Pileus Git - ~andy/linux/blob - 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 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 bool
 380 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 381                  const 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 bool
 399 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
 400                     const 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(const struct pnfs_layout_range *lseg_range,
 413                  const struct pnfs_layout_range *recall_range)
 414 {
 415         return (recall_range->iomode == IOMODE_ANY ||
 416                 lseg_range->iomode == recall_range->iomode) &&
 417                pnfs_lseg_range_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 (!nfs4_valid_open_stateid(open_state)) {
 722                 status = -EBADF;
 723         } else if (list_empty(&lo->plh_segs)) {
 724                 int seq;
 725
 726                 do {
 727                         seq = read_seqbegin(&open_state->seqlock);
 728                         nfs4_stateid_copy(dst, &open_state->stateid);
 729                 } while (read_seqretry(&open_state->seqlock, seq));
 730         } else
 731                 nfs4_stateid_copy(dst, &lo->plh_stateid);
 732         spin_unlock(&lo->plh_inode->i_lock);
 733         dprintk("<-- %s\n", __func__);
 734         return status;
 735 }
 736
 737 /*
 738 * Get layout from server.
 739 *    for now, assume that whole file layouts are requested.
 740 *    arg->offset: 0
 741 *    arg->length: all ones
 742 */
 743 static struct pnfs_layout_segment *
 744 send_layoutget(struct pnfs_layout_hdr *lo,
 745            struct nfs_open_context *ctx,
 746            struct pnfs_layout_range *range,
 747            gfp_t gfp_flags)
 748 {
 749         struct inode *ino = lo->plh_inode;
 750         struct nfs_server *server = NFS_SERVER(ino);
 751         struct nfs4_layoutget *lgp;
 752         struct pnfs_layout_segment *lseg;
 753
 754         dprintk("--> %s\n", __func__);
 755
 756         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 757         if (lgp == NULL)
 758                 return NULL;
 759
 760         lgp->args.minlength = PAGE_CACHE_SIZE;
 761         if (lgp->args.minlength > range->length)
 762                 lgp->args.minlength = range->length;
 763         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 764         lgp->args.range = *range;
 765         lgp->args.type = server->pnfs_curr_ld->id;
 766         lgp->args.inode = ino;
 767         lgp->args.ctx = get_nfs_open_context(ctx);
 768         lgp->gfp_flags = gfp_flags;
 769         lgp->cred = lo->plh_lc_cred;
 770
 771         /* Synchronously retrieve layout information from server and
 772          * store in lseg.
 773          */
 774         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
 775         if (IS_ERR(lseg)) {
 776                 switch (PTR_ERR(lseg)) {
 777                 case -ENOMEM:
 778                 case -ERESTARTSYS:
 779                         break;
 780                 default:
 781                         /* remember that LAYOUTGET failed and suspend trying */
 782                         pnfs_layout_io_set_failed(lo, range->iomode);
 783                 }
 784                 return NULL;
 785         }
 786
 787         return lseg;
 788 }
 789
 790 static void pnfs_clear_layoutcommit(struct inode *inode,
 791                 struct list_head *head)
 792 {
 793         struct nfs_inode *nfsi = NFS_I(inode);
 794         struct pnfs_layout_segment *lseg, *tmp;
 795
 796         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 797                 return;
 798         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 799                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 800                         continue;
 801                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 802         }
 803 }
 804
 805 /*
 806  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
 807  * when the layout segment list is empty.
 808  *
 809  * Note that a pnfs_layout_hdr can exist with an empty layout segment
 810  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
 811  * deviceid is marked invalid.
