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 #include "nfs4trace.h"
  37
  38 #define NFSDBG_FACILITY         NFSDBG_PNFS
  39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  40
  41 /* Locking:
  42  *
  43  * pnfs_spinlock:
  44  *      protects pnfs_modules_tbl.
  45  */
  46 static DEFINE_SPINLOCK(pnfs_spinlock);
  47
  48 /*
  49  * pnfs_modules_tbl holds all pnfs modules
  50  */
  51 static LIST_HEAD(pnfs_modules_tbl);
  52
  53 /* Return the registered pnfs layout driver module matching given id */
  54 static struct pnfs_layoutdriver_type *
  55 find_pnfs_driver_locked(u32 id)
  56 {
  57         struct pnfs_layoutdriver_type *local;
  58
  59         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  60                 if (local->id == id)
  61                         goto out;
  62         local = NULL;
  63 out:
  64         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  65         return local;
  66 }
  67
  68 static struct pnfs_layoutdriver_type *
  69 find_pnfs_driver(u32 id)
  70 {
  71         struct pnfs_layoutdriver_type *local;
  72
  73         spin_lock(&pnfs_spinlock);
  74         local = find_pnfs_driver_locked(id);
  75         if (local != NULL && !try_module_get(local->owner)) {
  76                 dprintk("%s: Could not grab reference on module\n", __func__);
  77                 local = NULL;
  78         }
  79         spin_unlock(&pnfs_spinlock);
  80         return local;
  81 }
  82
  83 void
  84 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  85 {
  86         if (nfss->pnfs_curr_ld) {
  87                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  88                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
  89                 /* Decrement the MDS count. Purge the deviceid cache if zero */
  90                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
  91                         nfs4_deviceid_purge_client(nfss->nfs_client);
  92                 module_put(nfss->pnfs_curr_ld->owner);
  93         }
  94         nfss->pnfs_curr_ld = NULL;
  95 }
  96
  97 /*
  98  * Try to set the server's pnfs module to the pnfs layout type specified by id.
  99  * Currently only one pNFS layout driver per filesystem is supported.
 100  *
 101  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
 102  */
 103 void
 104 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 105                       u32 id)
 106 {
 107         struct pnfs_layoutdriver_type *ld_type = NULL;
 108
 109         if (id == 0)
 110                 goto out_no_driver;
 111         if (!(server->nfs_client->cl_exchange_flags &
 112                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 113                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
 114                         __func__, id, server->nfs_client->cl_exchange_flags);
 115                 goto out_no_driver;
 116         }
 117         ld_type = find_pnfs_driver(id);
 118         if (!ld_type) {
 119                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
 120                 ld_type = find_pnfs_driver(id);
 121                 if (!ld_type) {
 122                         dprintk("%s: No pNFS module found for %u.\n",
 123                                 __func__, id);
 124                         goto out_no_driver;
 125                 }
 126         }
 127         server->pnfs_curr_ld = ld_type;
 128         if (ld_type->set_layoutdriver
 129             && ld_type->set_layoutdriver(server, mntfh)) {
 130                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 131                         "driver %u.\n", __func__, id);
 132                 module_put(ld_type->owner);
 133                 goto out_no_driver;
 134         }
 135         /* Bump the MDS count */
 136         atomic_inc(&server->nfs_client->cl_mds_count);
 137
 138         dprintk("%s: pNFS module for %u set\n", __func__, id);
 139         return;
 140
 141 out_no_driver:
 142         dprintk("%s: Using NFSv4 I/O\n", __func__);
 143         server->pnfs_curr_ld = NULL;
 144 }
 145
 146 int
 147 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 148 {
 149         int status = -EINVAL;
 150         struct pnfs_layoutdriver_type *tmp;
 151
 152         if (ld_type->id == 0) {
 153                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 154                 return status;
 155         }
 156         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 157                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 158                        "alloc_lseg and free_lseg.\n", __func__);
 159                 return status;
 160         }
 161
 162         spin_lock(&pnfs_spinlock);
 163         tmp = find_pnfs_driver_locked(ld_type->id);
 164         if (!tmp) {
 165                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 166                 status = 0;
 167                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 168                         ld_type->name);
 169         } else {
 170                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 171                         __func__, ld_type->id);
 172         }
 173         spin_unlock(&pnfs_spinlock);
 174
 175         return status;
 176 }
 177 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 178
 179 void
 180 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 181 {
 182         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 183         spin_lock(&pnfs_spinlock);
 184         list_del(&ld_type->pnfs_tblid);
 185         spin_unlock(&pnfs_spinlock);
 186 }
 187 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 188
 189 /*
 190  * pNFS client layout cache
 191  */
 192
 193 /* Need to hold i_lock if caller does not already hold reference */
 194 void
 195 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 196 {
 197         atomic_inc(&lo->plh_refcount);
 198 }
 199
 200 static struct pnfs_layout_hdr *
 201 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 202 {
 203         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 204         return ld->alloc_layout_hdr(ino, gfp_flags);
 205 }
 206
 207 static void
 208 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 209 {
 210         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 211         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 212
 213         if (!list_empty(&lo->plh_layouts)) {
 214                 struct nfs_client *clp = server->nfs_client;
 215
 216                 spin_lock(&clp->cl_lock);
 217                 list_del_init(&lo->plh_layouts);
 218                 spin_unlock(&clp->cl_lock);
 219         }
 220         put_rpccred(lo->plh_lc_cred);
 221         return ld->free_layout_hdr(lo);
 222 }
 223
 224 static void
 225 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 226 {
 227         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 228         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 229         nfsi->layout = NULL;
 230         /* Reset MDS Threshold I/O counters */
 231         nfsi->write_io = 0;
 232         nfsi->read_io = 0;
 233 }
 234
 235 void
 236 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 237 {
 238         struct inode *inode = lo->plh_inode;
 239
 240         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 241                 pnfs_detach_layout_hdr(lo);
 242                 spin_unlock(&inode->i_lock);
 243                 pnfs_free_layout_hdr(lo);
 244         }
 245 }
 246
 247 static int
 248 pnfs_iomode_to_fail_bit(u32 iomode)
 249 {
 250         return iomode == IOMODE_RW ?
