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 <linux/sort.h>
  34 #include "internal.h"
  35 #include "pnfs.h"
  36 #include "iostat.h"
  37 #include "nfs4trace.h"
  38 #include "delegation.h"
  39 #include "nfs42.h"
  40
  41 #define NFSDBG_FACILITY         NFSDBG_PNFS
  42 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
  43
  44 /* Locking:
  45  *
  46  * pnfs_spinlock:
  47  *      protects pnfs_modules_tbl.
  48  */
  49 static DEFINE_SPINLOCK(pnfs_spinlock);
  50
  51 /*
  52  * pnfs_modules_tbl holds all pnfs modules
  53  */
  54 static LIST_HEAD(pnfs_modules_tbl);
  55
  56 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
  57 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
  58                 struct list_head *free_me,
  59                 const struct pnfs_layout_range *range,
  60                 u32 seq);
  61 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
  62                                 struct list_head *tmp_list);
  63
  64 /* Return the registered pnfs layout driver module matching given id */
  65 static struct pnfs_layoutdriver_type *
  66 find_pnfs_driver_locked(u32 id)
  67 {
  68         struct pnfs_layoutdriver_type *local;
  69
  70         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
  71                 if (local->id == id)
  72                         goto out;
  73         local = NULL;
  74 out:
  75         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
  76         return local;
  77 }
  78
  79 static struct pnfs_layoutdriver_type *
  80 find_pnfs_driver(u32 id)
  81 {
  82         struct pnfs_layoutdriver_type *local;
  83
  84         spin_lock(&pnfs_spinlock);
  85         local = find_pnfs_driver_locked(id);
  86         if (local != NULL && !try_module_get(local->owner)) {
  87                 dprintk("%s: Could not grab reference on module\n", __func__);
  88                 local = NULL;
  89         }
  90         spin_unlock(&pnfs_spinlock);
  91         return local;
  92 }
  93
  94 void
  95 unset_pnfs_layoutdriver(struct nfs_server *nfss)
  96 {
  97         if (nfss->pnfs_curr_ld) {
  98                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
  99                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
 100                 /* Decrement the MDS count. Purge the deviceid cache if zero */
 101                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
 102                         nfs4_deviceid_purge_client(nfss->nfs_client);
 103                 module_put(nfss->pnfs_curr_ld->owner);
 104         }
 105         nfss->pnfs_curr_ld = NULL;
 106 }
 107
 108 /*
 109  * When the server sends a list of layout types, we choose one in the order
 110  * given in the list below.
 111  *
 112  * FIXME: should this list be configurable in some fashion? module param?
 113  *        mount option? something else?
 114  */
 115 static const u32 ld_prefs[] = {
 116         LAYOUT_SCSI,
 117         LAYOUT_BLOCK_VOLUME,
 118         LAYOUT_OSD2_OBJECTS,
 119         LAYOUT_FLEX_FILES,
 120         LAYOUT_NFSV4_1_FILES,
 121         0
 122 };
 123
 124 static int
 125 ld_cmp(const void *e1, const void *e2)
 126 {
 127         u32 ld1 = *((u32 *)e1);
 128         u32 ld2 = *((u32 *)e2);
 129         int i;
 130
 131         for (i = 0; ld_prefs[i] != 0; i++) {
 132                 if (ld1 == ld_prefs[i])
 133                         return -1;
 134
 135                 if (ld2 == ld_prefs[i])
 136                         return 1;
 137         }
 138         return 0;
 139 }
 140
 141 /*
 142  * Try to set the server's pnfs module to the pnfs layout type specified by id.
 143  * Currently only one pNFS layout driver per filesystem is supported.
 144  *
 145  * @ids array of layout types supported by MDS.
 146  */
 147 void
 148 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
 149                       struct nfs_fsinfo *fsinfo)
 150 {
 151         struct pnfs_layoutdriver_type *ld_type = NULL;
 152         u32 id;
 153         int i;
 154
 155         if (fsinfo->nlayouttypes == 0)
 156                 goto out_no_driver;
 157         if (!(server->nfs_client->cl_exchange_flags &
 158                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
 159                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
 160                         __func__, server->nfs_client->cl_exchange_flags);
 161                 goto out_no_driver;
 162         }
 163
 164         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
 165                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
 166
 167         for (i = 0; i < fsinfo->nlayouttypes; i++) {
 168                 id = fsinfo->layouttype[i];
 169                 ld_type = find_pnfs_driver(id);
 170                 if (!ld_type) {
 171                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
 172                                         id);
 173                         ld_type = find_pnfs_driver(id);
 174                 }
 175                 if (ld_type)
 176                         break;
 177         }
 178
 179         if (!ld_type) {
 180                 dprintk("%s: No pNFS module found!\n", __func__);
 181                 goto out_no_driver;
 182         }
 183
 184         server->pnfs_curr_ld = ld_type;
 185         if (ld_type->set_layoutdriver
 186             && ld_type->set_layoutdriver(server, mntfh)) {
 187                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
 188                         "driver %u.\n", __func__, id);
 189                 module_put(ld_type->owner);
 190                 goto out_no_driver;
 191         }
 192         /* Bump the MDS count */
 193         atomic_inc(&server->nfs_client->cl_mds_count);
 194
 195         dprintk("%s: pNFS module for %u set\n", __func__, id);
 196         return;
 197
 198 out_no_driver:
 199         dprintk("%s: Using NFSv4 I/O\n", __func__);
 200         server->pnfs_curr_ld = NULL;
 201 }
 202
 203 int
 204 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 205 {
 206         int status = -EINVAL;
 207         struct pnfs_layoutdriver_type *tmp;
 208
 209         if (ld_type->id == 0) {
 210                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
 211                 return status;
 212         }
 213         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
 214                 printk(KERN_ERR "NFS: %s Layout driver must provide "
 215                        "alloc_lseg and free_lseg.\n", __func__);
 216                 return status;
 217         }
 218
 219         spin_lock(&pnfs_spinlock);
 220         tmp = find_pnfs_driver_locked(ld_type->id);
 221         if (!tmp) {
 222                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
 223                 status = 0;
 224                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
 225                         ld_type->name);
 226         } else {
 227                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
 228                         __func__, ld_type->id);
 229         }
 230         spin_unlock(&pnfs_spinlock);
 231
 232         return status;
 233 }
 234 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
 235
 236 void
 237 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
 238 {
 239         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
 240         spin_lock(&pnfs_spinlock);
 241         list_del(&ld_type->pnfs_tblid);
 242         spin_unlock(&pnfs_spinlock);
 243 }
 244 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
 245
 246 /*
 247  * pNFS client layout cache
 248  */
 249
 250 /* Need to hold i_lock if caller does not already hold reference */
 251 void
 252 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
 253 {
 254         atomic_inc(&lo->plh_refcount);
 255 }
 256
 257 static struct pnfs_layout_hdr *
 258 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
 259 {
 260         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
 261         return ld->alloc_layout_hdr(ino, gfp_flags);
 262 }
 263
 264 static void
 265 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
 266 {
 267         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
 268         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
 269
 270         if (!list_empty(&lo->plh_layouts)) {
 271                 struct nfs_client *clp = server->nfs_client;
 272
 273                 spin_lock(&clp->cl_lock);
 274                 list_del_init(&lo->plh_layouts);
 275                 spin_unlock(&clp->cl_lock);
 276         }
 277         put_rpccred(lo->plh_lc_cred);
 278         return ld->free_layout_hdr(lo);
 279 }
 280
 281 static void
 282 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
 283 {
 284         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
 285         dprintk("%s: freeing layout cache %p\n", __func__, lo);
 286         nfsi->layout = NULL;
 287         /* Reset MDS Threshold I/O counters */
 288         nfsi->write_io = 0;
 289         nfsi->read_io = 0;
 290 }
 291
 292 void
 293 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
 294 {
 295         struct inode *inode = lo->plh_inode;
 296
 297         pnfs_layoutreturn_before_put_layout_hdr(lo);
 298
 299         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
 300                 if (!list_empty(&lo->plh_segs))
 301                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
 302                 pnfs_detach_layout_hdr(lo);
 303                 spin_unlock(&inode->i_lock);
 304                 pnfs_free_layout_hdr(lo);
 305         }
 306 }
 307
 308 static void
 309 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
 310                          u32 seq)
 311 {
 312         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
 313                 iomode = IOMODE_ANY;
 314         lo->plh_return_iomode = iomode;
 315         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 316         if (seq != 0) {
 317                 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
 318                 lo->plh_return_seq = seq;
 319         }
 320 }
 321
 322 static void
 323 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
 324 {
 325         struct pnfs_layout_segment *lseg;
 326         lo->plh_return_iomode = 0;
 327         lo->plh_return_seq = 0;
 328         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
 329         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
 330                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 331                         continue;
 332                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 333         }
 334 }
 335
 336 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
 337 {
 338         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
 339         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
 340         smp_mb__after_atomic();
 341         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
 342         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
 343 }
 344
 345 static void
 346 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
 347                 struct list_head *free_me)
 348 {
 349         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
 350         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
 351         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
 352                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 353         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 354                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
 355 }
 356
 357 /*
 358  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
 359  *
 360  * In order to continue using the pnfs_layout_hdr, a full recovery
 361  * is required.
 362  * Note that caller must hold inode->i_lock.
 363  */
 364 int
 365 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
 366                 struct list_head *lseg_list)
 367 {
 368         struct pnfs_layout_range range = {
 369                 .iomode = IOMODE_ANY,
 370                 .offset = 0,
 371                 .length = NFS4_MAX_UINT64,
 372         };
 373         struct pnfs_layout_segment *lseg, *next;
 374
 375         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 376         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 377                 pnfs_clear_lseg_state(lseg, lseg_list);
 378         pnfs_clear_layoutreturn_info(lo);
 379         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
 380         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
 381             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
 382                 pnfs_clear_layoutreturn_waitbit(lo);
 383         return !list_empty(&lo->plh_segs);
 384 }
 385
 386 static int
 387 pnfs_iomode_to_fail_bit(u32 iomode)
 388 {
 389         return iomode == IOMODE_RW ?
