 goto out;
1582         }
1583
1584         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1585                 goto out_free;
1586
1587         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1588                 if (!sync) {
1589                         status = -EAGAIN;
1590                         goto out_free;
1591                 }
1592                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1593                                         nfs_wait_bit_killable, TASK_KILLABLE);
1594                 if (status)
1595                         goto out_free;
1596         }
1597
1598         INIT_LIST_HEAD(&data->lseg_list);
1599         spin_lock(&inode->i_lock);
1600         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1601                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1602                 spin_unlock(&inode->i_lock);
1603                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1604                 goto out_free;
1605         }
1606
1607         pnfs_list_write_lseg(inode, &data->lseg_list);
1608
1609         end_pos = nfsi->layout->plh_lwb;
1610         nfsi->layout->plh_lwb = 0;
1611
1612         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1613         spin_unlock(&inode->i_lock);
1614
1615         data->args.inode = inode;
1616         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1617         nfs_fattr_init(&data->fattr);
1618         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1619         data->res.fattr = &data->fattr;
1620         data->args.lastbytewritten = end_pos - 1;
1621         data->res.server = NFS_SERVER(inode);
1622
1623         status = nfs4_proc_layoutcommit(data, sync);
1624 out:
1625         if (status)
1626                 mark_inode_dirty_sync(inode);
1627         dprintk("<-- %s status %d\n", __func__, status);
1628         return status;
1629 out_free:
1630         kfree(data);
1631         goto out;
1632 }
1633
1634 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1635 {
1636         struct nfs4_threshold *thp;
1637
1638         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1639         if (!thp) {
1640                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1641                 return NULL;
1642         }
1643         return thp;
1644 }