 251                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 252 }
 253
 254 static void
 255 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 256 {
 257         lo->plh_retry_timestamp = jiffies;
 258         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 259                 atomic_inc(&lo->plh_refcount);
 260 }
 261
 262 static void
 263 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 264 {
 265         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 266                 atomic_dec(&lo->plh_refcount);
 267 }
 268
 269 static void
 270 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 271 {
 272         struct inode *inode = lo->plh_inode;
 273         struct pnfs_layout_range range = {
 274                 .iomode = iomode,
 275                 .offset = 0,
 276                 .length = NFS4_MAX_UINT64,
 277         };
 278         LIST_HEAD(head);
 279
 280         spin_lock(&inode->i_lock);
 281         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 282         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
 283         spin_unlock(&inode->i_lock);
 284         pnfs_free_lseg_list(&head);
 285         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 286                         iomode == IOMODE_RW ?  "RW" : "READ");
 287 }
 288
 289 static bool
 290 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 291 {
 292         unsigned long start, end;
 293         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 294
 295         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 296                 return false;
 297         end = jiffies;
 298         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 299         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 300                 /* It is time to retry the failed layoutgets */
 301                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 302                 return false;
 303         }
 304         return true;
 305 }
 306
 307 static void
 308 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 309 {
 310         INIT_LIST_HEAD(&lseg->pls_list);
 311         INIT_LIST_HEAD(&lseg->pls_lc_list);
 312         atomic_set(&lseg->pls_refcount, 1);
 313         smp_mb();
 314         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 315         lseg->pls_layout = lo;
 316 }
 317
 318 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 319 {
 320         struct inode *ino = lseg->pls_layout->plh_inode;
 321
 322         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 323 }
 324
 325 static void
 326 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 327                 struct pnfs_layout_segment *lseg)
 328 {
 329         struct inode *inode = lo->plh_inode;
 330
 331         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 332         list_del_init(&lseg->pls_list);
 333         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 334         atomic_dec(&lo->plh_refcount);
 335         if (list_empty(&lo->plh_segs))
 336                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 337         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
 338 }
 339
 340 void
 341 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 342 {
 343         struct pnfs_layout_hdr *lo;
 344         struct inode *inode;
 345
 346         if (!lseg)
 347                 return;
 348
 349         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 350                 atomic_read(&lseg->pls_refcount),
 351                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 352         lo = lseg->pls_layout;
 353         inode = lo->plh_inode;
 354         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 355                 pnfs_get_layout_hdr(lo);
 356                 pnfs_layout_remove_lseg(lo, lseg);
 357                 spin_unlock(&inode->i_lock);
 358                 pnfs_free_lseg(lseg);
 359                 pnfs_put_layout_hdr(lo);
 360         }
 361 }
 362 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 363
 364 static u64
 365 end_offset(u64 start, u64 len)
 366 {
 367         u64 end;
 368
 369         end = start + len;
 370         return end >= start ? end : NFS4_MAX_UINT64;
 371 }
 372
 373 /*
 374  * is l2 fully contained in l1?
 375  *   start1                             end1
 376  *   [----------------------------------)
 377  *           start2           end2
 378  *           [----------------)
 379  */
 380 static bool
 381 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 382                  const struct pnfs_layout_range *l2)
 383 {
 384         u64 start1 = l1->offset;
 385         u64 end1 = end_offset(start1, l1->length);
 386         u64 start2 = l2->offset;
 387         u64 end2 = end_offset(start2, l2->length);
 388
 389         return (start1 <= start2) && (end1 >= end2);
 390 }
 391
 392 /*
 393  * is l1 and l2 intersecting?
 394  *   start1                             end1
 395  *   [----------------------------------)
 396  *                              start2           end2
 397  *                              [----------------)
 398  */
 399 static bool
 400 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 401                     const struct pnfs_layout_range *l2)
 402 {
 403         u64 start1 = l1->offset;
 404         u64 end1 = end_offset(start1, l1->length);
 405         u64 start2 = l2->offset;
 406         u64 end2 = end_offset(start2, l2->length);
 407
 408         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
 409                (end2 == NFS4_MAX_UINT64 || end2 > start1);
 410 }
 411
 412 static bool
 413 should_free_lseg(const struct pnfs_layout_range *lseg_range,
 414                  const struct pnfs_layout_range *recall_range)
 415 {
 416         return (recall_range->iomode == IOMODE_ANY ||
 417                 lseg_range->iomode == recall_range->iomode) &&
 418                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 419 }
 420
 421 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 422                 struct list_head *tmp_list)
 423 {
 424         if (!atomic_dec_and_test(&lseg->pls_refcount))
 425                 return false;
 426         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 427         list_add(&lseg->pls_list, tmp_list);
 428         return true;
 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 (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 445                         rv = 1;
 446         }
 447         return rv;
 448 }
 449
 450 /* Returns count of number of matching invalid lsegs remaining in list
 451  * after call.
 452  */
 453 int
 454 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 455                             struct list_head *tmp_list,
 456                             struct pnfs_layout_range *recall_range)
 457 {
 458         struct pnfs_layout_segment *lseg, *next;
 459         int invalid = 0, removed = 0;
 460
 461         dprintk("%s:Begin lo %p\n", __func__, lo);
 462
 463         if (list_empty(&lo->plh_segs))
 464                 return 0;
 465         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 466                 if (!recall_range ||
 467                     should_free_lseg(&lseg->pls_range, recall_range)) {
 468                         dprintk("%s: freeing lseg %p iomode %d "
 469                                 "offset %llu length %llu\n", __func__,
 470                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
 471                                 lseg->pls_range.length);
 472                         invalid++;
 473                         removed += mark_lseg_invalid(lseg, tmp_list);
 474                 }
 475         dprintk("%s:Return %i\n", __func__, invalid - removed);
 476         return invalid - removed;
 477 }
 478
 479 /* note free_me must contain lsegs from a single layout_hdr */
 480 void
 481 pnfs_free_lseg_list(struct list_head *free_me)
 482 {
 483         struct pnfs_layout_segment *lseg, *tmp;
 484
 485         if (list_empty(free_me))
 486                 return;
 487
 488         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 489                 list_del(&lseg->pls_list);
 490                 pnfs_free_lseg(lseg);
 491         }
 492 }
 493
 494 void
 495 pnfs_destroy_layout(struct nfs_inode *nfsi)
 496 {
 497         struct pnfs_layout_hdr *lo;
 498         LIST_HEAD(tmp_list);
 499
 500         spin_lock(&nfsi->vfs_inode.i_lock);
 501         lo = nfsi->layout;
 502         if (lo) {
 503                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 504                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 505                 pnfs_get_layout_hdr(lo);
 506                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 507                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 508                 spin_unlock(&nfsi->vfs_inode.i_lock);
 509                 pnfs_free_lseg_list(&tmp_list);