 390                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
 391 }
 392
 393 static void
 394 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 395 {
 396         lo->plh_retry_timestamp = jiffies;
 397         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
 398                 atomic_inc(&lo->plh_refcount);
 399 }
 400
 401 static void
 402 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
 403 {
 404         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
 405                 atomic_dec(&lo->plh_refcount);
 406 }
 407
 408 static void
 409 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 410 {
 411         struct inode *inode = lo->plh_inode;
 412         struct pnfs_layout_range range = {
 413                 .iomode = iomode,
 414                 .offset = 0,
 415                 .length = NFS4_MAX_UINT64,
 416         };
 417         LIST_HEAD(head);
 418
 419         spin_lock(&inode->i_lock);
 420         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
 421         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
 422         spin_unlock(&inode->i_lock);
 423         pnfs_free_lseg_list(&head);
 424         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
 425                         iomode == IOMODE_RW ?  "RW" : "READ");
 426 }
 427
 428 static bool
 429 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
 430 {
 431         unsigned long start, end;
 432         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
 433
 434         if (test_bit(fail_bit, &lo->plh_flags) == 0)
 435                 return false;
 436         end = jiffies;
 437         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
 438         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
 439                 /* It is time to retry the failed layoutgets */
 440                 pnfs_layout_clear_fail_bit(lo, fail_bit);
 441                 return false;
 442         }
 443         return true;
 444 }
 445
 446 static void
 447 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
 448                 const struct pnfs_layout_range *range,
 449                 const nfs4_stateid *stateid)
 450 {
 451         INIT_LIST_HEAD(&lseg->pls_list);
 452         INIT_LIST_HEAD(&lseg->pls_lc_list);
 453         atomic_set(&lseg->pls_refcount, 1);
 454         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 455         lseg->pls_layout = lo;
 456         lseg->pls_range = *range;
 457         lseg->pls_seq = be32_to_cpu(stateid->seqid);
 458 }
 459
 460 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
 461 {
 462         if (lseg != NULL) {
 463                 struct inode *inode = lseg->pls_layout->plh_inode;
 464                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
 465         }
 466 }
 467
 468 static void
 469 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
 470                 struct pnfs_layout_segment *lseg)
 471 {
 472         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 473         list_del_init(&lseg->pls_list);
 474         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
 475         atomic_dec(&lo->plh_refcount);
 476         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
 477                 return;
 478         if (list_empty(&lo->plh_segs) &&
 479             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
 480             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
 481                 if (atomic_read(&lo->plh_outstanding) == 0)
 482                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 483                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 484         }
 485 }
 486
 487 static bool
 488 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
 489                 struct pnfs_layout_segment *lseg)
 490 {
 491         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
 492             pnfs_layout_is_valid(lo)) {
 493                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
 494                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
 495                 return true;
 496         }
 497         return false;
 498 }
 499
 500 void
 501 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
 502 {
 503         struct pnfs_layout_hdr *lo;
 504         struct inode *inode;
 505
 506         if (!lseg)
 507                 return;
 508
 509         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
 510                 atomic_read(&lseg->pls_refcount),
 511                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
 512
 513         lo = lseg->pls_layout;
 514         inode = lo->plh_inode;
 515
 516         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
 517                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 518                         spin_unlock(&inode->i_lock);
 519                         return;
 520                 }
 521                 pnfs_get_layout_hdr(lo);
 522                 pnfs_layout_remove_lseg(lo, lseg);
 523                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
 524                         lseg = NULL;
 525                 spin_unlock(&inode->i_lock);
 526                 pnfs_free_lseg(lseg);
 527                 pnfs_put_layout_hdr(lo);
 528         }
 529 }
 530 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
 531
 532 /*
 533  * is l2 fully contained in l1?
 534  *   start1                             end1
 535  *   [----------------------------------)
 536  *           start2           end2
 537  *           [----------------)
 538  */
 539 static bool
 540 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
 541                  const struct pnfs_layout_range *l2)
 542 {
 543         u64 start1 = l1->offset;
 544         u64 end1 = pnfs_end_offset(start1, l1->length);
 545         u64 start2 = l2->offset;
 546         u64 end2 = pnfs_end_offset(start2, l2->length);
 547
 548         return (start1 <= start2) && (end1 >= end2);
 549 }
 550
 551 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
 552                 struct list_head *tmp_list)
 553 {
 554         if (!atomic_dec_and_test(&lseg->pls_refcount))
 555                 return false;
 556         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
 557         list_add(&lseg->pls_list, tmp_list);
 558         return true;
 559 }
 560
 561 /* Returns 1 if lseg is removed from list, 0 otherwise */
 562 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
 563                              struct list_head *tmp_list)
 564 {
 565         int rv = 0;
 566
 567         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
 568                 /* Remove the reference keeping the lseg in the
 569                  * list.  It will now be removed when all
 570                  * outstanding io is finished.
 571                  */
 572                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
 573                         atomic_read(&lseg->pls_refcount));
 574                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
 575                         rv = 1;
 576         }
 577         return rv;
 578 }
 579
 580 /*
 581  * Compare 2 layout stateid sequence ids, to see which is newer,
 582  * taking into account wraparound issues.
 583  */
 584 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
 585 {
 586         return (s32)(s1 - s2) > 0;
 587 }
 588
 589 static bool
 590 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
 591                  const struct pnfs_layout_range *recall_range)
 592 {
 593         return (recall_range->iomode == IOMODE_ANY ||
 594                 lseg_range->iomode == recall_range->iomode) &&
 595                pnfs_lseg_range_intersecting(lseg_range, recall_range);
 596 }
 597
 598 static bool
 599 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
 600                 const struct pnfs_layout_range *recall_range,
 601                 u32 seq)
 602 {
 603         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
 604                 return false;
 605         if (recall_range == NULL)
 606                 return true;
 607         return pnfs_should_free_range(&lseg->pls_range, recall_range);
 608 }
 609
 610 /**
 611  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
 612  * @lo: layout header containing the lsegs
 613  * @tmp_list: list head where doomed lsegs should go
 614  * @recall_range: optional recall range argument to match (may be NULL)
 615  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
 616  *
 617  * Walk the list of lsegs in the layout header, and tear down any that should
 618  * be destroyed. If "recall_range" is specified then the segment must match
 619  * that range. If "seq" is non-zero, then only match segments that were handed
 620  * out at or before that sequence.
 621  *
 622  * Returns number of matching invalid lsegs remaining in list after scanning
 623  * it and purging them.
 624  */
 625 int
 626 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
 627                             struct list_head *tmp_list,
 628                             const struct pnfs_layout_range *recall_range,
 629                             u32 seq)
 630 {
 631         struct pnfs_layout_segment *lseg, *next;
 632         int remaining = 0;
 633
 634         dprintk("%s:Begin lo %p\n", __func__, lo);
 635
 636         if (list_empty(&lo->plh_segs))
 637                 return 0;
 638         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
 639                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
 640                         dprintk("%s: freeing lseg %p iomode %d seq %u"
 641                                 "offset %llu length %llu\n", __func__,
 642                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
 643                                 lseg->pls_range.offset, lseg->pls_range.length);
 644                         if (!mark_lseg_invalid(lseg, tmp_list))
 645                                 remaining++;
 646                 }
 647         dprintk("%s:Return %i\n", __func__, remaining);
 648         return remaining;
 649 }
 650
 651 static void
 652 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
 653                 struct list_head *free_me,
 654                 const struct pnfs_layout_range *range,
 655                 u32 seq)
 656 {
 657         struct pnfs_layout_segment *lseg, *next;
 658
 659         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
 660                 if (pnfs_match_lseg_recall(lseg, range, seq))
 661                         list_move_tail(&lseg->pls_list, free_me);
 662         }
 663 }
 664
 665 /* note free_me must contain lsegs from a single layout_hdr */
 666 void
 667 pnfs_free_lseg_list(struct list_head *free_me)
 668 {
 669         struct pnfs_layout_segment *lseg, *tmp;
 670
 671         if (list_empty(free_me))
 672                 return;
 673
 674         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
 675                 list_del(&lseg->pls_list);
 676                 pnfs_free_lseg(lseg);
 677         }
 678 }
 679
 680 void
 681 pnfs_destroy_layout(struct nfs_inode *nfsi)
 682 {
 683         struct pnfs_layout_hdr *lo;
 684         LIST_HEAD(tmp_list);
 685
 686         spin_lock(&nfsi->vfs_inode.i_lock);
 687         lo = nfsi->layout;
 688         if (lo) {
 689                 pnfs_get_layout_hdr(lo);
 690                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
 691                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
 692                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
 693                 spin_unlock(&nfsi->vfs_inode.i_lock);
 694                 pnfs_free_lseg_list(&tmp_list);
 695                 nfs_commit_inode(&nfsi->vfs_inode, 0);
 696                 pnfs_put_layout_hdr(lo);
 697         } else
 698                 spin_unlock(&nfsi->vfs_inode.i_lock);
 699 }
 700 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
 701
 702 static bool
 703 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
 704                 struct list_head *layout_list)
 705 {
 706         struct pnfs_layout_hdr *lo;
 707         bool ret = false;
 708
 709         spin_lock(&inode->i_lock);
 710         lo = NFS_I(inode)->layout;
 711         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
 712                 pnfs_get_layout_hdr(lo);
 713                 list_add(&lo->plh_bulk_destroy, layout_list);
 714                 ret = true;
 715         }
 716         spin_unlock(&inode->i_lock);
 717         return ret;
 718 }
 719
 720 /* Caller must hold rcu_read_lock and clp->cl_lock */
 721 static int
 722 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
 723                 struct nfs_server *server,
 724                 struct list_head *layout_list)
 725 {
 726         struct pnfs_layout_hdr *lo, *next;
 727         struct inode *inode;
 728
 729         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
 730                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
 731                         continue;
 732                 inode = igrab(lo->plh_inode);
 733                 if (inode == NULL)
 734                         continue;
 735                 list_del_init(&lo->plh_layouts);
 736                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
 737                         continue;
 738                 rcu_read_unlock();
 739                 spin_unlock(&clp->cl_lock);
 740                 iput(inode);
 741                 spin_lock(&clp->cl_lock);
 742                 rcu_read_lock();
 743                 return -EAGAIN;
 744         }
 745         return 0;
 746 }
 747
 748 static int
 749 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
 750                 bool is_bulk_recall)
 751 {
