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 /*
 373  * is l2 fully contained in l1?
 374  *   start1                             end1
 375  *   [----------------------------------)
 376  *           start2           end2
 377  *           [----------------)
 378  */
 379 static inline int
 380 lo_seg_contained(struct pnfs_layout_range *l1,
 381                  struct pnfs_layout_range *l2)
 382 {
 383         u64 start1 = l1->offset;
 384         u64 end1 = end_offset(start1, l1->length);
 385         u64 start2 = l2->offset;
 386         u64 end2 = end_offset(start2, l2->length);
 387
 388         return (start1 <= start2) && (end1 >= end2);
 389 }
 390
 391 /*
 392  * is l1 and l2 intersecting?
 393  *   start1                             end1
 394  *   [----------------------------------)
 395  *                              start2           end2
 396  *                              [----------------)
 397  */
 398 static inline int
 399 lo_seg_intersecting(struct pnfs_layout_range *l1,
 400                     struct pnfs_layout_range *l2)
 401 {
 402         u64 start1 = l1->offset;
 403         u64 end1 = end_offset(start1, l1->length);
 404         u64 start2 = l2->offset;
 405         u64 end2 = end_offset(start2, l2->length);
 406
 407         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 408                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 409 }
 410
 411 static bool
 412 should_free_lseg(struct pnfs_layout_range *lseg_range,
 413                  struct pnfs_layout_range *recall_range)
 414 {
 415         return (recall_range->iomode == IOMODE_ANY ||
 416                 lseg_range->iomode == recall_range->iomode) &&
 417                lo_seg_intersecting(lseg_range, recall_range);
 418 }
 419
 420 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 421                 struct list_head *tmp_list)
 422 {
 423         if (!atomic_dec_and_test(&lseg->pls_refcount))
 424                 return false;
 425         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 426         list_add(&lseg->pls_list, tmp_list);
 427         return true;
 428 }
 429
 430 /* Returns 1 if lseg is removed from list, 0 otherwise */
 431 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 432                              struct list_head *tmp_list)
 433 {
 434         int rv = 0;
 435
 436         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 437                 /* Remove the reference keeping the lseg in the
 438                  * list.  It will now be removed when all
 439                  * outstanding io is finished.
 440                  */
 441                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 442                         atomic_read(&lseg->pls_refcount));
 443                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 444                         rv = 1;
 445         }
 446         return rv;
 447 }
 448
 449 /* Returns count of number of matching invalid lsegs remaining in list
 450  * after call.
 451  */
 452 int
 453 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 454                             struct list_head *tmp_list,
 455                             struct pnfs_layout_range *recall_range)
 456 {
 457         struct pnfs_layout_segment *lseg, *next;
 458         int invalid = 0, removed = 0;
 459
 460         dprintk("%s:Begin lo %p\n", __func__, lo);
 461
 462         if (list_empty(&lo->plh_segs))
 463                 return 0;
 464         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 465                 if (!recall_range ||
 466                     should_free_lseg(&lseg->pls_range, recall_range)) {
 467                         dprintk("%s: freeing lseg %p iomode %d "
 468                                 "offset %llu length %llu\n", __func__,
 469                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 470                                 lseg->pls_range.length);
 471                         invalid++;
 472                         removed += mark_lseg_invalid(lseg, tmp_list);
 473                 }
 474         dprintk("%s:Return %i\n", __func__, invalid - removed);
 475         return invalid - removed;
 476 }
 477
 478 /* note free_me must contain lsegs from a single layout_hdr */
 479 void
 480 pnfs_free_lseg_list(struct list_head *free_me)
 481 {
 482         struct pnfs_layout_segment *lseg, *tmp;
 483
 484         if (list_empty(free_me))
 485                 return;
 486
 487         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 488                 list_del(&lseg->pls_list);
 489                 pnfs_free_lseg(lseg);
 490         }
 491 }
 492
 493 void
 494 pnfs_destroy_layout(struct nfs_inode *nfsi)
 495 {
 496         struct pnfs_layout_hdr *lo;
 497         LIST_HEAD(tmp_list);
 498
 499         spin_lock(&nfsi->vfs_inode.i_lock);
 500         lo = nfsi->layout;
 501         if (lo) {
 502                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 503                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 504                 pnfs_get_layout_hdr(lo);
 505                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 506                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 507                 spin_unlock(&nfsi->vfs_inode.i_lock);
 508                 pnfs_free_lseg_list(&tmp_list);
 509                 pnfs_put_layout_hdr(lo);
 510         } else
 511                 spin_unlock(&nfsi->vfs_inode.i_lock);
 512 }
 513 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 514
 515 static bool
 516 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 517                 struct list_head *layout_list)
 518 {
 519         struct pnfs_layout_hdr *lo;
 520         bool ret = false;
 521
 522         spin_lock(&inode->i_lock);
 523         lo = NFS_I(inode)->layout;
 524         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 525                 pnfs_get_layout_hdr(lo);
 526                 list_add(&lo->plh_bulk_destroy, layout_list);
 527                 ret = true;
 528         }
 529         spin_unlock(&inode->i_lock);
 530         return ret;
 531 }
 532
 533 /* Caller must hold rcu_read_lock and clp->cl_lock */
 534 static int
 535 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 536                 struct nfs_server *server,
 537                 struct list_head *layout_list)
 538 {
 539         struct pnfs_layout_hdr *lo, *next;
 540         struct inode *inode;
 541
 542         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 543                 inode = igrab(lo->plh_inode);
 544                 if (inode == NULL)
 545                         continue;
 546                 list_del_init(&lo->plh_layouts);
 547                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 548                         continue;
 549                 rcu_read_unlock();
 550                 spin_unlock(&clp->cl_lock);
 551                 iput(inode);
 552                 spin_lock(&clp->cl_lock);
 553                 rcu_read_lock();
 554                 return -EAGAIN;
 555         }
 556         return 0;
 557 }
 558
 559 static int
 560 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 561                 bool is_bulk_recall)
 562 {
 563         struct pnfs_layout_hdr *lo;
 564         struct inode *inode;
 565         struct pnfs_layout_range range = {
 566                 .iomode = IOMODE_ANY,
 567                 .offset = 0,
 568                 .length = NFS4_MAX_UINT64,