 510                 pnfs_put_layout_hdr(lo);
 511         } else
 512                 spin_unlock(&nfsi->vfs_inode.i_lock);
 513 }
 514 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 515
 516 static bool
 517 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 518                 struct list_head *layout_list)
 519 {
 520         struct pnfs_layout_hdr *lo;
 521         bool ret = false;
 522
 523         spin_lock(&inode->i_lock);
 524         lo = NFS_I(inode)->layout;
 525         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 526                 pnfs_get_layout_hdr(lo);
 527                 list_add(&lo->plh_bulk_destroy, layout_list);
 528                 ret = true;
 529         }
 530         spin_unlock(&inode->i_lock);
 531         return ret;
 532 }
 533
 534 /* Caller must hold rcu_read_lock and clp->cl_lock */
 535 static int
 536 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 537                 struct nfs_server *server,
 538                 struct list_head *layout_list)
 539 {
 540         struct pnfs_layout_hdr *lo, *next;
 541         struct inode *inode;
 542
 543         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 544                 inode = igrab(lo->plh_inode);
 545                 if (inode == NULL)
 546                         continue;
 547                 list_del_init(&lo->plh_layouts);
 548                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 549                         continue;
 550                 rcu_read_unlock();
 551                 spin_unlock(&clp->cl_lock);
 552                 iput(inode);
 553                 spin_lock(&clp->cl_lock);
 554                 rcu_read_lock();
 555                 return -EAGAIN;
 556         }
 557         return 0;
 558 }
 559
 560 static int
 561 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 562                 bool is_bulk_recall)
 563 {
 564         struct pnfs_layout_hdr *lo;
 565         struct inode *inode;
 566         struct pnfs_layout_range range = {
 567                 .iomode = IOMODE_ANY,
 568                 .offset = 0,
 569                 .length = NFS4_MAX_UINT64,
 570         };
 571         LIST_HEAD(lseg_list);
 572         int ret = 0;
 573
 574         while (!list_empty(layout_list)) {
 575                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 576                                 plh_bulk_destroy);
 577                 dprintk("%s freeing layout for inode %lu\n", __func__,
 578                         lo->plh_inode->i_ino);
 579                 inode = lo->plh_inode;
 580                 spin_lock(&inode->i_lock);
 581                 list_del_init(&lo->plh_bulk_destroy);
 582                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
 583                 if (is_bulk_recall)
 584                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 585                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
 586                         ret = -EAGAIN;
 587                 spin_unlock(&inode->i_lock);
 588                 pnfs_free_lseg_list(&lseg_list);
 589                 pnfs_put_layout_hdr(lo);
 590                 iput(inode);
 591         }
 592         return ret;
 593 }
 594
 595 int
 596 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 597                 struct nfs_fsid *fsid,
 598                 bool is_recall)
 599 {
 600         struct nfs_server *server;
 601         LIST_HEAD(layout_list);
 602
 603         spin_lock(&clp->cl_lock);
 604         rcu_read_lock();
 605 restart:
 606         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 607                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 608                         continue;
 609                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 610                                 server,
 611                                 &layout_list) != 0)
 612                         goto restart;
 613         }
 614         rcu_read_unlock();
 615         spin_unlock(&clp->cl_lock);
 616
 617         if (list_empty(&layout_list))
 618                 return 0;
 619         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 620 }
 621
 622 int
 623 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 624                 bool is_recall)
 625 {
 626         struct nfs_server *server;
 627         LIST_HEAD(layout_list);
 628
 629         spin_lock(&clp->cl_lock);
 630         rcu_read_lock();
 631 restart:
 632         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 633                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 634                                         server,
 635                                         &layout_list) != 0)
 636                         goto restart;
 637         }
 638         rcu_read_unlock();
 639         spin_unlock(&clp->cl_lock);
 640
 641         if (list_empty(&layout_list))
 642                 return 0;
 643         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 644 }
 645
 646 /*
 647  * Called by the state manger to remove all layouts established under an
 648  * expired lease.
 649  */
 650 void
 651 pnfs_destroy_all_layouts(struct nfs_client *clp)
 652 {
 653         nfs4_deviceid_mark_client_invalid(clp);
 654         nfs4_deviceid_purge_client(clp);
 655
 656         pnfs_destroy_layouts_byclid(clp, false);
 657 }
 658
 659 /*
 660  * Compare 2 layout stateid sequence ids, to see which is newer,
 661  * taking into account wraparound issues.
 662  */
 663 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 664 {
 665         return (s32)(s1 - s2) > 0;
 666 }
 667
 668 static void
 669 pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
 670                 const nfs4_stateid *new,
 671                 struct list_head *free_me_list)
 672 {
 673         if (nfs4_stateid_match_other(&lo->plh_stateid, new))
 674                 return;
 675         /* Layout is new! Kill existing layout segments */
 676         pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
 677 }
 678
 679 /* update lo->plh_stateid with new if is more recent */
 680 void
 681 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 682                         bool update_barrier)
 683 {
 684         u32 oldseq, newseq, new_barrier;
 685         int empty = list_empty(&lo->plh_segs);
 686
 687         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 688         newseq = be32_to_cpu(new->seqid);
 689         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
 690                 nfs4_stateid_copy(&lo->plh_stateid, new);
 691                 if (update_barrier) {
 692                         new_barrier = be32_to_cpu(new->seqid);
 693                 } else {
 694                         /* Because of wraparound, we want to keep the barrier
 695                          * "close" to the current seqids.
 696                          */
 697                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 698                 }
 699                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 700                         lo->plh_barrier = new_barrier;
 701         }
 702 }
 703
 704 static bool
 705 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 706                 const nfs4_stateid *stateid)
 707 {
 708         u32 seqid = be32_to_cpu(stateid->seqid);
 709
 710         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 711 }
 712
 713 /* lget is set to 1 if called from inside send_layoutget call chain */
 714 static bool
 715 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
 716 {
 717         return lo->plh_block_lgets ||
 718                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
 719                 (list_empty(&lo->plh_segs) &&
 720                  (atomic_read(&lo->plh_outstanding) > lget));
 721 }
 722
 723 int
 724 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
 725                               struct nfs4_state *open_state)
 726 {
 727         int status = 0;
 728
 729         dprintk("--> %s\n", __func__);
 730         spin_lock(&lo->plh_inode->i_lock);
 731         if (pnfs_layoutgets_blocked(lo, 1)) {
 732                 status = -EAGAIN;
 733         } else if (!nfs4_valid_open_stateid(open_state)) {
 734                 status = -EBADF;
 735         } else if (list_empty(&lo->plh_segs)) {
 736                 int seq;
 737
 738                 do {
 739                         seq = read_seqbegin(&open_state->seqlock);
 740                         nfs4_stateid_copy(dst, &open_state->stateid);
 741                 } while (read_seqretry(&open_state->seqlock, seq));
 742         } else
 743                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 744         spin_unlock(&lo->plh_inode->i_lock);
 745         dprintk("<-- %s\n", __func__);
 746         return status;
 747 }
 748
 749 /*
 750 * Get layout from server.
 751 *    for now, assume that whole file layouts are requested.