 752         struct pnfs_layout_hdr *lo;
 753         struct inode *inode;
 754         LIST_HEAD(lseg_list);
 755         int ret = 0;
 756
 757         while (!list_empty(layout_list)) {
 758                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
 759                                 plh_bulk_destroy);
 760                 dprintk("%s freeing layout for inode %lu\n", __func__,
 761                         lo->plh_inode->i_ino);
 762                 inode = lo->plh_inode;
 763
 764                 pnfs_layoutcommit_inode(inode, false);
 765
 766                 spin_lock(&inode->i_lock);
 767                 list_del_init(&lo->plh_bulk_destroy);
 768                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
 769                         if (is_bulk_recall)
 770                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 771                         ret = -EAGAIN;
 772                 }
 773                 spin_unlock(&inode->i_lock);
 774                 pnfs_free_lseg_list(&lseg_list);
 775                 /* Free all lsegs that are attached to commit buckets */
 776                 nfs_commit_inode(inode, 0);
 777                 pnfs_put_layout_hdr(lo);
 778                 iput(inode);
 779         }
 780         return ret;
 781 }
 782
 783 int
 784 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
 785                 struct nfs_fsid *fsid,
 786                 bool is_recall)
 787 {
 788         struct nfs_server *server;
 789         LIST_HEAD(layout_list);
 790
 791         spin_lock(&clp->cl_lock);
 792         rcu_read_lock();
 793 restart:
 794         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 795                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
 796                         continue;
 797                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 798                                 server,
 799                                 &layout_list) != 0)
 800                         goto restart;
 801         }
 802         rcu_read_unlock();
 803         spin_unlock(&clp->cl_lock);
 804
 805         if (list_empty(&layout_list))
 806                 return 0;
 807         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 808 }
 809
 810 int
 811 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
 812                 bool is_recall)
 813 {
 814         struct nfs_server *server;
 815         LIST_HEAD(layout_list);
 816
 817         spin_lock(&clp->cl_lock);
 818         rcu_read_lock();
 819 restart:
 820         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
 821                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
 822                                         server,
 823                                         &layout_list) != 0)
 824                         goto restart;
 825         }
 826         rcu_read_unlock();
 827         spin_unlock(&clp->cl_lock);
 828
 829         if (list_empty(&layout_list))
 830                 return 0;
 831         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
 832 }
 833
 834 /*
 835  * Called by the state manger to remove all layouts established under an
 836  * expired lease.
 837  */
 838 void
 839 pnfs_destroy_all_layouts(struct nfs_client *clp)
 840 {
 841         nfs4_deviceid_mark_client_invalid(clp);
 842         nfs4_deviceid_purge_client(clp);
 843
 844         pnfs_destroy_layouts_byclid(clp, false);
 845 }
 846
 847 /* update lo->plh_stateid with new if is more recent */
 848 void
 849 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
 850                         bool update_barrier)
 851 {
 852         u32 oldseq, newseq, new_barrier = 0;
 853
 854         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
 855         newseq = be32_to_cpu(new->seqid);
 856
 857         if (!pnfs_layout_is_valid(lo)) {
 858                 nfs4_stateid_copy(&lo->plh_stateid, new);
 859                 lo->plh_barrier = newseq;
 860                 pnfs_clear_layoutreturn_info(lo);
 861                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
 862                 return;
 863         }
 864         if (pnfs_seqid_is_newer(newseq, oldseq)) {
 865                 nfs4_stateid_copy(&lo->plh_stateid, new);
 866                 /*
 867                  * Because of wraparound, we want to keep the barrier
 868                  * "close" to the current seqids.
 869                  */
 870                 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
 871         }
 872         if (update_barrier)
 873                 new_barrier = be32_to_cpu(new->seqid);
 874         else if (new_barrier == 0)
 875                 return;
 876         if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
 877                 lo->plh_barrier = new_barrier;
 878 }
 879
 880 static bool
 881 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
 882                 const nfs4_stateid *stateid)
 883 {
 884         u32 seqid = be32_to_cpu(stateid->seqid);
 885
 886         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
 887 }
 888
 889 /* lget is set to 1 if called from inside send_layoutget call chain */
 890 static bool
 891 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
 892 {
 893         return lo->plh_block_lgets ||
 894                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
 895 }
 896
 897 /*
 898  * Get layout from server.
 899  *    for now, assume that whole file layouts are requested.
 900  *    arg->offset: 0
 901  *    arg->length: all ones
 902  */
 903 static struct pnfs_layout_segment *
 904 send_layoutget(struct pnfs_layout_hdr *lo,
 905            struct nfs_open_context *ctx,
 906            nfs4_stateid *stateid,
 907            const struct pnfs_layout_range *range,
 908            long *timeout, gfp_t gfp_flags)
 909 {
 910         struct inode *ino = lo->plh_inode;
 911         struct nfs_server *server = NFS_SERVER(ino);
 912         struct nfs4_layoutget *lgp;
 913         loff_t i_size;
 914
 915         dprintk("--> %s\n", __func__);
 916
 917         /*
 918          * Synchronously retrieve layout information from server and
 919          * store in lseg. If we race with a concurrent seqid morphing
 920          * op, then re-send the LAYOUTGET.
 921          */
 922         lgp = kzalloc(sizeof(*lgp), gfp_flags);
 923         if (lgp == NULL)
 924                 return ERR_PTR(-ENOMEM);
 925
 926         i_size = i_size_read(ino);
 927
 928         lgp->args.minlength = PAGE_SIZE;
 929         if (lgp->args.minlength > range->length)
 930                 lgp->args.minlength = range->length;
 931         if (range->iomode == IOMODE_READ) {
 932                 if (range->offset >= i_size)
 933                         lgp->args.minlength = 0;
 934                 else if (i_size - range->offset < lgp->args.minlength)
 935                         lgp->args.minlength = i_size - range->offset;
 936         }
 937         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
 938         pnfs_copy_range(&lgp->args.range, range);
 939         lgp->args.type = server->pnfs_curr_ld->id;
 940         lgp->args.inode = ino;
 941         lgp->args.ctx = get_nfs_open_context(ctx);
 942         nfs4_stateid_copy(&lgp->args.stateid, stateid);
 943         lgp->gfp_flags = gfp_flags;
 944         lgp->cred = lo->plh_lc_cred;
 945
 946         return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
 947 }
 948
 949 static void pnfs_clear_layoutcommit(struct inode *inode,
 950                 struct list_head *head)
 951 {
 952         struct nfs_inode *nfsi = NFS_I(inode);
 953         struct pnfs_layout_segment *lseg, *tmp;
 954
 955         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
 956                 return;
 957         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
 958                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
 959                         continue;
 960                 pnfs_lseg_dec_and_remove_zero(lseg, head);
 961         }
 962 }
 963
 964 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
 965                 const nfs4_stateid *arg_stateid,
 966                 const struct pnfs_layout_range *range,
 967                 const nfs4_stateid *stateid)
 968 {
 969         struct inode *inode = lo->plh_inode;
 970         LIST_HEAD(freeme);
 971
 972         spin_lock(&inode->i_lock);
 973         if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
 974             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
 975                 goto out_unlock;
 976         if (stateid) {
 977                 u32 seq = be32_to_cpu(arg_stateid->seqid);
 978
 979                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
 980                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
 981                 pnfs_set_layout_stateid(lo, stateid, true);
 982         } else
 983                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
 984 out_unlock:
 985         pnfs_clear_layoutreturn_waitbit(lo);
 986         spin_unlock(&inode->i_lock);
 987         pnfs_free_lseg_list(&freeme);
 988
 989 }
 990
 991 static bool
 992 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
 993                 nfs4_stateid *stateid,
 994                 enum pnfs_iomode *iomode)
 995 {
 996         /* Serialise LAYOUTGET/LAYOUTRETURN */
 997         if (atomic_read(&lo->plh_outstanding) != 0)
 998                 return false;
 999         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1000                 return false;
1001         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1002         pnfs_get_layout_hdr(lo);
1003         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1004                 if (stateid != NULL) {
1005                         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1006                         if (lo->plh_return_seq != 0)
1007                                 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1008                 }
1009                 if (iomode != NULL)
1010                         *iomode = lo->plh_return_iomode;
1011                 pnfs_clear_layoutreturn_info(lo);
1012                 return true;
1013         }
1014         if (stateid != NULL)
1015                 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1016         if (iomode != NULL)
1017                 *iomode = IOMODE_ANY;
1018         return true;
1019 }
1020
1021 static void
1022 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1023                 struct pnfs_layout_hdr *lo,
1024                 const nfs4_stateid *stateid,
1025                 enum pnfs_iomode iomode)
1026 {
1027         struct inode *inode = lo->plh_inode;
1028
1029         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1030         args->inode = inode;
1031         args->range.iomode = iomode;
1032         args->range.offset = 0;
1033         args->range.length = NFS4_MAX_UINT64;
1034         args->layout = lo;
1035         nfs4_stateid_copy(&args->stateid, stateid);
1036 }
1037
1038 static int
1039 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1040                        enum pnfs_iomode iomode, bool sync)
1041 {
1042         struct inode *ino = lo->plh_inode;
1043         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1044         struct nfs4_layoutreturn *lrp;
1045         int status = 0;
1046
1047         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1048         if (unlikely(lrp == NULL)) {
1049                 status = -ENOMEM;
1050                 spin_lock(&ino->i_lock);
1051                 pnfs_clear_layoutreturn_waitbit(lo);
1052                 spin_unlock(&ino->i_lock);
1053                 pnfs_put_layout_hdr(lo);
1054                 goto out;
1055         }
1056
1057         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1058         lrp->args.ld_private = &lrp->ld_private;
1059         lrp->clp = NFS_SERVER(ino)->nfs_client;
1060         lrp->cred = lo->plh_lc_cred;
1061         if (ld->prepare_layoutreturn)
1062                 ld->prepare_layoutreturn(&lrp->args);
1063
1064         status = nfs4_proc_layoutreturn(lrp, sync);
1065 out:
1066         dprintk("<-- %s status: %d\n", __func__, status);
1067         return status;
1068 }
1069
1070 /* Return true if layoutreturn is needed */
1071 static bool
1072 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1073 {
1074         struct pnfs_layout_segment *s;
1075
1076         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1077                 return false;
1078
1079         /* Defer layoutreturn until all lsegs are done */
1080         list_for_each_entry(s, &lo->plh_segs, pls_list) {
1081                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1082                         return false;
1083         }
1084
1085         return true;
1086 }
1087
1088 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1089 {
1090         struct inode *inode= lo->plh_inode;
1091
1092         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1093                 return;
1094         spin_lock(&inode->i_lock);
1095         if (pnfs_layout_need_return(lo)) {
1096                 nfs4_stateid stateid;
1097                 enum pnfs_iomode iomode;
1098                 bool send;
1099
1100                 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1101                 spin_unlock(&inode->i_lock);
1102                 if (send) {
1103                         /* Send an async layoutreturn so we dont deadlock */
1104                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1105                 }
1106         } else
1107                 spin_unlock(&inode->i_lock);
1108 }
1109
1110 /*
1111  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1112  * when the layout segment list is empty.
