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 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  39
  40 /* Locking:
  41  *
  42  * pnfs_spinlock:
  43  *      protects pnfs_modules_tbl.
  44  */
  45 static DEFINE_SPINLOCK(pnfs_spinlock);
  46
  47 /*
  48  * pnfs_modules_tbl holds all pnfs modules
  49  */
  50 static LIST_HEAD(pnfs_modules_tbl);
  51
  52 /* Return the registered pnfs layout driver module matching given id */
  53 static struct pnfs_layoutdriver_type *
  54 find_pnfs_driver_locked(u32 id)
  55 {
  56         struct pnfs_layoutdriver_type *local;
  57
  58         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  59                 if (local->id == id)
  60                         goto out;
  61         local = NULL;
  62 out:
  63         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  64         return local;
  65 }
  66
  67 static struct pnfs_layoutdriver_type *
  68 find_pnfs_driver(u32 id)
  69 {
  70         struct pnfs_layoutdriver_type *local;
  71
  72         spin_lock(&pnfs_spinlock);
  73         local = find_pnfs_driver_locked(id);
  74         if (local != NULL && !try_module_get(local->owner)) {
  75                 dprintk("%s: Could not grab reference on module\n", __func__);
  76                 local = NULL;
  77         }
  78         spin_unlock(&pnfs_spinlock);
  79         return local;
  80 }
  81
  82 void
  83 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  84 {
  85         if (nfss->pnfs_curr_ld) {
  86                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  87                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  88                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  89                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  90                         nfs4_deviceid_purge_client(nfss->nfs_client);
  91                 module_put(nfss->pnfs_curr_ld->owner);
  92         }
  93         nfss->pnfs_curr_ld = NULL;
  94 }
  95
  96 /*
  97  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  98  * Currently only one pNFS layout driver per filesystem is supported.
  99  *
 100  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 101  */
 102 void
 103 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 104                       u32 id)
 105 {
 106         struct pnfs_layoutdriver_type *ld_type = NULL;
 107
 108         if (id == 0)
 109                 goto out_no_driver;
 110         if (!(server->nfs_client->cl_exchange_flags &
 111                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 112                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 113                         __func__, id, server->nfs_client->cl_exchange_flags);
 114                 goto out_no_driver;
 115         }
 116         ld_type = find_pnfs_driver(id);
 117         if (!ld_type) {
 118                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 119                 ld_type = find_pnfs_driver(id);
 120                 if (!ld_type) {
 121                         dprintk("%s: No pNFS module found for %u.\n",
 122                                 __func__, id);
 123                         goto out_no_driver;
 124                 }
 125         }
 126         server->pnfs_curr_ld = ld_type;
 127         if (ld_type->set_layoutdriver
 128             && ld_type->set_layoutdriver(server, mntfh)) {
 129                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 130                         "driver %u.\n", __func__, id);
 131                 module_put(ld_type->owner);
 132                 goto out_no_driver;
 133         }
 134         /* Bump the MDS count */
 135         atomic_inc(&server->nfs_client->cl_mds_count);
 136
 137         dprintk("%s: pNFS module for %u set\n", __func__, id);
 138         return;
 139
 140 out_no_driver:
 141         dprintk("%s: Using NFSv4 I/O\n", __func__);
 142         server->pnfs_curr_ld = NULL;
 143 }
 144
 145 int
 146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 147 {
 148         int status = -EINVAL;
 149         struct pnfs_layoutdriver_type *tmp;
 150
 151         if (ld_type->id == 0) {
 152                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 153                 return status;
 154         }
 155         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 156                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 157                        "alloc_lseg and free_lseg.\n", __func__);
 158                 return status;
 159         }
 160
 161         spin_lock(&pnfs_spinlock);
 162         tmp = find_pnfs_driver_locked(ld_type->id);
 163         if (!tmp) {
 164                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 165                 status = 0;
 166                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 167                         ld_type->name);
 168         } else {
 169                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 170                         __func__, ld_type->id);
 171         }
 172         spin_unlock(&pnfs_spinlock);
 173
 174         return status;
 175 }
 176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 177
 178 void
 179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 180 {
 181         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 182         spin_lock(&pnfs_spinlock);
 183         list_del(&ld_type->pnfs_tblid);
 184         spin_unlock(&pnfs_spinlock);
 185 }
 186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 187
 188 /*
 189  * pNFS client layout cache
 190  */
 191
 192 /* Need to hold i_lock if caller does not already hold reference */
 193 void
 194 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 195 {
 196         atomic_inc(&lo->plh_refcount);
 197 }
 198
 199 static struct pnfs_layout_hdr *
 200 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 201 {
 202         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 203         return ld->alloc_layout_hdr(ino, gfp_flags);
 204 }
 205
 206 static void
 207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 208 {
 209         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 210         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 211
 212         if (!list_empty(&lo->plh_layouts)) {
 213                 struct nfs_client *clp = server->nfs_client;
 214
 215                 spin_lock(&clp->cl_lock);
 216                 list_del_init(&lo->plh_layouts);
 217                 spin_unlock(&clp->cl_lock);
 218         }
 219         put_rpccred(lo->plh_lc_cred);
 220         return ld->free_layout_hdr(lo);
 221 }
 222
 223 static void
 224 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 225 {
 226         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 227         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 228         nfsi->layout = NULL;
 229         /* Reset MDS Threshold I/O counters */
 230         nfsi->write_io = 0;
 231         nfsi->read_io = 0;
 232 }
 233
 234 void
 235 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 236 {
 237         struct inode *inode = lo->plh_inode;
 238
 239         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 240                 pnfs_detach_layout_hdr(lo);
 241                 spin_unlock(&inode->i_lock);
 242                 pnfs_free_layout_hdr(lo);
 243         }
 244 }
 245
 246 static int
 247 pnfs_iomode_to_fail_bit(u32 iomode)
 248 {
 249         return iomode == IOMODE_RW ?