 752 *    arg->offset: 0
 753 *    arg->length: all ones
 754 */
 755 static struct pnfs_layout_segment *
 756 send_layoutget(struct pnfs_layout_hdr *lo,
 757            struct nfs_open_context *ctx,
 758            struct pnfs_layout_range *range,
 759            gfp_t gfp_flags)
 760 {
 761         struct inode *ino = lo->plh_inode;
 762         struct nfs_server *server = NFS_SERVER(ino);
 763         struct nfs4_layoutget *lgp;
 764         struct pnfs_layout_segment *lseg;
 765
 766         dprintk("--> %s\n", __func__);
 767
 768         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 769         if (lgp == NULL)
 770                 return NULL;
 771
 772         lgp->args.minlength = PAGE_CACHE_SIZE;
 773         if (lgp->args.minlength > range->length)
 774                 lgp->args.minlength = range->length;
 775         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 776         lgp->args.range = *range;
 777         lgp->args.type = server->pnfs_curr_ld->id;
 778         lgp->args.inode = ino;
 779         lgp->args.ctx = get_nfs_open_context(ctx);
 780         lgp->gfp_flags = gfp_flags;
 781         lgp->cred = lo->plh_lc_cred;
 782
 783         /* Synchronously retrieve layout information from server and
 784          * store in lseg.
 785          */
 786         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 787         if (IS_ERR(lseg)) {
 788                 switch (PTR_ERR(lseg)) {
 789                 case -ENOMEM:
 790                 case -ERESTARTSYS:
 791                         break;
 792                 default:
 793                         /* remember that LAYOUTGET failed and suspend trying */
 794                         pnfs_layout_io_set_failed(lo, range->iomode);
 795                 }
 796                 return NULL;
 797         }
 798
 799         return lseg;
 800 }
 801
 802 static void pnfs_clear_layoutcommit(struct inode *inode,
 803                 struct list_head *head)
 804 {
 805         struct nfs_inode *nfsi = NFS_I(inode);
 806         struct pnfs_layout_segment *lseg, *tmp;
 807
 808         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 809                 return;
 810         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 811                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 812                         continue;
 813                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 814         }
 815 }
 816
 817 /*
 818  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 819  * when the layout segment list is empty.
 820  *
 821  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 822  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 823  * deviceid is marked invalid.
 824  */
 825 int
 826 _pnfs_return_layout(struct inode *ino)
 827 {
 828         struct pnfs_layout_hdr *lo = NULL;
 829         struct nfs_inode *nfsi = NFS_I(ino);
 830         LIST_HEAD(tmp_list);
 831         struct nfs4_layoutreturn *lrp;
 832         nfs4_stateid stateid;
 833         int status = 0, empty;
 834
 835         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 836
 837         spin_lock(&ino->i_lock);
 838         lo = nfsi->layout;
 839         if (!lo) {
 840                 spin_unlock(&ino->i_lock);
 841                 dprintk("NFS: %s no layout to return\n", __func__);
 842                 goto out;
 843         }
 844         stateid = nfsi->layout->plh_stateid;
 845         /* Reference matched in nfs4_layoutreturn_release */
 846         pnfs_get_layout_hdr(lo);
 847         empty = list_empty(&lo->plh_segs);
 848         pnfs_clear_layoutcommit(ino, &tmp_list);
 849         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 850         /* Don't send a LAYOUTRETURN if list was initially empty */
 851         if (empty) {
 852                 spin_unlock(&ino->i_lock);
 853                 pnfs_put_layout_hdr(lo);
 854                 dprintk("NFS: %s no layout segments to return\n", __func__);
 855                 goto out;
 856         }
 857         lo->plh_block_lgets++;
 858         spin_unlock(&ino->i_lock);
 859         pnfs_free_lseg_list(&tmp_list);
 860
 861         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 862         if (unlikely(lrp == NULL)) {
 863                 status = -ENOMEM;
 864                 spin_lock(&ino->i_lock);
 865                 lo->plh_block_lgets--;
 866                 spin_unlock(&ino->i_lock);
 867                 pnfs_put_layout_hdr(lo);
 868                 goto out;
 869         }
 870
 871         lrp->args.stateid = stateid;
 872         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 873         lrp->args.inode = ino;
 874         lrp->args.layout = lo;
 875         lrp->clp = NFS_SERVER(ino)->nfs_client;
 876         lrp->cred = lo->plh_lc_cred;
 877
 878         status = nfs4_proc_layoutreturn(lrp);
 879 out:
 880         dprintk("<-- %s status: %d\n", __func__, status);
 881         return status;
 882 }
 883 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 884
 885 int
 886 pnfs_commit_and_return_layout(struct inode *inode)
 887 {
 888         struct pnfs_layout_hdr *lo;
 889         int ret;
 890
 891         spin_lock(&inode->i_lock);
 892         lo = NFS_I(inode)->layout;
 893         if (lo == NULL) {
 894                 spin_unlock(&inode->i_lock);
 895                 return 0;
 896         }
 897         pnfs_get_layout_hdr(lo);
 898         /* Block new layoutgets and read/write to ds */
 899         lo->plh_block_lgets++;
 900         spin_unlock(&inode->i_lock);
 901         filemap_fdatawait(inode->i_mapping);
 902         ret = pnfs_layoutcommit_inode(inode, true);
 903         if (ret == 0)
 904                 ret = _pnfs_return_layout(inode);
 905         spin_lock(&inode->i_lock);
 906         lo->plh_block_lgets--;
 907         spin_unlock(&inode->i_lock);
 908         pnfs_put_layout_hdr(lo);
 909         return ret;
 910 }
 911
 912 bool pnfs_roc(struct inode *ino)
 913 {
 914         struct pnfs_layout_hdr *lo;
 915         struct pnfs_layout_segment *lseg, *tmp;
 916         LIST_HEAD(tmp_list);
 917         bool found = false;
 918
 919         spin_lock(&ino->i_lock);
 920         lo = NFS_I(ino)->layout;
 921         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 922             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 923                 goto out_nolayout;
 924         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 925                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 926                         mark_lseg_invalid(lseg, &tmp_list);
 927                         found = true;
 928                 }
 929         if (!found)
 930                 goto out_nolayout;
 931         lo->plh_block_lgets++;
 932         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 933         spin_unlock(&ino->i_lock);
 934         pnfs_free_lseg_list(&tmp_list);
 935         return true;
 936
 937 out_nolayout:
 938         spin_unlock(&ino->i_lock);
 939         return false;
 940 }
 941
 942 void pnfs_roc_release(struct inode *ino)
 943 {
 944         struct pnfs_layout_hdr *lo;
 945
 946         spin_lock(&ino->i_lock);
 947         lo = NFS_I(ino)->layout;
 948         lo->plh_block_lgets--;
 949         if (atomic_dec_and_test(&lo->plh_refcount)) {
 950                 pnfs_detach_layout_hdr(lo);
 951                 spin_unlock(&ino->i_lock);
 952                 pnfs_free_layout_hdr(lo);
 953         } else
 954                 spin_unlock(&ino->i_lock);
 955 }
 956
 957 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 958 {
 959         struct pnfs_layout_hdr *lo;
 960
 961         spin_lock(&ino->i_lock);
 962         lo = NFS_I(ino)->layout;
 963         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 964                 lo->plh_barrier = barrier;
 965         spin_unlock(&ino->i_lock);
 966 }
 967
 968 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 969 {
 970         struct nfs_inode *nfsi = NFS_I(ino);
 971         struct pnfs_layout_hdr *lo;
 972         struct pnfs_layout_segment *lseg;
 973         u32 current_seqid;
 974         bool found = false;
 975
 976         spin_lock(&ino->i_lock);
 977         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 978                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 979                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 980                         found = true;
 981                         goto out;
 982                 }
 983         lo = nfsi->layout;
 984         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 985
 986         /* Since close does not return a layout stateid for use as
 987          * a barrier, we choose the worst-case barrier.