 812  */
 813 int
 814 _pnfs_return_layout(struct inode *ino)
 815 {
 816         struct pnfs_layout_hdr *lo = NULL;
 817         struct nfs_inode *nfsi = NFS_I(ino);
 818         LIST_HEAD(tmp_list);
 819         struct nfs4_layoutreturn *lrp;
 820         nfs4_stateid stateid;
 821         int status = 0, empty;
 822
 823         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
 824
 825         spin_lock(&ino->i_lock);
 826         lo = nfsi->layout;
 827         if (!lo) {
 828                 spin_unlock(&ino->i_lock);
 829                 dprintk("NFS: %s no layout to return\n", __func__);
 830                 goto out;
 831         }
 832         stateid = nfsi->layout->plh_stateid;
 833         /* Reference matched in nfs4_layoutreturn_release */
 834         pnfs_get_layout_hdr(lo);
 835         empty = list_empty(&lo->plh_segs);
 836         pnfs_clear_layoutcommit(ino, &tmp_list);
 837         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
 838         /* Don't send a LAYOUTRETURN if list was initially empty */
 839         if (empty) {
 840                 spin_unlock(&ino->i_lock);
 841                 pnfs_put_layout_hdr(lo);
 842                 dprintk("NFS: %s no layout segments to return\n", __func__);
 843                 goto out;
 844         }
 845         lo->plh_block_lgets++;
 846         spin_unlock(&ino->i_lock);
 847         pnfs_free_lseg_list(&tmp_list);
 848
 849         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
 850         if (unlikely(lrp == NULL)) {
 851                 status = -ENOMEM;
 852                 spin_lock(&ino->i_lock);
 853                 lo->plh_block_lgets--;
 854                 spin_unlock(&ino->i_lock);
 855                 pnfs_put_layout_hdr(lo);
 856                 goto out;
 857         }
 858
 859         lrp->args.stateid = stateid;
 860         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
 861         lrp->args.inode = ino;
 862         lrp->args.layout = lo;
 863         lrp->clp = NFS_SERVER(ino)->nfs_client;
 864         lrp->cred = lo->plh_lc_cred;
 865
 866         status = nfs4_proc_layoutreturn(lrp);
 867 out:
 868         dprintk("<-- %s status: %d\n", __func__, status);
 869         return status;
 870 }
 871 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
 872
 873 int
 874 pnfs_commit_and_return_layout(struct inode *inode)
 875 {
 876         struct pnfs_layout_hdr *lo;
 877         int ret;
 878
 879         spin_lock(&inode->i_lock);
 880         lo = NFS_I(inode)->layout;
 881         if (lo == NULL) {
 882                 spin_unlock(&inode->i_lock);
 883                 return 0;
 884         }
 885         pnfs_get_layout_hdr(lo);
 886         /* Block new layoutgets and read/write to ds */
 887         lo->plh_block_lgets++;
 888         spin_unlock(&inode->i_lock);
 889         filemap_fdatawait(inode->i_mapping);
 890         ret = pnfs_layoutcommit_inode(inode, true);
 891         if (ret == 0)
 892                 ret = _pnfs_return_layout(inode);
 893         spin_lock(&inode->i_lock);
 894         lo->plh_block_lgets--;
 895         spin_unlock(&inode->i_lock);
 896         pnfs_put_layout_hdr(lo);
 897         return ret;
 898 }
 899
 900 bool pnfs_roc(struct inode *ino)
 901 {
 902         struct pnfs_layout_hdr *lo;
 903         struct pnfs_layout_segment *lseg, *tmp;
 904         LIST_HEAD(tmp_list);
 905         bool found = false;
 906
 907         spin_lock(&ino->i_lock);
 908         lo = NFS_I(ino)->layout;
 909         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
 910             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
 911                 goto out_nolayout;
 912         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
 913                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 914                         mark_lseg_invalid(lseg, &tmp_list);
 915                         found = true;
 916                 }
 917         if (!found)
 918                 goto out_nolayout;
 919         lo->plh_block_lgets++;
 920         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
 921         spin_unlock(&ino->i_lock);
 922         pnfs_free_lseg_list(&tmp_list);
 923         return true;
 924
 925 out_nolayout:
 926         spin_unlock(&ino->i_lock);
 927         return false;
 928 }
 929
 930 void pnfs_roc_release(struct inode *ino)
 931 {
 932         struct pnfs_layout_hdr *lo;
 933
 934         spin_lock(&ino->i_lock);
 935         lo = NFS_I(ino)->layout;
 936         lo->plh_block_lgets--;
 937         if (atomic_dec_and_test(&lo->plh_refcount)) {
 938                 pnfs_detach_layout_hdr(lo);
 939                 spin_unlock(&ino->i_lock);
 940                 pnfs_free_layout_hdr(lo);
 941         } else
 942                 spin_unlock(&ino->i_lock);
 943 }
 944
 945 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
 946 {
 947         struct pnfs_layout_hdr *lo;
 948
 949         spin_lock(&ino->i_lock);
 950         lo = NFS_I(ino)->layout;
 951         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
 952                 lo->plh_barrier = barrier;
 953         spin_unlock(&ino->i_lock);
 954 }
 955
 956 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
 957 {
 958         struct nfs_inode *nfsi = NFS_I(ino);
 959         struct pnfs_layout_hdr *lo;
 960         struct pnfs_layout_segment *lseg;
 961         u32 current_seqid;
 962         bool found = false;
 963
 964         spin_lock(&ino->i_lock);
 965         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
 966                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
 967                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
 968                         found = true;
 969                         goto out;
 970                 }
 971         lo = nfsi->layout;
 972         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
 973
 974         /* Since close does not return a layout stateid for use as
 975          * a barrier, we choose the worst-case barrier.