 250                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 251 }
 252
 253 static void
 254 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 255 {
 256         lo->plh_retry_timestamp = jiffies;
 257         if (test_and_set_bit(fail_bit, &lo->plh_flags))
 258                 atomic_inc(&lo->plh_refcount);
 259 }
 260
 261 static void
 262 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 263 {
 264         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 265                 atomic_dec(&lo->plh_refcount);
 266 }
 267
 268 static void
 269 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 270 {
 271         struct inode *inode = lo->plh_inode;
 272         struct pnfs_layout_range range = {
 273                 .iomode = iomode,
 274                 .offset = 0,
 275                 .length = NFS4_MAX_UINT64,
 276         };
 277         LIST_HEAD(head);
 278
 279         spin_lock(&inode->i_lock);
 280         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 281         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 282         spin_unlock(&inode->i_lock);
 283         pnfs_free_lseg_list(&head);
 284         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 285                         iomode == IOMODE_RW ?  "RW" : "READ");
 286 }
 287
 288 static bool
 289 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 290 {
 291         unsigned long start, end;
 292         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 293
 294         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 295                 return false;
 296         end = jiffies;
 297         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 298         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 299                 /* It is time to retry the failed layoutgets */
 300                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 301                 return false;
 302         }
 303         return true;
 304 }
 305
 306 static void
 307 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 308 {
 309         INIT_LIST_HEAD(&lseg->pls_list);
 310         INIT_LIST_HEAD(&lseg->pls_lc_list);
 311         atomic_set(&lseg->pls_refcount, 1);
 312         smp_mb();
 313         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 314         lseg->pls_layout = lo;
 315 }
 316
 317 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 318 {
 319         struct inode *ino = lseg->pls_layout->plh_inode;
 320
 321         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 322 }
 323
 324 static void
 325 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 326                 struct pnfs_layout_segment *lseg)
 327 {
 328         struct inode *inode = lo->plh_inode;
 329
 330         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 331         list_del_init(&lseg->pls_list);
 332         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 333         atomic_dec(&lo->plh_refcount);
 334         if (list_empty(&lo->plh_segs))
 335                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 336         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 337 }
 338
 339 void
 340 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 341 {
 342         struct pnfs_layout_hdr *lo;
 343         struct inode *inode;
 344
 345         if (!lseg)
 346                 return;
 347
 348         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 349                 atomic_read(&lseg->pls_refcount),
 350                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 351         lo = lseg->pls_layout;
 352         inode = lo->plh_inode;
 353         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 354                 pnfs_get_layout_hdr(lo);
 355                 pnfs_layout_remove_lseg(lo, lseg);
 356                 spin_unlock(&inode->i_lock);
 357                 pnfs_free_lseg(lseg);
 358                 pnfs_put_layout_hdr(lo);
 359         }
 360 }
 361 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 362
 363 static inline u64
 364 end_offset(u64 start, u64 len)
 365 {
 366         u64 end;
 367
 368         end = start + len;
 369         return end >= start ? end : NFS4_MAX_UINT64;
 370 }
 371
 372 /* last octet in a range */
 373 static inline u64
 374 last_byte_offset(u64 start, u64 len)
 375 {
 376         u64 end;
 377
 378         BUG_ON(!len);
 379         end = start + len;
 380         return end > start ? end - 1 : NFS4_MAX_UINT64;
 381 }
 382
 383 /*
 384  * is l2 fully contained in l1?
 385  *   start1                             end1
 386  *   [----------------------------------)
 387  *           start2           end2
 388  *           [----------------)
 389  */
 390 static inline int
 391 lo_seg_contained(struct pnfs_layout_range *l1,
 392                  struct pnfs_layout_range *l2)
 393 {
 394         u64 start1 = l1->offset;
 395         u64 end1 = end_offset(start1, l1->length);
 396         u64 start2 = l2->offset;
 397         u64 end2 = end_offset(start2, l2->length);
 398
 399         return (start1 <= start2) && (end1 >= end2);
 400 }
 401
 402 /*
 403  * is l1 and l2 intersecting?
 404  *   start1                             end1
 405  *   [----------------------------------)
 406  *                              start2           end2
 407  *                              [----------------)
 408  */
 409 static inline int
 410 lo_seg_intersecting(struct pnfs_layout_range *l1,
 411                     struct pnfs_layout_range *l2)
 412 {
 413         u64 start1 = l1->offset;
 414         u64 end1 = end_offset(start1, l1->length);
 415         u64 start2 = l2->offset;
 416         u64 end2 = end_offset(start2, l2->length);
 417
 418         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 419                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 420 }
 421
 422 static bool
 423 should_free_lseg(struct pnfs_layout_range *lseg_range,
 424                  struct pnfs_layout_range *recall_range)
 425 {
 426         return (recall_range->iomode == IOMODE_ANY ||
 427                 lseg_range->iomode == recall_range->iomode) &&
 428                lo_seg_intersecting(lseg_range, recall_range);
 429 }
 430
 431 /* Returns 1 if lseg is removed from list, 0 otherwise */
 432 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 433                              struct list_head *tmp_list)
 434 {
 435         int rv = 0;
 436
 437         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 438                 /* Remove the reference keeping the lseg in the
 439                  * list.  It will now be removed when all
 440                  * outstanding io is finished.
 441                  */
 442                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 443                         atomic_read(&lseg->pls_refcount));
 444                 if (atomic_dec_and_test(&lseg->pls_refcount)) {
 445                         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 446                         list_add(&lseg->pls_list, tmp_list);
 447                         rv = 1;
 448                 }
 449         }
 450         return rv;
 451 }
 452
 453 /* Returns count of number of matching invalid lsegs remaining in list
 454  * after call.
 455  */
 456 int
 457 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 458                             struct list_head *tmp_list,
 459                             struct pnfs_layout_range *recall_range)
 460 {
 461         struct pnfs_layout_segment *lseg, *next;
 462         int invalid = 0, removed = 0;
 463
 464         dprintk("%s:Begin lo %p\n", __func__, lo);
 465
 466         if (list_empty(&lo->plh_segs))
 467                 return 0;
 468         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 469                 if (!recall_range ||
 470                     should_free_lseg(&lseg->pls_range, recall_range)) {
 471                         dprintk("%s: freeing lseg %p iomode %d "
 472                                 "offset %llu length %llu\n", __func__,
 473                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 474                                 lseg->pls_range.length);
 475                         invalid++;
 476                         removed += mark_lseg_invalid(lseg, tmp_list);
 477                 }
 478         dprintk("%s:Return %i\n", __func__, invalid - removed);
 479         return invalid - removed;
 480 }
 481
 482 /* note free_me must contain lsegs from a single layout_hdr */
 483 void
 484 pnfs_free_lseg_list(struct list_head *free_me)
 485 {
 486         struct pnfs_layout_segment *lseg, *tmp;
 487
 488         if (list_empty(free_me))
 489                 return;
 490
 491         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 492                 list_del(&lseg->pls_list);
 493                 pnfs_free_lseg(lseg);
 494         }
 495 }
 496
 497 void
 498 pnfs_destroy_layout(struct nfs_inode *nfsi)
 499 {
 500         struct pnfs_layout_hdr *lo;
 501         LIST_HEAD(tmp_list);
 502
 503         spin_lock(&nfsi->vfs_inode.i_lock);
 504         lo = nfsi->layout;
 505         if (lo) {
 506                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 507                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 508                 pnfs_get_layout_hdr(lo);
 509                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 510                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 511                 spin_unlock(&nfsi->vfs_inode.i_lock);
 512                 pnfs_free_lseg_list(&tmp_list);
 513                 pnfs_put_layout_hdr(lo);
 514         } else
 515                 spin_unlock(&nfsi->vfs_inode.i_lock);
 516 }
 517 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 518
 519 /*
 520  * Called by the state manger to remove all layouts established under an
 521  * expired lease.
