fs/nfs/pnfs.c

   1 /*
   2  *  pNFS functions to call and manage layout drivers.
   3  *
   4  *  Copyright (c) 2002 [year of first publication]
   5  *  The Regents of the University of Michigan
   6  *  All Rights Reserved
   7  *
   8  *  Dean Hildebrand <dhildebz@umich.edu>
   9  *
  10  *  Permission is granted to use, copy, create derivative works, and
  11  *  redistribute this software and such derivative works for any purpose,
  12  *  so long as the name of the University of Michigan is not used in
  13  *  any advertising or publicity pertaining to the use or distribution
  14  *  of this software without specific, written prior authorization. If
  15  *  the above copyright notice or any other identification of the
  16  *  University of Michigan is included in any copy of any portion of
  17  *  this software, then the disclaimer below must also be included.
  18  *
  19  *  This software is provided as is, without representation or warranty
  20  *  of any kind either express or implied, including without limitation
  21  *  the implied warranties of merchantability, fitness for a particular
  22  *  purpose, or noninfringement.  The Regents of the University of
  23  *  Michigan shall not be liable for any damages, including special,
  24  *  indirect, incidental, or consequential damages, with respect to any
  25  *  claim arising out of or in connection with the use of the software,
  26  *  even if it has been or is hereafter advised of the possibility of
  27  *  such damages.
  28  */
  29
  30 #include <linux/nfs_fs.h>
  31 #include <linux/nfs_page.h>
  32 #include <linux/module.h>
  33 #include "internal.h"
  34 #include "pnfs.h"
  35 #include "iostat.h"
  36
  37 #define NFSDBG_FACILITY         NFSDBG_PNFS
  38
  39 /* Locking:
  40  *
  41  * pnfs_spinlock:
  42  *      protects pnfs_modules_tbl.
  43  */
  44 static DEFINE_SPINLOCK(pnfs_spinlock);
  45
  46 /*
  47  * pnfs_modules_tbl holds all pnfs modules
  48  */
  49 static LIST_HEAD(pnfs_modules_tbl);
  50
  51 /* Return the registered pnfs layout driver module matching given id */
  52 static struct pnfs_layoutdriver_type *
  53 find_pnfs_driver_locked(u32 id)
  54 {
  55         struct pnfs_layoutdriver_type *local;
  56
  57         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  58                 if (local->id == id)
  59                         goto out;
  60         local = NULL;
  61 out:
  62         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  63         return local;
  64 }
  65
  66 static struct pnfs_layoutdriver_type *
  67 find_pnfs_driver(u32 id)
  68 {
  69         struct pnfs_layoutdriver_type *local;
  70
  71         spin_lock(&pnfs_spinlock);
  72         local = find_pnfs_driver_locked(id);
  73         if (local != NULL && !try_module_get(local->owner)) {
  74                 dprintk("%s: Could not grab reference on module\n", __func__);
  75                 local = NULL;
  76         }
  77         spin_unlock(&pnfs_spinlock);
  78         return local;
  79 }
  80
  81 void
  82 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  83 {
  84         if (nfss->pnfs_curr_ld) {
  85                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  86                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  87                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  88                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  89                         nfs4_deviceid_purge_client(nfss->nfs_client);
  90                 module_put(nfss->pnfs_curr_ld->owner);
  91         }
  92         nfss->pnfs_curr_ld = NULL;
  93 }
  94
  95 /*
  96  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  97  * Currently only one pNFS layout driver per filesystem is supported.
  98  *
  99  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 100  */
 101 void
 102 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 103                       u32 id)
 104 {
 105         struct pnfs_layoutdriver_type *ld_type = NULL;
 106
 107         if (id == 0)
 108                 goto out_no_driver;
 109         if (!(server->nfs_client->cl_exchange_flags &
 110                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 111                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 112                         __func__, id, server->nfs_client->cl_exchange_flags);
 113                 goto out_no_driver;
 114         }
 115         ld_type = find_pnfs_driver(id);
 116         if (!ld_type) {
 117                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 118                 ld_type = find_pnfs_driver(id);
 119                 if (!ld_type) {
 120                         dprintk("%s: No pNFS module found for %u.\n",
 121                                 __func__, id);
 122                         goto out_no_driver;
 123                 }
 124         }
 125         server->pnfs_curr_ld = ld_type;
 126         if (ld_type->set_layoutdriver
 127             && ld_type->set_layoutdriver(server, mntfh)) {
 128                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 129                         "driver %u.\n", __func__, id);
 130                 module_put(ld_type->owner);
 131                 goto out_no_driver;
 132         }
 133         /* Bump the MDS count */
 134         atomic_inc(&server->nfs_client->cl_mds_count);
 135
 136         dprintk("%s: pNFS module for %u set\n", __func__, id);
 137         return;
 138
 139 out_no_driver:
 140         dprintk("%s: Using NFSv4 I/O\n", __func__);
 141         server->pnfs_curr_ld = NULL;
 142 }
 143
 144 int
 145 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 146 {
 147         int status = -EINVAL;
 148         struct pnfs_layoutdriver_type *tmp;
 149
 150         if (ld_type->id == 0) {
 151                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 152                 return status;
 153         }
 154         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 155                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 156                        "alloc_lseg and free_lseg.\n", __func__);
 157                 return status;
 158         }
 159
 160         spin_lock(&pnfs_spinlock);
 161         tmp = find_pnfs_driver_locked(ld_type->id);
 162         if (!tmp) {
 163                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 164                 status = 0;
 165                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 166                         ld_type->name);
 167         } else {
 168                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 169                         __func__, ld_type->id);
 170         }
 171         spin_unlock(&pnfs_spinlock);
 172
 173         return status;
 174 }
 175 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 176
 177 void
 178 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 179 {
 180         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 181         spin_lock(&pnfs_spinlock);
 182         list_del(&ld_type->pnfs_tblid);
 183         spin_unlock(&pnfs_spinlock);
 184 }