 988          */
 989         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 990 out:
 991         spin_unlock(&ino->i_lock);
 992         return found;
 993 }
 994
 995 /*
 996  * Compare two layout segments for sorting into layout cache.
 997  * We want to preferentially return RW over RO layouts, so ensure those
 998  * are seen first.
 999  */
1000 static s64
1001 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1002            const struct pnfs_layout_range *l2)
1003 {
1004         s64 d;
1005
1006         /* high offset > low offset */
1007         d = l1->offset - l2->offset;
1008         if (d)
1009                 return d;
1010
1011         /* short length > long length */
1012         d = l2->length - l1->length;
1013         if (d)
1014                 return d;
1015
1016         /* read > read/write */
1017         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1018 }
1019
1020 static void
1021 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1022                    struct pnfs_layout_segment *lseg)
1023 {
1024         struct pnfs_layout_segment *lp;
1025
1026         dprintk("%s:Begin\n", __func__);
1027
1028         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1029                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1030                         continue;
1031                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1032                 dprintk("%s: inserted lseg %p "
1033                         "iomode %d offset %llu length %llu before "
1034                         "lp %p iomode %d offset %llu length %llu\n",
1035                         __func__, lseg, lseg->pls_range.iomode,
1036                         lseg->pls_range.offset, lseg->pls_range.length,
1037                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1038                         lp->pls_range.length);
1039                 goto out;
1040         }
1041         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1042         dprintk("%s: inserted lseg %p "
1043                 "iomode %d offset %llu length %llu at tail\n",
1044                 __func__, lseg, lseg->pls_range.iomode,
1045                 lseg->pls_range.offset, lseg->pls_range.length);
1046 out:
1047         pnfs_get_layout_hdr(lo);
1048
1049         dprintk("%s:Return\n", __func__);
1050 }
1051
1052 static struct pnfs_layout_hdr *
1053 alloc_init_layout_hdr(struct inode *ino,
1054                       struct nfs_open_context *ctx,
1055                       gfp_t gfp_flags)
1056 {
1057         struct pnfs_layout_hdr *lo;
1058
1059         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1060         if (!lo)
1061                 return NULL;
1062         atomic_set(&lo->plh_refcount, 1);
1063         INIT_LIST_HEAD(&lo->plh_layouts);
1064         INIT_LIST_HEAD(&lo->plh_segs);
1065         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1066         lo->plh_inode = ino;
1067         lo->plh_lc_cred = get_rpccred(ctx->cred);
1068         return lo;
1069 }
1070
1071 static struct pnfs_layout_hdr *
1072 pnfs_find_alloc_layout(struct inode *ino,
1073                        struct nfs_open_context *ctx,
1074                        gfp_t gfp_flags)
1075 {
1076         struct nfs_inode *nfsi = NFS_I(ino);
1077         struct pnfs_layout_hdr *new = NULL;
1078
1079         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1080
1081         if (nfsi->layout != NULL)
1082                 goto out_existing;
1083         spin_unlock(&ino->i_lock);
1084         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1085         spin_lock(&ino->i_lock);
1086
1087         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1088                 nfsi->layout = new;
1089                 return new;
1090         } else if (new != NULL)
1091                 pnfs_free_layout_hdr(new);
1092 out_existing:
1093         pnfs_get_layout_hdr(nfsi->layout);
1094         return nfsi->layout;
1095 }
1096
1097 /*
1098  * iomode matching rules:
1099  * iomode       lseg    match
1100  * -----        -----   -----
1101  * ANY          READ    true
1102  * ANY          RW      true
1103  * RW           READ    false
1104  * RW           RW      true
1105  * READ         READ    true
1106  * READ         RW      true
1107  */
1108 static bool
1109 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1110                  const struct pnfs_layout_range *range)
1111 {
1112         struct pnfs_layout_range range1;
1113
1114         if ((range->iomode == IOMODE_RW &&
1115              ls_range->iomode != IOMODE_RW) ||
1116             !pnfs_lseg_range_intersecting(ls_range, range))
1117                 return 0;
1118
1119         /* range1 covers only the first byte in the range */
1120         range1 = *range;
1121         range1.length = 1;
1122         return pnfs_lseg_range_contained(ls_range, &range1);
1123 }
1124
1125 /*
1126  * lookup range in layout
1127  */
1128 static struct pnfs_layout_segment *
1129 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1130                 struct pnfs_layout_range *range)
1131 {
1132         struct pnfs_layout_segment *lseg, *ret = NULL;
1133
1134         dprintk("%s:Begin\n", __func__);
1135
1136         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1137                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1138                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1139                         ret = pnfs_get_lseg(lseg);
1140                         break;
1141                 }
1142                 if (lseg->pls_range.offset > range->offset)
1143                         break;
1144         }
1145
1146         dprintk("%s:Return lseg %p ref %d\n",
1147                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1148         return ret;
1149 }
1150
1151 /*
1152  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1153  * to the MDS or over pNFS
1154  *
1155  * The nfs_inode read_io and write_io fields are cumulative counters reset
1156  * when there are no layout segments. Note that in pnfs_update_layout iomode
1157  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1158  * WRITE request.
1159  *
1160  * A return of true means use MDS I/O.
1161  *
1162  * From rfc 5661:
1163  * If a file's size is smaller than the file size threshold, data accesses
1164  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1165  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1166  * server.  If both file size and I/O size are provided, the client SHOULD
1167  * reach or exceed  both thresholds before sending its read or write
1168  * requests to the data server.