 522  */
 523 void
 524 pnfs_destroy_all_layouts(struct nfs_client *clp)
 525 {
 526         struct nfs_server *server;
 527         struct pnfs_layout_hdr *lo;
 528         LIST_HEAD(tmp_list);
 529
 530         nfs4_deviceid_mark_client_invalid(clp);
 531         nfs4_deviceid_purge_client(clp);
 532
 533         spin_lock(&clp->cl_lock);
 534         rcu_read_lock();
 535         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 536                 if (!list_empty(&server->layouts))
 537                         list_splice_init(&server->layouts, &tmp_list);
 538         }
 539         rcu_read_unlock();
 540         spin_unlock(&clp->cl_lock);
 541
 542         while (!list_empty(&tmp_list)) {
 543                 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
 544                                 plh_layouts);
 545                 dprintk("%s freeing layout for inode %lu\n", __func__,
 546                         lo->plh_inode->i_ino);
 547                 list_del_init(&lo->plh_layouts);
 548                 pnfs_destroy_layout(NFS_I(lo->plh_inode));
 549         }
 550 }
 551
 552 /*
 553  * Compare 2 layout stateid sequence ids, to see which is newer,
 554  * taking into account wraparound issues.
 555  */
 556 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 557 {
 558         return (s32)s1 - (s32)s2 > 0;
 559 }
 560
 561 /* update lo->plh_stateid with new if is more recent */
 562 void
 563 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 564                         bool update_barrier)
 565 {
 566         u32 oldseq, newseq, new_barrier;
 567         int empty = list_empty(&lo->plh_segs);
 568
 569         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 570         newseq = be32_to_cpu(new->seqid);
 571         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 572                 nfs4_stateid_copy(&lo->plh_stateid, new);
 573                 if (update_barrier) {
 574                         new_barrier = be32_to_cpu(new->seqid);
 575                 } else {
 576                         /* Because of wraparound, we want to keep the barrier
 577                          * "close" to the current seqids.
 578                          */
 579                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 580                 }
 581                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 582                         lo->plh_barrier = new_barrier;
 583         }
 584 }
 585
 586 static bool
 587 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 588                 const nfs4_stateid *stateid)
 589 {
 590         u32 seqid = be32_to_cpu(stateid->seqid);
 591
 592         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 593 }
 594
 595 /* lget is set to 1 if called from inside send_layoutget call chain */
 596 static bool
 597 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 598 {
 599         return lo->plh_block_lgets ||
 600                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 601                 (list_empty(&lo->plh_segs) &&
 602                  (atomic_read(&lo->plh_outstanding) > lget));
 603 }
 604
 605 int
 606 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 607                               struct nfs4_state *open_state)
 608 {
 609         int status = 0;
 610
 611         dprintk("--> %s\n", __func__);
 612         spin_lock(&lo->plh_inode->i_lock);
 613         if (pnfs_layoutgets_blocked(lo, 1)) {
 614                 status = -EAGAIN;
 615         } else if (list_empty(&lo->plh_segs)) {
 616                 int seq;
 617
 618                 do {
 619                         seq = read_seqbegin(&open_state->seqlock);
 620                         nfs4_stateid_copy(dst, &open_state->stateid);
 621                 } while (read_seqretry(&open_state->seqlock, seq));
 622         } else
 623                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 624         spin_unlock(&lo->plh_inode->i_lock);
 625         dprintk("<-- %s\n", __func__);
 626         return status;
 627 }
 628
 629 /*
 630 * Get layout from server.
 631 *    for now, assume that whole file layouts are requested.
 632 *    arg->offset: 0
 633 *    arg->length: all ones
 634 */
 635 static struct pnfs_layout_segment *
 636 send_layoutget(struct pnfs_layout_hdr *lo,
 637            struct nfs_open_context *ctx,
 638            struct pnfs_layout_range *range,
 639            gfp_t gfp_flags)
 640 {
 641         struct inode *ino = lo->plh_inode;
 642         struct nfs_server *server = NFS_SERVER(ino);
 643         struct nfs4_layoutget *lgp;
 644         struct pnfs_layout_segment *lseg;
 645
 646         dprintk("--> %s\n", __func__);
 647
 648         BUG_ON(ctx == NULL);
 649         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 650         if (lgp == NULL)
 651                 return NULL;
 652
 653         lgp->args.minlength = PAGE_CACHE_SIZE;
 654         if (lgp->args.minlength > range->length)
 655                 lgp->args.minlength = range->length;
 656         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 657         lgp->args.range = *range;
 658         lgp->args.type = server->pnfs_curr_ld->id;
 659         lgp->args.inode = ino;
 660         lgp->args.ctx = get_nfs_open_context(ctx);
 661         lgp->gfp_flags = gfp_flags;
 662
 663         /* Synchronously retrieve layout information from server and
 664          * store in lseg.
 665          */
 666         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 667         if (IS_ERR(lseg)) {
 668                 switch (PTR_ERR(lseg)) {
 669                 case -ENOMEM:
 670                 case -ERESTARTSYS:
 671                         break;
 672                 default:
 673                         /* remember that LAYOUTGET failed and suspend trying */
 674                         pnfs_layout_io_set_failed(lo, range->iomode);
 675                 }
 676                 return NULL;
 677         }
 678
 679         return lseg;
 680 }
 681
 682 /*
 683  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 684  * when the layout segment list is empty.
 685  *
 686  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 687  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 688  * deviceid is marked invalid.