1113  *
1114  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1115  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1116  * deviceid is marked invalid.
1117  */
1118 int
1119 _pnfs_return_layout(struct inode *ino)
1120 {
1121         struct pnfs_layout_hdr *lo = NULL;
1122         struct nfs_inode *nfsi = NFS_I(ino);
1123         LIST_HEAD(tmp_list);
1124         nfs4_stateid stateid;
1125         int status = 0;
1126         bool send;
1127
1128         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1129
1130         spin_lock(&ino->i_lock);
1131         lo = nfsi->layout;
1132         if (!lo) {
1133                 spin_unlock(&ino->i_lock);
1134                 dprintk("NFS: %s no layout to return\n", __func__);
1135                 goto out;
1136         }
1137         /* Reference matched in nfs4_layoutreturn_release */
1138         pnfs_get_layout_hdr(lo);
1139         /* Is there an outstanding layoutreturn ? */
1140         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1141                 spin_unlock(&ino->i_lock);
1142                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1143                                         TASK_UNINTERRUPTIBLE))
1144                         goto out_put_layout_hdr;
1145                 spin_lock(&ino->i_lock);
1146         }
1147         pnfs_clear_layoutcommit(ino, &tmp_list);
1148         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1149
1150         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1151                 struct pnfs_layout_range range = {
1152                         .iomode         = IOMODE_ANY,
1153                         .offset         = 0,
1154                         .length         = NFS4_MAX_UINT64,
1155                 };
1156                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1157         }
1158
1159         /* Don't send a LAYOUTRETURN if list was initially empty */
1160         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1161                 spin_unlock(&ino->i_lock);
1162                 dprintk("NFS: %s no layout segments to return\n", __func__);
1163                 goto out_put_layout_hdr;
1164         }
1165
1166         send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1167         spin_unlock(&ino->i_lock);
1168         if (send)
1169                 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1170 out_put_layout_hdr:
1171         pnfs_free_lseg_list(&tmp_list);
1172         pnfs_put_layout_hdr(lo);
1173 out:
1174         dprintk("<-- %s status: %d\n", __func__, status);
1175         return status;
1176 }
1177
1178 int
1179 pnfs_commit_and_return_layout(struct inode *inode)
1180 {
1181         struct pnfs_layout_hdr *lo;
1182         int ret;
1183
1184         spin_lock(&inode->i_lock);
1185         lo = NFS_I(inode)->layout;
1186         if (lo == NULL) {
1187                 spin_unlock(&inode->i_lock);
1188                 return 0;
1189         }
1190         pnfs_get_layout_hdr(lo);
1191         /* Block new layoutgets and read/write to ds */
1192         lo->plh_block_lgets++;
1193         spin_unlock(&inode->i_lock);
1194         filemap_fdatawait(inode->i_mapping);
1195         ret = pnfs_layoutcommit_inode(inode, true);
1196         if (ret == 0)
1197                 ret = _pnfs_return_layout(inode);
1198         spin_lock(&inode->i_lock);
1199         lo->plh_block_lgets--;
1200         spin_unlock(&inode->i_lock);
1201         pnfs_put_layout_hdr(lo);
1202         return ret;
1203 }
1204
1205 bool pnfs_roc(struct inode *ino,
1206                 struct nfs4_layoutreturn_args *args,
1207                 struct nfs4_layoutreturn_res *res,
1208                 const struct rpc_cred *cred)
1209 {
1210         struct nfs_inode *nfsi = NFS_I(ino);
1211         struct nfs_open_context *ctx;
1212         struct nfs4_state *state;
1213         struct pnfs_layout_hdr *lo;
1214         struct pnfs_layout_segment *lseg, *next;
1215         nfs4_stateid stateid;
1216         enum pnfs_iomode iomode = 0;
1217         bool layoutreturn = false, roc = false;
1218         bool skip_read = false;
1219
1220         if (!nfs_have_layout(ino))
1221                 return false;
1222 retry:
1223         spin_lock(&ino->i_lock);
1224         lo = nfsi->layout;
1225         if (!lo || !pnfs_layout_is_valid(lo) ||
1226             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1227                 goto out_noroc;
1228         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1229                 pnfs_get_layout_hdr(lo);
1230                 spin_unlock(&ino->i_lock);
1231                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1232                                 TASK_UNINTERRUPTIBLE);
1233                 pnfs_put_layout_hdr(lo);
1234                 goto retry;
1235         }
1236
1237         /* no roc if we hold a delegation */
1238         if (nfs4_check_delegation(ino, FMODE_READ)) {
1239                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1240                         goto out_noroc;
1241                 skip_read = true;
1242         }
1243
1244         list_for_each_entry(ctx, &nfsi->open_files, list) {
1245                 state = ctx->state;
1246                 if (state == NULL)
1247                         continue;
1248                 /* Don't return layout if there is open file state */
1249                 if (state->state & FMODE_WRITE)
1250                         goto out_noroc;
1251                 if (state->state & FMODE_READ)
1252                         skip_read = true;
1253         }
1254
1255
1256         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1257                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1258                         continue;
1259                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1260                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1261                         continue;
1262                 /*
1263                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1264                  * doesn't invalidate the layout for us.
1265                  */
1266                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1267                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1268                         continue;
1269                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1270         }
1271
1272         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1273                 goto out_noroc;
1274
1275         /* ROC in two conditions:
1276          * 1. there are ROC lsegs
1277          * 2. we don't send layoutreturn
1278          */
1279         /* lo ref dropped in pnfs_roc_release() */
1280         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1281         /* If the creds don't match, we can't compound the layoutreturn */
1282         if (!layoutreturn || cred != lo->plh_lc_cred)
1283                 goto out_noroc;
1284
1285         roc = layoutreturn;
1286         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1287         res->lrs_present = 0;
1288         layoutreturn = false;
1289
1290 out_noroc:
1291         spin_unlock(&ino->i_lock);
1292         pnfs_layoutcommit_inode(ino, true);
1293         if (roc) {
1294                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1295                 if (ld->prepare_layoutreturn)
1296                         ld->prepare_layoutreturn(args);
1297                 return true;
1298         }
1299         if (layoutreturn)
1300                 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1301         return false;
1302 }
1303
1304 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1305                 struct nfs4_layoutreturn_res *res,
1306                 int ret)
1307 {
1308         struct pnfs_layout_hdr *lo = args->layout;
1309         const nfs4_stateid *arg_stateid = NULL;
1310         const nfs4_stateid *res_stateid = NULL;
1311         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1312
1313         if (ret == 0) {
1314                 arg_stateid = &args->stateid;
1315                 if (res->lrs_present)
1316                         res_stateid = &res->stateid;
1317         }
1318         pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1319                         res_stateid);
1320         if (ld_private && ld_private->ops && ld_private->ops->free)
1321                 ld_private->ops->free(ld_private);
1322         pnfs_put_layout_hdr(lo);
1323         trace_nfs4_layoutreturn_on_close(args->inode, 0);
1324 }
1325
1326 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1327 {
1328         struct nfs_inode *nfsi = NFS_I(ino);
1329         struct pnfs_layout_hdr *lo;
1330         bool sleep = false;
1331
1332         /* we might not have grabbed lo reference. so need to check under
1333          * i_lock */
1334         spin_lock(&ino->i_lock);
1335         lo = nfsi->layout;
1336         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1337                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1338                 sleep = true;
1339         }
1340         spin_unlock(&ino->i_lock);
1341         return sleep;
1342 }
1343
1344 /*
1345  * Compare two layout segments for sorting into layout cache.
1346  * We want to preferentially return RW over RO layouts, so ensure those
1347  * are seen first.