1169  */
1170 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1171                                      struct inode *ino, int iomode)
1172 {
1173         struct nfs4_threshold *t = ctx->mdsthreshold;
1174         struct nfs_inode *nfsi = NFS_I(ino);
1175         loff_t fsize = i_size_read(ino);
1176         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1177
1178         if (t == NULL)
1179                 return ret;
1180
1181         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1182                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1183
1184         switch (iomode) {
1185         case IOMODE_READ:
1186                 if (t->bm & THRESHOLD_RD) {
1187                         dprintk("%s fsize %llu\n", __func__, fsize);
1188                         size_set = true;
1189                         if (fsize < t->rd_sz)
1190                                 size = true;
1191                 }
1192                 if (t->bm & THRESHOLD_RD_IO) {
1193                         dprintk("%s nfsi->read_io %llu\n", __func__,
1194                                 nfsi->read_io);
1195                         io_set = true;
1196                         if (nfsi->read_io < t->rd_io_sz)
1197                                 io = true;
1198                 }
1199                 break;
1200         case IOMODE_RW:
1201                 if (t->bm & THRESHOLD_WR) {
1202                         dprintk("%s fsize %llu\n", __func__, fsize);
1203                         size_set = true;
1204                         if (fsize < t->wr_sz)
1205                                 size = true;
1206                 }
1207                 if (t->bm & THRESHOLD_WR_IO) {
1208                         dprintk("%s nfsi->write_io %llu\n", __func__,
1209                                 nfsi->write_io);
1210                         io_set = true;
1211                         if (nfsi->write_io < t->wr_io_sz)
1212                                 io = true;
1213                 }
1214                 break;
1215         }
1216         if (size_set && io_set) {
1217                 if (size && io)
1218                         ret = true;
1219         } else if (size || io)
1220                 ret = true;
1221
1222         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1223         return ret;
1224 }
1225
1226 /*
1227  * Layout segment is retreived from the server if not cached.
1228  * The appropriate layout segment is referenced and returned to the caller.
1229  */
1230 struct pnfs_layout_segment *
1231 pnfs_update_layout(struct inode *ino,
1232                    struct nfs_open_context *ctx,
1233                    loff_t pos,
1234                    u64 count,
1235                    enum pnfs_iomode iomode,
1236                    gfp_t gfp_flags)
1237 {
1238         struct pnfs_layout_range arg = {
1239                 .iomode = iomode,
1240                 .offset = pos,
1241                 .length = count,
1242         };
1243         unsigned pg_offset;
1244         struct nfs_server *server = NFS_SERVER(ino);
1245         struct nfs_client *clp = server->nfs_client;
1246         struct pnfs_layout_hdr *lo;
1247         struct pnfs_layout_segment *lseg = NULL;
1248         bool first;
1249
1250         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1251                 goto out;
1252
1253         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1254                 goto out;
1255
1256         spin_lock(&ino->i_lock);
1257         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1258         if (lo == NULL) {
1259                 spin_unlock(&ino->i_lock);
1260                 goto out;
1261         }
1262
1263         /* Do we even need to bother with this? */
1264         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1265                 dprintk("%s matches recall, use MDS\n", __func__);
1266                 goto out_unlock;
1267         }
1268
1269         /* if LAYOUTGET already failed once we don't try again */
1270         if (pnfs_layout_io_test_failed(lo, iomode))
1271                 goto out_unlock;
1272
1273         /* Check to see if the layout for the given range already exists */
1274         lseg = pnfs_find_lseg(lo, &arg);
1275         if (lseg)
1276                 goto out_unlock;
1277
1278         if (pnfs_layoutgets_blocked(lo, 0))
1279                 goto out_unlock;
1280         atomic_inc(&lo->plh_outstanding);
1281
1282         first = list_empty(&lo->plh_layouts) ? true : false;
1283         spin_unlock(&ino->i_lock);
1284
1285         if (first) {
1286                 /* The lo must be on the clp list if there is any
1287                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1288                  */
1289                 spin_lock(&clp->cl_lock);
1290                 list_add_tail(&lo->plh_layouts, &server->layouts);
1291                 spin_unlock(&clp->cl_lock);
1292         }
1293
1294         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1295         if (pg_offset) {
1296                 arg.offset -= pg_offset;
1297                 arg.length += pg_offset;
1298         }
1299         if (arg.length != NFS4_MAX_UINT64)
1300                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1301
1302         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1303         atomic_dec(&lo->plh_outstanding);
1304 out_put_layout_hdr:
1305         pnfs_put_layout_hdr(lo);
1306 out:
1307         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1308                         "(%s, offset: %llu, length: %llu)\n",
1309                         __func__, ino->i_sb->s_id,
1310                         (unsigned long long)NFS_FILEID(ino),
1311                         lseg == NULL ? "not found" : "found",
1312                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1313                         (unsigned long long)pos,
1314                         (unsigned long long)count);
1315         return lseg;
1316 out_unlock:
1317         spin_unlock(&ino->i_lock);
1318         goto out_put_layout_hdr;
1319 }
1320 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1321
1322 struct pnfs_layout_segment *
1323 pnfs_layout_process(struct nfs4_layoutget *lgp)
1324 {
1325         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1326         struct nfs4_layoutget_res *res = &lgp->res;
1327         struct pnfs_layout_segment *lseg;
1328         struct inode *ino = lo->plh_inode;
1329         LIST_HEAD(free_me);
1330         int status = 0;
1331
1332         /* Inject layout blob into I/O device driver */
1333         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1334         if (!lseg || IS_ERR(lseg)) {
1335                 if (!lseg)
1336                         status = -ENOMEM;
1337                 else
1338                         status = PTR_ERR(lseg);
1339                 dprintk("%s: Could not allocate layout: error %d\n",
1340                        __func__, status);
1341                 goto out;
1342         }
1343
1344         spin_lock(&ino->i_lock);
1345         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1346                 dprintk("%s forget reply due to recall\n", __func__);
1347                 goto out_forget_reply;
1348         }
1349
1350         if (pnfs_layoutgets_blocked(lo, 1) ||
1351             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1352                 dprintk("%s forget reply due to state\n", __func__);
1353                 goto out_forget_reply;
1354         }
1355
1356         /* Check that the new stateid matches the old stateid */
1357         pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
1358         /* Done processing layoutget. Set the layout stateid */
1359         pnfs_set_layout_stateid(lo, &res->stateid, false);
1360
1361         init_lseg(lo, lseg);
1362         lseg->pls_range = res->range;
1363         pnfs_get_lseg(lseg);
1364         pnfs_layout_insert_lseg(lo, lseg);
1365
1366         if (res->return_on_close) {
1367                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1368                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1369         }
1370
1371         spin_unlock(&ino->i_lock);
1372         pnfs_free_lseg_list(&free_me);
1373         return lseg;
1374 out:
1375         return ERR_PTR(status);
1376
1377 out_forget_reply:
1378         spin_unlock(&ino->i_lock);
1379         lseg->pls_layout = lo;
1380         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1381         goto out;
1382 }
1383
1384 void
1385 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1386 {
1387         u64 rd_size = req->wb_bytes;
1388
1389         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1390
1391         if (pgio->pg_dreq == NULL)
1392                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1393         else
1394                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1395
1396         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1397                                            req->wb_context,
1398                                            req_offset(req),
1399                                            rd_size,
1400                                            IOMODE_READ,
1401                                            GFP_KERNEL);
1402         /* If no lseg, fall back to read through mds */
1403         if (pgio->pg_lseg == NULL)
1404                 nfs_pageio_reset_read_mds(pgio);
1405
1406 }
1407 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1408
1409 void
1410 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1411                            struct nfs_page *req, u64 wb_size)
1412 {
1413         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1414
1415         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1416                                            req->wb_context,
1417                                            req_offset(req),
1418                                            wb_size,
1419                                            IOMODE_RW,
1420                                            GFP_NOFS);
1421         /* If no lseg, fall back to write through mds */
1422         if (pgio->pg_lseg == NULL)
1423                 nfs_pageio_reset_write_mds(pgio);
1424 }
1425 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1426
1427 /*
1428  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1429  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1430  */
1431 size_t
1432 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1433                      struct nfs_page *req)
1434 {
1435         unsigned int size;
1436         u64 seg_end, req_start, seg_left;
1437
1438         size = nfs_generic_pg_test(pgio, prev, req);
1439         if (!size)
1440                 return 0;
1441
1442         /*
1443          * 'size' contains the number of bytes left in the current page (up
1444          * to the original size asked for in @req->wb_bytes).