 185 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 186
 187 /*
 188  * pNFS client layout cache
 189  */
 190
 191 /* Need to hold i_lock if caller does not already hold reference */
 192 void
 193 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 194 {
 195         atomic_inc(&lo->plh_refcount);
 196 }
 197
 198 static struct pnfs_layout_hdr *
 199 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 200 {
 201         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 202         return ld->alloc_layout_hdr ? ld->alloc_layout_hdr(ino, gfp_flags) :
 203                 kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
 204 }
 205
 206 static void
 207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 208 {
 209         struct pnfs_layoutdriver_type *ld = NFS_SERVER(lo->plh_inode)->pnfs_curr_ld;
 210         put_rpccred(lo->plh_lc_cred);
 211         return ld->alloc_layout_hdr ? ld->free_layout_hdr(lo) : kfree(lo);
 212 }
 213
 214 static void
 215 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
 216 {
 217         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 218         BUG_ON(!list_empty(&lo->plh_layouts));
 219         NFS_I(lo->plh_inode)->layout = NULL;
 220         pnfs_free_layout_hdr(lo);
 221 }
 222
 223 static void
 224 pnfs_put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
 225 {
 226         if (atomic_dec_and_test(&lo->plh_refcount))
 227                 destroy_layout_hdr(lo);
 228 }
 229
 230 void
 231 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 232 {
 233         struct inode *inode = lo->plh_inode;
 234
 235         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 236                 destroy_layout_hdr(lo);
 237                 spin_unlock(&inode->i_lock);
 238         }
 239 }
 240
 241 static void
 242 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 243 {
 244         INIT_LIST_HEAD(&lseg->pls_list);
 245         INIT_LIST_HEAD(&lseg->pls_lc_list);
 246         atomic_set(&lseg->pls_refcount, 1);
 247         smp_mb();
 248         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 249         lseg->pls_layout = lo;
 250 }
 251
 252 static void free_lseg(struct pnfs_layout_segment *lseg)
 253 {
 254         struct inode *ino = lseg->pls_layout->plh_inode;
 255
 256         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 257         /* Matched by pnfs_get_layout_hdr in pnfs_insert_layout */
 258         pnfs_put_layout_hdr(NFS_I(ino)->layout);
 259 }
 260
 261 static void
 262 pnfs_put_lseg_common(struct pnfs_layout_segment *lseg)
 263 {
 264         struct inode *inode = lseg->pls_layout->plh_inode;
 265
 266         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 267         list_del_init(&lseg->pls_list);
 268         if (list_empty(&lseg->pls_layout->plh_segs)) {
 269                 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
 270                 /* Matched by initial refcount set in alloc_init_layout_hdr */
 271                 pnfs_put_layout_hdr_locked(lseg->pls_layout);
 272         }
 273         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 274 }
 275
 276 void
 277 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 278 {
 279         struct inode *inode;
 280
 281         if (!lseg)
 282                 return;
 283
 284         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 285                 atomic_read(&lseg->pls_refcount),
 286                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 287         inode = lseg->pls_layout->plh_inode;
 288         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 289                 LIST_HEAD(free_me);
 290
 291                 pnfs_put_lseg_common(lseg);
 292                 list_add(&lseg->pls_list, &free_me);
 293                 spin_unlock(&inode->i_lock);
 294                 pnfs_free_lseg_list(&free_me);
 295         }
 296 }
 297 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 298
 299 static inline u64
 300 end_offset(u64 start, u64 len)
 301 {
 302         u64 end;
 303
 304         end = start + len;
 305         return end >= start ? end : NFS4_MAX_UINT64;
 306 }
 307
 308 /* last octet in a range */
 309 static inline u64
 310 last_byte_offset(u64 start, u64 len)
 311 {
 312         u64 end;
 313
 314         BUG_ON(!len);
 315         end = start + len;
 316         return end > start ? end - 1 : NFS4_MAX_UINT64;
 317 }
 318
 319 /*
 320  * is l2 fully contained in l1?
 321  *   start1                             end1
 322  *   [----------------------------------)
 323  *           start2           end2
 324  *           [----------------)
 325  */
 326 static inline int
 327 lo_seg_contained(struct pnfs_layout_range *l1,
 328                  struct pnfs_layout_range *l2)
 329 {
 330         u64 start1 = l1->offset;
 331         u64 end1 = end_offset(start1, l1->length);
 332         u64 start2 = l2->offset;
 333         u64 end2 = end_offset(start2, l2->length);
 334
 335         return (start1 <= start2) && (end1 >= end2);
 336 }
 337
 338 /*
 339  * is l1 and l2 intersecting?
 340  *   start1                             end1
 341  *   [----------------------------------)
 342  *                              start2           end2
 343  *                              [----------------)
 344  */
 345 static inline int
 346 lo_seg_intersecting(struct pnfs_layout_range *l1,
 347                     struct pnfs_layout_range *l2)
 348 {
 349         u64 start1 = l1->offset;
 350         u64 end1 = end_offset(start1, l1->length);
 351         u64 start2 = l2->offset;
 352         u64 end2 = end_offset(start2, l2->length);
 353
 354         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 355                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 356 }
 357
 358 static bool
 359 should_free_lseg(struct pnfs_layout_range *lseg_range,
 360                  struct pnfs_layout_range *recall_range)
 361 {
 362         return (recall_range->iomode == IOMODE_ANY ||
 363                 lseg_range->iomode == recall_range->iomode) &&
 364                lo_seg_intersecting(lseg_range, recall_range);
 365 }
 366
 367 /* Returns 1 if lseg is removed from list, 0 otherwise */
 368 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 369                              struct list_head *tmp_list)
 370 {
 371         int rv = 0;
 372
 373         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 374                 /* Remove the reference keeping the lseg in the
 375                  * list.  It will now be removed when all
 376                  * outstanding io is finished.
 377                  */
 378                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 379                         atomic_read(&lseg->pls_refcount));
 380                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 381                         pnfs_put_lseg_common(lseg);
 382                         list_add(&lseg->pls_list, tmp_list);
 383                         rv = 1;
 384                 }
 385         }
 386         return rv;
 387 }
 388
 389 /* Returns count of number of matching invalid lsegs remaining in list
 390  * after call.