 569         };
 570         LIST_HEAD(lseg_list);
 571         int ret = 0;
 572
 573         while (!list_empty(layout_list)) {
 574                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 575                                 plh_bulk_destroy);
 576                 dprintk("%s freeing layout for inode %lu\n", __func__,
 577                         lo->plh_inode->i_ino);
 578                 inode = lo->plh_inode;
 579                 spin_lock(&inode->i_lock);
 580                 list_del_init(&lo->plh_bulk_destroy);
 581                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 582                 if (is_bulk_recall)
 583                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 584                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 585                         ret = -EAGAIN;
 586                 spin_unlock(&inode->i_lock);
 587                 pnfs_free_lseg_list(&lseg_list);
 588                 pnfs_put_layout_hdr(lo);
 589                 iput(inode);
 590         }
 591         return ret;
 592 }
 593
 594 int
 595 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 596                 struct nfs_fsid *fsid,
 597                 bool is_recall)
 598 {
 599         struct nfs_server *server;
 600         LIST_HEAD(layout_list);
 601
 602         spin_lock(&clp->cl_lock);
 603         rcu_read_lock();
 604 restart:
 605         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 606                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 607                         continue;
 608                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 609                                 server,
 610                                 &layout_list) != 0)
 611                         goto restart;
 612         }
 613         rcu_read_unlock();
 614         spin_unlock(&clp->cl_lock);
 615
 616         if (list_empty(&layout_list))
 617                 return 0;
 618         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 619 }
 620
 621 int
 622 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 623                 bool is_recall)
 624 {
 625         struct nfs_server *server;
 626         LIST_HEAD(layout_list);
 627
 628         spin_lock(&clp->cl_lock);
 629         rcu_read_lock();
 630 restart:
 631         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 632                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 633                                         server,
 634                                         &layout_list) != 0)
 635                         goto restart;
 636         }
 637         rcu_read_unlock();
 638         spin_unlock(&clp->cl_lock);
 639
 640         if (list_empty(&layout_list))
 641                 return 0;
 642         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 643 }
 644
 645 /*
 646  * Called by the state manger to remove all layouts established under an
 647  * expired lease.
 648  */
 649 void
 650 pnfs_destroy_all_layouts(struct nfs_client *clp)
 651 {
 652         nfs4_deviceid_mark_client_invalid(clp);
 653         nfs4_deviceid_purge_client(clp);
 654
 655         pnfs_destroy_layouts_byclid(clp, false);
 656 }
 657
 658 /*
 659  * Compare 2 layout stateid sequence ids, to see which is newer,
 660  * taking into account wraparound issues.
 661  */
 662 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 663 {
 664         return (s32)s1 - (s32)s2 > 0;
 665 }
 666
 667 /* update lo->plh_stateid with new if is more recent */
 668 void
 669 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 670                         bool update_barrier)
 671 {
 672         u32 oldseq, newseq, new_barrier;
 673         int empty = list_empty(&lo->plh_segs);
 674
 675         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 676         newseq = be32_to_cpu(new->seqid);
 677         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 678                 nfs4_stateid_copy(&lo->plh_stateid, new);
 679                 if (update_barrier) {
 680                         new_barrier = be32_to_cpu(new->seqid);
 681                 } else {
 682                         /* Because of wraparound, we want to keep the barrier
 683                          * "close" to the current seqids.
 684                          */
 685                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 686                 }
 687                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 688                         lo->plh_barrier = new_barrier;
 689         }
 690 }
 691
 692 static bool
 693 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 694                 const nfs4_stateid *stateid)
 695 {
 696         u32 seqid = be32_to_cpu(stateid->seqid);
 697
 698         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 699 }
 700
 701 /* lget is set to 1 if called from inside send_layoutget call chain */
 702 static bool
 703 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 704 {
 705         return lo->plh_block_lgets ||
 706                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 707                 (list_empty(&lo->plh_segs) &&
 708                  (atomic_read(&lo->plh_outstanding) > lget));
 709 }
 710
 711 int
 712 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 713                               struct nfs4_state *open_state)
 714 {
 715         int status = 0;
 716
 717         dprintk("--> %s\n", __func__);
 718         spin_lock(&lo->plh_inode->i_lock);
 719         if (pnfs_layoutgets_blocked(lo, 1)) {
 720                 status = -EAGAIN;
 721         } else if (list_empty(&lo->plh_segs)) {
 722                 int seq;
 723
 724                 do {
 725                         seq = read_seqbegin(&open_state->seqlock);
 726                         nfs4_stateid_copy(dst, &open_state->stateid);
 727                 } while (read_seqretry(&open_state->seqlock, seq));
 728         } else
 729                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 730         spin_unlock(&lo->plh_inode->i_lock);
 731         dprintk("<-- %s\n", __func__);
 732         return status;
 733 }
 734
 735 /*
 736 * Get layout from server.
 737 *    for now, assume that whole file layouts are requested.
 738 *    arg->offset: 0
 739 *    arg->length: all ones
 740 */
 741 static struct pnfs_layout_segment *
 742 send_layoutget(struct pnfs_layout_hdr *lo,
 743            struct nfs_open_context *ctx,
 744            struct pnfs_layout_range *range,
 745            gfp_t gfp_flags)
 746 {
 747         struct inode *ino = lo->plh_inode;
 748         struct nfs_server *server = NFS_SERVER(ino);
 749         struct nfs4_layoutget *lgp;
 750         struct pnfs_layout_segment *lseg;
 751
 752         dprintk("--> %s\n", __func__);
 753
 754         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 755         if (lgp == NULL)
 756                 return NULL;
 757
 758         lgp->args.minlength = PAGE_CACHE_SIZE;
 759         if (lgp->args.minlength > range->length)
 760                 lgp->args.minlength = range->length;
 761         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 762         lgp->args.range = *range;
 763         lgp->args.type = server->pnfs_curr_ld->id;
 764         lgp->args.inode = ino;
 765         lgp->args.ctx = get_nfs_open_context(ctx);
 766         lgp->gfp_flags = gfp_flags;
 767
 768         /* Synchronously retrieve layout information from server and
 769          * store in lseg.