1348  */
1349 static s64
1350 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1351            const struct pnfs_layout_range *l2)
1352 {
1353         s64 d;
1354
1355         /* high offset > low offset */
1356         d = l1->offset - l2->offset;
1357         if (d)
1358                 return d;
1359
1360         /* short length > long length */
1361         d = l2->length - l1->length;
1362         if (d)
1363                 return d;
1364
1365         /* read > read/write */
1366         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1367 }
1368
1369 static bool
1370 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1371                 const struct pnfs_layout_range *l2)
1372 {
1373         return pnfs_lseg_range_cmp(l1, l2) > 0;
1374 }
1375
1376 static bool
1377 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1378                 struct pnfs_layout_segment *old)
1379 {
1380         return false;
1381 }
1382
1383 void
1384 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1385                    struct pnfs_layout_segment *lseg,
1386                    bool (*is_after)(const struct pnfs_layout_range *,
1387                            const struct pnfs_layout_range *),
1388                    bool (*do_merge)(struct pnfs_layout_segment *,
1389                            struct pnfs_layout_segment *),
1390                    struct list_head *free_me)
1391 {
1392         struct pnfs_layout_segment *lp, *tmp;
1393
1394         dprintk("%s:Begin\n", __func__);
1395
1396         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1397                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1398                         continue;
1399                 if (do_merge(lseg, lp)) {
1400                         mark_lseg_invalid(lp, free_me);
1401                         continue;
1402                 }
1403                 if (is_after(&lseg->pls_range, &lp->pls_range))
1404                         continue;
1405                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1406                 dprintk("%s: inserted lseg %p "
1407                         "iomode %d offset %llu length %llu before "
1408                         "lp %p iomode %d offset %llu length %llu\n",
1409                         __func__, lseg, lseg->pls_range.iomode,
1410                         lseg->pls_range.offset, lseg->pls_range.length,
1411                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1412                         lp->pls_range.length);
1413                 goto out;
1414         }
1415         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1416         dprintk("%s: inserted lseg %p "
1417                 "iomode %d offset %llu length %llu at tail\n",
1418                 __func__, lseg, lseg->pls_range.iomode,
1419                 lseg->pls_range.offset, lseg->pls_range.length);
1420 out:
1421         pnfs_get_layout_hdr(lo);
1422
1423         dprintk("%s:Return\n", __func__);
1424 }
1425 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1426
1427 static void
1428 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1429                    struct pnfs_layout_segment *lseg,
1430                    struct list_head *free_me)
1431 {
1432         struct inode *inode = lo->plh_inode;
1433         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1434
1435         if (ld->add_lseg != NULL)
1436                 ld->add_lseg(lo, lseg, free_me);
1437         else
1438                 pnfs_generic_layout_insert_lseg(lo, lseg,
1439                                 pnfs_lseg_range_is_after,
1440                                 pnfs_lseg_no_merge,
1441                                 free_me);
1442 }
1443
1444 static struct pnfs_layout_hdr *
1445 alloc_init_layout_hdr(struct inode *ino,
1446                       struct nfs_open_context *ctx,
1447                       gfp_t gfp_flags)
1448 {
1449         struct pnfs_layout_hdr *lo;
1450
1451         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1452         if (!lo)
1453                 return NULL;
1454         atomic_set(&lo->plh_refcount, 1);
1455         INIT_LIST_HEAD(&lo->plh_layouts);
1456         INIT_LIST_HEAD(&lo->plh_segs);
1457         INIT_LIST_HEAD(&lo->plh_return_segs);
1458         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1459         lo->plh_inode = ino;
1460         lo->plh_lc_cred = get_rpccred(ctx->cred);
1461         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1462         return lo;
1463 }
1464
1465 static struct pnfs_layout_hdr *
1466 pnfs_find_alloc_layout(struct inode *ino,
1467                        struct nfs_open_context *ctx,
1468                        gfp_t gfp_flags)
1469         __releases(&ino->i_lock)
1470         __acquires(&ino->i_lock)
1471 {
1472         struct nfs_inode *nfsi = NFS_I(ino);
1473         struct pnfs_layout_hdr *new = NULL;
1474
1475         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1476
1477         if (nfsi->layout != NULL)
1478                 goto out_existing;
1479         spin_unlock(&ino->i_lock);
1480         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1481         spin_lock(&ino->i_lock);
1482
1483         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1484                 nfsi->layout = new;
1485                 return new;
1486         } else if (new != NULL)
1487                 pnfs_free_layout_hdr(new);
1488 out_existing:
1489         pnfs_get_layout_hdr(nfsi->layout);
1490         return nfsi->layout;
1491 }
1492
1493 /*
1494  * iomode matching rules:
1495  * iomode       lseg    strict match
1496  *                      iomode
1497  * -----        -----   ------ -----
1498  * ANY          READ    N/A    true
1499  * ANY          RW      N/A    true
1500  * RW           READ    N/A    false
1501  * RW           RW      N/A    true
1502  * READ         READ    N/A    true
1503  * READ         RW      true   false
1504  * READ         RW      false  true
1505  */
1506 static bool
1507 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1508                  const struct pnfs_layout_range *range,
1509                  bool strict_iomode)
1510 {
1511         struct pnfs_layout_range range1;
1512
1513         if ((range->iomode == IOMODE_RW &&
1514              ls_range->iomode != IOMODE_RW) ||
1515             (range->iomode != ls_range->iomode &&
1516              strict_iomode == true) ||
1517             !pnfs_lseg_range_intersecting(ls_range, range))
1518                 return 0;
1519
1520         /* range1 covers only the first byte in the range */
1521         range1 = *range;
1522         range1.length = 1;
1523         return pnfs_lseg_range_contained(ls_range, &range1);
1524 }
1525
1526 /*
1527  * lookup range in layout
1528  */
1529 static struct pnfs_layout_segment *
1530 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1531                 struct pnfs_layout_range *range,
1532                 bool strict_iomode)
1533 {
1534         struct pnfs_layout_segment *lseg, *ret = NULL;
1535
1536         dprintk("%s:Begin\n", __func__);
1537
1538         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1539                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1540                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1541                     pnfs_lseg_range_match(&lseg->pls_range, range,
1542                                           strict_iomode)) {
1543                         ret = pnfs_get_lseg(lseg);
1544                         break;
1545                 }
1546         }
1547
1548         dprintk("%s:Return lseg %p ref %d\n",
1549                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1550         return ret;
1551 }
1552
1553 /*
1554  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1555  * to the MDS or over pNFS
1556  *
1557  * The nfs_inode read_io and write_io fields are cumulative counters reset
1558  * when there are no layout segments. Note that in pnfs_update_layout iomode
1559  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1560  * WRITE request.
1561  *
1562  * A return of true means use MDS I/O.
1563  *
1564  * From rfc 5661:
1565  * If a file's size is smaller than the file size threshold, data accesses
1566  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1567  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1568  * server.  If both file size and I/O size are provided, the client SHOULD
1569  * reach or exceed  both thresholds before sending its read or write
1570  * requests to the data server.
1571  */
1572 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1573                                      struct inode *ino, int iomode)
1574 {
1575         struct nfs4_threshold *t = ctx->mdsthreshold;
1576         struct nfs_inode *nfsi = NFS_I(ino);
1577         loff_t fsize = i_size_read(ino);
1578         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1579
1580         if (t == NULL)
1581                 return ret;
1582
1583         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1584                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1585
1586         switch (iomode) {
1587         case IOMODE_READ:
1588                 if (t->bm & THRESHOLD_RD) {
1589                         dprintk("%s fsize %llu\n", __func__, fsize);
1590                         size_set = true;
1591                         if (fsize < t->rd_sz)
1592                                 size = true;
1593                 }
1594                 if (t->bm & THRESHOLD_RD_IO) {
1595                         dprintk("%s nfsi->read_io %llu\n", __func__,
1596                                 nfsi->read_io);
1597                         io_set = true;
1598                         if (nfsi->read_io < t->rd_io_sz)
1599                                 io = true;
1600                 }
1601                 break;
1602         case IOMODE_RW:
1603                 if (t->bm & THRESHOLD_WR) {
1604                         dprintk("%s fsize %llu\n", __func__, fsize);
1605                         size_set = true;
1606                         if (fsize < t->wr_sz)
1607                                 size = true;
1608                 }
1609                 if (t->bm & THRESHOLD_WR_IO) {
1610                         dprintk("%s nfsi->write_io %llu\n", __func__,
1611                                 nfsi->write_io);
1612                         io_set = true;
1613                         if (nfsi->write_io < t->wr_io_sz)
1614                                 io = true;
1615                 }
1616                 break;
1617         }
1618         if (size_set && io_set) {
1619                 if (size && io)
1620                         ret = true;
1621         } else if (size || io)
1622                 ret = true;
1623
1624         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1625         return ret;
1626 }
1627
1628 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1629 {
1630         /*
1631          * send layoutcommit as it can hold up layoutreturn due to lseg
1632          * reference
1633          */
1634         pnfs_layoutcommit_inode(lo->plh_inode, false);
1635         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1636                                    nfs_wait_bit_killable,
1637                                    TASK_UNINTERRUPTIBLE);
1638 }
1639
1640 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1641 {
1642         unsigned long *bitlock = &lo->plh_flags;
1643
1644         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1645         smp_mb__after_atomic();
1646         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1647 }
1648
1649 /*
1650  * Layout segment is retreived from the server if not cached.
1651  * The appropriate layout segment is referenced and returned to the caller.