 391  */
 392 int
 393 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 394                             struct list_head *tmp_list,
 395                             struct pnfs_layout_range *recall_range)
 396 {
 397         struct pnfs_layout_segment *lseg, *next;
 398         int invalid = 0, removed = 0;
 399
 400         dprintk("%s:Begin lo %p\n", __func__, lo);
 401
 402         if (list_empty(&lo->plh_segs)) {
 403                 /* Reset MDS Threshold I/O counters */
 404                 NFS_I(lo->plh_inode)->write_io = 0;
 405                 NFS_I(lo->plh_inode)->read_io = 0;
 406                 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
 407                         pnfs_put_layout_hdr_locked(lo);
 408                 return 0;
 409         }
 410         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 411                 if (!recall_range ||
 412                     should_free_lseg(&lseg->pls_range, recall_range)) {
 413                         dprintk("%s: freeing lseg %p iomode %d "
 414                                 "offset %llu length %llu\n", __func__,
 415                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 416                                 lseg->pls_range.length);
 417                         invalid++;
 418                         removed += mark_lseg_invalid(lseg, tmp_list);
 419                 }
 420         dprintk("%s:Return %i\n", __func__, invalid - removed);
 421         return invalid - removed;
 422 }
 423
 424 /* note free_me must contain lsegs from a single layout_hdr */
 425 void
 426 pnfs_free_lseg_list(struct list_head *free_me)
 427 {
 428         struct pnfs_layout_segment *lseg, *tmp;
 429         struct pnfs_layout_hdr *lo;
 430
 431         if (list_empty(free_me))
 432                 return;
 433
 434         lo = list_first_entry(free_me, struct pnfs_layout_segment,
 435                               pls_list)->pls_layout;
 436
 437         if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
 438                 struct nfs_client *clp;
 439
 440                 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
 441                 spin_lock(&clp->cl_lock);
 442                 list_del_init(&lo->plh_layouts);
 443                 spin_unlock(&clp->cl_lock);
 444         }
 445         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 446                 list_del(&lseg->pls_list);
 447                 free_lseg(lseg);
 448         }
 449 }
 450
 451 void
 452 pnfs_destroy_layout(struct nfs_inode *nfsi)
 453 {
 454         struct pnfs_layout_hdr *lo;
 455         LIST_HEAD(tmp_list);
 456
 457         spin_lock(&nfsi->vfs_inode.i_lock);
 458         lo = nfsi->layout;
 459         if (lo) {
 460                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 461                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 462         }
 463         spin_unlock(&nfsi->vfs_inode.i_lock);
 464         pnfs_free_lseg_list(&tmp_list);
 465 }
 466 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 467
 468 /*
 469  * Called by the state manger to remove all layouts established under an
 470  * expired lease.
 471  */
 472 void
 473 pnfs_destroy_all_layouts(struct nfs_client *clp)
 474 {
 475         struct nfs_server *server;
 476         struct pnfs_layout_hdr *lo;
 477         LIST_HEAD(tmp_list);
 478
 479         nfs4_deviceid_mark_client_invalid(clp);
 480         nfs4_deviceid_purge_client(clp);
 481
 482         spin_lock(&clp->cl_lock);
 483         rcu_read_lock();
 484         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 485                 if (!list_empty(&server->layouts))
 486                         list_splice_init(&server->layouts, &tmp_list);
 487         }
 488         rcu_read_unlock();
 489         spin_unlock(&clp->cl_lock);
 490
 491         while (!list_empty(&tmp_list)) {
 492                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 493                                 plh_layouts);
 494                 dprintk("%s freeing layout for inode %lu\n", __func__,
 495                         lo->plh_inode->i_ino);
 496                 list_del_init(&lo->plh_layouts);
 497                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 498         }
 499 }
 500
 501 /* update lo->plh_stateid with new if is more recent */
 502 void
 503 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 504                         bool update_barrier)
 505 {
 506         u32 oldseq, newseq;
 507
 508         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 509         newseq = be32_to_cpu(new->seqid);
 510         if ((int)(newseq - oldseq) > 0) {
 511                 nfs4_stateid_copy(&lo->plh_stateid, new);
 512                 if (update_barrier) {
 513                         u32 new_barrier = be32_to_cpu(new->seqid);
 514
 515                         if ((int)(new_barrier - lo->plh_barrier))
 516                                 lo->plh_barrier = new_barrier;
 517                 } else {
 518                         /* Because of wraparound, we want to keep the barrier
 519                          * "close" to the current seqids.  It needs to be
 520                          * within 2**31 to count as "behind", so if it
 521                          * gets too near that limit, give us a litle leeway
 522                          * and bring it to within 2**30.
 523                          * NOTE - and yes, this is all unsigned arithmetic.
 524                          */
 525                         if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
 526                                 lo->plh_barrier = newseq - (1 << 30);
 527                 }
 528         }
 529 }
 530
 531 /* lget is set to 1 if called from inside send_layoutget call chain */
 532 static bool
 533 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
 534                         int lget)
 535 {
 536         if ((stateid) &&
 537             (int)(lo->plh_barrier - be32_to_cpu(stateid->seqid)) >= 0)
 538                 return true;
 539         return lo->plh_block_lgets ||
 540                 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
 541                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 542                 (list_empty(&lo->plh_segs) &&
 543                  (atomic_read(&lo->plh_outstanding) > lget));
 544 }
 545
 546 int
 547 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 548                               struct nfs4_state *open_state)
 549 {
 550         int status = 0;
 551
 552         dprintk("--> %s\n", __func__);
 553         spin_lock(&lo->plh_inode->i_lock);
 554         if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
 555                 status = -EAGAIN;
 556         } else if (list_empty(&lo->plh_segs)) {
 557                 int seq;
 558
 559                 do {
 560                         seq = read_seqbegin(&open_state->seqlock);
 561                         nfs4_stateid_copy(dst, &open_state->stateid);
 562                 } while (read_seqretry(&open_state->seqlock, seq));
 563         } else
 564                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 565         spin_unlock(&lo->plh_inode->i_lock);
 566         dprintk("<-- %s\n", __func__);
 567         return status;
 568 }
 569
 570 /*
 571 * Get layout from server.
 572 *    for now, assume that whole file layouts are requested.
 573 *    arg->offset: 0
 574 *    arg->length: all ones
 575 */
 576 static struct pnfs_layout_segment *
 577 send_layoutget(struct pnfs_layout_hdr *lo,
 578            struct nfs_open_context *ctx,
 579            struct pnfs_layout_range *range,
 580            gfp_t gfp_flags)
 581 {
 582         struct inode *ino = lo->plh_inode;
 583         struct nfs_server *server = NFS_SERVER(ino);
 584         struct nfs4_layoutget *lgp;
 585         struct pnfs_layout_segment *lseg;
 586
 587         dprintk("--> %s\n", __func__);
 588
 589         BUG_ON(ctx == NULL);
 590         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 591         if (lgp == NULL)
 592                 return NULL;
 593
 594         lgp->args.minlength = PAGE_CACHE_SIZE;
 595         if (lgp->args.minlength > range->length)
 596                 lgp->args.minlength = range->length;
 597         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 598         lgp->args.range = *range;
 599         lgp->args.type = server->pnfs_curr_ld->id;
 600         lgp->args.inode = ino;
 601         lgp->args.ctx = get_nfs_open_context(ctx);
 602         lgp->gfp_flags = gfp_flags;
 603
 604         /* Synchronously retrieve layout information from server and
 605          * store in lseg.
 606          */
 607         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 608         if (IS_ERR(lseg)) {
 609                 switch (PTR_ERR(lseg)) {
 610                 case -ENOMEM:
 611                 case -ERESTARTSYS:
 612                         break;
 613                 default:
 614                         /* remember that LAYOUTGET failed and suspend trying */
 615                         set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
 616                 }
 617                 return NULL;
 618         }
 619
 620         return lseg;
 621 }
 622
 623 /*
 624  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 625  * when the layout segment list is empty.