 770          */
 771         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 772         if (IS_ERR(lseg)) {
 773                 switch (PTR_ERR(lseg)) {
 774                 case -ENOMEM:
 775                 case -ERESTARTSYS:
 776                         break;
 777                 default:
 778                         /* remember that LAYOUTGET failed and suspend trying */
 779                         pnfs_layout_io_set_failed(lo, range->iomode);
 780                 }
 781                 return NULL;
 782         }
 783
 784         return lseg;
 785 }
 786
 787 static void pnfs_clear_layoutcommit(struct inode *inode,
 788                 struct list_head *head)
 789 {
 790         struct nfs_inode *nfsi = NFS_I(inode);
 791         struct pnfs_layout_segment *lseg, *tmp;
 792
 793         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 794                 return;
 795         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 796                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 797                         continue;
 798                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 799         }
 800 }
 801
 802 /*
 803  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 804  * when the layout segment list is empty.
 805  *
 806  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 807  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 808  * deviceid is marked invalid.
 809  */
 810 int
 811 _pnfs_return_layout(struct inode *ino)
 812 {
 813         struct pnfs_layout_hdr *lo = NULL;
 814         struct nfs_inode *nfsi = NFS_I(ino);
 815         LIST_HEAD(tmp_list);
 816         struct nfs4_layoutreturn *lrp;
 817         nfs4_stateid stateid;
 818         int status = 0, empty;
 819
 820         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 821
 822         spin_lock(&ino->i_lock);
 823         lo = nfsi->layout;
 824         if (!lo) {
 825                 spin_unlock(&ino->i_lock);
 826                 dprintk("NFS: %s no layout to return\n", __func__);
 827                 goto out;
 828         }
 829         stateid = nfsi->layout->plh_stateid;
 830         /* Reference matched in nfs4_layoutreturn_release */
 831         pnfs_get_layout_hdr(lo);
 832         empty = list_empty(&lo->plh_segs);
 833         pnfs_clear_layoutcommit(ino, &tmp_list);
 834         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 835         /* Don't send a LAYOUTRETURN if list was initially empty */
 836         if (empty) {
 837                 spin_unlock(&ino->i_lock);
 838                 pnfs_put_layout_hdr(lo);
 839                 dprintk("NFS: %s no layout segments to return\n", __func__);
 840                 goto out;
 841         }
 842         lo->plh_block_lgets++;
 843         spin_unlock(&ino->i_lock);
 844         pnfs_free_lseg_list(&tmp_list);
 845
 846         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 847         if (unlikely(lrp == NULL)) {
 848                 status = -ENOMEM;
 849                 spin_lock(&ino->i_lock);
 850                 lo->plh_block_lgets--;
 851                 spin_unlock(&ino->i_lock);
 852                 pnfs_put_layout_hdr(lo);
 853                 goto out;
 854         }
 855
 856         lrp->args.stateid = stateid;
 857         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 858         lrp->args.inode = ino;
 859         lrp->args.layout = lo;
 860         lrp->clp = NFS_SERVER(ino)->nfs_client;
 861
 862         status = nfs4_proc_layoutreturn(lrp);
 863 out:
 864         dprintk("<-- %s status: %d\n", __func__, status);
 865         return status;
 866 }
 867 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 868
 869 int
 870 pnfs_commit_and_return_layout(struct inode *inode)
 871 {
 872         struct pnfs_layout_hdr *lo;
 873         int ret;
 874
 875         spin_lock(&inode->i_lock);
 876         lo = NFS_I(inode)->layout;
 877         if (lo == NULL) {
 878                 spin_unlock(&inode->i_lock);
 879                 return 0;
 880         }
 881         pnfs_get_layout_hdr(lo);
 882         /* Block new layoutgets and read/write to ds */
 883         lo->plh_block_lgets++;
 884         spin_unlock(&inode->i_lock);
 885         filemap_fdatawait(inode->i_mapping);
 886         ret = pnfs_layoutcommit_inode(inode, true);
 887         if (ret == 0)
 888                 ret = _pnfs_return_layout(inode);
 889         spin_lock(&inode->i_lock);
 890         lo->plh_block_lgets--;
 891         spin_unlock(&inode->i_lock);
 892         pnfs_put_layout_hdr(lo);
 893         return ret;
 894 }
 895
 896 bool pnfs_roc(struct inode *ino)
 897 {
 898         struct pnfs_layout_hdr *lo;
 899         struct pnfs_layout_segment *lseg, *tmp;
 900         LIST_HEAD(tmp_list);
 901         bool found = false;
 902
 903         spin_lock(&ino->i_lock);
 904         lo = NFS_I(ino)->layout;
 905         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 906             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 907                 goto out_nolayout;
 908         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 909                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 910                         mark_lseg_invalid(lseg, &tmp_list);
 911                         found = true;
 912                 }
 913         if (!found)
 914                 goto out_nolayout;
 915         lo->plh_block_lgets++;
 916         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 917         spin_unlock(&ino->i_lock);
 918         pnfs_free_lseg_list(&tmp_list);
 919         return true;
 920
 921 out_nolayout:
 922         spin_unlock(&ino->i_lock);
 923         return false;
 924 }
 925
 926 void pnfs_roc_release(struct inode *ino)
 927 {
 928         struct pnfs_layout_hdr *lo;
 929
 930         spin_lock(&ino->i_lock);
 931         lo = NFS_I(ino)->layout;
 932         lo->plh_block_lgets--;
 933         if (atomic_dec_and_test(&lo->plh_refcount)) {
 934                 pnfs_detach_layout_hdr(lo);
 935                 spin_unlock(&ino->i_lock);
 936                 pnfs_free_layout_hdr(lo);
 937         } else
 938                 spin_unlock(&ino->i_lock);
 939 }
 940
 941 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 942 {
 943         struct pnfs_layout_hdr *lo;
 944
 945         spin_lock(&ino->i_lock);
 946         lo = NFS_I(ino)->layout;
 947         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 948                 lo->plh_barrier = barrier;
 949         spin_unlock(&ino->i_lock);
 950 }
 951
 952 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 953 {
 954         struct nfs_inode *nfsi = NFS_I(ino);
 955         struct pnfs_layout_hdr *lo;
 956         struct pnfs_layout_segment *lseg;
 957         u32 current_seqid;
 958         bool found = false;
 959
 960         spin_lock(&ino->i_lock);
 961         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 962                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 963                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 964                         found = true;
 965                         goto out;
 966                 }
 967         lo = nfsi->layout;
 968         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 969
 970         /* Since close does not return a layout stateid for use as
 971          * a barrier, we choose the worst-case barrier.