1652  */
1653 struct pnfs_layout_segment *
1654 pnfs_update_layout(struct inode *ino,
1655                    struct nfs_open_context *ctx,
1656                    loff_t pos,
1657                    u64 count,
1658                    enum pnfs_iomode iomode,
1659                    bool strict_iomode,
1660                    gfp_t gfp_flags)
1661 {
1662         struct pnfs_layout_range arg = {
1663                 .iomode = iomode,
1664                 .offset = pos,
1665                 .length = count,
1666         };
1667         unsigned pg_offset;
1668         struct nfs_server *server = NFS_SERVER(ino);
1669         struct nfs_client *clp = server->nfs_client;
1670         struct pnfs_layout_hdr *lo = NULL;
1671         struct pnfs_layout_segment *lseg = NULL;
1672         nfs4_stateid stateid;
1673         long timeout = 0;
1674         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1675         bool first;
1676
1677         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1678                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1679                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1680                 goto out;
1681         }
1682
1683         if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1684                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1685                                  PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1686                 goto out;
1687         }
1688
1689         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1690                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1691                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1692                 goto out;
1693         }
1694
1695 lookup_again:
1696         nfs4_client_recover_expired_lease(clp);
1697         first = false;
1698         spin_lock(&ino->i_lock);
1699         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1700         if (lo == NULL) {
1701                 spin_unlock(&ino->i_lock);
1702                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1703                                  PNFS_UPDATE_LAYOUT_NOMEM);
1704                 goto out;
1705         }
1706
1707         /* Do we even need to bother with this? */
1708         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1709                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1710                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
1711                 dprintk("%s matches recall, use MDS\n", __func__);
1712                 goto out_unlock;
1713         }
1714
1715         /* if LAYOUTGET already failed once we don't try again */
1716         if (pnfs_layout_io_test_failed(lo, iomode)) {
1717                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1718                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1719                 goto out_unlock;
1720         }
1721
1722         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1723         if (lseg) {
1724                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1725                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1726                 goto out_unlock;
1727         }
1728
1729         if (!nfs4_valid_open_stateid(ctx->state)) {
1730                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1731                                 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1732                 goto out_unlock;
1733         }
1734
1735         /*
1736          * Choose a stateid for the LAYOUTGET. If we don't have a layout
1737          * stateid, or it has been invalidated, then we must use the open
1738          * stateid.
1739          */
1740         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1741
1742                 /*
1743                  * The first layoutget for the file. Need to serialize per
1744                  * RFC 5661 Errata 3208.
1745                  */
1746                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1747                                      &lo->plh_flags)) {
1748                         spin_unlock(&ino->i_lock);
1749                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1750                                     TASK_UNINTERRUPTIBLE);
1751                         pnfs_put_layout_hdr(lo);
1752                         dprintk("%s retrying\n", __func__);
1753                         goto lookup_again;
1754                 }
1755
1756                 first = true;
1757                 if (nfs4_select_rw_stateid(ctx->state,
1758                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1759                                         NULL, &stateid, NULL) != 0) {
1760                         trace_pnfs_update_layout(ino, pos, count,
1761                                         iomode, lo, lseg,
1762                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1763                         goto out_unlock;
1764                 }
1765         } else {
1766                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1767         }
1768
1769         /*
1770          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1771          * for LAYOUTRETURN even if first is true.
1772          */
1773         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1774                 spin_unlock(&ino->i_lock);
1775                 dprintk("%s wait for layoutreturn\n", __func__);
1776                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1777                         if (first)
1778                                 pnfs_clear_first_layoutget(lo);
1779                         pnfs_put_layout_hdr(lo);
1780                         dprintk("%s retrying\n", __func__);
1781                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1782                                         lseg, PNFS_UPDATE_LAYOUT_RETRY);
1783                         goto lookup_again;
1784                 }
1785                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1786                                 PNFS_UPDATE_LAYOUT_RETURN);
1787                 goto out_put_layout_hdr;
1788         }
1789
1790         if (pnfs_layoutgets_blocked(lo)) {
1791                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1792                                 PNFS_UPDATE_LAYOUT_BLOCKED);
1793                 goto out_unlock;
1794         }
1795         atomic_inc(&lo->plh_outstanding);
1796         spin_unlock(&ino->i_lock);
1797
1798         if (list_empty(&lo->plh_layouts)) {
1799                 /* The lo must be on the clp list if there is any
1800                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1801                  */
1802                 spin_lock(&clp->cl_lock);
1803                 if (list_empty(&lo->plh_layouts))
1804                         list_add_tail(&lo->plh_layouts, &server->layouts);
1805                 spin_unlock(&clp->cl_lock);
1806         }
1807
1808         pg_offset = arg.offset & ~PAGE_MASK;
1809         if (pg_offset) {
1810                 arg.offset -= pg_offset;
1811                 arg.length += pg_offset;
1812         }
1813         if (arg.length != NFS4_MAX_UINT64)
1814                 arg.length = PAGE_ALIGN(arg.length);
1815
1816         lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1817         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1818                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1819         atomic_dec(&lo->plh_outstanding);
1820         if (IS_ERR(lseg)) {
1821                 switch(PTR_ERR(lseg)) {
1822                 case -EBUSY:
1823                         if (time_after(jiffies, giveup))
1824                                 lseg = NULL;
1825                         break;
1826                 case -ERECALLCONFLICT:
1827                         /* Huh? We hold no layouts, how is there a recall? */
1828                         if (first) {
1829                                 lseg = NULL;
1830                                 break;
1831                         }
1832                         /* Destroy the existing layout and start over */
1833                         if (time_after(jiffies, giveup))
1834                                 pnfs_destroy_layout(NFS_I(ino));
1835                         /* Fallthrough */
1836                 case -EAGAIN:
1837                         break;
1838                 default:
1839                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1840                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1841                                 lseg = NULL;
1842                         }
1843                         goto out_put_layout_hdr;
1844                 }
1845                 if (lseg) {
1846                         if (first)
1847                                 pnfs_clear_first_layoutget(lo);
1848                         trace_pnfs_update_layout(ino, pos, count,
1849                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1850                         pnfs_put_layout_hdr(lo);
1851                         goto lookup_again;
1852                 }
1853         } else {
1854                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1855         }
1856
1857 out_put_layout_hdr:
1858         if (first)
1859                 pnfs_clear_first_layoutget(lo);
1860         pnfs_put_layout_hdr(lo);
1861 out:
1862         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1863                         "(%s, offset: %llu, length: %llu)\n",
1864                         __func__, ino->i_sb->s_id,
1865                         (unsigned long long)NFS_FILEID(ino),
1866                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1867                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1868                         (unsigned long long)pos,
1869                         (unsigned long long)count);
1870         return lseg;
1871 out_unlock:
1872         spin_unlock(&ino->i_lock);
1873         goto out_put_layout_hdr;
1874 }
1875 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1876
1877 static bool
1878 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1879 {
1880         switch (range->iomode) {
1881         case IOMODE_READ:
1882         case IOMODE_RW:
1883                 break;
1884         default:
1885                 return false;
1886         }
1887         if (range->offset == NFS4_MAX_UINT64)
1888                 return false;
1889         if (range->length == 0)
1890                 return false;
1891         if (range->length != NFS4_MAX_UINT64 &&
1892             range->length > NFS4_MAX_UINT64 - range->offset)
1893                 return false;
1894         return true;
1895 }
1896
1897 struct pnfs_layout_segment *
1898 pnfs_layout_process(struct nfs4_layoutget *lgp)
1899 {
1900         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1901         struct nfs4_layoutget_res *res = &lgp->res;
1902         struct pnfs_layout_segment *lseg;
1903         struct inode *ino = lo->plh_inode;
1904         LIST_HEAD(free_me);
1905
1906         if (!pnfs_sanity_check_layout_range(&res->range))
1907                 return ERR_PTR(-EINVAL);
1908
1909         /* Inject layout blob into I/O device driver */
1910         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1911         if (IS_ERR_OR_NULL(lseg)) {
1912                 if (!lseg)
1913                         lseg = ERR_PTR(-ENOMEM);
1914
1915                 dprintk("%s: Could not allocate layout: error %ld\n",
1916                        __func__, PTR_ERR(lseg));
1917                 return lseg;
1918         }
1919
1920         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1921
1922         spin_lock(&ino->i_lock);
1923         if (pnfs_layoutgets_blocked(lo)) {
1924                 dprintk("%s forget reply due to state\n", __func__);
1925                 goto out_forget;
1926         }
1927
1928         if (!pnfs_layout_is_valid(lo)) {
1929                 /* We have a completely new layout */
1930                 pnfs_set_layout_stateid(lo, &res->stateid, true);
1931         } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1932                 /* existing state ID, make sure the sequence number matches. */
1933                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1934                         dprintk("%s forget reply due to sequence\n", __func__);
1935                         goto out_forget;
1936                 }
1937                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1938         } else {
1939                 /*
1940                  * We got an entirely new state ID.  Mark all segments for the
1941                  * inode invalid, and retry the layoutget
1942                  */
1943                 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1944                 goto out_forget;
1945         }
1946
1947         pnfs_get_lseg(lseg);
1948         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1949
1950
1951         if (res->return_on_close)
1952                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1953
1954         spin_unlock(&ino->i_lock);
1955         pnfs_free_lseg_list(&free_me);
1956         return lseg;
1957
1958 out_forget:
1959         spin_unlock(&ino->i_lock);
1960         lseg->pls_layout = lo;
1961         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1962         if (!pnfs_layout_is_valid(lo))
1963                 nfs_commit_inode(ino, 0);
1964         return ERR_PTR(-EAGAIN);
1965 }
1966
1967 /**
1968  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1969  * @lo: pointer to layout header
1970  * @tmp_list: list header to be used with pnfs_free_lseg_list()
1971  * @return_range: describe layout segment ranges to be returned
1972  *
1973  * This function is mainly intended for use by layoutrecall. It attempts
1974  * to free the layout segment immediately, or else to mark it for return
1975  * as soon as its reference count drops to zero.