 626  *
 627  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 628  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 629  * deviceid is marked invalid.
 630  */
 631 int
 632 _pnfs_return_layout(struct inode *ino)
 633 {
 634         struct pnfs_layout_hdr *lo = NULL;
 635         struct nfs_inode *nfsi = NFS_I(ino);
 636         LIST_HEAD(tmp_list);
 637         struct nfs4_layoutreturn *lrp;
 638         nfs4_stateid stateid;
 639         int status = 0, empty;
 640
 641         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 642
 643         spin_lock(&ino->i_lock);
 644         lo = nfsi->layout;
 645         if (!lo || pnfs_test_layout_returned(lo)) {
 646                 spin_unlock(&ino->i_lock);
 647                 dprintk("NFS: %s no layout to return\n", __func__);
 648                 goto out;
 649         }
 650         stateid = nfsi->layout->plh_stateid;
 651         /* Reference matched in nfs4_layoutreturn_release */
 652         pnfs_get_layout_hdr(lo);
 653         empty = list_empty(&lo->plh_segs);
 654         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 655         /* Don't send a LAYOUTRETURN if list was initially empty */
 656         if (empty) {
 657                 spin_unlock(&ino->i_lock);
 658                 pnfs_put_layout_hdr(lo);
 659                 dprintk("NFS: %s no layout segments to return\n", __func__);
 660                 goto out;
 661         }
 662         lo->plh_block_lgets++;
 663         pnfs_mark_layout_returned(lo);
 664         spin_unlock(&ino->i_lock);
 665         pnfs_free_lseg_list(&tmp_list);
 666
 667         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 668
 669         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 670         if (unlikely(lrp == NULL)) {
 671                 status = -ENOMEM;
 672                 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
 673                 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
 674                 pnfs_clear_layout_returned(lo);
 675                 pnfs_put_layout_hdr(lo);
 676                 goto out;
 677         }
 678
 679         lrp->args.stateid = stateid;
 680         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 681         lrp->args.inode = ino;
 682         lrp->args.layout = lo;
 683         lrp->clp = NFS_SERVER(ino)->nfs_client;
 684
 685         status = nfs4_proc_layoutreturn(lrp);
 686 out:
 687         dprintk("<-- %s status: %d\n", __func__, status);
 688         return status;
 689 }
 690 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 691
 692 bool pnfs_roc(struct inode *ino)
 693 {
 694         struct pnfs_layout_hdr *lo;
 695         struct pnfs_layout_segment *lseg, *tmp;
 696         LIST_HEAD(tmp_list);
 697         bool found = false;
 698
 699         spin_lock(&ino->i_lock);
 700         lo = NFS_I(ino)->layout;
 701         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 702             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 703                 goto out_nolayout;
 704         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 705                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 706                         mark_lseg_invalid(lseg, &tmp_list);
 707                         found = true;
 708                 }
 709         if (!found)
 710                 goto out_nolayout;
 711         lo->plh_block_lgets++;
 712         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 713         spin_unlock(&ino->i_lock);
 714         pnfs_free_lseg_list(&tmp_list);
 715         return true;
 716
 717 out_nolayout:
 718         spin_unlock(&ino->i_lock);
 719         return false;
 720 }
 721
 722 void pnfs_roc_release(struct inode *ino)
 723 {
 724         struct pnfs_layout_hdr *lo;
 725
 726         spin_lock(&ino->i_lock);
 727         lo = NFS_I(ino)->layout;
 728         lo->plh_block_lgets--;
 729         pnfs_put_layout_hdr_locked(lo);
 730         spin_unlock(&ino->i_lock);
 731 }
 732
 733 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 734 {
 735         struct pnfs_layout_hdr *lo;
 736
 737         spin_lock(&ino->i_lock);
 738         lo = NFS_I(ino)->layout;
 739         if ((int)(barrier - lo->plh_barrier) > 0)
 740                 lo->plh_barrier = barrier;
 741         spin_unlock(&ino->i_lock);
 742 }
 743
 744 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
 745 {
 746         struct nfs_inode *nfsi = NFS_I(ino);
 747         struct pnfs_layout_segment *lseg;
 748         bool found = false;
 749
 750         spin_lock(&ino->i_lock);
 751         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 752                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 753                         found = true;
 754                         break;
 755                 }
 756         if (!found) {
 757                 struct pnfs_layout_hdr *lo = nfsi->layout;
 758                 u32 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 759
 760                 /* Since close does not return a layout stateid for use as
 761                  * a barrier, we choose the worst-case barrier.
 762                  */
 763                 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 764         }
 765         spin_unlock(&ino->i_lock);
 766         return found;
 767 }
 768
 769 /*
 770  * Compare two layout segments for sorting into layout cache.
 771  * We want to preferentially return RW over RO layouts, so ensure those
 772  * are seen first.