 972          */
 973         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 974 out:
 975         spin_unlock(&ino->i_lock);
 976         return found;
 977 }
 978
 979 /*
 980  * Compare two layout segments for sorting into layout cache.
 981  * We want to preferentially return RW over RO layouts, so ensure those
 982  * are seen first.
 983  */
 984 static s64
 985 cmp_layout(struct pnfs_layout_range *l1,
 986            struct pnfs_layout_range *l2)
 987 {
 988         s64 d;
 989
 990         /* high offset > low offset */
 991         d = l1->offset - l2->offset;
 992         if (d)
 993                 return d;
 994
 995         /* short length > long length */
 996         d = l2->length - l1->length;
 997         if (d)
 998                 return d;
 999
1000         /* read > read/write */
1001         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1002 }
1003
1004 static void
1005 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1006                    struct pnfs_layout_segment *lseg)
1007 {
1008         struct pnfs_layout_segment *lp;
1009
1010         dprintk("%s:Begin\n", __func__);
1011
1012         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1013                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
1014                         continue;
1015                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1016                 dprintk("%s: inserted lseg %p "
1017                         "iomode %d offset %llu length %llu before "
1018                         "lp %p iomode %d offset %llu length %llu\n",
1019                         __func__, lseg, lseg->pls_range.iomode,
1020                         lseg->pls_range.offset, lseg->pls_range.length,
1021                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1022                         lp->pls_range.length);
1023                 goto out;
1024         }
1025         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1026         dprintk("%s: inserted lseg %p "
1027                 "iomode %d offset %llu length %llu at tail\n",
1028                 __func__, lseg, lseg->pls_range.iomode,
1029                 lseg->pls_range.offset, lseg->pls_range.length);
1030 out:
1031         pnfs_get_layout_hdr(lo);
1032
1033         dprintk("%s:Return\n", __func__);
1034 }
1035
1036 static struct pnfs_layout_hdr *
1037 alloc_init_layout_hdr(struct inode *ino,
1038                       struct nfs_open_context *ctx,
1039                       gfp_t gfp_flags)
1040 {
1041         struct pnfs_layout_hdr *lo;
1042
1043         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1044         if (!lo)
1045                 return NULL;
1046         atomic_set(&lo->plh_refcount, 1);
1047         INIT_LIST_HEAD(&lo->plh_layouts);
1048         INIT_LIST_HEAD(&lo->plh_segs);
1049         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1050         lo->plh_inode = ino;
1051         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
1052         return lo;
1053 }
1054
1055 static struct pnfs_layout_hdr *
1056 pnfs_find_alloc_layout(struct inode *ino,
1057                        struct nfs_open_context *ctx,
1058                        gfp_t gfp_flags)
1059 {
1060         struct nfs_inode *nfsi = NFS_I(ino);
1061         struct pnfs_layout_hdr *new = NULL;
1062
1063         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1064
1065         if (nfsi->layout != NULL)
1066                 goto out_existing;
1067         spin_unlock(&ino->i_lock);
1068         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1069         spin_lock(&ino->i_lock);
1070
1071         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1072                 nfsi->layout = new;
1073                 return new;
1074         } else if (new != NULL)
1075                 pnfs_free_layout_hdr(new);
1076 out_existing:
1077         pnfs_get_layout_hdr(nfsi->layout);
1078         return nfsi->layout;
1079 }
1080
1081 /*
1082  * iomode matching rules:
1083  * iomode       lseg    match
1084  * -----        -----   -----
1085  * ANY          READ    true
1086  * ANY          RW      true
1087  * RW           READ    false
1088  * RW           RW      true
1089  * READ         READ    true
1090  * READ         RW      true
1091  */
1092 static int
1093 is_matching_lseg(struct pnfs_layout_range *ls_range,
1094                  struct pnfs_layout_range *range)
1095 {
1096         struct pnfs_layout_range range1;
1097
1098         if ((range->iomode == IOMODE_RW &&
1099              ls_range->iomode != IOMODE_RW) ||
1100             !lo_seg_intersecting(ls_range, range))
1101                 return 0;
1102
1103         /* range1 covers only the first byte in the range */
1104         range1 = *range;
1105         range1.length = 1;
1106         return lo_seg_contained(ls_range, &range1);
1107 }
1108
1109 /*
1110  * lookup range in layout
1111  */
1112 static struct pnfs_layout_segment *
1113 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1114                 struct pnfs_layout_range *range)
1115 {
1116         struct pnfs_layout_segment *lseg, *ret = NULL;
1117
1118         dprintk("%s:Begin\n", __func__);
1119
1120         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1121                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1122                     is_matching_lseg(&lseg->pls_range, range)) {
1123                         ret = pnfs_get_lseg(lseg);
1124                         break;
1125                 }
1126                 if (lseg->pls_range.offset > range->offset)
1127                         break;
1128         }
1129
1130         dprintk("%s:Return lseg %p ref %d\n",
1131                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1132         return ret;
1133 }
1134
1135 /*
1136  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1137  * to the MDS or over pNFS
1138  *
1139  * The nfs_inode read_io and write_io fields are cumulative counters reset
1140  * when there are no layout segments. Note that in pnfs_update_layout iomode
1141  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1142  * WRITE request.