1976  */
1977 int
1978 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1979                                 struct list_head *tmp_list,
1980                                 const struct pnfs_layout_range *return_range,
1981                                 u32 seq)
1982 {
1983         struct pnfs_layout_segment *lseg, *next;
1984         int remaining = 0;
1985
1986         dprintk("%s:Begin lo %p\n", __func__, lo);
1987
1988         if (list_empty(&lo->plh_segs))
1989                 return 0;
1990
1991         assert_spin_locked(&lo->plh_inode->i_lock);
1992
1993         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1994                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
1995                         dprintk("%s: marking lseg %p iomode %d "
1996                                 "offset %llu length %llu\n", __func__,
1997                                 lseg, lseg->pls_range.iomode,
1998                                 lseg->pls_range.offset,
1999                                 lseg->pls_range.length);
2000                         if (mark_lseg_invalid(lseg, tmp_list))
2001                                 continue;
2002                         remaining++;
2003                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2004                 }
2005
2006         if (remaining)
2007                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2008
2009         return remaining;
2010 }
2011
2012 void pnfs_error_mark_layout_for_return(struct inode *inode,
2013                                        struct pnfs_layout_segment *lseg)
2014 {
2015         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2016         struct pnfs_layout_range range = {
2017                 .iomode = lseg->pls_range.iomode,
2018                 .offset = 0,
2019                 .length = NFS4_MAX_UINT64,
2020         };
2021         bool return_now = false;
2022
2023         spin_lock(&inode->i_lock);
2024         if (!pnfs_layout_is_valid(lo)) {
2025                 spin_unlock(&inode->i_lock);
2026                 return;
2027         }
2028         pnfs_set_plh_return_info(lo, range.iomode, 0);
2029         /*
2030          * mark all matching lsegs so that we are sure to have no live
2031          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2032          * for how it works.
2033          */
2034         if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2035                 nfs4_stateid stateid;
2036                 enum pnfs_iomode iomode;
2037
2038                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2039                 spin_unlock(&inode->i_lock);
2040                 if (return_now)
2041                         pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2042         } else {
2043                 spin_unlock(&inode->i_lock);
2044                 nfs_commit_inode(inode, 0);
2045         }
2046 }
2047 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2048
2049 void
2050 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2051 {
2052         if (pgio->pg_lseg == NULL ||
2053             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2054                 return;
2055         pnfs_put_lseg(pgio->pg_lseg);
2056         pgio->pg_lseg = NULL;
2057 }
2058 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2059
2060 /*
2061  * Check for any intersection between the request and the pgio->pg_lseg,
2062  * and if none, put this pgio->pg_lseg away.
2063  */
2064 static void
2065 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2066 {
2067         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2068                 pnfs_put_lseg(pgio->pg_lseg);
2069                 pgio->pg_lseg = NULL;
2070         }
2071 }
2072
2073 void
2074 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2075 {
2076         u64 rd_size = req->wb_bytes;
2077
2078         pnfs_generic_pg_check_layout(pgio);
2079         pnfs_generic_pg_check_range(pgio, req);
2080         if (pgio->pg_lseg == NULL) {
2081                 if (pgio->pg_dreq == NULL)
2082                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2083                 else
2084                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2085
2086                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2087                                                    req->wb_context,
2088                                                    req_offset(req),
2089                                                    rd_size,
2090                                                    IOMODE_READ,
2091                                                    false,
2092                                                    GFP_KERNEL);
2093                 if (IS_ERR(pgio->pg_lseg)) {
2094                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2095                         pgio->pg_lseg = NULL;
2096                         return;
2097                 }
2098         }
2099         /* If no lseg, fall back to read through mds */
2100         if (pgio->pg_lseg == NULL)
2101                 nfs_pageio_reset_read_mds(pgio);
2102
2103 }
2104 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2105
2106 void
2107 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2108                            struct nfs_page *req, u64 wb_size)
2109 {
2110         pnfs_generic_pg_check_layout(pgio);
2111         pnfs_generic_pg_check_range(pgio, req);
2112         if (pgio->pg_lseg == NULL) {
2113                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2114                                                    req->wb_context,
2115                                                    req_offset(req),
2116                                                    wb_size,
2117                                                    IOMODE_RW,
2118                                                    false,
2119                                                    GFP_NOFS);
2120                 if (IS_ERR(pgio->pg_lseg)) {
2121                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2122                         pgio->pg_lseg = NULL;
2123                         return;
2124                 }
2125         }
2126         /* If no lseg, fall back to write through mds */
2127         if (pgio->pg_lseg == NULL)
2128                 nfs_pageio_reset_write_mds(pgio);
2129 }
2130 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2131
2132 void
2133 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2134 {
2135         if (desc->pg_lseg) {
2136                 pnfs_put_lseg(desc->pg_lseg);
2137                 desc->pg_lseg = NULL;
2138         }
2139 }
2140 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2141
2142 /*
2143  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2144  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2145  */
2146 size_t
2147 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2148                      struct nfs_page *prev, struct nfs_page *req)
2149 {
2150         unsigned int size;
2151         u64 seg_end, req_start, seg_left;
2152
2153         size = nfs_generic_pg_test(pgio, prev, req);
2154         if (!size)
2155                 return 0;
2156
2157         /*
2158          * 'size' contains the number of bytes left in the current page (up
2159          * to the original size asked for in @req->wb_bytes).
2160          *
2161          * Calculate how many bytes are left in the layout segment
2162          * and if there are less bytes than 'size', return that instead.
2163          *
2164          * Please also note that 'end_offset' is actually the offset of the
2165          * first byte that lies outside the pnfs_layout_range. FIXME?
2166          *
2167          */
2168         if (pgio->pg_lseg) {
2169                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2170                                      pgio->pg_lseg->pls_range.length);
2171                 req_start = req_offset(req);
2172
2173                 /* start of request is past the last byte of this segment */
2174                 if (req_start >= seg_end)
2175                         return 0;
2176
2177                 /* adjust 'size' iff there are fewer bytes left in the
2178                  * segment than what nfs_generic_pg_test returned */
2179                 seg_left = seg_end - req_start;
2180                 if (seg_left < size)
2181                         size = (unsigned int)seg_left;
2182         }
2183
2184         return size;
2185 }
2186 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2187
2188 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2189 {
2190         struct nfs_pageio_descriptor pgio;
2191
2192         /* Resend all requests through the MDS */
2193         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2194                               hdr->completion_ops);
2195         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2196         return nfs_pageio_resend(&pgio, hdr);
2197 }
2198 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2199
2200 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2201 {
2202
2203         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2204         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2205             PNFS_LAYOUTRET_ON_ERROR) {
2206                 pnfs_return_layout(hdr->inode);
2207         }
2208         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2209                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2210 }
2211
2212 /*
2213  * Called by non rpc-based layout drivers
2214  */
2215 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2216 {
2217         if (likely(!hdr->pnfs_error)) {
2218                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2219                                 hdr->mds_offset + hdr->res.count);
2220                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2221         }
2222         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2223         if (unlikely(hdr->pnfs_error))
2224                 pnfs_ld_handle_write_error(hdr);
2225         hdr->mds_ops->rpc_release(hdr);
2226 }
2227 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2228
2229 static void
2230 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2231                 struct nfs_pgio_header *hdr)
2232 {
2233         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2234
2235         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2236                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2237                 nfs_pageio_reset_write_mds(desc);
2238                 mirror->pg_recoalesce = 1;
2239         }
2240         hdr->release(hdr);
2241 }
2242
2243 static enum pnfs_try_status
2244 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2245                         const struct rpc_call_ops *call_ops,
2246                         struct pnfs_layout_segment *lseg,
2247                         int how)
2248 {
2249         struct inode *inode = hdr->inode;
2250         enum pnfs_try_status trypnfs;
2251         struct nfs_server *nfss = NFS_SERVER(inode);
2252
2253         hdr->mds_ops = call_ops;
2254
2255         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2256                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2257         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2258         if (trypnfs != PNFS_NOT_ATTEMPTED)
2259                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2260         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2261         return trypnfs;
2262 }
2263
2264 static void
2265 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2266               struct nfs_pgio_header *hdr, int how)
2267 {
2268         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2269         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2270         enum pnfs_try_status trypnfs;
2271
2272         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2273         switch (trypnfs) {
2274         case PNFS_NOT_ATTEMPTED:
2275                 pnfs_write_through_mds(desc, hdr);
2276         case PNFS_ATTEMPTED:
2277                 break;
2278         case PNFS_TRY_AGAIN:
2279                 /* cleanup hdr and prepare to redo pnfs */
2280                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2281                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2282                         list_splice_init(&hdr->pages, &mirror->pg_list);
2283                         mirror->pg_recoalesce = 1;
2284                 }
2285                 hdr->mds_ops->rpc_release(hdr);
2286         }
2287 }
2288
2289 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2290 {
2291         pnfs_put_lseg(hdr->lseg);
2292         nfs_pgio_header_free(hdr);
2293 }
2294
2295 int
2296 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2297 {
2298         struct nfs_pgio_header *hdr;
2299         int ret;
2300
2301         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2302         if (!hdr) {
2303                 desc->pg_error = -ENOMEM;
2304                 return desc->pg_error;
2305         }
2306         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2307
2308         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2309         ret = nfs_generic_pgio(desc, hdr);
2310         if (!ret)
2311                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2312
2313         return ret;
2314 }
2315 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2316
2317 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2318 {
2319         struct nfs_pageio_descriptor pgio;
2320
2321         /* Resend all requests through the MDS */
2322         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2323         return nfs_pageio_resend(&pgio, hdr);
2324 }
2325 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2326
2327 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2328 {
2329         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2330         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2331             PNFS_LAYOUTRET_ON_ERROR) {
2332                 pnfs_return_layout(hdr->inode);
2333         }
2334         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2335                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2336 }
2337
2338 /*
2339  * Called by non rpc-based layout drivers
2340  */
2341 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2342 {
2343         if (likely(!hdr->pnfs_error))
2344                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2345         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2346         if (unlikely(hdr->pnfs_error))
2347                 pnfs_ld_handle_read_error(hdr);
2348         hdr->mds_ops->rpc_release(hdr);
2349 }
2350 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2351
2352 static void
2353 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2354                 struct nfs_pgio_header *hdr)
2355 {
2356         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2357
2358         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2359                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2360                 nfs_pageio_reset_read_mds(desc);
2361                 mirror->pg_recoalesce = 1;
2362         }
2363         hdr->release(hdr);
2364 }
2365
2366 /*
2367  * Call the appropriate parallel I/O subsystem read function.