 976          */
 977         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
 978 out:
 979         spin_unlock(&ino->i_lock);
 980         return found;
 981 }
 982
 983 /*
 984  * Compare two layout segments for sorting into layout cache.
 985  * We want to preferentially return RW over RO layouts, so ensure those
 986  * are seen first.
 987  */
 988 static s64
 989 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
 990            const struct pnfs_layout_range *l2)
 991 {
 992         s64 d;
 993
 994         /* high offset > low offset */
 995         d = l1->offset - l2->offset;
 996         if (d)
 997                 return d;
 998
 999         /* short length > long length */
1000         d = l2->length - l1->length;
1001         if (d)
1002                 return d;
1003
1004         /* read > read/write */
1005         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1006 }
1007
1008 static void
1009 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1010                    struct pnfs_layout_segment *lseg)
1011 {
1012         struct pnfs_layout_segment *lp;
1013
1014         dprintk("%s:Begin\n", __func__);
1015
1016         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1017                 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1018                         continue;
1019                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1020                 dprintk("%s: inserted lseg %p "
1021                         "iomode %d offset %llu length %llu before "
1022                         "lp %p iomode %d offset %llu length %llu\n",
1023                         __func__, lseg, lseg->pls_range.iomode,
1024                         lseg->pls_range.offset, lseg->pls_range.length,
1025                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1026                         lp->pls_range.length);
1027                 goto out;
1028         }
1029         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1030         dprintk("%s: inserted lseg %p "
1031                 "iomode %d offset %llu length %llu at tail\n",
1032                 __func__, lseg, lseg->pls_range.iomode,
1033                 lseg->pls_range.offset, lseg->pls_range.length);
1034 out:
1035         pnfs_get_layout_hdr(lo);
1036
1037         dprintk("%s:Return\n", __func__);
1038 }
1039
1040 static struct pnfs_layout_hdr *
1041 alloc_init_layout_hdr(struct inode *ino,
1042                       struct nfs_open_context *ctx,
1043                       gfp_t gfp_flags)
1044 {
1045         struct pnfs_layout_hdr *lo;
1046
1047         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1048         if (!lo)
1049                 return NULL;
1050         atomic_set(&lo->plh_refcount, 1);
1051         INIT_LIST_HEAD(&lo->plh_layouts);
1052         INIT_LIST_HEAD(&lo->plh_segs);
1053         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1054         lo->plh_inode = ino;
1055         lo->plh_lc_cred = get_rpccred(ctx->cred);
1056         return lo;
1057 }
1058
1059 static struct pnfs_layout_hdr *
1060 pnfs_find_alloc_layout(struct inode *ino,
1061                        struct nfs_open_context *ctx,
1062                        gfp_t gfp_flags)
1063 {
1064         struct nfs_inode *nfsi = NFS_I(ino);
1065         struct pnfs_layout_hdr *new = NULL;
1066
1067         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1068
1069         if (nfsi->layout != NULL)
1070                 goto out_existing;
1071         spin_unlock(&ino->i_lock);
1072         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1073         spin_lock(&ino->i_lock);
1074
1075         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1076                 nfsi->layout = new;
1077                 return new;
1078         } else if (new != NULL)
1079                 pnfs_free_layout_hdr(new);
1080 out_existing:
1081         pnfs_get_layout_hdr(nfsi->layout);
1082         return nfsi->layout;
1083 }
1084
1085 /*
1086  * iomode matching rules:
1087  * iomode       lseg    match
1088  * -----        -----   -----
1089  * ANY          READ    true
1090  * ANY          RW      true
1091  * RW           READ    false
1092  * RW           RW      true
1093  * READ         READ    true
1094  * READ         RW      true
1095  */
1096 static bool
1097 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1098                  const struct pnfs_layout_range *range)
1099 {
1100         struct pnfs_layout_range range1;
1101
1102         if ((range->iomode == IOMODE_RW &&
1103              ls_range->iomode != IOMODE_RW) ||
1104             !pnfs_lseg_range_intersecting(ls_range, range))
1105                 return 0;
1106
1107         /* range1 covers only the first byte in the range */
1108         range1 = *range;
1109         range1.length = 1;
1110         return pnfs_lseg_range_contained(ls_range, &range1);
1111 }
1112
1113 /*
1114  * lookup range in layout
1115  */
1116 static struct pnfs_layout_segment *
1117 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1118                 struct pnfs_layout_range *range)
1119 {
1120         struct pnfs_layout_segment *lseg, *ret = NULL;
1121
1122         dprintk("%s:Begin\n", __func__);
1123
1124         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1125                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1126                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1127                         ret = pnfs_get_lseg(lseg);
1128                         break;
1129                 }
1130                 if (lseg->pls_range.offset > range->offset)
1131                         break;
1132         }
1133
1134         dprintk("%s:Return lseg %p ref %d\n",
1135                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1136         return ret;
1137 }
1138
1139 /*
1140  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1141  * to the MDS or over pNFS
1142  *
1143  * The nfs_inode read_io and write_io fields are cumulative counters reset
1144  * when there are no layout segments. Note that in pnfs_update_layout iomode
1145  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1146  * WRITE request.