1143  *
1144  * A return of true means use MDS I/O.
1145  *
1146  * From rfc 5661:
1147  * If a file's size is smaller than the file size threshold, data accesses
1148  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1149  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1150  * server.  If both file size and I/O size are provided, the client SHOULD
1151  * reach or exceed  both thresholds before sending its read or write
1152  * requests to the data server.
1153  */
1154 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1155                                      struct inode *ino, int iomode)
1156 {
1157         struct nfs4_threshold *t = ctx->mdsthreshold;
1158         struct nfs_inode *nfsi = NFS_I(ino);
1159         loff_t fsize = i_size_read(ino);
1160         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1161
1162         if (t == NULL)
1163                 return ret;
1164
1165         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1166                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1167
1168         switch (iomode) {
1169         case IOMODE_READ:
1170                 if (t->bm & THRESHOLD_RD) {
1171                         dprintk("%s fsize %llu\n", __func__, fsize);
1172                         size_set = true;
1173                         if (fsize < t->rd_sz)
1174                                 size = true;
1175                 }
1176                 if (t->bm & THRESHOLD_RD_IO) {
1177                         dprintk("%s nfsi->read_io %llu\n", __func__,
1178                                 nfsi->read_io);
1179                         io_set = true;
1180                         if (nfsi->read_io < t->rd_io_sz)
1181                                 io = true;
1182                 }
1183                 break;
1184         case IOMODE_RW:
1185                 if (t->bm & THRESHOLD_WR) {
1186                         dprintk("%s fsize %llu\n", __func__, fsize);
1187                         size_set = true;
1188                         if (fsize < t->wr_sz)
1189                                 size = true;
1190                 }
1191                 if (t->bm & THRESHOLD_WR_IO) {
1192                         dprintk("%s nfsi->write_io %llu\n", __func__,
1193                                 nfsi->write_io);
1194                         io_set = true;
1195                         if (nfsi->write_io < t->wr_io_sz)
1196                                 io = true;
1197                 }
1198                 break;
1199         }
1200         if (size_set && io_set) {
1201                 if (size && io)
1202                         ret = true;
1203         } else if (size || io)
1204                 ret = true;
1205
1206         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1207         return ret;
1208 }
1209
1210 /*
1211  * Layout segment is retreived from the server if not cached.
1212  * The appropriate layout segment is referenced and returned to the caller.
1213  */
1214 struct pnfs_layout_segment *
1215 pnfs_update_layout(struct inode *ino,
1216                    struct nfs_open_context *ctx,
1217                    loff_t pos,
1218                    u64 count,
1219                    enum pnfs_iomode iomode,
1220                    gfp_t gfp_flags)
1221 {
1222         struct pnfs_layout_range arg = {
1223                 .iomode = iomode,
1224                 .offset = pos,
1225                 .length = count,
1226         };
1227         unsigned pg_offset;
1228         struct nfs_server *server = NFS_SERVER(ino);
1229         struct nfs_client *clp = server->nfs_client;
1230         struct pnfs_layout_hdr *lo;
1231         struct pnfs_layout_segment *lseg = NULL;
1232         bool first;
1233
1234         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1235                 goto out;
1236
1237         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1238                 goto out;
1239
1240         spin_lock(&ino->i_lock);
1241         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1242         if (lo == NULL) {
1243                 spin_unlock(&ino->i_lock);
1244                 goto out;
1245         }
1246
1247         /* Do we even need to bother with this? */
1248         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1249                 dprintk("%s matches recall, use MDS\n", __func__);
1250                 goto out_unlock;
1251         }
1252
1253         /* if LAYOUTGET already failed once we don't try again */
1254         if (pnfs_layout_io_test_failed(lo, iomode))
1255                 goto out_unlock;
1256
1257         /* Check to see if the layout for the given range already exists */
1258         lseg = pnfs_find_lseg(lo, &arg);
1259         if (lseg)
1260                 goto out_unlock;
1261
1262         if (pnfs_layoutgets_blocked(lo, 0))
1263                 goto out_unlock;
1264         atomic_inc(&lo->plh_outstanding);
1265
1266         first = list_empty(&lo->plh_layouts) ? true : false;
1267         spin_unlock(&ino->i_lock);
1268
1269         if (first) {
1270                 /* The lo must be on the clp list if there is any
1271                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1272                  */
1273                 spin_lock(&clp->cl_lock);
1274                 list_add_tail(&lo->plh_layouts, &server->layouts);
1275                 spin_unlock(&clp->cl_lock);
1276         }
1277
1278         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1279         if (pg_offset) {
1280                 arg.offset -= pg_offset;
1281                 arg.length += pg_offset;
1282         }
1283         if (arg.length != NFS4_MAX_UINT64)
1284                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1285
1286         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1287         atomic_dec(&lo->plh_outstanding);
1288 out_put_layout_hdr:
1289         pnfs_put_layout_hdr(lo);
1290 out:
1291         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1292                         "(%s, offset: %llu, length: %llu)\n",
1293                         __func__, ino->i_sb->s_id,
1294                         (unsigned long long)NFS_FILEID(ino),
1295                         lseg == NULL ? "not found" : "found",
1296                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1297                         (unsigned long long)pos,
1298                         (unsigned long long)count);
1299         return lseg;
1300 out_unlock:
1301         spin_unlock(&ino->i_lock);
1302         goto out_put_layout_hdr;
1303 }
1304 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1305
1306 struct pnfs_layout_segment *
1307 pnfs_layout_process(struct nfs4_layoutget *lgp)
1308 {
1309         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1310         struct nfs4_layoutget_res *res = &lgp->res;
1311         struct pnfs_layout_segment *lseg;
1312         struct inode *ino = lo->plh_inode;