1445          *
1446          * Calculate how many bytes are left in the layout segment
1447          * and if there are less bytes than 'size', return that instead.
1448          *
1449          * Please also note that 'end_offset' is actually the offset of the
1450          * first byte that lies outside the pnfs_layout_range. FIXME?
1451          *
1452          */
1453         if (pgio->pg_lseg) {
1454                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1455                                      pgio->pg_lseg->pls_range.length);
1456                 req_start = req_offset(req);
1457                 WARN_ON_ONCE(req_start > seg_end);
1458                 /* start of request is past the last byte of this segment */
1459                 if (req_start >= seg_end)
1460                         return 0;
1461
1462                 /* adjust 'size' iff there are fewer bytes left in the
1463                  * segment than what nfs_generic_pg_test returned */
1464                 seg_left = seg_end - req_start;
1465                 if (seg_left < size)
1466                         size = (unsigned int)seg_left;
1467         }
1468
1469         return size;
1470 }
1471 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1472
1473 int pnfs_write_done_resend_to_mds(struct inode *inode,
1474                                 struct list_head *head,
1475                                 const struct nfs_pgio_completion_ops *compl_ops,
1476                                 struct nfs_direct_req *dreq)
1477 {
1478         struct nfs_pageio_descriptor pgio;
1479         LIST_HEAD(failed);
1480
1481         /* Resend all requests through the MDS */
1482         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, true, compl_ops);
1483         pgio.pg_dreq = dreq;
1484         while (!list_empty(head)) {
1485                 struct nfs_page *req = nfs_list_entry(head->next);
1486
1487                 nfs_list_remove_request(req);
1488                 if (!nfs_pageio_add_request(&pgio, req))
1489                         nfs_list_add_request(req, &failed);
1490         }
1491         nfs_pageio_complete(&pgio);
1492
1493         if (!list_empty(&failed)) {
1494                 /* For some reason our attempt to resend pages. Mark the
1495                  * overall send request as having failed, and let
1496                  * nfs_writeback_release_full deal with the error.
1497                  */
1498                 list_move(&failed, head);
1499                 return -EIO;
1500         }
1501         return 0;
1502 }
1503 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1504
1505 static void pnfs_ld_handle_write_error(struct nfs_pgio_data *data)
1506 {
1507         struct nfs_pgio_header *hdr = data->header;
1508
1509         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1510         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1511             PNFS_LAYOUTRET_ON_ERROR) {
1512                 pnfs_return_layout(hdr->inode);
1513         }
1514         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1515                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1516                                                         &hdr->pages,
1517                                                         hdr->completion_ops,
1518                                                         hdr->dreq);
1519 }
1520
1521 /*
1522  * Called by non rpc-based layout drivers
1523  */
1524 void pnfs_ld_write_done(struct nfs_pgio_data *data)
1525 {
1526         struct nfs_pgio_header *hdr = data->header;
1527
1528         trace_nfs4_pnfs_write(data, hdr->pnfs_error);
1529         if (!hdr->pnfs_error) {
1530                 pnfs_set_layoutcommit(data);
1531                 hdr->mds_ops->rpc_call_done(&data->task, data);
1532         } else
1533                 pnfs_ld_handle_write_error(data);
1534         hdr->mds_ops->rpc_release(data);
1535 }
1536 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1537
1538 static void
1539 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1540                 struct nfs_pgio_data *data)
1541 {
1542         struct nfs_pgio_header *hdr = data->header;
1543
1544         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1545                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1546                 nfs_pageio_reset_write_mds(desc);
1547                 desc->pg_recoalesce = 1;
1548         }
1549         nfs_pgio_data_release(data);
1550 }
1551
1552 static enum pnfs_try_status
1553 pnfs_try_to_write_data(struct nfs_pgio_data *wdata,
1554                         const struct rpc_call_ops *call_ops,
1555                         struct pnfs_layout_segment *lseg,
1556                         int how)
1557 {
1558         struct nfs_pgio_header *hdr = wdata->header;
1559         struct inode *inode = hdr->inode;
1560         enum pnfs_try_status trypnfs;
1561         struct nfs_server *nfss = NFS_SERVER(inode);
1562
1563         hdr->mds_ops = call_ops;
1564
1565         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1566                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1567         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1568         if (trypnfs != PNFS_NOT_ATTEMPTED)
1569                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1570         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1571         return trypnfs;
1572 }
1573
1574 static void
1575 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1576               struct nfs_pgio_header *hdr, int how)
1577 {
1578         struct nfs_pgio_data *data = hdr->data;
1579         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1580         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1581         enum pnfs_try_status trypnfs;
1582
1583         desc->pg_lseg = NULL;
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         pnfs_put_lseg(lseg);
1588 }
1589
1590 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1591 {
1592         pnfs_put_lseg(hdr->lseg);
1593         nfs_rw_header_free(hdr);
1594 }
1595 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1596
1597 int
1598 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1599 {
1600         struct nfs_rw_header *whdr;
1601         struct nfs_pgio_header *hdr;
1602         int ret;
1603
1604         whdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1605         if (!whdr) {
1606                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1607                 pnfs_put_lseg(desc->pg_lseg);
1608                 desc->pg_lseg = NULL;
1609                 return -ENOMEM;
1610         }
1611         hdr = &whdr->header;
1612         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1613         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1614         atomic_inc(&hdr->refcnt);
1615         ret = nfs_generic_pgio(desc, hdr);
1616         if (ret != 0) {
1617                 pnfs_put_lseg(desc->pg_lseg);
1618                 desc->pg_lseg = NULL;
1619         } else
1620                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1621         if (atomic_dec_and_test(&hdr->refcnt))
1622                 hdr->completion_ops->completion(hdr);
1623         return ret;
1624 }
1625 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1626
1627 int pnfs_read_done_resend_to_mds(struct inode *inode,
1628                                 struct list_head *head,
1629                                 const struct nfs_pgio_completion_ops *compl_ops,
1630                                 struct nfs_direct_req *dreq)
1631 {
1632         struct nfs_pageio_descriptor pgio;
1633         LIST_HEAD(failed);
1634
1635         /* Resend all requests through the MDS */