1147  *
1148  * A return of true means use MDS I/O.
1149  *
1150  * From rfc 5661:
1151  * If a file's size is smaller than the file size threshold, data accesses
1152  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1153  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1154  * server.  If both file size and I/O size are provided, the client SHOULD
1155  * reach or exceed  both thresholds before sending its read or write
1156  * requests to the data server.
1157  */
1158 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1159                                      struct inode *ino, int iomode)
1160 {
1161         struct nfs4_threshold *t = ctx->mdsthreshold;
1162         struct nfs_inode *nfsi = NFS_I(ino);
1163         loff_t fsize = i_size_read(ino);
1164         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1165
1166         if (t == NULL)
1167                 return ret;
1168
1169         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1170                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1171
1172         switch (iomode) {
1173         case IOMODE_READ:
1174                 if (t->bm & THRESHOLD_RD) {
1175                         dprintk("%s fsize %llu\n", __func__, fsize);
1176                         size_set = true;
1177                         if (fsize < t->rd_sz)
1178                                 size = true;
1179                 }
1180                 if (t->bm & THRESHOLD_RD_IO) {
1181                         dprintk("%s nfsi->read_io %llu\n", __func__,
1182                                 nfsi->read_io);
1183                         io_set = true;
1184                         if (nfsi->read_io < t->rd_io_sz)
1185                                 io = true;
1186                 }
1187                 break;
1188         case IOMODE_RW:
1189                 if (t->bm & THRESHOLD_WR) {
1190                         dprintk("%s fsize %llu\n", __func__, fsize);
1191                         size_set = true;
1192                         if (fsize < t->wr_sz)
1193                                 size = true;
1194                 }
1195                 if (t->bm & THRESHOLD_WR_IO) {
1196                         dprintk("%s nfsi->write_io %llu\n", __func__,
1197                                 nfsi->write_io);
1198                         io_set = true;
1199                         if (nfsi->write_io < t->wr_io_sz)
1200                                 io = true;
1201                 }
1202                 break;
1203         }
1204         if (size_set && io_set) {
1205                 if (size && io)
1206                         ret = true;
1207         } else if (size || io)
1208                 ret = true;
1209
1210         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1211         return ret;
1212 }
1213
1214 /*
1215  * Layout segment is retreived from the server if not cached.
1216  * The appropriate layout segment is referenced and returned to the caller.
1217  */
1218 struct pnfs_layout_segment *
1219 pnfs_update_layout(struct inode *ino,
1220                    struct nfs_open_context *ctx,
1221                    loff_t pos,
1222                    u64 count,
1223                    enum pnfs_iomode iomode,
1224                    gfp_t gfp_flags)
1225 {
1226         struct pnfs_layout_range arg = {
1227                 .iomode = iomode,
1228                 .offset = pos,
1229                 .length = count,
1230         };
1231         unsigned pg_offset;
1232         struct nfs_server *server = NFS_SERVER(ino);
1233         struct nfs_client *clp = server->nfs_client;
1234         struct pnfs_layout_hdr *lo;
1235         struct pnfs_layout_segment *lseg = NULL;
1236         bool first;
1237
1238         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1239                 goto out;
1240
1241         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1242                 goto out;
1243
1244         spin_lock(&ino->i_lock);
1245         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1246         if (lo == NULL) {
1247                 spin_unlock(&ino->i_lock);
1248                 goto out;
1249         }
1250
1251         /* Do we even need to bother with this? */
1252         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1253                 dprintk("%s matches recall, use MDS\n", __func__);
1254                 goto out_unlock;
1255         }
1256
1257         /* if LAYOUTGET already failed once we don't try again */
1258         if (pnfs_layout_io_test_failed(lo, iomode))
1259                 goto out_unlock;
1260
1261         /* Check to see if the layout for the given range already exists */
1262         lseg = pnfs_find_lseg(lo, &arg);
1263         if (lseg)
1264                 goto out_unlock;
1265
1266         if (pnfs_layoutgets_blocked(lo, 0))
1267                 goto out_unlock;
1268         atomic_inc(&lo->plh_outstanding);
1269
1270         first = list_empty(&lo->plh_layouts) ? true : false;
1271         spin_unlock(&ino->i_lock);
1272
1273         if (first) {
1274                 /* The lo must be on the clp list if there is any
1275                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1276                  */
1277                 spin_lock(&clp->cl_lock);
1278                 list_add_tail(&lo->plh_layouts, &server->layouts);
1279                 spin_unlock(&clp->cl_lock);
1280         }
1281
1282         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1283         if (pg_offset) {
1284                 arg.offset -= pg_offset;
1285                 arg.length += pg_offset;
1286         }
1287         if (arg.length != NFS4_MAX_UINT64)
1288                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1289
1290         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1291         atomic_dec(&lo->plh_outstanding);
1292 out_put_layout_hdr:
1293         pnfs_put_layout_hdr(lo);
1294 out:
1295         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1296                         "(%s, offset: %llu, length: %llu)\n",
1297                         __func__, ino->i_sb->s_id,
1298                         (unsigned long long)NFS_FILEID(ino),
1299                         lseg == NULL ? "not found" : "found",