 689  */
 690 int
 691 _pnfs_return_layout(struct inode *ino)
 692 {
 693         struct pnfs_layout_hdr *lo = NULL;
 694         struct nfs_inode *nfsi = NFS_I(ino);
 695         LIST_HEAD(tmp_list);
 696         struct nfs4_layoutreturn *lrp;
 697         nfs4_stateid stateid;
 698         int status = 0, empty;
 699
 700         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 701
 702         spin_lock(&ino->i_lock);
 703         lo = nfsi->layout;
 704         if (!lo) {
 705                 spin_unlock(&ino->i_lock);
 706                 dprintk("NFS: %s no layout to return\n", __func__);
 707                 goto out;
 708         }
 709         stateid = nfsi->layout->plh_stateid;
 710         /* Reference matched in nfs4_layoutreturn_release */
 711         pnfs_get_layout_hdr(lo);
 712         empty = list_empty(&lo->plh_segs);
 713         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 714         /* Don't send a LAYOUTRETURN if list was initially empty */
 715         if (empty) {
 716                 spin_unlock(&ino->i_lock);
 717                 pnfs_put_layout_hdr(lo);
 718                 dprintk("NFS: %s no layout segments to return\n", __func__);
 719                 goto out;
 720         }
 721         lo->plh_block_lgets++;
 722         spin_unlock(&ino->i_lock);
 723         pnfs_free_lseg_list(&tmp_list);
 724
 725         WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
 726
 727         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 728         if (unlikely(lrp == NULL)) {
 729                 status = -ENOMEM;
 730                 spin_lock(&ino->i_lock);
 731                 lo->plh_block_lgets--;
 732                 spin_unlock(&ino->i_lock);
 733                 pnfs_put_layout_hdr(lo);
 734                 goto out;
 735         }
 736
 737         lrp->args.stateid = stateid;
 738         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 739         lrp->args.inode = ino;
 740         lrp->args.layout = lo;
 741         lrp->clp = NFS_SERVER(ino)->nfs_client;
 742
 743         status = nfs4_proc_layoutreturn(lrp);
 744 out:
 745         dprintk("<-- %s status: %d\n", __func__, status);
 746         return status;
 747 }
 748 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 749
 750 bool pnfs_roc(struct inode *ino)
 751 {
 752         struct pnfs_layout_hdr *lo;
 753         struct pnfs_layout_segment *lseg, *tmp;
 754         LIST_HEAD(tmp_list);
 755         bool found = false;
 756
 757         spin_lock(&ino->i_lock);
 758         lo = NFS_I(ino)->layout;
 759         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 760             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 761                 goto out_nolayout;
 762         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 763                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 764                         mark_lseg_invalid(lseg, &tmp_list);
 765                         found = true;
 766                 }
 767         if (!found)
 768                 goto out_nolayout;
 769         lo->plh_block_lgets++;
 770         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 771         spin_unlock(&ino->i_lock);
 772         pnfs_free_lseg_list(&tmp_list);
 773         return true;
 774
 775 out_nolayout:
 776         spin_unlock(&ino->i_lock);
 777         return false;
 778 }
 779
 780 void pnfs_roc_release(struct inode *ino)
 781 {
 782         struct pnfs_layout_hdr *lo;
 783
 784         spin_lock(&ino->i_lock);
 785         lo = NFS_I(ino)->layout;
 786         lo->plh_block_lgets--;
 787         if (atomic_dec_and_test(&lo->plh_refcount)) {
 788                 pnfs_detach_layout_hdr(lo);
 789                 spin_unlock(&ino->i_lock);
 790                 pnfs_free_layout_hdr(lo);
 791         } else
 792                 spin_unlock(&ino->i_lock);
 793 }
 794
 795 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 796 {
 797         struct pnfs_layout_hdr *lo;
 798
 799         spin_lock(&ino->i_lock);
 800         lo = NFS_I(ino)->layout;
 801         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 802                 lo->plh_barrier = barrier;
 803         spin_unlock(&ino->i_lock);
 804 }
 805
 806 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 807 {
 808         struct nfs_inode *nfsi = NFS_I(ino);
 809         struct pnfs_layout_hdr *lo;
 810         struct pnfs_layout_segment *lseg;
 811         u32 current_seqid;
 812         bool found = false;
 813
 814         spin_lock(&ino->i_lock);
 815         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 816                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 817                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 818                         found = true;
 819                         goto out;
 820                 }
 821         lo = nfsi->layout;
 822         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 823
 824         /* Since close does not return a layout stateid for use as
 825          * a barrier, we choose the worst-case barrier.
 826          */
 827         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 828 out:
 829         spin_unlock(&ino->i_lock);
 830         return found;
 831 }
 832
 833 /*
 834  * Compare two layout segments for sorting into layout cache.
 835  * We want to preferentially return RW over RO layouts, so ensure those
 836  * are seen first.
 837  */
 838 static s64
 839 cmp_layout(struct pnfs_layout_range *l1,
 840            struct pnfs_layout_range *l2)
 841 {
 842         s64 d;
 843
 844         /* high offset > low offset */
 845         d = l1->offset - l2->offset;
 846         if (d)
 847                 return d;
 848
 849         /* short length > long length */
 850         d = l2->length - l1->length;
 851         if (d)
 852                 return d;
 853
 854         /* read > read/write */
 855         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
 856 }
 857
 858 static void
 859 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
 860                    struct pnfs_layout_segment *lseg)
 861 {
 862         struct pnfs_layout_segment *lp;
 863
 864         dprintk("%s:Begin\n", __func__);
 865
 866         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
 867                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
 868                         continue;
 869                 list_add_tail(&lseg->pls_list, &lp->pls_list);
 870                 dprintk("%s: inserted lseg %p "
 871                         "iomode %d offset %llu length %llu before "
 872                         "lp %p iomode %d offset %llu length %llu\n",
 873                         __func__, lseg, lseg->pls_range.iomode,
 874                         lseg->pls_range.offset, lseg->pls_range.length,
 875                         lp, lp->pls_range.iomode, lp->pls_range.offset,
 876                         lp->pls_range.length);
 877                 goto out;
 878         }
 879         list_add_tail(&lseg->pls_list, &lo->plh_segs);