 773  */
 774 static s64
 775 cmp_layout(struct pnfs_layout_range *l1,
 776            struct pnfs_layout_range *l2)
 777 {
 778         s64 d;
 779
 780         /* high offset > low offset */
 781         d = l1->offset - l2->offset;
 782         if (d)
 783                 return d;
 784
 785         /* short length > long length */
 786         d = l2->length - l1->length;
 787         if (d)
 788                 return d;
 789
 790         /* read > read/write */
 791         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 792 }
 793
 794 static void
 795 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
 796                    struct pnfs_layout_segment *lseg)
 797 {
 798         struct pnfs_layout_segment *lp;
 799
 800         dprintk("%s:Begin\n", __func__);
 801
 802         assert_spin_locked(&lo->plh_inode->i_lock);
 803         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 804                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 805                         continue;
 806                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 807                 dprintk("%s: inserted lseg %p "
 808                         "iomode %d offset %llu length %llu before "
 809                         "lp %p iomode %d offset %llu length %llu\n",
 810                         __func__, lseg, lseg->pls_range.iomode,
 811                         lseg->pls_range.offset, lseg->pls_range.length,
 812                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 813                         lp->pls_range.length);
 814                 goto out;
 815         }
 816         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 817         dprintk("%s: inserted lseg %p "
 818                 "iomode %d offset %llu length %llu at tail\n",
 819                 __func__, lseg, lseg->pls_range.iomode,
 820                 lseg->pls_range.offset, lseg->pls_range.length);
 821 out:
 822         pnfs_get_layout_hdr(lo);
 823
 824         dprintk("%s:Return\n", __func__);
 825 }
 826
 827 static struct pnfs_layout_hdr *
 828 alloc_init_layout_hdr(struct inode *ino,
 829                       struct nfs_open_context *ctx,
 830                       gfp_t gfp_flags)
 831 {
 832         struct pnfs_layout_hdr *lo;
 833
 834         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 835         if (!lo)
 836                 return NULL;
 837         atomic_set(&lo->plh_refcount, 1);
 838         INIT_LIST_HEAD(&lo->plh_layouts);
 839         INIT_LIST_HEAD(&lo->plh_segs);
 840         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 841         lo->plh_inode = ino;
 842         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 843         return lo;
 844 }
 845
 846 static struct pnfs_layout_hdr *
 847 pnfs_find_alloc_layout(struct inode *ino,
 848                        struct nfs_open_context *ctx,
 849                        gfp_t gfp_flags)
 850 {
 851         struct nfs_inode *nfsi = NFS_I(ino);
 852         struct pnfs_layout_hdr *new = NULL;
 853
 854         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 855
 856         assert_spin_locked(&ino->i_lock);
 857         if (nfsi->layout) {
 858                 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
 859                         return NULL;
 860                 else
 861                         return nfsi->layout;
 862         }
 863         spin_unlock(&ino->i_lock);
 864         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 865         spin_lock(&ino->i_lock);
 866
 867         if (likely(nfsi->layout == NULL))       /* Won the race? */
 868                 nfsi->layout = new;
 869         else
 870                 pnfs_free_layout_hdr(new);
 871         return nfsi->layout;
 872 }
 873
 874 /*
 875  * iomode matching rules:
 876  * iomode       lseg    match
 877  * -----        -----   -----
 878  * ANY          READ    true
 879  * ANY          RW      true
 880  * RW           READ    false
 881  * RW           RW      true
 882  * READ         READ    true
 883  * READ         RW      true
 884  */
 885 static int
 886 is_matching_lseg(struct pnfs_layout_range *ls_range,
 887                  struct pnfs_layout_range *range)
 888 {
 889         struct pnfs_layout_range range1;
 890
 891         if ((range->iomode == IOMODE_RW &&
 892              ls_range->iomode != IOMODE_RW) ||
 893             !lo_seg_intersecting(ls_range, range))
 894                 return 0;
 895
 896         /* range1 covers only the first byte in the range */
 897         range1 = *range;
 898         range1.length = 1;
 899         return lo_seg_contained(ls_range, &range1);
 900 }
 901
 902 /*
 903  * lookup range in layout
 904  */
 905 static struct pnfs_layout_segment *
 906 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 907                 struct pnfs_layout_range *range)
 908 {
 909         struct pnfs_layout_segment *lseg, *ret = NULL;
 910
 911         dprintk("%s:Begin\n", __func__);
 912
 913         assert_spin_locked(&lo->plh_inode->i_lock);
 914         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 915                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 916                     is_matching_lseg(&lseg->pls_range, range)) {
 917                         ret = pnfs_get_lseg(lseg);
 918                         break;
 919                 }
 920                 if (lseg->pls_range.offset > range->offset)
 921                         break;
 922         }
 923
 924         dprintk("%s:Return lseg %p ref %d\n",
 925                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 926         return ret;
 927 }
 928
 929 /*
 930  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 931  * to the MDS or over pNFS
 932  *
 933  * The nfs_inode read_io and write_io fields are cumulative counters reset
 934  * when there are no layout segments. Note that in pnfs_update_layout iomode
 935  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 936  * WRITE request.
 937  *
 938  * A return of true means use MDS I/O.
 939  *
 940  * From rfc 5661:
 941  * If a file's size is smaller than the file size threshold, data accesses
 942  * SHOULD be sent to the metadata server.  If an I/O request has a length that
 943  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
 944  * server.  If both file size and I/O size are provided, the client SHOULD
 945  * reach or exceed  both thresholds before sending its read or write
 946  * requests to the data server.
 947  */
 948 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
 949                                      struct inode *ino, int iomode)
 950 {
 951         struct nfs4_threshold *t = ctx->mdsthreshold;
 952         struct nfs_inode *nfsi = NFS_I(ino);
 953         loff_t fsize = i_size_read(ino);
 954         bool size = false, size_set = false, io = false, io_set = false, ret = false;
 955
 956         if (t == NULL)
 957                 return ret;
 958
 959         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
 960                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
 961
 962         switch (iomode) {
 963         case IOMODE_READ:
 964                 if (t->bm & THRESHOLD_RD) {
 965                         dprintk("%s fsize %llu\n", __func__, fsize);
 966                         size_set = true;
 967                         if (fsize < t->rd_sz)
 968                                 size = true;
 969                 }
 970                 if (t->bm & THRESHOLD_RD_IO) {
 971                         dprintk("%s nfsi->read_io %llu\n", __func__,
 972                                 nfsi->read_io);
 973                         io_set = true;
 974                         if (nfsi->read_io < t->rd_io_sz)
 975                                 io = true;
 976                 }
 977                 break;
 978         case IOMODE_RW:
 979                 if (t->bm & THRESHOLD_WR) {
 980                         dprintk("%s fsize %llu\n", __func__, fsize);
 981                         size_set = true;
 982                         if (fsize < t->wr_sz)
 983                                 size = true;
 984                 }
 985                 if (t->bm & THRESHOLD_WR_IO) {
 986                         dprintk("%s nfsi->write_io %llu\n", __func__,
 987                                 nfsi->write_io);
 988                         io_set = true;
 989                         if (nfsi->write_io < t->wr_io_sz)
 990                                 io = true;
 991                 }
 992                 break;
 993         }
 994         if (size_set && io_set) {
 995                 if (size && io)
 996                         ret = true;
 997         } else if (size || io)
 998                 ret = true;
 999
1000         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1001         return ret;
1002 }
1003
1004 /*
1005  * Layout segment is retreived from the server if not cached.
1006  * The appropriate layout segment is referenced and returned to the caller.