1300                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1301                         (unsigned long long)pos,
1302                         (unsigned long long)count);
1303         return lseg;
1304 out_unlock:
1305         spin_unlock(&ino->i_lock);
1306         goto out_put_layout_hdr;
1307 }
1308 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1309
1310 struct pnfs_layout_segment *
1311 pnfs_layout_process(struct nfs4_layoutget *lgp)
1312 {
1313         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1314         struct nfs4_layoutget_res *res = &lgp->res;
1315         struct pnfs_layout_segment *lseg;
1316         struct inode *ino = lo->plh_inode;
1317         int status = 0;
1318
1319         /* Inject layout blob into I/O device driver */
1320         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1321         if (!lseg || IS_ERR(lseg)) {
1322                 if (!lseg)
1323                         status = -ENOMEM;
1324                 else
1325                         status = PTR_ERR(lseg);
1326                 dprintk("%s: Could not allocate layout: error %d\n",
1327                        __func__, status);
1328                 goto out;
1329         }
1330
1331         spin_lock(&ino->i_lock);
1332         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1333                 dprintk("%s forget reply due to recall\n", __func__);
1334                 goto out_forget_reply;
1335         }
1336
1337         if (pnfs_layoutgets_blocked(lo, 1) ||
1338             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1339                 dprintk("%s forget reply due to state\n", __func__);
1340                 goto out_forget_reply;
1341         }
1342
1343         /* Done processing layoutget. Set the layout stateid */
1344         pnfs_set_layout_stateid(lo, &res->stateid, false);
1345
1346         init_lseg(lo, lseg);
1347         lseg->pls_range = res->range;
1348         pnfs_get_lseg(lseg);
1349         pnfs_layout_insert_lseg(lo, lseg);
1350
1351         if (res->return_on_close) {
1352                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1353                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1354         }
1355
1356         spin_unlock(&ino->i_lock);
1357         return lseg;
1358 out:
1359         return ERR_PTR(status);
1360
1361 out_forget_reply:
1362         spin_unlock(&ino->i_lock);
1363         lseg->pls_layout = lo;
1364         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1365         goto out;
1366 }
1367
1368 void
1369 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1370 {
1371         u64 rd_size = req->wb_bytes;
1372
1373         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1374
1375         if (req->wb_offset != req->wb_pgbase) {
1376                 nfs_pageio_reset_read_mds(pgio);
1377                 return;
1378         }
1379
1380         if (pgio->pg_dreq == NULL)
1381                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1382         else
1383                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1384
1385         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1386                                            req->wb_context,
1387                                            req_offset(req),
1388                                            rd_size,
1389                                            IOMODE_READ,
1390                                            GFP_KERNEL);
1391         /* If no lseg, fall back to read through mds */
1392         if (pgio->pg_lseg == NULL)
1393                 nfs_pageio_reset_read_mds(pgio);
1394
1395 }
1396 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1397
1398 void
1399 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1400                            struct nfs_page *req, u64 wb_size)
1401 {
1402         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1403
1404         if (req->wb_offset != req->wb_pgbase) {
1405                 nfs_pageio_reset_write_mds(pgio);
1406                 return;
1407         }
1408
1409         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1410                                            req->wb_context,
1411                                            req_offset(req),
1412                                            wb_size,
1413                                            IOMODE_RW,
1414                                            GFP_NOFS);
1415         /* If no lseg, fall back to write through mds */
1416         if (pgio->pg_lseg == NULL)
1417                 nfs_pageio_reset_write_mds(pgio);
1418 }
1419 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1420
1421 void
1422 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1423                       const struct nfs_pgio_completion_ops *compl_ops)
1424 {
1425         struct nfs_server *server = NFS_SERVER(inode);
1426         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1427
1428         if (ld == NULL)
1429                 nfs_pageio_init_read(pgio, inode, compl_ops);
1430         else
1431                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1432 }
1433
1434 void
1435 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1436                        int ioflags,
1437                        const struct nfs_pgio_completion_ops *compl_ops)
1438 {
1439         struct nfs_server *server = NFS_SERVER(inode);
1440         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1441
1442         if (ld == NULL)
1443                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1444         else
1445                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1446 }
1447
1448 bool
1449 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1450                      struct nfs_page *req)
1451 {
1452         if (pgio->pg_lseg == NULL)
1453                 return nfs_generic_pg_test(pgio, prev, req);
1454
1455         /*
1456          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1457          * Note that this test makes several assumptions:
1458          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1459          *   is known to lie within the range.