1313         int status = 0;
1314
1315         /* Inject layout blob into I/O device driver */
1316         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1317         if (!lseg || IS_ERR(lseg)) {
1318                 if (!lseg)
1319                         status = -ENOMEM;
1320                 else
1321                         status = PTR_ERR(lseg);
1322                 dprintk("%s: Could not allocate layout: error %d\n",
1323                        __func__, status);
1324                 goto out;
1325         }
1326
1327         spin_lock(&ino->i_lock);
1328         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1329                 dprintk("%s forget reply due to recall\n", __func__);
1330                 goto out_forget_reply;
1331         }
1332
1333         if (pnfs_layoutgets_blocked(lo, 1) ||
1334             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1335                 dprintk("%s forget reply due to state\n", __func__);
1336                 goto out_forget_reply;
1337         }
1338
1339         /* Done processing layoutget. Set the layout stateid */
1340         pnfs_set_layout_stateid(lo, &res->stateid, false);
1341
1342         init_lseg(lo, lseg);
1343         lseg->pls_range = res->range;
1344         pnfs_get_lseg(lseg);
1345         pnfs_layout_insert_lseg(lo, lseg);
1346
1347         if (res->return_on_close) {
1348                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1349                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1350         }
1351
1352         spin_unlock(&ino->i_lock);
1353         return lseg;
1354 out:
1355         return ERR_PTR(status);
1356
1357 out_forget_reply:
1358         spin_unlock(&ino->i_lock);
1359         lseg->pls_layout = lo;
1360         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1361         goto out;
1362 }
1363
1364 void
1365 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1366 {
1367         u64 rd_size = req->wb_bytes;
1368
1369         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1370
1371         if (req->wb_offset != req->wb_pgbase) {
1372                 nfs_pageio_reset_read_mds(pgio);
1373                 return;
1374         }
1375
1376         if (pgio->pg_dreq == NULL)
1377                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1378         else
1379                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1380
1381         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1382                                            req->wb_context,
1383                                            req_offset(req),
1384                                            rd_size,
1385                                            IOMODE_READ,
1386                                            GFP_KERNEL);
1387         /* If no lseg, fall back to read through mds */
1388         if (pgio->pg_lseg == NULL)
1389                 nfs_pageio_reset_read_mds(pgio);
1390
1391 }
1392 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1393
1394 void
1395 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1396                            struct nfs_page *req, u64 wb_size)
1397 {
1398         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1399
1400         if (req->wb_offset != req->wb_pgbase) {
1401                 nfs_pageio_reset_write_mds(pgio);
1402                 return;
1403         }
1404
1405         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1406                                            req->wb_context,
1407                                            req_offset(req),
1408                                            wb_size,
1409                                            IOMODE_RW,
1410                                            GFP_NOFS);
1411         /* If no lseg, fall back to write through mds */
1412         if (pgio->pg_lseg == NULL)
1413                 nfs_pageio_reset_write_mds(pgio);
1414 }
1415 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1416
1417 void
1418 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1419                       const struct nfs_pgio_completion_ops *compl_ops)
1420 {
1421         struct nfs_server *server = NFS_SERVER(inode);
1422         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1423
1424         if (ld == NULL)
1425                 nfs_pageio_init_read(pgio, inode, compl_ops);
1426         else
1427                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1428 }
1429
1430 void
1431 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1432                        int ioflags,
1433                        const struct nfs_pgio_completion_ops *compl_ops)
1434 {
1435         struct nfs_server *server = NFS_SERVER(inode);
1436         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1437
1438         if (ld == NULL)
1439                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1440         else
1441                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1442 }
1443
1444 bool
1445 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1446                      struct nfs_page *req)
1447 {
1448         if (pgio->pg_lseg == NULL)
1449                 return nfs_generic_pg_test(pgio, prev, req);
1450
1451         /*
1452          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1453          * Note that this test makes several assumptions:
1454          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1455          *   is known to lie within the range.
1456          *   - that the nfs_page being tested is known to be contiguous with the
1457          *   previous nfs_page.
1458          *   - Layout ranges are page aligned, so we only have to test the
1459          *   start offset of the request.
1460          *
1461          * Please also note that 'end_offset' is actually the offset of the
1462          * first byte that lies outside the pnfs_layout_range. FIXME?
1463          *
1464          */
1465         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1466                                          pgio->pg_lseg->pls_range.length);
1467 }
1468 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1469
1470 int pnfs_write_done_resend_to_mds(struct inode *inode,
1471                                 struct list_head *head,
1472                                 const struct nfs_pgio_completion_ops *compl_ops,
1473                                 struct nfs_direct_req *dreq)
1474 {
1475         struct nfs_pageio_descriptor pgio;
1476         LIST_HEAD(failed);
1477
1478         /* Resend all requests through the MDS */
1479         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1480         pgio.pg_dreq = dreq;
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                 /* For some reason our attempt to resend pages. Mark the
1492                  * overall send request as having failed, and let
1493                  * nfs_writeback_release_full deal with the error.