1007  */
1008 struct pnfs_layout_segment *
1009 pnfs_update_layout(struct inode *ino,
1010                    struct nfs_open_context *ctx,
1011                    loff_t pos,
1012                    u64 count,
1013                    enum pnfs_iomode iomode,
1014                    gfp_t gfp_flags)
1015 {
1016         struct pnfs_layout_range arg = {
1017                 .iomode = iomode,
1018                 .offset = pos,
1019                 .length = count,
1020         };
1021         unsigned pg_offset;
1022         struct nfs_inode *nfsi = NFS_I(ino);
1023         struct nfs_server *server = NFS_SERVER(ino);
1024         struct nfs_client *clp = server->nfs_client;
1025         struct pnfs_layout_hdr *lo;
1026         struct pnfs_layout_segment *lseg = NULL;
1027         bool first = false;
1028
1029         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1030                 goto out;
1031
1032         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1033                 goto out;
1034
1035         spin_lock(&ino->i_lock);
1036         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1037         if (lo == NULL)
1038                 goto out_unlock;
1039
1040         /* Do we even need to bother with this? */
1041         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1042                 dprintk("%s matches recall, use MDS\n", __func__);
1043                 goto out_unlock;
1044         }
1045
1046         /* if LAYOUTGET already failed once we don't try again */
1047         if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
1048                 goto out_unlock;
1049
1050         /* Check to see if the layout for the given range already exists */
1051         lseg = pnfs_find_lseg(lo, &arg);
1052         if (lseg)
1053                 goto out_unlock;
1054
1055         if (pnfs_layoutgets_blocked(lo, NULL, 0))
1056                 goto out_unlock;
1057         atomic_inc(&lo->plh_outstanding);
1058
1059         pnfs_get_layout_hdr(lo);
1060         if (list_empty(&lo->plh_segs))
1061                 first = true;
1062
1063         /* Enable LAYOUTRETURNs */
1064         pnfs_clear_layout_returned(lo);
1065
1066         spin_unlock(&ino->i_lock);
1067         if (first) {
1068                 /* The lo must be on the clp list if there is any
1069                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1070                  */
1071                 spin_lock(&clp->cl_lock);
1072                 BUG_ON(!list_empty(&lo->plh_layouts));
1073                 list_add_tail(&lo->plh_layouts, &server->layouts);
1074                 spin_unlock(&clp->cl_lock);
1075         }
1076
1077         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1078         if (pg_offset) {
1079                 arg.offset -= pg_offset;
1080                 arg.length += pg_offset;
1081         }
1082         if (arg.length != NFS4_MAX_UINT64)
1083                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1084
1085         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1086         if (!lseg && first) {
1087                 spin_lock(&clp->cl_lock);
1088                 list_del_init(&lo->plh_layouts);
1089                 spin_unlock(&clp->cl_lock);
1090         }
1091         atomic_dec(&lo->plh_outstanding);
1092         pnfs_put_layout_hdr(lo);
1093 out:
1094         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1095                         "(%s, offset: %llu, length: %llu)\n",
1096                         __func__, ino->i_sb->s_id,
1097                         (unsigned long long)NFS_FILEID(ino),
1098                         lseg == NULL ? "not found" : "found",
1099                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1100                         (unsigned long long)pos,
1101                         (unsigned long long)count);
1102         return lseg;
1103 out_unlock:
1104         spin_unlock(&ino->i_lock);
1105         goto out;
1106 }
1107 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1108
1109 struct pnfs_layout_segment *
1110 pnfs_layout_process(struct nfs4_layoutget *lgp)
1111 {
1112         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1113         struct nfs4_layoutget_res *res = &lgp->res;
1114         struct pnfs_layout_segment *lseg;
1115         struct inode *ino = lo->plh_inode;
1116         int status = 0;
1117
1118         /* Inject layout blob into I/O device driver */
1119         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1120         if (!lseg || IS_ERR(lseg)) {
1121                 if (!lseg)
1122                         status = -ENOMEM;
1123                 else
1124                         status = PTR_ERR(lseg);
1125                 dprintk("%s: Could not allocate layout: error %d\n",
1126                        __func__, status);
1127                 goto out;
1128         }
1129
1130         spin_lock(&ino->i_lock);
1131         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1132                 dprintk("%s forget reply due to recall\n", __func__);
1133                 goto out_forget_reply;
1134         }
1135
1136         if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1137                 dprintk("%s forget reply due to state\n", __func__);
1138                 goto out_forget_reply;
1139         }
1140         init_lseg(lo, lseg);
1141         lseg->pls_range = res->range;
1142         pnfs_get_lseg(lseg);
1143         pnfs_insert_layout(lo, lseg);
1144
1145         if (res->return_on_close) {
1146                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1147                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1148         }
1149
1150         /* Done processing layoutget. Set the layout stateid */
1151         pnfs_set_layout_stateid(lo, &res->stateid, false);
1152         spin_unlock(&ino->i_lock);
1153         return lseg;
1154 out:
1155         return ERR_PTR(status);
1156
1157 out_forget_reply:
1158         spin_unlock(&ino->i_lock);
1159         lseg->pls_layout = lo;
1160         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1161         goto out;
1162 }
1163
1164 void
1165 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1166 {
1167         BUG_ON(pgio->pg_lseg != NULL);
1168
1169         if (req->wb_offset != req->wb_pgbase) {
1170                 nfs_pageio_reset_read_mds(pgio);
1171                 return;
1172         }
1173         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1174                                            req->wb_context,
1175                                            req_offset(req),
1176                                            req->wb_bytes,
1177                                            IOMODE_READ,
1178                                            GFP_KERNEL);
1179         /* If no lseg, fall back to read through mds */
1180         if (pgio->pg_lseg == NULL)
1181                 nfs_pageio_reset_read_mds(pgio);
1182
1183 }
1184 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1185
1186 void
1187 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1188 {
1189         BUG_ON(pgio->pg_lseg != NULL);
1190
1191         if (req->wb_offset != req->wb_pgbase) {
1192                 nfs_pageio_reset_write_mds(pgio);
1193                 return;
1194         }
1195         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1196                                            req->wb_context,
1197                                            req_offset(req),
1198                                            req->wb_bytes,
1199                                            IOMODE_RW,
1200                                            GFP_NOFS);
1201         /* If no lseg, fall back to write through mds */
1202         if (pgio->pg_lseg == NULL)
1203                 nfs_pageio_reset_write_mds(pgio);
1204 }
1205 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1206
1207 void
1208 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1209                       const struct nfs_pgio_completion_ops *compl_ops)
1210 {
1211         struct nfs_server *server = NFS_SERVER(inode);
1212         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1213
1214         if (ld == NULL)
1215                 nfs_pageio_init_read(pgio, inode, compl_ops);
1216         else
1217                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1218 }
1219
1220 void
1221 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1222                        int ioflags,
1223                        const struct nfs_pgio_completion_ops *compl_ops)
1224 {
1225         struct nfs_server *server = NFS_SERVER(inode);
1226         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1227
1228         if (ld == NULL)
1229                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1230         else
1231                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1232 }
1233
1234 bool
1235 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1236                      struct nfs_page *req)
1237 {
1238         if (pgio->pg_lseg == NULL)
1239                 return nfs_generic_pg_test(pgio, prev, req);
1240
1241         /*
1242          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1243          * Note that this test makes several assumptions:
1244          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1245          *   is known to lie within the range.
1246          *   - that the nfs_page being tested is known to be contiguous with the
1247          *   previous nfs_page.
1248          *   - Layout ranges are page aligned, so we only have to test the
1249          *   start offset of the request.
1250          *
1251          * Please also note that 'end_offset' is actually the offset of the
1252          * first byte that lies outside the pnfs_layout_range. FIXME?