1460          *   - that the nfs_page being tested is known to be contiguous with the
1461          *   previous nfs_page.
1462          *   - Layout ranges are page aligned, so we only have to test the
1463          *   start offset of the request.
1464          *
1465          * Please also note that 'end_offset' is actually the offset of the
1466          * first byte that lies outside the pnfs_layout_range. FIXME?
1467          *
1468          */
1469         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1470                                          pgio->pg_lseg->pls_range.length);
1471 }
1472 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1473
1474 int pnfs_write_done_resend_to_mds(struct inode *inode,
1475                                 struct list_head *head,
1476                                 const struct nfs_pgio_completion_ops *compl_ops,
1477                                 struct nfs_direct_req *dreq)
1478 {
1479         struct nfs_pageio_descriptor pgio;
1480         LIST_HEAD(failed);
1481
1482         /* Resend all requests through the MDS */
1483         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1484         pgio.pg_dreq = dreq;
1485         while (!list_empty(head)) {
1486                 struct nfs_page *req = nfs_list_entry(head->next);
1487
1488                 nfs_list_remove_request(req);
1489                 if (!nfs_pageio_add_request(&pgio, req))
1490                         nfs_list_add_request(req, &failed);
1491         }
1492         nfs_pageio_complete(&pgio);
1493
1494         if (!list_empty(&failed)) {
1495                 /* For some reason our attempt to resend pages. Mark the
1496                  * overall send request as having failed, and let
1497                  * nfs_writeback_release_full deal with the error.
1498                  */
1499                 list_move(&failed, head);
1500                 return -EIO;
1501         }
1502         return 0;
1503 }
1504 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1505
1506 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1507 {
1508         struct nfs_pgio_header *hdr = data->header;
1509
1510         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1511         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1512             PNFS_LAYOUTRET_ON_ERROR) {
1513                 pnfs_return_layout(hdr->inode);
1514         }
1515         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1516                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1517                                                         &hdr->pages,
1518                                                         hdr->completion_ops,
1519                                                         hdr->dreq);
1520 }
1521
1522 /*
1523  * Called by non rpc-based layout drivers
1524  */
1525 void pnfs_ld_write_done(struct nfs_write_data *data)
1526 {
1527         struct nfs_pgio_header *hdr = data->header;
1528
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_write_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_writedata_release(data);
1550 }
1551
1552 static enum pnfs_try_status
1553 pnfs_try_to_write_data(struct nfs_write_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_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1576 {
1577         struct nfs_write_data *data;
1578         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1579         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1580
1581         desc->pg_lseg = NULL;
1582         while (!list_empty(head)) {
1583                 enum pnfs_try_status trypnfs;
1584
1585                 data = list_first_entry(head, struct nfs_write_data, list);
1586                 list_del_init(&data->list);
1587
1588                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1589                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1590                         pnfs_write_through_mds(desc, data);
1591         }
1592         pnfs_put_lseg(lseg);
1593 }
1594
1595 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1596 {
1597         pnfs_put_lseg(hdr->lseg);
1598         nfs_writehdr_free(hdr);
1599 }
1600 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1601
1602 int
1603 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1604 {
1605         struct nfs_write_header *whdr;
1606         struct nfs_pgio_header *hdr;
1607         int ret;
1608
1609         whdr = nfs_writehdr_alloc();
1610         if (!whdr) {
1611                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1612                 pnfs_put_lseg(desc->pg_lseg);
1613                 desc->pg_lseg = NULL;
1614                 return -ENOMEM;
1615         }
1616         hdr = &whdr->header;
1617         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1618         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1619         atomic_inc(&hdr->refcnt);
1620         ret = nfs_generic_flush(desc, hdr);
1621         if (ret != 0) {
1622                 pnfs_put_lseg(desc->pg_lseg);
1623                 desc->pg_lseg = NULL;
1624         } else
1625                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1626         if (atomic_dec_and_test(&hdr->refcnt))
1627                 hdr->completion_ops->completion(hdr);
1628         return ret;
1629 }
1630 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1631
1632 int pnfs_read_done_resend_to_mds(struct inode *inode,
1633                                 struct list_head *head,
1634                                 const struct nfs_pgio_completion_ops *compl_ops,
1635                                 struct nfs_direct_req *dreq)
1636 {
1637         struct nfs_pageio_descriptor pgio;
1638         LIST_HEAD(failed);
1639