1636         nfs_pageio_init_read(&pgio, inode, true, compl_ops);
1637         pgio.pg_dreq = dreq;
1638         while (!list_empty(head)) {
1639                 struct nfs_page *req = nfs_list_entry(head->next);
1640
1641                 nfs_list_remove_request(req);
1642                 if (!nfs_pageio_add_request(&pgio, req))
1643                         nfs_list_add_request(req, &failed);
1644         }
1645         nfs_pageio_complete(&pgio);
1646
1647         if (!list_empty(&failed)) {
1648                 list_move(&failed, head);
1649                 return -EIO;
1650         }
1651         return 0;
1652 }
1653 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1654
1655 static void pnfs_ld_handle_read_error(struct nfs_pgio_data *data)
1656 {
1657         struct nfs_pgio_header *hdr = data->header;
1658
1659         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1660         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1661             PNFS_LAYOUTRET_ON_ERROR) {
1662                 pnfs_return_layout(hdr->inode);
1663         }
1664         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1665                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1666                                                         &hdr->pages,
1667                                                         hdr->completion_ops,
1668                                                         hdr->dreq);
1669 }
1670
1671 /*
1672  * Called by non rpc-based layout drivers
1673  */
1674 void pnfs_ld_read_done(struct nfs_pgio_data *data)
1675 {
1676         struct nfs_pgio_header *hdr = data->header;
1677
1678         trace_nfs4_pnfs_read(data, hdr->pnfs_error);
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_pgio_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_pgio_data_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_pgio_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_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1729 {
1730         struct nfs_pgio_data *data = hdr->data;
1731         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1732         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1733         enum pnfs_try_status trypnfs;
1734
1735         desc->pg_lseg = NULL;
1736         trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1737         if (trypnfs == PNFS_NOT_ATTEMPTED)
1738                 pnfs_read_through_mds(desc, data);
1739         pnfs_put_lseg(lseg);
1740 }
1741
1742 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1743 {
1744         pnfs_put_lseg(hdr->lseg);
1745         nfs_rw_header_free(hdr);
1746 }
1747 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1748
1749 int
1750 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1751 {
1752         struct nfs_rw_header *rhdr;
1753         struct nfs_pgio_header *hdr;
1754         int ret;
1755
1756         rhdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1757         if (!rhdr) {
1758                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1759                 ret = -ENOMEM;
1760                 pnfs_put_lseg(desc->pg_lseg);
1761                 desc->pg_lseg = NULL;
1762                 return ret;
1763         }
1764         hdr = &rhdr->header;
1765         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1766         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1767         atomic_inc(&hdr->refcnt);
1768         ret = nfs_generic_pgio(desc, hdr);
1769         if (ret != 0) {
1770                 pnfs_put_lseg(desc->pg_lseg);
1771                 desc->pg_lseg = NULL;
1772         } else
1773                 pnfs_do_read(desc, hdr);
1774         if (atomic_dec_and_test(&hdr->refcnt))
1775                 hdr->completion_ops->completion(hdr);
1776         return ret;
1777 }
1778 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1779
1780 static void pnfs_clear_layoutcommitting(struct inode *inode)
1781 {
1782         unsigned long *bitlock = &NFS_I(inode)->flags;
1783
1784         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1785         smp_mb__after_atomic();
1786         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1787 }
1788
1789 /*
1790  * There can be multiple RW segments.
1791  */
1792 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1793 {
1794         struct pnfs_layout_segment *lseg;
1795
1796         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1797                 if (lseg->pls_range.iomode == IOMODE_RW &&
1798                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1799                         list_add(&lseg->pls_lc_list, listp);
1800         }
1801 }
1802
1803 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1804 {
1805         struct pnfs_layout_segment *lseg, *tmp;
1806
1807         /* Matched by references in pnfs_set_layoutcommit */
1808         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1809                 list_del_init(&lseg->pls_lc_list);
1810                 pnfs_put_lseg(lseg);
1811         }
1812
1813         pnfs_clear_layoutcommitting(inode);
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_pgio_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;
1878
1879         if (!pnfs_layoutcommit_outstanding(inode))
1880                 return 0;
1881
1882         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1883
1884         status = -EAGAIN;
1885         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1886                 if (!sync)
1887                         goto out;
1888                 status = wait_on_bit_lock_action(&nfsi->flags,
1889                                 NFS_INO_LAYOUTCOMMITTING,
1890                                 nfs_wait_bit_killable,
1891                                 TASK_KILLABLE);
1892                 if (status)
1893                         goto out;
1894         }
1895
1896         status = -ENOMEM;
1897         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1898         data = kzalloc(sizeof(*data), GFP_NOFS);
1899         if (!data)
1900                 goto clear_layoutcommitting;
1901
1902         status = 0;
1903         spin_lock(&inode->i_lock);
1904         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1905                 goto out_unlock;
1906
1907         INIT_LIST_HEAD(&data->lseg_list);
1908         pnfs_list_write_lseg(inode, &data->lseg_list);
1909
1910         end_pos = nfsi->layout->plh_lwb;
1911         nfsi->layout->plh_lwb = 0;
1912
1913         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1914         spin_unlock(&inode->i_lock);
1915
1916         data->args.inode = inode;
1917         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1918         nfs_fattr_init(&data->fattr);
1919         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1920         data->res.fattr = &data->fattr;
1921         data->args.lastbytewritten = end_pos - 1;
1922         data->res.server = NFS_SERVER(inode);
1923
1924         status = nfs4_proc_layoutcommit(data, sync);
1925 out:
1926         if (status)
1927                 mark_inode_dirty_sync(inode);
1928         dprintk("<-- %s status %d\n", __func__, status);
1929         return status;
1930 out_unlock:
1931         spin_unlock(&inode->i_lock);
1932         kfree(data);
1933 clear_layoutcommitting:
1934         pnfs_clear_layoutcommitting(inode);
1935         goto out;
1936 }
1937
1938 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1939 {
1940         struct nfs4_threshold *thp;
1941
1942         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1943         if (!thp) {
1944                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1945                 return NULL;
1946         }
1947         return thp;
1948 }