1494                  */
1495                 list_move(&failed, head);
1496                 return -EIO;
1497         }
1498         return 0;
1499 }
1500 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1501
1502 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1503 {
1504         struct nfs_pgio_header *hdr = data->header;
1505
1506         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1507         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1508             PNFS_LAYOUTRET_ON_ERROR) {
1509                 pnfs_return_layout(hdr->inode);
1510         }
1511         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1512                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1513                                                         &hdr->pages,
1514                                                         hdr->completion_ops,
1515                                                         hdr->dreq);
1516 }
1517
1518 /*
1519  * Called by non rpc-based layout drivers
1520  */
1521 void pnfs_ld_write_done(struct nfs_write_data *data)
1522 {
1523         struct nfs_pgio_header *hdr = data->header;
1524
1525         if (!hdr->pnfs_error) {
1526                 pnfs_set_layoutcommit(data);
1527                 hdr->mds_ops->rpc_call_done(&data->task, data);
1528         } else
1529                 pnfs_ld_handle_write_error(data);
1530         hdr->mds_ops->rpc_release(data);
1531 }
1532 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1533
1534 static void
1535 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1536                 struct nfs_write_data *data)
1537 {
1538         struct nfs_pgio_header *hdr = data->header;
1539
1540         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1541                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1542                 nfs_pageio_reset_write_mds(desc);
1543                 desc->pg_recoalesce = 1;
1544         }
1545         nfs_writedata_release(data);
1546 }
1547
1548 static enum pnfs_try_status
1549 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1550                         const struct rpc_call_ops *call_ops,
1551                         struct pnfs_layout_segment *lseg,
1552                         int how)
1553 {
1554         struct nfs_pgio_header *hdr = wdata->header;
1555         struct inode *inode = hdr->inode;
1556         enum pnfs_try_status trypnfs;
1557         struct nfs_server *nfss = NFS_SERVER(inode);
1558
1559         hdr->mds_ops = call_ops;
1560
1561         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1562                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1563         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1564         if (trypnfs != PNFS_NOT_ATTEMPTED)
1565                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1566         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1567         return trypnfs;
1568 }
1569
1570 static void
1571 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1572 {
1573         struct nfs_write_data *data;
1574         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1575         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1576
1577         desc->pg_lseg = NULL;
1578         while (!list_empty(head)) {
1579                 enum pnfs_try_status trypnfs;
1580
1581                 data = list_first_entry(head, struct nfs_write_data, list);
1582                 list_del_init(&data->list);
1583
1584                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1585                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1586                         pnfs_write_through_mds(desc, data);
1587         }
1588         pnfs_put_lseg(lseg);
1589 }
1590
1591 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1592 {
1593         pnfs_put_lseg(hdr->lseg);
1594         nfs_writehdr_free(hdr);
1595 }
1596 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1597
1598 int
1599 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1600 {
1601         struct nfs_write_header *whdr;
1602         struct nfs_pgio_header *hdr;
1603         int ret;
1604
1605         whdr = nfs_writehdr_alloc();
1606         if (!whdr) {
1607                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1608                 pnfs_put_lseg(desc->pg_lseg);
1609                 desc->pg_lseg = NULL;
1610                 return -ENOMEM;
1611         }
1612         hdr = &whdr->header;
1613         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1614         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1615         atomic_inc(&hdr->refcnt);
1616         ret = nfs_generic_flush(desc, hdr);
1617         if (ret != 0) {
1618                 pnfs_put_lseg(desc->pg_lseg);
1619                 desc->pg_lseg = NULL;
1620         } else
1621                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
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_writepages);
1627
1628 int pnfs_read_done_resend_to_mds(struct inode *inode,
1629                                 struct list_head *head,
1630                                 const struct nfs_pgio_completion_ops *compl_ops,
1631                                 struct nfs_direct_req *dreq)
1632 {
1633         struct nfs_pageio_descriptor pgio;
1634         LIST_HEAD(failed);
1635
1636         /* Resend all requests through the MDS */
1637         nfs_pageio_init_read(&pgio, inode, compl_ops);
1638         pgio.pg_dreq = dreq;
1639         while (!list_empty(head)) {
1640                 struct nfs_page *req = nfs_list_entry(head->next);
1641
1642                 nfs_list_remove_request(req);
1643                 if (!nfs_pageio_add_request(&pgio, req))
1644                         nfs_list_add_request(req, &failed);
1645         }
1646         nfs_pageio_complete(&pgio);
1647
1648         if (!list_empty(&failed)) {
1649                 list_move(&failed, head);
1650                 return -EIO;
1651         }
1652         return 0;
1653 }
1654 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1655
1656 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1657 {
1658         struct nfs_pgio_header *hdr = data->header;
1659
1660         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1661         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1662             PNFS_LAYOUTRET_ON_ERROR) {
1663                 pnfs_return_layout(hdr->inode);
1664         }
1665         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1666                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1667                                                         &hdr->pages,
1668                                                         hdr->completion_ops,
1669                                                         hdr->dreq);
1670 }
1671
1672 /*
1673  * Called by non rpc-based layout drivers
1674  */
1675 void pnfs_ld_read_done(struct nfs_read_data *data)
1676 {
1677         struct nfs_pgio_header *hdr = data->header;
1678
1679         if (likely(!hdr->pnfs_error)) {
1680                 __nfs4_read_done_cb(data);
1681                 hdr->mds_ops->rpc_call_done(&data->task, data);
1682         } else
1683                 pnfs_ld_handle_read_error(data);
1684         hdr->mds_ops->rpc_release(data);
1685 }
1686 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1687
1688 static void
1689 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1690                 struct nfs_read_data *data)
1691 {
1692         struct nfs_pgio_header *hdr = data->header;
1693
1694         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1695                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1696                 nfs_pageio_reset_read_mds(desc);
1697                 desc->pg_recoalesce = 1;
1698         }
1699         nfs_readdata_release(data);
1700 }
1701
1702 /*
1703  * Call the appropriate parallel I/O subsystem read function.