 880         dprintk("%s: inserted lseg %p "
 881                 "iomode %d offset %llu length %llu at tail\n",
 882                 __func__, lseg, lseg->pls_range.iomode,
 883                 lseg->pls_range.offset, lseg->pls_range.length);
 884 out:
 885         pnfs_get_layout_hdr(lo);
 886
 887         dprintk("%s:Return\n", __func__);
 888 }
 889
 890 static struct pnfs_layout_hdr *
 891 alloc_init_layout_hdr(struct inode *ino,
 892                       struct nfs_open_context *ctx,
 893                       gfp_t gfp_flags)
 894 {
 895         struct pnfs_layout_hdr *lo;
 896
 897         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
 898         if (!lo)
 899                 return NULL;
 900         atomic_set(&lo->plh_refcount, 1);
 901         INIT_LIST_HEAD(&lo->plh_layouts);
 902         INIT_LIST_HEAD(&lo->plh_segs);
 903         INIT_LIST_HEAD(&lo->plh_bulk_recall);
 904         lo->plh_inode = ino;
 905         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
 906         return lo;
 907 }
 908
 909 static struct pnfs_layout_hdr *
 910 pnfs_find_alloc_layout(struct inode *ino,
 911                        struct nfs_open_context *ctx,
 912                        gfp_t gfp_flags)
 913 {
 914         struct nfs_inode *nfsi = NFS_I(ino);
 915         struct pnfs_layout_hdr *new = NULL;
 916
 917         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 918
 919         if (nfsi->layout != NULL)
 920                 goto out_existing;
 921         spin_unlock(&ino->i_lock);
 922         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
 923         spin_lock(&ino->i_lock);
 924
 925         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
 926                 nfsi->layout = new;
 927                 return new;
 928         }
 929         pnfs_free_layout_hdr(new);
 930 out_existing:
 931         pnfs_get_layout_hdr(nfsi->layout);
 932         return nfsi->layout;
 933 }
 934
 935 /*
 936  * iomode matching rules:
 937  * iomode       lseg    match
 938  * -----        -----   -----
 939  * ANY          READ    true
 940  * ANY          RW      true
 941  * RW           READ    false
 942  * RW           RW      true
 943  * READ         READ    true
 944  * READ         RW      true
 945  */
 946 static int
 947 is_matching_lseg(struct pnfs_layout_range *ls_range,
 948                  struct pnfs_layout_range *range)
 949 {
 950         struct pnfs_layout_range range1;
 951
 952         if ((range->iomode == IOMODE_RW &&
 953              ls_range->iomode != IOMODE_RW) ||
 954             !lo_seg_intersecting(ls_range, range))
 955                 return 0;
 956
 957         /* range1 covers only the first byte in the range */
 958         range1 = *range;
 959         range1.length = 1;
 960         return lo_seg_contained(ls_range, &range1);
 961 }
 962
 963 /*
 964  * lookup range in layout
 965  */
 966 static struct pnfs_layout_segment *
 967 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
 968                 struct pnfs_layout_range *range)
 969 {
 970         struct pnfs_layout_segment *lseg, *ret = NULL;
 971
 972         dprintk("%s:Begin\n", __func__);
 973
 974         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 975                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
 976                     is_matching_lseg(&lseg->pls_range, range)) {
 977                         ret = pnfs_get_lseg(lseg);
 978                         break;
 979                 }
 980                 if (lseg->pls_range.offset > range->offset)
 981                         break;
 982         }
 983
 984         dprintk("%s:Return lseg %p ref %d\n",
 985                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 986         return ret;
 987 }
 988
 989 /*
 990  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
 991  * to the MDS or over pNFS
 992  *
 993  * The nfs_inode read_io and write_io fields are cumulative counters reset
 994  * when there are no layout segments. Note that in pnfs_update_layout iomode
 995  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
 996  * WRITE request.
 997  *
 998  * A return of true means use MDS I/O.
 999  *
1000  * From rfc 5661:
1001  * If a file's size is smaller than the file size threshold, data accesses
1002  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1003  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1004  * server.  If both file size and I/O size are provided, the client SHOULD
1005  * reach or exceed  both thresholds before sending its read or write
1006  * requests to the data server.
1007  */
1008 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1009                                      struct inode *ino, int iomode)
1010 {
1011         struct nfs4_threshold *t = ctx->mdsthreshold;
1012         struct nfs_inode *nfsi = NFS_I(ino);
1013         loff_t fsize = i_size_read(ino);
1014         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1015
1016         if (t == NULL)
1017                 return ret;
1018
1019         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1020                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1021
1022         switch (iomode) {
1023         case IOMODE_READ:
1024                 if (t->bm & THRESHOLD_RD) {
1025                         dprintk("%s fsize %llu\n", __func__, fsize);
1026                         size_set = true;
1027                         if (fsize < t->rd_sz)
1028                                 size = true;
1029                 }
1030                 if (t->bm & THRESHOLD_RD_IO) {
1031                         dprintk("%s nfsi->read_io %llu\n", __func__,
1032                                 nfsi->read_io);
1033                         io_set = true;
1034                         if (nfsi->read_io < t->rd_io_sz)
1035                                 io = true;
1036                 }
1037                 break;
1038         case IOMODE_RW:
1039                 if (t->bm & THRESHOLD_WR) {
1040                         dprintk("%s fsize %llu\n", __func__, fsize);
1041                         size_set = true;
1042                         if (fsize < t->wr_sz)
1043                                 size = true;
1044                 }
1045                 if (t->bm & THRESHOLD_WR_IO) {
1046                         dprintk("%s nfsi->write_io %llu\n", __func__,
1047                                 nfsi->write_io);
1048                         io_set = true;
1049                         if (nfsi->write_io < t->wr_io_sz)
1050                                 io = true;
1051                 }
1052                 break;
1053         }
1054         if (size_set && io_set) {
1055                 if (size && io)
1056                         ret = true;
1057         } else if (size || io)
1058                 ret = true;
1059
1060         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1061         return ret;
1062 }
1063
1064 /*
1065  * Layout segment is retreived from the server if not cached.
1066  * The appropriate layout segment is referenced and returned to the caller.