2368  */
2369 static enum pnfs_try_status
2370 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2371                        const struct rpc_call_ops *call_ops,
2372                        struct pnfs_layout_segment *lseg)
2373 {
2374         struct inode *inode = hdr->inode;
2375         struct nfs_server *nfss = NFS_SERVER(inode);
2376         enum pnfs_try_status trypnfs;
2377
2378         hdr->mds_ops = call_ops;
2379
2380         dprintk("%s: Reading ino:%lu %u@%llu\n",
2381                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2382
2383         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2384         if (trypnfs != PNFS_NOT_ATTEMPTED)
2385                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2386         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2387         return trypnfs;
2388 }
2389
2390 /* Resend all requests through pnfs. */
2391 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2392 {
2393         struct nfs_pageio_descriptor pgio;
2394
2395         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2396                 /* Prevent deadlocks with layoutreturn! */
2397                 pnfs_put_lseg(hdr->lseg);
2398                 hdr->lseg = NULL;
2399
2400                 nfs_pageio_init_read(&pgio, hdr->inode, false,
2401                                         hdr->completion_ops);
2402                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2403         }
2404 }
2405 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2406
2407 static void
2408 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2409 {
2410         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2411         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2412         enum pnfs_try_status trypnfs;
2413
2414         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2415         switch (trypnfs) {
2416         case PNFS_NOT_ATTEMPTED:
2417                 pnfs_read_through_mds(desc, hdr);
2418         case PNFS_ATTEMPTED:
2419                 break;
2420         case PNFS_TRY_AGAIN:
2421                 /* cleanup hdr and prepare to redo pnfs */
2422                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2423                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2424                         list_splice_init(&hdr->pages, &mirror->pg_list);
2425                         mirror->pg_recoalesce = 1;
2426                 }
2427                 hdr->mds_ops->rpc_release(hdr);
2428         }
2429 }
2430
2431 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2432 {
2433         pnfs_put_lseg(hdr->lseg);
2434         nfs_pgio_header_free(hdr);
2435 }
2436
2437 int
2438 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2439 {
2440         struct nfs_pgio_header *hdr;
2441         int ret;
2442
2443         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2444         if (!hdr) {
2445                 desc->pg_error = -ENOMEM;
2446                 return desc->pg_error;
2447         }
2448         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2449         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2450         ret = nfs_generic_pgio(desc, hdr);
2451         if (!ret)
2452                 pnfs_do_read(desc, hdr);
2453         return ret;
2454 }
2455 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2456
2457 static void pnfs_clear_layoutcommitting(struct inode *inode)
2458 {
2459         unsigned long *bitlock = &NFS_I(inode)->flags;
2460
2461         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2462         smp_mb__after_atomic();
2463         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2464 }
2465
2466 /*
2467  * There can be multiple RW segments.
2468  */
2469 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2470 {
2471         struct pnfs_layout_segment *lseg;
2472
2473         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2474                 if (lseg->pls_range.iomode == IOMODE_RW &&
2475                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2476                         list_add(&lseg->pls_lc_list, listp);
2477         }
2478 }
2479
2480 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2481 {
2482         struct pnfs_layout_segment *lseg, *tmp;
2483
2484         /* Matched by references in pnfs_set_layoutcommit */
2485         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2486                 list_del_init(&lseg->pls_lc_list);
2487                 pnfs_put_lseg(lseg);
2488         }
2489
2490         pnfs_clear_layoutcommitting(inode);
2491 }
2492
2493 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2494 {
2495         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2496 }
2497 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2498
2499 void
2500 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2501                 loff_t end_pos)
2502 {
2503         struct nfs_inode *nfsi = NFS_I(inode);
2504         bool mark_as_dirty = false;
2505
2506         spin_lock(&inode->i_lock);
2507         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2508                 nfsi->layout->plh_lwb = end_pos;
2509                 mark_as_dirty = true;
2510                 dprintk("%s: Set layoutcommit for inode %lu ",
2511                         __func__, inode->i_ino);
2512         } else if (end_pos > nfsi->layout->plh_lwb)
2513                 nfsi->layout->plh_lwb = end_pos;
2514         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2515                 /* references matched in nfs4_layoutcommit_release */
2516                 pnfs_get_lseg(lseg);
2517         }
2518         spin_unlock(&inode->i_lock);
2519         dprintk("%s: lseg %p end_pos %llu\n",
2520                 __func__, lseg, nfsi->layout->plh_lwb);
2521
2522         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2523          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2524         if (mark_as_dirty)
2525                 mark_inode_dirty_sync(inode);
2526 }
2527 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2528
2529 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2530 {
2531         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2532
2533         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2534                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2535         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2536 }
2537
2538 /*
2539  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2540  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2541  * data to disk to allow the server to recover the data if it crashes.
2542  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2543  * is off, and a COMMIT is sent to a data server, or
2544  * if WRITEs to a data server return NFS_DATA_SYNC.
2545  */
2546 int
2547 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2548 {
2549         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2550         struct nfs4_layoutcommit_data *data;
2551         struct nfs_inode *nfsi = NFS_I(inode);
2552         loff_t end_pos;
2553         int status;
2554
2555         if (!pnfs_layoutcommit_outstanding(inode))
2556                 return 0;
2557
2558         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2559
2560         status = -EAGAIN;
2561         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2562                 if (!sync)
2563                         goto out;
2564                 status = wait_on_bit_lock_action(&nfsi->flags,
2565                                 NFS_INO_LAYOUTCOMMITTING,
2566                                 nfs_wait_bit_killable,
2567                                 TASK_KILLABLE);
2568                 if (status)
2569                         goto out;
2570         }
2571
2572         status = -ENOMEM;
2573         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2574         data = kzalloc(sizeof(*data), GFP_NOFS);
2575         if (!data)
2576                 goto clear_layoutcommitting;
2577
2578         status = 0;
2579         spin_lock(&inode->i_lock);
2580         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2581                 goto out_unlock;
2582
2583         INIT_LIST_HEAD(&data->lseg_list);
2584         pnfs_list_write_lseg(inode, &data->lseg_list);
2585
2586         end_pos = nfsi->layout->plh_lwb;
2587
2588         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2589         spin_unlock(&inode->i_lock);
2590
2591         data->args.inode = inode;
2592         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2593         nfs_fattr_init(&data->fattr);
2594         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2595         data->res.fattr = &data->fattr;
2596         if (end_pos != 0)
2597                 data->args.lastbytewritten = end_pos - 1;
2598         else
2599                 data->args.lastbytewritten = U64_MAX;
2600         data->res.server = NFS_SERVER(inode);
2601
2602         if (ld->prepare_layoutcommit) {
2603                 status = ld->prepare_layoutcommit(&data->args);
2604                 if (status) {
2605                         put_rpccred(data->cred);
2606                         spin_lock(&inode->i_lock);
2607                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2608                         if (end_pos > nfsi->layout->plh_lwb)
2609                                 nfsi->layout->plh_lwb = end_pos;
2610                         goto out_unlock;
2611                 }
2612         }
2613
2614
2615         status = nfs4_proc_layoutcommit(data, sync);
2616 out:
2617         if (status)
2618                 mark_inode_dirty_sync(inode);
2619         dprintk("<-- %s status %d\n", __func__, status);
2620         return status;
2621 out_unlock:
2622         spin_unlock(&inode->i_lock);
2623         kfree(data);
2624 clear_layoutcommitting:
2625         pnfs_clear_layoutcommitting(inode);
2626         goto out;
2627 }
2628 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2629
2630 int
2631 pnfs_generic_sync(struct inode *inode, bool datasync)
2632 {
2633         return pnfs_layoutcommit_inode(inode, true);
2634 }
2635 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2636
2637 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2638 {
2639         struct nfs4_threshold *thp;
2640
2641         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2642         if (!thp) {
2643                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2644                 return NULL;
2645         }
2646         return thp;
2647 }
2648
2649 #if IS_ENABLED(CONFIG_NFS_V4_2)
2650 int
2651 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2652 {
2653         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2654         struct nfs_server *server = NFS_SERVER(inode);
2655         struct nfs_inode *nfsi = NFS_I(inode);
2656         struct nfs42_layoutstat_data *data;
2657         struct pnfs_layout_hdr *hdr;
2658         int status = 0;
2659
2660         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2661                 goto out;
2662
2663         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2664                 goto out;
2665
2666         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2667                 goto out;
2668
2669         spin_lock(&inode->i_lock);
2670         if (!NFS_I(inode)->layout) {
2671                 spin_unlock(&inode->i_lock);
2672                 goto out_clear_layoutstats;
2673         }
2674         hdr = NFS_I(inode)->layout;
2675         pnfs_get_layout_hdr(hdr);
2676         spin_unlock(&inode->i_lock);
2677
2678         data = kzalloc(sizeof(*data), gfp_flags);
2679         if (!data) {
2680                 status = -ENOMEM;
2681                 goto out_put;
2682         }
2683
2684         data->args.fh = NFS_FH(inode);
2685         data->args.inode = inode;
2686         status = ld->prepare_layoutstats(&data->args);
2687         if (status)
2688                 goto out_free;
2689
2690         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2691
2692 out:
2693         dprintk("%s returns %d\n", __func__, status);
2694         return status;
2695
2696 out_free:
2697         kfree(data);
2698 out_put:
2699         pnfs_put_layout_hdr(hdr);
2700 out_clear_layoutstats:
2701         smp_mb__before_atomic();
2702         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2703         smp_mb__after_atomic();
2704         goto out;
2705 }
2706 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2707 #endif
2708
2709 unsigned int layoutstats_timer;
2710 module_param(layoutstats_timer, uint, 0644);
2711 EXPORT_SYMBOL_GPL(layoutstats_timer);