1704  */
1705 static enum pnfs_try_status
1706 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1707                        const struct rpc_call_ops *call_ops,
1708                        struct pnfs_layout_segment *lseg)
1709 {
1710         struct nfs_pgio_header *hdr = rdata->header;
1711         struct inode *inode = hdr->inode;
1712         struct nfs_server *nfss = NFS_SERVER(inode);
1713         enum pnfs_try_status trypnfs;
1714
1715         hdr->mds_ops = call_ops;
1716
1717         dprintk("%s: Reading ino:%lu %u@%llu\n",
1718                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1719
1720         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1721         if (trypnfs != PNFS_NOT_ATTEMPTED)
1722                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1723         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1724         return trypnfs;
1725 }
1726
1727 static void
1728 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1729 {
1730         struct nfs_read_data *data;
1731         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1732         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1733
1734         desc->pg_lseg = NULL;
1735         while (!list_empty(head)) {
1736                 enum pnfs_try_status trypnfs;
1737
1738                 data = list_first_entry(head, struct nfs_read_data, list);
1739                 list_del_init(&data->list);
1740
1741                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1742                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1743                         pnfs_read_through_mds(desc, data);
1744         }
1745         pnfs_put_lseg(lseg);
1746 }
1747
1748 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1749 {
1750         pnfs_put_lseg(hdr->lseg);
1751         nfs_readhdr_free(hdr);
1752 }
1753 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1754
1755 int
1756 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1757 {
1758         struct nfs_read_header *rhdr;
1759         struct nfs_pgio_header *hdr;
1760         int ret;
1761
1762         rhdr = nfs_readhdr_alloc();
1763         if (!rhdr) {
1764                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1765                 ret = -ENOMEM;
1766                 pnfs_put_lseg(desc->pg_lseg);
1767                 desc->pg_lseg = NULL;
1768                 return ret;
1769         }
1770         hdr = &rhdr->header;
1771         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1772         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1773         atomic_inc(&hdr->refcnt);
1774         ret = nfs_generic_pagein(desc, hdr);
1775         if (ret != 0) {
1776                 pnfs_put_lseg(desc->pg_lseg);
1777                 desc->pg_lseg = NULL;
1778         } else
1779                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1780         if (atomic_dec_and_test(&hdr->refcnt))
1781                 hdr->completion_ops->completion(hdr);
1782         return ret;
1783 }
1784 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1785
1786 /*
1787  * There can be multiple RW segments.
1788  */
1789 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1790 {
1791         struct pnfs_layout_segment *lseg;
1792
1793         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1794                 if (lseg->pls_range.iomode == IOMODE_RW &&
1795                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1796                         list_add(&lseg->pls_lc_list, listp);
1797         }
1798 }
1799
1800 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1801 {
1802         struct pnfs_layout_segment *lseg, *tmp;
1803         unsigned long *bitlock = &NFS_I(inode)->flags;
1804
1805         /* Matched by references in pnfs_set_layoutcommit */
1806         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1807                 list_del_init(&lseg->pls_lc_list);
1808                 pnfs_put_lseg(lseg);
1809         }
1810
1811         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1812         smp_mb__after_clear_bit();
1813         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1814 }
1815
1816 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1817 {
1818         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1819 }
1820 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1821
1822 void
1823 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1824 {
1825         struct nfs_pgio_header *hdr = wdata->header;
1826         struct inode *inode = hdr->inode;
1827         struct nfs_inode *nfsi = NFS_I(inode);
1828         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1829         bool mark_as_dirty = false;
1830
1831         spin_lock(&inode->i_lock);
1832         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1833                 mark_as_dirty = true;
1834                 dprintk("%s: Set layoutcommit for inode %lu ",
1835                         __func__, inode->i_ino);
1836         }
1837         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1838                 /* references matched in nfs4_layoutcommit_release */
1839                 pnfs_get_lseg(hdr->lseg);
1840         }
1841         if (end_pos > nfsi->layout->plh_lwb)
1842                 nfsi->layout->plh_lwb = end_pos;
1843         spin_unlock(&inode->i_lock);
1844         dprintk("%s: lseg %p end_pos %llu\n",
1845                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1846
1847         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1848          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1849         if (mark_as_dirty)
1850                 mark_inode_dirty_sync(inode);
1851 }
1852 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1853
1854 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1855 {
1856         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1857
1858         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1859                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1860         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1861 }
1862
1863 /*
1864  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1865  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1866  * data to disk to allow the server to recover the data if it crashes.
1867  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1868  * is off, and a COMMIT is sent to a data server, or
1869  * if WRITEs to a data server return NFS_DATA_SYNC.
1870  */
1871 int
1872 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1873 {
1874         struct nfs4_layoutcommit_data *data;
1875         struct nfs_inode *nfsi = NFS_I(inode);
1876         loff_t end_pos;
1877         int status = 0;
1878
1879         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1880
1881         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1882                 return 0;
1883
1884         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1885         data = kzalloc(sizeof(*data), GFP_NOFS);
1886         if (!data) {
1887                 status = -ENOMEM;
1888                 goto out;
1889         }
1890
1891         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1892                 goto out_free;
1893
1894         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1895                 if (!sync) {
1896                         status = -EAGAIN;
1897                         goto out_free;
1898                 }
1899                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1900                                         nfs_wait_bit_killable, TASK_KILLABLE);
1901                 if (status)
1902                         goto out_free;
1903         }
1904
1905         INIT_LIST_HEAD(&data->lseg_list);
1906         spin_lock(&inode->i_lock);
1907         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1908                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1909                 spin_unlock(&inode->i_lock);
1910                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1911                 goto out_free;
1912         }
1913
1914         pnfs_list_write_lseg(inode, &data->lseg_list);
1915
1916         end_pos = nfsi->layout->plh_lwb;
1917         nfsi->layout->plh_lwb = 0;
1918
1919         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1920         spin_unlock(&inode->i_lock);
1921
1922         data->args.inode = inode;
1923         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1924         nfs_fattr_init(&data->fattr);
1925         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1926         data->res.fattr = &data->fattr;
1927         data->args.lastbytewritten = end_pos - 1;
1928         data->res.server = NFS_SERVER(inode);
1929
1930         status = nfs4_proc_layoutcommit(data, sync);
1931 out:
1932         if (status)
1933                 mark_inode_dirty_sync(inode);
1934         dprintk("<-- %s status %d\n", __func__, status);
1935         return status;
1936 out_free:
1937         kfree(data);
1938         goto out;
1939 }
1940
1941 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1942 {
1943         struct nfs4_threshold *thp;
1944
1945         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1946         if (!thp) {
1947                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1948                 return NULL;
1949         }
1950         return thp;
1951 }