1067  */
1068 struct pnfs_layout_segment *
1069 pnfs_update_layout(struct inode *ino,
1070                    struct nfs_open_context *ctx,
1071                    loff_t pos,
1072                    u64 count,
1073                    enum pnfs_iomode iomode,
1074                    gfp_t gfp_flags)
1075 {
1076         struct pnfs_layout_range arg = {
1077                 .iomode = iomode,
1078                 .offset = pos,
1079                 .length = count,
1080         };
1081         unsigned pg_offset;
1082         struct nfs_server *server = NFS_SERVER(ino);
1083         struct nfs_client *clp = server->nfs_client;
1084         struct pnfs_layout_hdr *lo;
1085         struct pnfs_layout_segment *lseg = NULL;
1086         bool first = false;
1087
1088         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1089                 goto out;
1090
1091         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1092                 goto out;
1093
1094         spin_lock(&ino->i_lock);
1095         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1096         if (lo == NULL) {
1097                 spin_unlock(&ino->i_lock);
1098                 goto out;
1099         }
1100
1101         /* Do we even need to bother with this? */
1102         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1103                 dprintk("%s matches recall, use MDS\n", __func__);
1104                 goto out_unlock;
1105         }
1106
1107         /* if LAYOUTGET already failed once we don't try again */
1108         if (pnfs_layout_io_test_failed(lo, iomode))
1109                 goto out_unlock;
1110
1111         /* Check to see if the layout for the given range already exists */
1112         lseg = pnfs_find_lseg(lo, &arg);
1113         if (lseg)
1114                 goto out_unlock;
1115
1116         if (pnfs_layoutgets_blocked(lo, 0))
1117                 goto out_unlock;
1118         atomic_inc(&lo->plh_outstanding);
1119
1120         if (list_empty(&lo->plh_segs))
1121                 first = true;
1122
1123         spin_unlock(&ino->i_lock);
1124         if (first) {
1125                 /* The lo must be on the clp list if there is any
1126                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1127                  */
1128                 spin_lock(&clp->cl_lock);
1129                 BUG_ON(!list_empty(&lo->plh_layouts));
1130                 list_add_tail(&lo->plh_layouts, &server->layouts);
1131                 spin_unlock(&clp->cl_lock);
1132         }
1133
1134         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1135         if (pg_offset) {
1136                 arg.offset -= pg_offset;
1137                 arg.length += pg_offset;
1138         }
1139         if (arg.length != NFS4_MAX_UINT64)
1140                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1141
1142         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1143         atomic_dec(&lo->plh_outstanding);
1144 out_put_layout_hdr:
1145         pnfs_put_layout_hdr(lo);
1146 out:
1147         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1148                         "(%s, offset: %llu, length: %llu)\n",
1149                         __func__, ino->i_sb->s_id,
1150                         (unsigned long long)NFS_FILEID(ino),
1151                         lseg == NULL ? "not found" : "found",
1152                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1153                         (unsigned long long)pos,
1154                         (unsigned long long)count);
1155         return lseg;
1156 out_unlock:
1157         spin_unlock(&ino->i_lock);
1158         goto out_put_layout_hdr;
1159 }
1160 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1161
1162 struct pnfs_layout_segment *
1163 pnfs_layout_process(struct nfs4_layoutget *lgp)
1164 {
1165         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1166         struct nfs4_layoutget_res *res = &lgp->res;
1167         struct pnfs_layout_segment *lseg;
1168         struct inode *ino = lo->plh_inode;
1169         int status = 0;
1170
1171         /* Inject layout blob into I/O device driver */
1172         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1173         if (!lseg || IS_ERR(lseg)) {
1174                 if (!lseg)
1175                         status = -ENOMEM;
1176                 else
1177                         status = PTR_ERR(lseg);
1178                 dprintk("%s: Could not allocate layout: error %d\n",
1179                        __func__, status);
1180                 goto out;
1181         }
1182
1183         spin_lock(&ino->i_lock);
1184         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1185                 dprintk("%s forget reply due to recall\n", __func__);
1186                 goto out_forget_reply;
1187         }
1188
1189         if (pnfs_layoutgets_blocked(lo, 1) ||
1190             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1191                 dprintk("%s forget reply due to state\n", __func__);
1192                 goto out_forget_reply;
1193         }
1194
1195         /* Done processing layoutget. Set the layout stateid */
1196         pnfs_set_layout_stateid(lo, &res->stateid, false);
1197
1198         init_lseg(lo, lseg);
1199         lseg->pls_range = res->range;
1200         pnfs_get_lseg(lseg);
1201         pnfs_layout_insert_lseg(lo, lseg);
1202
1203         if (res->return_on_close) {
1204                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1205                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1206         }
1207
1208         spin_unlock(&ino->i_lock);
1209         return lseg;
1210 out:
1211         return ERR_PTR(status);
1212
1213 out_forget_reply:
1214         spin_unlock(&ino->i_lock);
1215         lseg->pls_layout = lo;
1216         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1217         goto out;
1218 }
1219
1220 void
1221 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1222 {
1223         BUG_ON(pgio->pg_lseg != NULL);
1224
1225         if (req->wb_offset != req->wb_pgbase) {
1226                 nfs_pageio_reset_read_mds(pgio);
1227                 return;
1228         }
1229         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1230                                            req->wb_context,
1231                                            req_offset(req),
1232                                            req->wb_bytes,
1233                                            IOMODE_READ,
1234                                            GFP_KERNEL);
1235         /* If no lseg, fall back to read through mds */
1236         if (pgio->pg_lseg == NULL)
1237                 nfs_pageio_reset_read_mds(pgio);
1238
1239 }
1240 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1241
1242 void
1243 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1244 {
1245         BUG_ON(pgio->pg_lseg != NULL);
1246
1247         if (req->wb_offset != req->wb_pgbase) {
1248                 nfs_pageio_reset_write_mds(pgio);
1249                 return;
1250         }
1251         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1252                                            req->wb_context,
1253                                            req_offset(req),
1254                                            req->wb_bytes,
1255                                            IOMODE_RW,
1256                                            GFP_NOFS);
1257         /* If no lseg, fall back to write through mds */
1258         if (pgio->pg_lseg == NULL)
1259                 nfs_pageio_reset_write_mds(pgio);
1260 }
1261 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1262
1263 void
1264 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1265                       const struct nfs_pgio_completion_ops *compl_ops)
1266 {
1267         struct nfs_server *server = NFS_SERVER(inode);
1268         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1269
1270         if (ld == NULL)
1271                 nfs_pageio_init_read(pgio, inode, compl_ops);
1272         else
1273                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1274 }
1275
1276 void
1277 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1278                        int ioflags,
1279                        const struct nfs_pgio_completion_ops *compl_ops)
1280 {
1281         struct nfs_server *server = NFS_SERVER(inode);
1282         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1283
1284         if (ld == NULL)
1285                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1286         else
1287                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1288 }
1289
1290 bool
1291 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1292                      struct nfs_page *req)
1293 {
1294         if (pgio->pg_lseg == NULL)
1295                 return nfs_generic_pg_test(pgio, prev, req);
1296
1297         /*
1298          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1299          * Note that this test makes several assumptions:
1300          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1301          *   is known to lie within the range.
1302          *   - that the nfs_page being tested is known to be contiguous with the
1303          *   previous nfs_page.
1304          *   - Layout ranges are page aligned, so we only have to test the
1305          *   start offset of the request.
1306          *
1307          * Please also note that 'end_offset' is actually the offset of the
1308          * first byte that lies outside the pnfs_layout_range. FIXME?