1253          *
1254          */
1255         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1256                                          pgio->pg_lseg->pls_range.length);
1257 }
1258 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1259
1260 int pnfs_write_done_resend_to_mds(struct inode *inode,
1261                                 struct list_head *head,
1262                                 const struct nfs_pgio_completion_ops *compl_ops)
1263 {
1264         struct nfs_pageio_descriptor pgio;
1265         LIST_HEAD(failed);
1266
1267         /* Resend all requests through the MDS */
1268         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1269         while (!list_empty(head)) {
1270                 struct nfs_page *req = nfs_list_entry(head->next);
1271
1272                 nfs_list_remove_request(req);
1273                 if (!nfs_pageio_add_request(&pgio, req))
1274                         nfs_list_add_request(req, &failed);
1275         }
1276         nfs_pageio_complete(&pgio);
1277
1278         if (!list_empty(&failed)) {
1279                 /* For some reason our attempt to resend pages. Mark the
1280                  * overall send request as having failed, and let
1281                  * nfs_writeback_release_full deal with the error.
1282                  */
1283                 list_move(&failed, head);
1284                 return -EIO;
1285         }
1286         return 0;
1287 }
1288 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1289
1290 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1291 {
1292         struct nfs_pgio_header *hdr = data->header;
1293
1294         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1295         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1296             PNFS_LAYOUTRET_ON_ERROR) {
1297                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1298                 pnfs_return_layout(hdr->inode);
1299         }
1300         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1301                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1302                                                         &hdr->pages,
1303                                                         hdr->completion_ops);
1304 }
1305
1306 /*
1307  * Called by non rpc-based layout drivers
1308  */
1309 void pnfs_ld_write_done(struct nfs_write_data *data)
1310 {
1311         struct nfs_pgio_header *hdr = data->header;
1312
1313         if (!hdr->pnfs_error) {
1314                 pnfs_set_layoutcommit(data);
1315                 hdr->mds_ops->rpc_call_done(&data->task, data);
1316         } else
1317                 pnfs_ld_handle_write_error(data);
1318         hdr->mds_ops->rpc_release(data);
1319 }
1320 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1321
1322 static void
1323 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1324                 struct nfs_write_data *data)
1325 {
1326         struct nfs_pgio_header *hdr = data->header;
1327
1328         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1329                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1330                 nfs_pageio_reset_write_mds(desc);
1331                 desc->pg_recoalesce = 1;
1332         }
1333         nfs_writedata_release(data);
1334 }
1335
1336 static enum pnfs_try_status
1337 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1338                         const struct rpc_call_ops *call_ops,
1339                         struct pnfs_layout_segment *lseg,
1340                         int how)
1341 {
1342         struct nfs_pgio_header *hdr = wdata->header;
1343         struct inode *inode = hdr->inode;
1344         enum pnfs_try_status trypnfs;
1345         struct nfs_server *nfss = NFS_SERVER(inode);
1346
1347         hdr->mds_ops = call_ops;
1348
1349         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1350                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1351         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1352         if (trypnfs != PNFS_NOT_ATTEMPTED)
1353                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1354         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1355         return trypnfs;
1356 }
1357
1358 static void
1359 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1360 {
1361         struct nfs_write_data *data;
1362         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1363         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1364
1365         desc->pg_lseg = NULL;
1366         while (!list_empty(head)) {
1367                 enum pnfs_try_status trypnfs;
1368
1369                 data = list_first_entry(head, struct nfs_write_data, list);
1370                 list_del_init(&data->list);
1371
1372                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1373                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1374                         pnfs_write_through_mds(desc, data);
1375         }
1376         pnfs_put_lseg(lseg);
1377 }
1378
1379 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1380 {
1381         pnfs_put_lseg(hdr->lseg);
1382         nfs_writehdr_free(hdr);
1383 }
1384 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1385
1386 int
1387 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1388 {
1389         struct nfs_write_header *whdr;
1390         struct nfs_pgio_header *hdr;
1391         int ret;
1392
1393         whdr = nfs_writehdr_alloc();
1394         if (!whdr) {
1395                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1396                 pnfs_put_lseg(desc->pg_lseg);
1397                 desc->pg_lseg = NULL;
1398                 return -ENOMEM;
1399         }
1400         hdr = &whdr->header;
1401         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1402         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1403         atomic_inc(&hdr->refcnt);
1404         ret = nfs_generic_flush(desc, hdr);
1405         if (ret != 0) {
1406                 pnfs_put_lseg(desc->pg_lseg);
1407                 desc->pg_lseg = NULL;
1408         } else
1409                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1410         if (atomic_dec_and_test(&hdr->refcnt))
1411                 hdr->completion_ops->completion(hdr);
1412         return ret;
1413 }
1414 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1415
1416 int pnfs_read_done_resend_to_mds(struct inode *inode,
1417                                 struct list_head *head,
1418                                 const struct nfs_pgio_completion_ops *compl_ops)
1419 {
1420         struct nfs_pageio_descriptor pgio;
1421         LIST_HEAD(failed);
1422
1423         /* Resend all requests through the MDS */
1424         nfs_pageio_init_read(&pgio, inode, compl_ops);
1425         while (!list_empty(head)) {
1426                 struct nfs_page *req = nfs_list_entry(head->next);
1427
1428                 nfs_list_remove_request(req);
1429                 if (!nfs_pageio_add_request(&pgio, req))
1430                         nfs_list_add_request(req, &failed);
1431         }
1432         nfs_pageio_complete(&pgio);
1433
1434         if (!list_empty(&failed)) {
1435                 list_move(&failed, head);
1436                 return -EIO;
1437         }
1438         return 0;
1439 }
1440 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1441
1442 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1443 {
1444         struct nfs_pgio_header *hdr = data->header;
1445
1446         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1447         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1448             PNFS_LAYOUTRET_ON_ERROR) {
1449                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1450                 pnfs_return_layout(hdr->inode);
1451         }
1452         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1453                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1454                                                         &hdr->pages,
1455                                                         hdr->completion_ops);
1456 }
1457
1458 /*
1459  * Called by non rpc-based layout drivers
1460  */
1461 void pnfs_ld_read_done(struct nfs_read_data *data)
1462 {
1463         struct nfs_pgio_header *hdr = data->header;
1464
1465         if (likely(!hdr->pnfs_error)) {
1466                 __nfs4_read_done_cb(data);
1467                 hdr->mds_ops->rpc_call_done(&data->task, data);
1468         } else
1469                 pnfs_ld_handle_read_error(data);
1470         hdr->mds_ops->rpc_release(data);
1471 }
1472 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1473
1474 static void
1475 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1476                 struct nfs_read_data *data)
1477 {
1478         struct nfs_pgio_header *hdr = data->header;
1479
1480         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1481                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1482                 nfs_pageio_reset_read_mds(desc);
1483                 desc->pg_recoalesce = 1;
1484         }
1485         nfs_readdata_release(data);
1486 }
1487
1488 /*
1489  * Call the appropriate parallel I/O subsystem read function.