1640         /* Resend all requests through the MDS */
1641         nfs_pageio_init_read(&pgio, inode, compl_ops);
1642         pgio.pg_dreq = dreq;
1643         while (!list_empty(head)) {
1644                 struct nfs_page *req = nfs_list_entry(head->next);
1645
1646                 nfs_list_remove_request(req);
1647                 if (!nfs_pageio_add_request(&pgio, req))
1648                         nfs_list_add_request(req, &failed);
1649         }
1650         nfs_pageio_complete(&pgio);
1651
1652         if (!list_empty(&failed)) {
1653                 list_move(&failed, head);
1654                 return -EIO;
1655         }
1656         return 0;
1657 }
1658 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1659
1660 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1661 {
1662         struct nfs_pgio_header *hdr = data->header;
1663
1664         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1665         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1666             PNFS_LAYOUTRET_ON_ERROR) {
1667                 pnfs_return_layout(hdr->inode);
1668         }
1669         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1670                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1671                                                         &hdr->pages,
1672                                                         hdr->completion_ops,
1673                                                         hdr->dreq);
1674 }
1675
1676 /*
1677  * Called by non rpc-based layout drivers
1678  */
1679 void pnfs_ld_read_done(struct nfs_read_data *data)
1680 {
1681         struct nfs_pgio_header *hdr = data->header;
1682
1683         if (likely(!hdr->pnfs_error)) {
1684                 __nfs4_read_done_cb(data);
1685                 hdr->mds_ops->rpc_call_done(&data->task, data);
1686         } else
1687                 pnfs_ld_handle_read_error(data);
1688         hdr->mds_ops->rpc_release(data);
1689 }
1690 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1691
1692 static void
1693 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1694                 struct nfs_read_data *data)
1695 {
1696         struct nfs_pgio_header *hdr = data->header;
1697
1698         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1699                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1700                 nfs_pageio_reset_read_mds(desc);
1701                 desc->pg_recoalesce = 1;
1702         }
1703         nfs_readdata_release(data);
1704 }
1705
1706 /*
1707  * Call the appropriate parallel I/O subsystem read function.
1708  */
1709 static enum pnfs_try_status
1710 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1711                        const struct rpc_call_ops *call_ops,
1712                        struct pnfs_layout_segment *lseg)
1713 {
1714         struct nfs_pgio_header *hdr = rdata->header;
1715         struct inode *inode = hdr->inode;
1716         struct nfs_server *nfss = NFS_SERVER(inode);
1717         enum pnfs_try_status trypnfs;
1718
1719         hdr->mds_ops = call_ops;
1720
1721         dprintk("%s: Reading ino:%lu %u@%llu\n",
1722                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1723
1724         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1725         if (trypnfs != PNFS_NOT_ATTEMPTED)
1726                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1727         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1728         return trypnfs;
1729 }
1730
1731 static void
1732 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1733 {
1734         struct nfs_read_data *data;
1735         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1736         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1737
1738         desc->pg_lseg = NULL;
1739         while (!list_empty(head)) {
1740                 enum pnfs_try_status trypnfs;
1741
1742                 data = list_first_entry(head, struct nfs_read_data, list);
1743                 list_del_init(&data->list);
1744
1745                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1746                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1747                         pnfs_read_through_mds(desc, data);
1748         }
1749         pnfs_put_lseg(lseg);
1750 }
1751
1752 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1753 {
1754         pnfs_put_lseg(hdr->lseg);
1755         nfs_readhdr_free(hdr);
1756 }
1757 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1758
1759 int
1760 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1761 {
1762         struct nfs_read_header *rhdr;
1763         struct nfs_pgio_header *hdr;
1764         int ret;
1765
1766         rhdr = nfs_readhdr_alloc();
1767         if (!rhdr) {
1768                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1769                 ret = -ENOMEM;
1770                 pnfs_put_lseg(desc->pg_lseg);
1771                 desc->pg_lseg = NULL;
1772                 return ret;
1773         }
1774         hdr = &rhdr->header;
1775         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1776         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1777         atomic_inc(&hdr->refcnt);
1778         ret = nfs_generic_pagein(desc, hdr);
1779         if (ret != 0) {
1780                 pnfs_put_lseg(desc->pg_lseg);
1781                 desc->pg_lseg = NULL;
1782         } else
1783                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1784         if (atomic_dec_and_test(&hdr->refcnt))
1785                 hdr->completion_ops->completion(hdr);
1786         return ret;
1787 }
1788 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1789
1790 /*
1791  * There can be multiple RW segments.