1309          *
1310          */
1311         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1312                                          pgio->pg_lseg->pls_range.length);
1313 }
1314 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1315
1316 int pnfs_write_done_resend_to_mds(struct inode *inode,
1317                                 struct list_head *head,
1318                                 const struct nfs_pgio_completion_ops *compl_ops)
1319 {
1320         struct nfs_pageio_descriptor pgio;
1321         LIST_HEAD(failed);
1322
1323         /* Resend all requests through the MDS */
1324         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1325         while (!list_empty(head)) {
1326                 struct nfs_page *req = nfs_list_entry(head->next);
1327
1328                 nfs_list_remove_request(req);
1329                 if (!nfs_pageio_add_request(&pgio, req))
1330                         nfs_list_add_request(req, &failed);
1331         }
1332         nfs_pageio_complete(&pgio);
1333
1334         if (!list_empty(&failed)) {
1335                 /* For some reason our attempt to resend pages. Mark the
1336                  * overall send request as having failed, and let
1337                  * nfs_writeback_release_full deal with the error.
1338                  */
1339                 list_move(&failed, head);
1340                 return -EIO;
1341         }
1342         return 0;
1343 }
1344 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1345
1346 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1347 {
1348         struct nfs_pgio_header *hdr = data->header;
1349
1350         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1351         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1352             PNFS_LAYOUTRET_ON_ERROR) {
1353                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1354                 pnfs_return_layout(hdr->inode);
1355         }
1356         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1357                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1358                                                         &hdr->pages,
1359                                                         hdr->completion_ops);
1360 }
1361
1362 /*
1363  * Called by non rpc-based layout drivers
1364  */
1365 void pnfs_ld_write_done(struct nfs_write_data *data)
1366 {
1367         struct nfs_pgio_header *hdr = data->header;
1368
1369         if (!hdr->pnfs_error) {
1370                 pnfs_set_layoutcommit(data);
1371                 hdr->mds_ops->rpc_call_done(&data->task, data);
1372         } else
1373                 pnfs_ld_handle_write_error(data);
1374         hdr->mds_ops->rpc_release(data);
1375 }
1376 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1377
1378 static void
1379 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1380                 struct nfs_write_data *data)
1381 {
1382         struct nfs_pgio_header *hdr = data->header;
1383
1384         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1385                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1386                 nfs_pageio_reset_write_mds(desc);
1387                 desc->pg_recoalesce = 1;
1388         }
1389         nfs_writedata_release(data);
1390 }
1391
1392 static enum pnfs_try_status
1393 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1394                         const struct rpc_call_ops *call_ops,
1395                         struct pnfs_layout_segment *lseg,
1396                         int how)
1397 {
1398         struct nfs_pgio_header *hdr = wdata->header;
1399         struct inode *inode = hdr->inode;
1400         enum pnfs_try_status trypnfs;
1401         struct nfs_server *nfss = NFS_SERVER(inode);
1402
1403         hdr->mds_ops = call_ops;
1404
1405         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1406                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1407         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1408         if (trypnfs != PNFS_NOT_ATTEMPTED)
1409                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1410         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1411         return trypnfs;
1412 }
1413
1414 static void
1415 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1416 {
1417         struct nfs_write_data *data;
1418         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1419         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1420
1421         desc->pg_lseg = NULL;
1422         while (!list_empty(head)) {
1423                 enum pnfs_try_status trypnfs;
1424
1425                 data = list_first_entry(head, struct nfs_write_data, list);
1426                 list_del_init(&data->list);
1427
1428                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1429                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1430                         pnfs_write_through_mds(desc, data);
1431         }
1432         pnfs_put_lseg(lseg);
1433 }
1434
1435 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1436 {
1437         pnfs_put_lseg(hdr->lseg);
1438         nfs_writehdr_free(hdr);
1439 }
1440 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1441
1442 int
1443 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1444 {
1445         struct nfs_write_header *whdr;
1446         struct nfs_pgio_header *hdr;
1447         int ret;
1448
1449         whdr = nfs_writehdr_alloc();
1450         if (!whdr) {
1451                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1452                 pnfs_put_lseg(desc->pg_lseg);
1453                 desc->pg_lseg = NULL;
1454                 return -ENOMEM;
1455         }
1456         hdr = &whdr->header;
1457         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1458         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1459         atomic_inc(&hdr->refcnt);
1460         ret = nfs_generic_flush(desc, hdr);
1461         if (ret != 0) {
1462                 pnfs_put_lseg(desc->pg_lseg);
1463                 desc->pg_lseg = NULL;
1464         } else
1465                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1466         if (atomic_dec_and_test(&hdr->refcnt))
1467                 hdr->completion_ops->completion(hdr);
1468         return ret;
1469 }
1470 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1471
1472 int pnfs_read_done_resend_to_mds(struct inode *inode,
1473                                 struct list_head *head,
1474                                 const struct nfs_pgio_completion_ops *compl_ops)
1475 {
1476         struct nfs_pageio_descriptor pgio;
1477         LIST_HEAD(failed);
1478
1479         /* Resend all requests through the MDS */
1480         nfs_pageio_init_read(&pgio, inode, compl_ops);
1481         while (!list_empty(head)) {
1482                 struct nfs_page *req = nfs_list_entry(head->next);
1483
1484                 nfs_list_remove_request(req);
1485                 if (!nfs_pageio_add_request(&pgio, req))
1486                         nfs_list_add_request(req, &failed);
1487         }
1488         nfs_pageio_complete(&pgio);
1489
1490         if (!list_empty(&failed)) {
1491                 list_move(&failed, head);
1492                 return -EIO;
1493         }
1494         return 0;
1495 }
1496 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1497
1498 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1499 {
1500         struct nfs_pgio_header *hdr = data->header;
1501
1502         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1503         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1504             PNFS_LAYOUTRET_ON_ERROR) {
1505                 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1506                 pnfs_return_layout(hdr->inode);
1507         }
1508         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1509                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1510                                                         &hdr->pages,
1511                                                         hdr->completion_ops);
1512 }
1513
1514 /*
1515  * Called by non rpc-based layout drivers
1516  */
1517 void pnfs_ld_read_done(struct nfs_read_data *data)
1518 {
1519         struct nfs_pgio_header *hdr = data->header;
1520
1521         if (likely(!hdr->pnfs_error)) {
1522                 __nfs4_read_done_cb(data);
1523                 hdr->mds_ops->rpc_call_done(&data->task, data);
1524         } else
1525                 pnfs_ld_handle_read_error(data);
1526         hdr->mds_ops->rpc_release(data);
1527 }
1528 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1529
1530 static void
1531 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1532                 struct nfs_read_data *data)
1533 {
1534         struct nfs_pgio_header *hdr = data->header;
1535
1536         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1537                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1538                 nfs_pageio_reset_read_mds(desc);
1539                 desc->pg_recoalesce = 1;
1540         }
1541         nfs_readdata_release(data);
1542 }
1543
1544 /*
1545  * Call the appropriate parallel I/O subsystem read function.