1490  */
1491 static enum pnfs_try_status
1492 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1493                        const struct rpc_call_ops *call_ops,
1494                        struct pnfs_layout_segment *lseg)
1495 {
1496         struct nfs_pgio_header *hdr = rdata->header;
1497         struct inode *inode = hdr->inode;
1498         struct nfs_server *nfss = NFS_SERVER(inode);
1499         enum pnfs_try_status trypnfs;
1500
1501         hdr->mds_ops = call_ops;
1502
1503         dprintk("%s: Reading ino:%lu %u@%llu\n",
1504                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1505
1506         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1507         if (trypnfs != PNFS_NOT_ATTEMPTED)
1508                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1509         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1510         return trypnfs;
1511 }
1512
1513 static void
1514 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1515 {
1516         struct nfs_read_data *data;
1517         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1518         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1519
1520         desc->pg_lseg = NULL;
1521         while (!list_empty(head)) {
1522                 enum pnfs_try_status trypnfs;
1523
1524                 data = list_first_entry(head, struct nfs_read_data, list);
1525                 list_del_init(&data->list);
1526
1527                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1528                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1529                         pnfs_read_through_mds(desc, data);
1530         }
1531         pnfs_put_lseg(lseg);
1532 }
1533
1534 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1535 {
1536         pnfs_put_lseg(hdr->lseg);
1537         nfs_readhdr_free(hdr);
1538 }
1539 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1540
1541 int
1542 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1543 {
1544         struct nfs_read_header *rhdr;
1545         struct nfs_pgio_header *hdr;
1546         int ret;
1547
1548         rhdr = nfs_readhdr_alloc();
1549         if (!rhdr) {
1550                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1551                 ret = -ENOMEM;
1552                 pnfs_put_lseg(desc->pg_lseg);
1553                 desc->pg_lseg = NULL;
1554                 return ret;
1555         }
1556         hdr = &rhdr->header;
1557         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1558         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1559         atomic_inc(&hdr->refcnt);
1560         ret = nfs_generic_pagein(desc, hdr);
1561         if (ret != 0) {
1562                 pnfs_put_lseg(desc->pg_lseg);
1563                 desc->pg_lseg = NULL;
1564         } else
1565                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1566         if (atomic_dec_and_test(&hdr->refcnt))
1567                 hdr->completion_ops->completion(hdr);
1568         return ret;
1569 }
1570 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1571
1572 /*
1573  * There can be multiple RW segments.
1574  */
1575 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1576 {
1577         struct pnfs_layout_segment *lseg;
1578
1579         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1580                 if (lseg->pls_range.iomode == IOMODE_RW &&
1581                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1582                         list_add(&lseg->pls_lc_list, listp);
1583         }
1584 }
1585
1586 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1587 {
1588         if (lseg->pls_range.iomode == IOMODE_RW) {
1589                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
1590                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
1591         } else {
1592                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
1593                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
1594         }
1595 }
1596 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1597
1598 void
1599 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1600 {
1601         struct nfs_pgio_header *hdr = wdata->header;
1602         struct inode *inode = hdr->inode;
1603         struct nfs_inode *nfsi = NFS_I(inode);
1604         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1605         bool mark_as_dirty = false;
1606
1607         spin_lock(&inode->i_lock);
1608         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1609                 mark_as_dirty = true;
1610                 dprintk("%s: Set layoutcommit for inode %lu ",
1611                         __func__, inode->i_ino);
1612         }
1613         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1614                 /* references matched in nfs4_layoutcommit_release */
1615                 pnfs_get_lseg(hdr->lseg);
1616         }
1617         if (end_pos > nfsi->layout->plh_lwb)
1618                 nfsi->layout->plh_lwb = end_pos;
1619         spin_unlock(&inode->i_lock);
1620         dprintk("%s: lseg %p end_pos %llu\n",
1621                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1622
1623         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1624          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1625         if (mark_as_dirty)
1626                 mark_inode_dirty_sync(inode);
1627 }
1628 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1629
1630 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1631 {
1632         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1633
1634         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1635                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1636 }
1637
1638 /*
1639  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1640  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1641  * data to disk to allow the server to recover the data if it crashes.
1642  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1643  * is off, and a COMMIT is sent to a data server, or
1644  * if WRITEs to a data server return NFS_DATA_SYNC.
1645  */
1646 int
1647 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1648 {
1649         struct nfs4_layoutcommit_data *data;
1650         struct nfs_inode *nfsi = NFS_I(inode);
1651         loff_t end_pos;
1652         int status = 0;
1653
1654         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1655
1656         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1657                 return 0;
1658
1659         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1660         data = kzalloc(sizeof(*data), GFP_NOFS);
1661         if (!data) {
1662                 status = -ENOMEM;
1663                 goto out;
1664         }
1665
1666         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1667                 goto out_free;
1668
1669         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1670                 if (!sync) {
1671                         status = -EAGAIN;
1672                         goto out_free;
1673                 }
1674                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1675                                         nfs_wait_bit_killable, TASK_KILLABLE);
1676                 if (status)
1677                         goto out_free;
1678         }
1679
1680         INIT_LIST_HEAD(&data->lseg_list);
1681         spin_lock(&inode->i_lock);
1682         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1683                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1684                 spin_unlock(&inode->i_lock);
1685                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1686                 goto out_free;
1687         }
1688
1689         pnfs_list_write_lseg(inode, &data->lseg_list);
1690
1691         end_pos = nfsi->layout->plh_lwb;
1692         nfsi->layout->plh_lwb = 0;
1693
1694         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1695         spin_unlock(&inode->i_lock);
1696
1697         data->args.inode = inode;
1698         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1699         nfs_fattr_init(&data->fattr);
1700         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1701         data->res.fattr = &data->fattr;
1702         data->args.lastbytewritten = end_pos - 1;
1703         data->res.server = NFS_SERVER(inode);
1704
1705         status = nfs4_proc_layoutcommit(data, sync);
1706 out:
1707         if (status)
1708                 mark_inode_dirty_sync(inode);
1709         dprintk("<-- %s status %d\n", __func__, status);
1710         return status;
1711 out_free:
1712         kfree(data);
1713         goto out;
1714 }
1715
1716 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1717 {
1718         struct nfs4_threshold *thp;
1719
1720         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1721         if (!thp) {
1722                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1723                 return NULL;
1724         }
1725         return thp;
1726 }