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