1546  */
1547 static enum pnfs_try_status
1548 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1549                        const struct rpc_call_ops *call_ops,
1550                        struct pnfs_layout_segment *lseg)
1551 {
1552         struct nfs_pgio_header *hdr = rdata->header;
1553         struct inode *inode = hdr->inode;
1554         struct nfs_server *nfss = NFS_SERVER(inode);
1555         enum pnfs_try_status trypnfs;
1556
1557         hdr->mds_ops = call_ops;
1558
1559         dprintk("%s: Reading ino:%lu %u@%llu\n",
1560                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1561
1562         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1563         if (trypnfs != PNFS_NOT_ATTEMPTED)
1564                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1565         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1566         return trypnfs;
1567 }
1568
1569 static void
1570 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1571 {
1572         struct nfs_read_data *data;
1573         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1574         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1575
1576         desc->pg_lseg = NULL;
1577         while (!list_empty(head)) {
1578                 enum pnfs_try_status trypnfs;
1579
1580                 data = list_first_entry(head, struct nfs_read_data, list);
1581                 list_del_init(&data->list);
1582
1583                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1584                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1585                         pnfs_read_through_mds(desc, data);
1586         }
1587         pnfs_put_lseg(lseg);
1588 }
1589
1590 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1591 {
1592         pnfs_put_lseg(hdr->lseg);
1593         nfs_readhdr_free(hdr);
1594 }
1595 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1596
1597 int
1598 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1599 {
1600         struct nfs_read_header *rhdr;
1601         struct nfs_pgio_header *hdr;
1602         int ret;
1603
1604         rhdr = nfs_readhdr_alloc();
1605         if (!rhdr) {
1606                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1607                 ret = -ENOMEM;
1608                 pnfs_put_lseg(desc->pg_lseg);
1609                 desc->pg_lseg = NULL;
1610                 return ret;
1611         }
1612         hdr = &rhdr->header;
1613         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1614         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1615         atomic_inc(&hdr->refcnt);
1616         ret = nfs_generic_pagein(desc, hdr);
1617         if (ret != 0) {
1618                 pnfs_put_lseg(desc->pg_lseg);
1619                 desc->pg_lseg = NULL;
1620         } else
1621                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1622         if (atomic_dec_and_test(&hdr->refcnt))
1623                 hdr->completion_ops->completion(hdr);
1624         return ret;
1625 }
1626 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1627
1628 /*
1629  * There can be multiple RW segments.
1630  */
1631 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1632 {
1633         struct pnfs_layout_segment *lseg;
1634
1635         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1636                 if (lseg->pls_range.iomode == IOMODE_RW &&
1637                     test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1638                         list_add(&lseg->pls_lc_list, listp);
1639         }
1640 }
1641
1642 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1643 {
1644         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1645 }
1646 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1647
1648 void
1649 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1650 {
1651         struct nfs_pgio_header *hdr = wdata->header;
1652         struct inode *inode = hdr->inode;
1653         struct nfs_inode *nfsi = NFS_I(inode);
1654         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1655         bool mark_as_dirty = false;
1656
1657         spin_lock(&inode->i_lock);
1658         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1659                 mark_as_dirty = true;
1660                 dprintk("%s: Set layoutcommit for inode %lu ",
1661                         __func__, inode->i_ino);
1662         }
1663         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1664                 /* references matched in nfs4_layoutcommit_release */
1665                 pnfs_get_lseg(hdr->lseg);
1666         }
1667         if (end_pos > nfsi->layout->plh_lwb)
1668                 nfsi->layout->plh_lwb = end_pos;
1669         spin_unlock(&inode->i_lock);
1670         dprintk("%s: lseg %p end_pos %llu\n",
1671                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1672
1673         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1674          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1675         if (mark_as_dirty)
1676                 mark_inode_dirty_sync(inode);
1677 }
1678 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1679
1680 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1681 {
1682         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1683
1684         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1685                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1686 }
1687
1688 /*
1689  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1690  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1691  * data to disk to allow the server to recover the data if it crashes.
1692  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1693  * is off, and a COMMIT is sent to a data server, or
1694  * if WRITEs to a data server return NFS_DATA_SYNC.
1695  */
1696 int
1697 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1698 {
1699         struct nfs4_layoutcommit_data *data;
1700         struct nfs_inode *nfsi = NFS_I(inode);
1701         loff_t end_pos;
1702         int status = 0;
1703
1704         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1705
1706         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1707                 return 0;
1708
1709         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1710         data = kzalloc(sizeof(*data), GFP_NOFS);
1711         if (!data) {
1712                 status = -ENOMEM;
1713                 goto out;
1714         }
1715
1716         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1717                 goto out_free;
1718
1719         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1720                 if (!sync) {
1721                         status = -EAGAIN;
1722                         goto out_free;
1723                 }
1724                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1725                                         nfs_wait_bit_killable, TASK_KILLABLE);
1726                 if (status)
1727                         goto out_free;
1728         }
1729
1730         INIT_LIST_HEAD(&data->lseg_list);
1731         spin_lock(&inode->i_lock);
1732         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1733                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1734                 spin_unlock(&inode->i_lock);
1735                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1736                 goto out_free;
1737         }
1738
1739         pnfs_list_write_lseg(inode, &data->lseg_list);
1740
1741         end_pos = nfsi->layout->plh_lwb;
1742         nfsi->layout->plh_lwb = 0;
1743
1744         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1745         spin_unlock(&inode->i_lock);
1746
1747         data->args.inode = inode;
1748         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1749         nfs_fattr_init(&data->fattr);
1750         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1751         data->res.fattr = &data->fattr;
1752         data->args.lastbytewritten = end_pos - 1;
1753         data->res.server = NFS_SERVER(inode);
1754
1755         status = nfs4_proc_layoutcommit(data, sync);
1756 out:
1757         if (status)
1758                 mark_inode_dirty_sync(inode);
1759         dprintk("<-- %s status %d\n", __func__, status);
1760         return status;
1761 out_free:
1762         kfree(data);
1763         goto out;
1764 }
1765
1766 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1767 {
1768         struct nfs4_threshold *thp;
1769
1770         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1771         if (!thp) {
1772                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1773                 return NULL;
1774         }
1775         return thp;
1776 }