1792  */
1793 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1794 {
1795         struct pnfs_layout_segment *lseg;
1796
1797         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1798                 if (lseg->pls_range.iomode == IOMODE_RW &&
1799                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1800                         list_add(&lseg->pls_lc_list, listp);
1801         }
1802 }
1803
1804 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1805 {
1806         struct pnfs_layout_segment *lseg, *tmp;
1807         unsigned long *bitlock = &NFS_I(inode)->flags;
1808
1809         /* Matched by references in pnfs_set_layoutcommit */
1810         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1811                 list_del_init(&lseg->pls_lc_list);
1812                 pnfs_put_lseg(lseg);
1813         }
1814
1815         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1816         smp_mb__after_clear_bit();
1817         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1818 }
1819
1820 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1821 {
1822         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1823 }
1824 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1825
1826 void
1827 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1828 {
1829         struct nfs_pgio_header *hdr = wdata->header;
1830         struct inode *inode = hdr->inode;
1831         struct nfs_inode *nfsi = NFS_I(inode);
1832         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1833         bool mark_as_dirty = false;
1834
1835         spin_lock(&inode->i_lock);
1836         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1837                 mark_as_dirty = true;
1838                 dprintk("%s: Set layoutcommit for inode %lu ",
1839                         __func__, inode->i_ino);
1840         }
1841         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1842                 /* references matched in nfs4_layoutcommit_release */
1843                 pnfs_get_lseg(hdr->lseg);
1844         }
1845         if (end_pos > nfsi->layout->plh_lwb)
1846                 nfsi->layout->plh_lwb = end_pos;
1847         spin_unlock(&inode->i_lock);
1848         dprintk("%s: lseg %p end_pos %llu\n",
1849                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1850
1851         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1852          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1853         if (mark_as_dirty)
1854                 mark_inode_dirty_sync(inode);
1855 }
1856 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1857
1858 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1859 {
1860         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1861
1862         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1863                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1864         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1865 }
1866
1867 /*
1868  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1869  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1870  * data to disk to allow the server to recover the data if it crashes.
1871  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1872  * is off, and a COMMIT is sent to a data server, or
1873  * if WRITEs to a data server return NFS_DATA_SYNC.
1874  */
1875 int
1876 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1877 {
1878         struct nfs4_layoutcommit_data *data;
1879         struct nfs_inode *nfsi = NFS_I(inode);
1880         loff_t end_pos;
1881         int status = 0;
1882
1883         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1884
1885         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1886                 return 0;
1887
1888         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1889         data = kzalloc(sizeof(*data), GFP_NOFS);
1890         if (!data) {
1891                 status = -ENOMEM;
1892                 goto out;
1893         }
1894
1895         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1896                 goto out_free;
1897
1898         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1899                 if (!sync) {
1900                         status = -EAGAIN;
1901                         goto out_free;
1902                 }
1903                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1904                                         nfs_wait_bit_killable, TASK_KILLABLE);
1905                 if (status)
1906                         goto out_free;
1907         }
1908
1909         INIT_LIST_HEAD(&data->lseg_list);
1910         spin_lock(&inode->i_lock);
1911         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1912                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1913                 spin_unlock(&inode->i_lock);
1914                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1915                 goto out_free;
1916         }
1917
1918         pnfs_list_write_lseg(inode, &data->lseg_list);
1919
1920         end_pos = nfsi->layout->plh_lwb;
1921         nfsi->layout->plh_lwb = 0;
1922
1923         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1924         spin_unlock(&inode->i_lock);
1925
1926         data->args.inode = inode;
1927         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1928         nfs_fattr_init(&data->fattr);
1929         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1930         data->res.fattr = &data->fattr;
1931         data->args.lastbytewritten = end_pos - 1;
1932         data->res.server = NFS_SERVER(inode);
1933
1934         status = nfs4_proc_layoutcommit(data, sync);
1935 out:
1936         if (status)
1937                 mark_inode_dirty_sync(inode);
1938         dprintk("<-- %s status %d\n", __func__, status);
1939         return status;
1940 out_free:
1941         kfree(data);
1942         goto out;
1943 }
1944
1945 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1946 {
1947         struct nfs4_threshold *thp;
1948
1949         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1950         if (!thp) {
1951                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1952                 return NULL;
1953         }
1954         return thp;
1955 }